RESEARCH DOSSIER
Measurement-Based Care and Individualized Treatment Selection
Abstract. Measurement-based care uses repeated, interpretable assessments of symptoms, functioning, safety, and patient-defined priorities to inform clinical decisions over time. Individualized treatment selection uses evidence about expected outcomes under alternative clinically appropriate options and the person's goals, constraints, and previous responses. This fundamental research monograph examines their conceptual history, psychometric foundations, causal logic, longitudinal decision rules, adaptive trial designs, comparative experiments, patient participation, implementation barriers, and the emerging use of prediction models and digital decision support. Its critical narrative synthesis distinguishes improvement in measurement from improvement in treatment, prognostic accuracy from treatment-effect heterogeneity, and statistical prediction from demonstrable clinical utility. The evidence spans PHQ-9 and GAD-7 validation, STAR*D, IMPACT, the Guo and Husain randomized trials, the Cochrane PROM review, the personalized advantage index, PETRUSHKA, SMART methodology, and contemporary implementation research. Findings support carefully implemented measurement-and-action pathways in particular clinical contexts, while showing that passive monitoring, individualized algorithms, and short-term symptom changes require separate evaluation against functioning, harms, fairness, long-term outcomes, and patient preferences. The synthesis develops an evidence-appraisal agenda for scientifically accountable, participatory, dynamically adaptive care.
Keywords: measurement-based care; individualized treatment selection; routine outcome monitoring; patient-reported outcome measures; PHQ-9; GAD-7; depression; clinical decision making; adaptive interventions; sequential multiple assignment randomized trials; SMART; treatment-effect heterogeneity; personalized advantage index; precision mental health; clinical prediction; calibration; external validation; shared decision making; mental health services; clinical utility; implementation science; safety monitoring; Artificial Era
Introduction: The Scientific Problem of Treatment Decisions Over Time
Clinical treatment is a sequence of decisions under uncertainty. A person enters care with a history, a fluctuating condition, goals, constraints, and an incompletely observed response to available interventions. A clinician must decide what to offer, what to measure, when to reassess, and whether to continue, adapt, combine, switch, intensify, or end an intervention. Each decision changes the conditions under which subsequent observations are made. A baseline diagnostic label identifies a broad problem; it cannot by itself determine whether an individual's current therapy is beneficial, tolerable, feasible, or aligned with a changing life. Measurement-based care and individualized treatment selection address complementary parts of this longitudinal decision problem Fortney et al. (2017) Cohen & DeRubeis (2018).
Measurement-based care, abbreviated MBC, is the routine collection, interpretation, sharing, and clinical use of structured information about a person's condition and treatment response. Its defining feature is a consequential feedback connection between evidence and an actual care decision. Repeated questionnaires stored without review constitute data collection; a score displayed without an agreed response pathway may provide information without improving care. Individualized treatment selection addresses a different question: among clinically appropriate alternatives, which option offers this particular person the most favorable expected combination of benefit, harm, burden, and acceptability? A scientifically credible answer requires evidence about alternatives and, ideally, variation in their comparative effects Resnick (2025) Cohen & DeRubeis (2018).
The two domains interact throughout treatment. At baseline, prognostic and prescriptive information can shape an initial choice; after treatment begins, repeated measures characterize response, emerging adverse effects, functional recovery, and patient preference; subsequent decisions use the observed history to revise the care plan. These processes operate within resource constraints, access barriers, cultural interpretation, measurement error, and clinician–patient relationships. Research must therefore distinguish the properties of instruments, the behavior of the decision-maker, the causal effects of an intervention, and the performance of the service implementing them. A scale can be valid without the clinical pathway built around it being effective Lutz et al. (2022) Burr et al. (2026).
This monograph investigates the conceptual and historical origins of MBC, clinical psychometrics, informative repeated measurement, evidence-producing treatment algorithms, experimental studies of MBC, individualized prediction, randomized comparison of decision support, sequential treatment strategies, implementation and safety, equity, and research opportunities. Its principal Science membership is Clinical Psychology and Psychopathology; connected domains include Research Methods and Psychometrics and Counseling Psychology and Psychotherapy. It concentrates on cross-modality treatment decisions, escalation, and switching. Psychotherapy session feedback as an intervention in its own right and generic measurement theory retain independent research boundaries. The primary question is how psychological knowledge becomes a justified action under individual uncertainty.
The method is a critical structured narrative synthesis of original experiments, psychometric studies, observational investigations, methodological contributions, clinical guidance, and systematic reviews available through November 3, 2026. Sources were selected purposively for construct definition, landmark evidence, causal interpretation, contradictory outcomes, and current relevance; this work claims neither an exhaustive search nor a registered systematic review. Study-specific numerical results are attributed to identifiable publications, and the distinction between a study's directly accessible report and a fuller inaccessible analysis is preserved. The review treats scientific rigor as a property of explicit claims and evidence rather than a claim generated by the size of the bibliography.
Historical Formation: From Clinical Rating to Learning Care Systems
Clinical medicine and psychiatry have long employed observation, descriptive case histories, severity judgments, and subsequent reassessment, but the systematic use of standardized patient measures to direct treatment is a more specific research program. The late twentieth century established methods for determining when an individual's observed change exceeded the error expected of an assessment and when it approached a meaningful clinical state. Jacobson and Truax (1991) articulated reliable and clinically significant change as separate evaluative questions, highlighting the difference between a numerical improvement and a meaningful transition in a person's functioning Jacobson & Truax (1991). This distinction remains essential whenever a modern dashboard announces progress.
Modern self-report instruments made frequent monitoring feasible outside specialist research units. The Patient Health Questionnaire-9 (PHQ-9) was validated as a brief depression severity measure by Kroenke, Spitzer, and Williams (2001), and later work examined its responsiveness to treatment-related change Kroenke et al. (2001) Löwe et al. (2004). The GAD-7 introduced a concise assessment of generalized anxiety symptom severity Spitzer et al. (2006). A longer methodological trajectory produced item banks and computerized approaches that allow broad domains of health and functioning to be measured with improved precision; the PROMIS program is an important illustration Cella et al. (2010). These instruments made repeated numerical observation more practicable but did not establish the benefits of any particular treatment algorithm.
A second historical stream developed organized, stepped, and collaborative treatment. In the IMPACT randomized trial, collaborative care for late-life depression in primary care combined systematic follow-up and symptom monitoring with care management and specialist support Unützer et al. (2002). The multi-step STAR*D research program subsequently investigated outcomes across successive treatment strategies after an initial antidepressant trial Rush et al. (2006). These programs exposed the temporal structure of treatment: initial choices often need revision, and real-world performance depends on coordinated execution. Their intervention packages were complex, so improvements cannot be assigned uniquely to measurement rather than the additional clinical attention, treatment access, and decision support embedded in the package.
In the following decades, a distinct implementation literature argued for integrating measures into ordinary clinical conversations and workflows. Fortney and colleagues (2017) described MBC as a potential turning point in mental health services, while Lutz and colleagues (2022) connected the wider movement to data-informed psychological care Fortney et al. (2017) Lutz et al. (2022). The worldwide uptake of digital questionnaires, electronic records, telehealth, and decision dashboards made monitoring technologically easier, yet the persistently uneven implementation described in more recent syntheses shows that technological availability and routine clinical utilization are different outcomes Burr et al. (2026).
The historical record thus contains several partly independent developments: instruments capable of measuring change, concepts of meaningful individual improvement, service models capable of following patients, trials comparing treatment pathways, and methods estimating differential benefit. Their integration supports contemporary MBC and personalized selection. It also explains the field's conceptual ambiguity. The same term is sometimes applied to a symptom checklist, an algorithm-driven prescribing program, a collaborative-care service, or a probabilistic treatment selector. Historical continuity is best established through the actual measurements and decision rules used in each study rather than by retrospective terminology.
Operational Definitions and the Architecture of Clinical Action
A minimal MBC process has at least four empirically distinct components. First, the care system obtains an interpretable baseline estimate of severity, functioning, risks, goals, or treatment burden. Second, it repeats relevant measures at clinically meaningful intervals, with clear responsibility for completion and review. Third, the patient and clinician interpret change against the patient's narrative, measurement uncertainty, treatment exposure, and contextual circumstances. Fourth, that interpretation has the ability to modify an agreed care plan. Resnick (2025) summarizes this interaction as Collect, Share, Act, a concise operational formulation Resnick (2025). The fourth step defines the difference between an observational archive and a care process.
Individualized treatment selection is the use of individual-level characteristics, preferences, and credible comparative evidence to choose among specific therapeutic options. A prediction of the likelihood of recovery under one treatment is prognostic. A prediction of the difference between outcomes under two treatments is prescriptive, assuming comparability can be established. The distinction is crucial: a variable that predicts a worse outcome for everyone can identify urgency without identifying which treatment is better. A treatment-selection rule becomes clinically meaningful when it compares feasible options on outcomes that matter and is prospectively evaluated Cohen & DeRubeis (2018) DeRubeis et al. (2014).
The unit of clinical decision is neither the instrument nor the average treatment effect in isolation, but a decision occasion defined by the person's history and available alternatives. A person's symptom score at week four depends on initial severity, what treatment was received, the intensity and quality of implementation, external events, and the interval since the previous assessment. A treatment change can itself affect subsequent response measurement. A clinically useful model must therefore register baseline state, chosen intervention, adherence or engagement, measurement timing, intermediate outcome, adverse effects, and contextual shocks rather than compress the history into a single last-observed score Murphy (2005).
Three distinct evaluation targets follow. Measurement validity asks whether the collected number represents the intended construct, with acceptable reliability and comparability across time and groups. Decision validity asks whether the thresholds and interpretations lead to defensible treatment choices. Effectiveness asks whether implementing those choices improves patient-important outcomes compared with a specified alternative. A system may succeed at one target and fail at another. A more precise questionnaire cannot overcome a decision rule that changes treatment prematurely; a technically accurate risk model can be clinically unhelpful when its proposed action is unavailable or unacceptable Vickers & Elkin (2006) Collins et al. (2024).
Personalized care additionally requires a normative and relational layer. Choices may be informed by patient preference, prior positive and negative experiences, expected adverse effects, accessibility, affordability, family obligations, disability-related access, and willingness to participate in monitoring. NICE guidance for adult depression situates repeated outcome monitoring within comprehensive assessment and collaborative treatment planning rather than treating numerical scores as a self-sufficient diagnosis NICE (2022/2026). These considerations also determine which outcomes deserve measurement. The same symptom improvement may coincide with quite different experiences of functioning, agency, or treatment burden.
The Measurement–Interpretation–Action Cycle
Each MBC cycle begins with the selection of a question that a repeated assessment could help answer. Is the condition improving at the expected rate? Is the current plan tolerable? Is the person able to carry out agreed treatment? Has day-to-day functioning recovered? Are adverse effects or immediate safety concerns emerging? Measures should be chosen because their answers bear on a decision. Habitual accumulation of every available symptom scale increases workload, creates multiple comparisons, and may obscure rather than clarify the clinically salient trajectory Fortney et al. (2017) Dey et al. (2025).
Collection is only partly a technical operation. Responses depend on recall periods, question wording, privacy, literacy, disability access, language, respondent fatigue, timing relative to treatment, and the possibility that a person does not feel safe disclosing particular concerns. Repeated measures should provide a stable frame for comparison while allowing the clinical meaning of scores to be revisited when circumstances change. When measures are administered inconsistently, apparent trajectories may be produced by changing contexts and respondent composition. A completion rate is an important process measure but cannot be equated with care quality or a therapeutic outcome Cella et al. (2010) Burr et al. (2026).
Sharing is the interpretive step at which observations become clinically legible. A patient should be able to recognize what the instrument asked, how the score changed, and how that information fits or conflicts with lived experience. The clinician must review the trajectory, investigate deterioration and discordance, identify whether a change is larger than likely measurement noise, and compare it with functional and safety information. Presentation matters: raw scores, uncertainty, and visible trends can stimulate useful dialogue; inflexible red–amber–green categories may conceal ambiguity and imply unwarranted diagnostic precision Jacobson & Truax (1991) Löwe et al. (2004).
Action requires an accountable decision pathway. Possible actions include continuing a helpful course, clarifying expectations, addressing barriers, reviewing tolerability, revisiting diagnosis, obtaining specialist consultation, selecting a different modality, or intensifying support when indicated. This list describes decision categories for scientific analysis, not a universal prescribing schedule. The role of clinical judgment is to determine which option is appropriate given the full case. The mere occurrence of an action does not prove that it was the right action: avoidable switching and delayed responses are both potential harms. Clinical services should record what decision followed the measurements and why NICE (2022/2026) Resnick (2025).
The cycle is closed only when subsequent assessment evaluates the consequences of the previous decision. This creates a learning loop in which data have temporal meaning: a score is both an observation of current state and a possible input into the next action. The system can be evaluated at several bottlenecks—completion, timely feedback, patient discussion, action taken, safety detection, outcome change, and equity across subgroups. A service that documents repeated measurement without evidence of shared interpretation or action is implementing an assessment program, not necessarily effective MBC Burr et al. (2026) Fortney et al. (2017).
Clinical Constructs, Measurement Levels, and Choice of Outcomes
Symptom severity is a central but incomplete target. Depression and anxiety measures can help characterize change in frequently reported experiences, while distinct assessments may be required for mania, psychosis, trauma-related symptoms, substance use, or neurocognitive difficulties. A generic depression score cannot replace condition-specific assessment or differential diagnosis. Research designs should identify the intended construct, period of recall, interpretive direction, and clinical population for every outcome. The PHQ-9 and GAD-7 demonstrate the scientific importance of developing and validating concise instruments, and the necessity of testing their use in each relevant setting Kroenke et al. (2001) Spitzer et al. (2006).
Functioning has partial independence from symptom intensity. A person may report fewer symptoms yet remain unable to resume occupational, educational, family, or social participation. Conversely, meaningful functional gains may occur while residual symptoms persist. A measurement plan that aims to evaluate recovery should therefore include disability, role participation, or patient-defined functional goals where appropriate. PROMIS research illustrates efforts to measure physical, mental, and social health domains on coordinated scales Cella et al. (2010). Its generality also requires attention to disease-specific features that generic item banks may not capture.
Safety is a separate construct with different temporal demands. The presence of suicidal thoughts, severe agitation, emerging mood elevation, treatment intolerance, or rapid functional deterioration may require clinical assessment outside the interval used for routine symptom scoring. Standardized item endorsement can alert a service to possible concern, but it is neither a complete risk assessment nor a reliable numerical forecast of a rare event. NICE explicitly calls for direct questioning about suicidal ideation and clinical response to risk, particularly around changes in depression treatment NICE (2022/2026). Research on MBC should report detection pathways and adverse events as outcomes rather than relegating them to implementation footnotes.
Treatment acceptability and burden are also outcomes. A patient may discontinue an intervention because of adverse effects, cost, schedule incompatibility, stigma, transportation barriers, or changing preferences. Some of these factors reflect an unsuccessful plan rather than deficient motivation. Engagement, appointment attendance, pharmacy dispensing, and self-reported adherence measure related but noninterchangeable behaviors. Their interpretation must remain linked to an agreed treatment and its clinical appropriateness. The PETRUSHKA trial's focus on antidepressant discontinuation demonstrates why a proximal measure of tolerability or acceptability can be a meaningful primary endpoint distinct from symptom reduction Cipriani et al. (2026).
A coherent outcome architecture consequently distinguishes symptoms, function, safety, experience, treatment participation, and longer-term recovery. Primary and secondary endpoints should be declared before analysis where possible, and thresholds should be chosen for the specific decision rather than retrospectively to maximize apparent success. Composite outcomes may be useful when each component is specified and weighted transparently, but a single aggregate can conceal worsening in a domain a patient considers decisive. The core question of measurement is what the number is capable of representing in relation to clinical action.
Psychometric Foundations: Reliability, Validity, and Responsiveness
A longitudinal care system is only as informative as the observations on which it depends. Reliability concerns the consistency or precision of measurement under defined conditions; construct validity concerns whether an instrument's scores can support a proposed interpretation; responsiveness concerns the ability to detect change in the construct across the interval of interest. These properties are related but empirically distinct. A scale developed for cross-sectional screening may also be useful for monitoring, yet that application needs evidence that score changes meaningfully reflect clinical change rather than shifts in item interpretation, temporary circumstances, or measurement error Kroenke et al. (2001) Löwe et al. (2004).
The PHQ-9 illustrates both utility and limits. It uses nine items reflecting symptoms included in the depressive syndrome, has an interpretable numerical range, and can be administered repeatedly with relatively modest burden. The original validation study supported its use as a brief depression measure, while subsequent investigations examined changes during treatment Kroenke et al. (2001) Löwe et al. (2004). These findings establish a plausible monitoring instrument; they do not make the same raw-score difference equally significant for every baseline severity, population, intervention, or interval. A person's baseline score, symptom pattern, duration, and functional history remain relevant to interpretation.
GAD-7 similarly provides a short measure of anxiety symptoms, introduced through a validation investigation in primary care Spitzer et al. (2006). Using both instruments in patients with co-occurring anxious and depressive symptoms can reveal divergent trajectories and prevent a single label from obscuring another clinically relevant burden. However, anxiety item responses are influenced by medical conditions, environmental threats, and changing interpretations of symptoms. Screening scores should be interpreted within the full differential and contextual assessment, particularly when treatment decisions have potentially serious consequences.
The PROMIS program provides a different design logic: item banks calibrated using item-response theory aim to measure relevant health domains efficiently across a range of respondent abilities Cella et al. (2010). Such instruments may improve precision, facilitate comparison, and reduce the number of questions needed for some purposes. Adaptive questionnaire administration can still change the sequence of items and the uncertainty attached to a score. A clinician-facing representation should retain the metric, measurement precision, and the meaning of change. A technically elegant scoring algorithm offers limited value when patients and clinicians cannot understand the reported construct.
Clinical validation is a continuing obligation. The population in which a measure was developed may differ from the one where it is deployed in language, literacy, age, disability, severity, comorbidity, setting, and service access. Measurement equivalence across these groups and over time must be examined rather than assumed. Differential item functioning or a response shift can make nominally identical scores represent different experiences. When measurement invariance is weak, threshold-driven algorithms may magnify inequity; when evidence is uncertain, qualitative interpretation and human review become more important, not less.
Interpreting Longitudinal Change and Clinical Significance
Repeated scores invite two tempting simplifications: every numerical movement appears to be improvement or decline, and every crossing of a cut point appears to represent a new clinical state. Both interpretations can be false. Each observation contains measurement noise and within-person fluctuation; the sampling interval changes what can be inferred about the slope; regression to the mean may produce change after selection of people with unusually high baseline scores. Longitudinal interpretation should ask whether observed movement is larger than plausible error, sustained across repeated assessments, corroborated by other observations, and meaningful to the patient Jacobson & Truax (1991) Löwe et al. (2004).
Jacobson and Truax (1991) distinguish reliable change from clinically significant change. Reliable change evaluates whether a difference is large enough to exceed expected measurement variability under relevant assumptions; clinically significant change concerns whether the person's condition enters a meaningfully different clinical range Jacobson & Truax (1991). The criteria address separate dimensions. A reliable but small improvement may matter if it changes functioning; a threshold crossing may be too unstable to interpret if a person sits close to the boundary. Subsequent monitoring should therefore preserve uncertainty instead of representing categories as abrupt changes in underlying reality.
Longitudinal change requires attention to treatment timing. A few days after a psychosocial intervention begins, a score may capture expectancy, crisis resolution, or unrelated events rather than its sustained mechanism. A later plateau may reflect residual impairment, insufficient treatment exposure, diminished engagement, or a ceiling in the instrument. Linear extrapolation from early observations is especially unreliable when recovery is episodic or delayed. Clinically interpretable graphs need visible assessment dates, intervention changes, missing intervals, and potentially relevant adverse events. Models can summarize trajectories, but a visually persuasive trend remains an inference from incomplete observations.
Population-level minimum important differences and individual-level reliable change are not interchangeable. The former often describes an average difference considered meaningful under a particular study method; the latter addresses whether a specific observed difference plausibly exceeds measurement error. The PHQ-9 change literature illustrates how different anchor-based and distribution-based methods produce related yet distinct estimates Löwe et al. (2004). A cut point selected in one diagnostic setting should not be reified as a universal rule for changing medications or psychological treatment. Decisions also depend on baseline severity, safety, patient burden, response velocity, and the consequences of delay.
Meaningful clinical change ultimately has more than one axis: symptom reduction, valued functioning, reduction of suffering, increased autonomy, and improved safety. A person can cross a symptom threshold while losing tolerability or functioning; another may remain above a nominal remission boundary while gaining meaningful stability and participation. An evidence-based monitoring conversation explicitly asks which changes matter, what remains unresolved, and whether the current treatment plan is still acceptable. This approach preserves the strengths of quantification without confusing the map of a phenomenon with the phenomenon in its entirety.
Decision Thresholds, Escalation, Switching, and Continuation
A decision threshold links an observed state to a possible action. In research, thresholds can be derived from prognostic studies, treatment trials, clinical guidelines, cost or harm tradeoffs, and patient preferences. Their validity depends on the specific decision: the threshold for considering a routine review need not equal that for urgent safety intervention, referral, or change of treatment modality. Thresholds are thus properties of a decision problem and its consequences, not natural divisions in the score distribution. A threshold cannot be justified merely because an instrument's manual supplies a conventional severity category Vickers & Elkin (2006) NICE (2022/2026).
Escalation means adding support or intensity when an initial course has produced inadequate benefit under adequate opportunity for response. Switching entails moving to a different option, potentially with distinct adverse effects, treatment burdens, and delayed benefits. Continuation preserves an acceptable course when improvement remains plausible. In all cases, apparent nonresponse must be investigated in relation to diagnosis, therapeutic dose or exposure when relevant, engagement, comorbid conditions, psychosocial context, adverse events, and measurement validity. A rule based exclusively on a score at a fixed week can make a patient appear refractory when the main problem is access or feasibility Rush et al. (2006) NICE (2022/2026).
STAR*D provides a major historical example of sequential care across several treatment steps Rush et al. (2006). The study demonstrated the practical importance of following outcomes beyond the first chosen antidepressant. Its sequence does not provide a universal individualized algorithm and does not eliminate selection effects between later treatment steps. Participants who continue differ from those who discontinue, and a later observed response may reflect history, attrition, or treatment combinations. Inference from successive steps must be framed within the actual design and the population remaining under observation.
Clinical guidance supplies one current example of a structured yet adaptable framework. NICE's adult depression guideline recommends review between two and four weeks after starting treatment, monitoring treatment concordance, side effects, and suicidal ideation, and considering appropriate validated outcome measures NICE (2022/2026). It also calls for attention to previous treatment experience, comorbidity, functioning, and preferences. These are context-specific recommendations from a defined health system and should not be misread as an automatic dosing or switching directive. Their theoretical significance is the integration of measurement into a broader clinical decision, not numerical control for its own sake.
Decision thresholds should be evaluated by the outcomes their use causes, including overlooked deterioration, unnecessary escalation, time to acceptable improvement, harm, treatment burden, and resource distribution. The same score may justify different responses for two patients because risks, constraints, and alternatives differ. An individualized strategy may also include planned periods of watchful waiting, active support, or additional diagnostic assessment. The scientific standard is prospective comparison of the decision strategy, with explicit recording of exceptions and reasons, rather than retrospective praise for frequent treatment changes.
Multi-Informant, Multi-Method, and Missing-Data Problems
Different measurement methods observe different traces of a clinical process. Self-reports capture personally experienced symptoms and goals; clinician ratings combine interview and observation; caregivers may report visible behavior or function; administrative records document visits and treatment changes; device-derived observations may quantify aspects of activity or sleep. Agreement among methods can strengthen an inference, but disagreement may be informative because constructs, recall windows, vantage points, and incentives differ. Method choice must follow the research question, ethical constraints, and scientific validation rather than a hierarchy that automatically privileges technological sensors over lived reports Cella et al. (2010) Spitzer et al. (2006).
Missing assessments often carry information. People who stop responding may have improved and no longer wish to complete surveys, may be deteriorating or disengaged, may lack connectivity, or may experience severe competing demands. Under these conditions missingness can depend on unobserved clinical state and violate common assumptions used in longitudinal estimation. Ignoring missing time points or imputing all of them with the last observed score can bias estimates of improvement and mask inequities in who remains represented. Trials should describe follow-up denominators, reasons for missingness when known, and sensitivity analyses under plausible departures from missing-at-random assumptions Kendrick et al. (2016) Cipriani et al. (2026).
Clinician-rated and patient-reported data are also vulnerable to expectation, social desirability, demand characteristics, and differences in the frequency of contact. Unblinded intervention studies may improve measured results partly through increased attention to symptoms or changed reporting. Blind assessors can mitigate one form of detection bias, as in the Guo trial, but cannot remove all forms of performance bias from an open clinical care pathway Guo et al. (2015). It is therefore useful to supplement primary symptom outcomes with functioning, tolerability, service use, and independently assessed indicators when feasible.
Technology introduces further missingness: forms may remain unsent, interfaces may fail, translations may be inappropriate, and wearable devices may have variable accuracy across contexts. A high rate of automatic data transmission can create false confidence about representativeness. Systems need clear definitions of completion, late response, unavailable response, and intentional refusal. They must also specify whether the absence of a measure itself triggers outreach, and whether outreach respects privacy and autonomy. These procedures are part of the intervention and belong in trial reporting and implementation evaluation.
A sophisticated longitudinal model can handle irregular observation times, measurement error, and correlated responses, but model complexity cannot recover information that was never collected without assumptions. Transparent sensitivity analyses and simple displays of observed versus missing data are therefore essential for interpretable research. In clinical practice, the same logic argues for displaying the last assessment date and uncertainty alongside the latest score. The apparent precision of a plotted line must never substitute for an honest account of incomplete observation.
Research Designs: What Counts as Evidence That Monitoring Works?
The causal claim that measurement-based care improves outcomes requires a comparison between alternative ways of organizing care. A before–after series showing symptom improvement in a monitored clinic is insufficient because clinical conditions can improve over time, because patients may enter treatment at moments of unusually high distress, and because other parts of care may change simultaneously. Observational cohorts can describe feasibility, trajectories, adherence, and associations; randomized trials can estimate the effect of assignment to a specified monitoring-and-action package relative to its comparator. The relevant estimand must state what components were assigned, the target population, the follow-up interval, and the outcomes used Kendrick et al. (2016) Zhu et al. (2021).
An MBC trial can compare structured monitoring with less formal assessment, an integrated decision algorithm with clinician-selected care, frequent versus sparse feedback, or a fully organized collaborative pathway with usual services. These interventions differ in causal content. When both groups complete the same measures but clinicians only receive scores in one arm, the contrast emphasizes feedback and its use. When one arm includes additional case management, scheduled follow-up, specialist consultation, and stepped treatment, improvements cannot be attributed to measurement independently. Reviews pooling these configurations can describe broad patterns but may obscure mechanisms and transportability Unützer et al. (2002) Fortney et al. (2017).
Core design threats include differences in co-interventions, contamination across clinicians, unblinded outcome assessment, differential treatment intensity, attrition, switching or nonadherence, baseline imbalances, and selective reporting of favorable time points. Randomization reduces expected baseline confounding in its defined comparison; it cannot guarantee comparable measurement after dropout. Cluster randomization may be needed when clinicians deliver the intervention, but introduces service-level variation and statistical correlation. Research should report assignment procedures, assessor blinding where feasible, fidelity, missingness, predefined endpoints, adverse events, and confidence intervals Bickman et al. (2011) Kendrick et al. (2016).
Claims about individualized treatment selection additionally require a stronger counterfactual comparison: the outcome for each person under an alternative selection rule. It is possible for a model to predict risk accurately without delivering better clinical outcomes than standard practice. A prospective trial comparing care under model-supported selection with care under a clearly specified comparator estimates the utility of the whole decision system, not the independent causal correctness of every one of its predictions Cipriani et al. (2026) Vickers & Elkin (2006).
The evidence base is therefore interpreted in three layers. Instrument studies establish whether observations are usable. Trials and quasi-experiments estimate the effects of specific decision-support and care processes. Prediction studies examine whether models discriminate, calibrate, and generalize, while implementation studies explain whether the care process can be maintained in real clinical institutions. Each layer answers a different question and should remain separately legible in scientific synthesis.
Collaborative Care and Sequenced Treatment: Foundational Trials
The IMPACT trial of late-life depression in primary care examined an organized collaborative-care intervention rather than a simple questionnaire schedule. Unützer and colleagues (2002) randomized 1,801 older adults across participating primary-care clinics to intervention or usual care. The intervention combined depression care management, treatment monitoring, structured follow-up, and specialist consultation. At 12 months, the intervention group showed a substantially higher proportion with a reported depression response than usual care. This experiment is an important demonstration that continuous measurement can function within a coordinated care system Unützer et al. (2002). It should not be used to assign the entire benefit to scales alone.
IMPACT also shows why delivery architecture matters. If a symptom report reveals continuing distress, the resulting information must reach someone with responsibility and capacity to respond. Access to consultation, coordination between professionals, care manager follow-up, and the practical possibility of adjusting treatment can determine whether measurement produces any benefit. A health service with identical instruments but no response capacity is implementing a different causal intervention. Transfer of the IMPACT results to a resource-constrained setting requires explicit consideration of which supporting functions are available and what alternatives patients would otherwise receive.
The STAR*D program provides complementary longitudinal evidence about depression treatment sequences. Rush and colleagues (2006) reported acute and longer-term outcomes across successive steps of treatment among depressed outpatients who required one or several courses Rush et al. (2006). Its principal scientific contribution here is recognition that many initial treatments produce incomplete benefit and that subsequent options need systematic assessment. A successive-step design also highlights the changing population at each stage: people eligible for later steps are those who did not reach a satisfactory outcome, tolerated earlier care, and remained engaged.
An individualized rule cannot be derived by reading later-stage outcome proportions as though patients had been randomly assigned to every conceivable treatment history from the outset. Selection into subsequent stages and other design details affect causal interpretation. The work instead motivates research designs that randomize alternative next steps at multiple decision points, with relevant history measured before each decision. In this sense STAR*D helped articulate a scientific problem to which later adaptive-treatment methodology offers a formal experimental response Murphy (2005).
Together IMPACT and STAR*D illuminate the institutional and temporal dimensions of MBC: repeated assessment, accountable follow-up, and revision of treatment are essential problems for clinical science. They are foundational precedents, and their specific populations, interventions, and outcomes remain the proper limits of inference. Their relevance extends across therapy and pharmacotherapy, but no single cross-modality prescription can be inferred from them.
The Guo Trial: Algorithm-Guided Depression Treatment
Guo and colleagues (2015) directly compared a structured measurement-based approach with standard clinician-directed decisions for adults with major depression in Beijing Guo et al. (2015). Patients with moderate to severe depression were randomized to 24 weeks of care; 61 were assigned to measurement-based, guideline- and rating-scale-informed decisions and 59 to standard care. The investigators used blinded raters for outcome assessment, reducing some forms of differential outcome ascertainment. Both groups had pharmacotherapy limited to paroxetine or mirtazapine, an important constraint when generalizing the findings to contemporary, diverse treatment pathways.
The published trial reported response in 86.9% of the MBC group versus 62.7% in standard treatment and remission in 73.8% versus 28.8% at the study's assessment endpoint Guo et al. (2015). These differences favor the structured pathway under the tested conditions. Their magnitude warrants attention because the comparator still received active clinical treatment, yet estimates based on a study of 120 people are subject to sampling uncertainty and local workflow effects. Differences in treatment adjustments and contact can mediate the effect; the data do not show that the questionnaires themselves are therapeutic agents.
The trial addresses a clinically meaningful question: can systematic reassessment and decision rules overcome delays or variability in medication management? In this setting the answer favored the structured intervention. Interpretation must include the restricted medications, care context, participant characteristics, study duration, and real-world ability to reproduce the intervention with comparable fidelity. The finding does not automatically extend to complex multimorbidity, bipolar depression, psychotic conditions, child populations, or choices between medication and psychotherapy.
Unblinded clinicians and patients remain a potential source of performance effects even when outcome raters are blind. The trial's response and remission definitions derive from specified clinician-administered scales; alternative patient-important outcomes such as long-term functioning, adverse-effect burden, service utilization, or sustained recovery would complement the results. It is methodologically inappropriate to translate the reported percentages into a probability that an individual reader would benefit from MBC. Clinical decisions require patient-level assessment and current evidence across available options.
The Guo experiment establishes a concrete example in which linking repeated measurements to an explicit adjustment process improved a measured depression outcome. It also illustrates the importance of describing that process in enough detail to distinguish replication of the intervention from routine use of its questionnaires.
The Husain Trial: Earlier Response and the Limits of Endpoint Claims
A later randomized clinical trial in Pakistan investigated measurement-based antidepressant treatment for major depressive disorder. Husain and colleagues (2025) enrolled 154 adults, assigning 76 to MBC and 78 to standard care, with 24-week follow-up Husain et al. (2025). The intervention used repeated depression assessments to inform decisions; the control group received clinician-directed treatment without the same structured repeated clinical measurements. The setting matters for the field because much MBC evidence has emerged from higher-resource systems, whereas barriers to follow-up and specialist access may shape the value of a structured pathway.
The original report found median time to response of two weeks with MBC versus four weeks with standard care, and median time to remission of four versus eight weeks under the study's definitions Husain et al. (2025). At week 24, however, response and remission rates were not significantly different between groups. This distinction is fundamental. Earlier improvement may be important to patients, services, and functional recovery, but acceleration and durable superiority are separate hypotheses that require separate evidence. A trial should not be summarized as demonstrating long-term remission superiority when its final categorical endpoints do not support that conclusion.
The reported improvements should be examined in relation to the drugs available within the study, follow-up and treatment-intensity differences, outcome-rater methods, participant socioeconomic conditions, and handling of incomplete assessments. The study was corrected on July 14, 2026, for values in the final three rows of its baseline demographic table; that correction does not justify substituting unsupported baseline numbers into a secondary synthesis Husain et al. (2025). The scientific lesson is that source version control is part of evidence appraisal, particularly when manuscripts and online tables can change after original publication.
Clinically, shorter time to improvement and sustained recovery are distinct benefits. A pathway that identifies inadequate early response more quickly may reduce prolonged suffering even if cumulative outcomes converge later. Conversely, a more intensive monitoring system carries costs and may be burdensome; its net value depends on whether gains in time to response, patient experience, and safety justify those costs. Longer observation is required to determine whether early advantages change relapse rates, functioning, and health-service demand.
Taken alongside Guo and earlier research, the Husain trial supports contextual evaluation rather than a single universal effect size for MBC. Differences in comparator care, intervention definitions, populations, and outcome timing make a simple numerical average an incomplete basis for practice. The appropriate question is which components of the measurement-and-action pathway cause benefit, in which settings, at what cost and with what durability.
Systematic Reviews, Discordant Findings, and Evidence Integration
Evidence synthesis provides a check on the selective visibility of positive trials. Kendrick and colleagues (2016), in a Cochrane review of routine patient-reported outcome monitoring for common mental health disorders, identified 17 studies with 8,787 participants Kendrick et al. (2016). A pooled analysis of 12 studies using sufficiently comparable measures included 3,696 participants and found a standardized mean difference of −0.07 (95% confidence interval −0.16 to 0.01) for symptoms between feedback and no-feedback groups. The authors judged the evidence low quality, with widespread risk-of-bias concerns and limited information on functioning, adverse effects, and costs.
The Cochrane finding is a substantive warning against equating assessment feedback with clinical improvement. Trials differed in setting, intervention design, information delivered, and whether the feedback altered care. The review did not establish that all structured measurement-and-action systems are ineffective; it showed that the evidence available for its specified routine monitoring comparisons did not support confident general claims. An evidence-integrating monograph must preserve this uncertainty even when later positive trials address different implementations.
Zhu and colleagues (2021) subsequently conducted a systematic review and meta-analysis targeting MBC for depressive disorders, with a search through July 2020 Zhu et al. (2021). Differences between this review and the earlier Cochrane work arise partly from eligibility criteria and what counts as MBC: routine administration plus practitioner and patient review aimed at decisions is a narrower operationalization than generic PROM feedback. Such choices change the included experiments and estimands. Apparent disagreement between reviews may therefore reflect different interventions rather than a simple change in underlying scientific truth.
Evidence in youth and psychotherapy settings adds another dimension. Bickman and colleagues (2011) tested frequent routine feedback in community youth mental health services and reported improved trajectories relative to less frequent clinician feedback Bickman et al. (2011). The design supports research on information frequency and clinical adaptation, but its population, treatment type, and service structure restrict transfer to adult prescribing. Reviews of data-informed psychological therapy describe feedback systems and prediction of deterioration in terms that partly overlap with MBC Lutz et al. (2022). Session-level psychotherapy feedback remains an adjacent, distinguishable research subject.
Across this literature, an evidence-weighted synthesis supports the following conditional conclusion: structured measurement can improve selected outcomes when its results are timely, intelligible, discussed, and linked to feasible treatment decisions; collecting routine scores alone has not consistently demonstrated clinically important benefit. Stronger claims require prospective trials that isolate or explicitly describe the action pathway, use appropriate comparators, track safety and functioning, and examine durability. A cross-study label is never a substitute for identifying the intervention being tested.
Individualized Treatment Selection: Prognosis, Moderation, and Counterfactual Benefit
Personalized treatment selection begins from a specific causal question: how would a person's outcomes differ under two or more feasible treatments? For a clinical characteristic to justify selecting one therapy over another, it must help identify a difference in comparative outcomes, not simply predict recovery in the aggregate. Age, symptom severity, socioeconomic context, personality features, and past treatment history can all be prognostic without being prescriptive. Cohen and DeRubeis (2018) organize this distinction as a central problem for precision treatment in depression Cohen & DeRubeis (2018).
In the language of potential outcomes, each person has conceptually defined outcomes under each alternative intervention, although only the outcome under the received intervention is observed. Randomized comparisons can identify average effects under assumptions such as consistency, appropriate assignment, and adequate follow-up. Estimating individual treatment advantages requires additional modeling of how comparative effects vary with pre-treatment characteristics. These conditional estimates are uncertain, and an exact individual counterfactual cannot be directly verified. A model that generates two numerical forecasts is making an inferential claim requiring independent validation DeRubeis et al. (2014) Collins et al. (2024).
Effect modification can be investigated by interaction terms or more flexible models, but the search for many candidate moderators risks overfitting. A statistically significant predictor within one treatment arm does not establish a treatment-by-characteristic interaction. The relevant test compares outcomes across arms at comparable values of the proposed modifier, ideally from randomized data and with attention to multiplicity. Features selected after exploratory searching may have impressive apparent performance in the discovery sample and weak performance elsewhere. The resulting personalized recommendation can be unstable when small model changes reverse which therapy appears preferable.
Personalization also requires clear definition of the outcome used for optimization. A model minimizing symptom scores may make a different recommendation from one maximizing sustainable participation, quality of life, or patient preference. Treatment burden, adverse effects, and access determine what alternatives are practically feasible. An individualized estimate is therefore specific to the treatments and outcome functions included in the model. It should not be presented as a universal ranking of clinical care options Vickers & Elkin (2006) Cipriani et al. (2026).
The scientific aspiration is to move from average efficacy to evidence about differential advantage under explicit uncertainty. The appropriate standard is improved decision quality or patient outcomes in prospective comparisons, with evidence that predicted advantages remain calibrated across settings and population groups. Personalization becomes a clinical achievement only when choosing according to the model actually improves a consequential endpoint compared with a credible alternative decision process.
The Personalized Advantage Index and the Logic of Treatment Matching
The Personalized Advantage Index (PAI) introduced by DeRubeis and colleagues (2014) provides a concrete demonstration of a potential-outcomes approach to treatment matching DeRubeis et al. (2014). Using randomized trial data from 154 patients receiving antidepressant medication (104) or cognitive behavioral therapy (50), the researchers modeled expected post-treatment depression outcomes under each therapy. Five pre-randomization variables associated with differential response contributed to predictions. Each person's observations were excluded from the model generating their own treatment forecasts, an important step toward limiting in-sample optimism.
For a particular baseline profile, the PAI is the predicted difference between outcomes under the two treatment options. The option with the lower predicted depression score is the model-preferred treatment, while the magnitude of the predicted difference expresses the estimated advantage. This formalizes a question often addressed informally in clinical conversations: whether the person's pre-treatment features suggest one otherwise reasonable option may be preferable. Because the two original treatments had comparable average outcomes, identifying differential benefit would have practical significance if the pattern generalized.
The demonstration was retrospective with respect to treatment assignment and recommendation. It did not randomize new patients to receive or disregard the PAI recommendation and then compare outcomes. Cross-validation within a trial is more informative than fitting and reporting on the same observations, but it is still less demanding than external validation in independent systems, populations, and treatment protocols. Clinical utility would require proving that use of the PAI changes decisions appropriately, that recommended treatments can actually be obtained, and that patient-important outcomes improve DeRubeis et al. (2014) Cohen & DeRubeis (2018).
One methodological danger is conflating the predicted advantage with a causal individual treatment effect of known magnitude. Model estimates depend on measured covariates, the assumption that the trial treatment contrast applies to new patients, and the reliability of all interactions included. The numerical precision of the index may conceal substantial uncertainty from the original sample. If a person's predicted advantages are small relative to uncertainty, preference and accessibility could reasonably dominate the decision rather than being treated as secondary constraints.
The PAI remains historically influential because it made treatment matching mathematically explicit and experimentally testable. Its contribution is the translation of differential-response research into a prospective decision hypothesis, not proof that the identified variables define immutable treatment-responsive types. The next scientific step is independent replication and pragmatic prospective testing of policies that follow such recommendations.
Sequential Decisions, Dynamic Regimes, and SMART Designs
Treatment personalization is dynamic. A person who has not improved after an adequate trial, who experiences unacceptable adverse effects, or whose circumstances change may need a different choice than the one favored at baseline. The relevant history includes previous treatments, response trajectories, adherence or engagement, adverse events, changes in goals, and time since each decision. A dynamic treatment regime maps this evolving history to a rule for selecting an intervention at a sequence of occasions. Murphy (2005) developed experimental methods aimed at identifying such adaptive strategies Murphy (2005).
Sequential multiple assignment randomized trials, commonly called SMART designs, may randomize individuals at baseline and re-randomize specified subgroups at later decision points based on observed response or another protocol-defined criterion. This generates evidence about complete sequences rather than merely the isolated efficacy of a first-line option. It permits comparisons such as what follow-up strategy is most beneficial for early nonresponders after a particular initial intervention, under the trial's eligibility, randomization, and outcome assumptions Murphy (2005).
SMART studies require clear definitions of response, when it is measured, what history determines eligibility for re-randomization, and which adaptive strategies are embedded in the design. The definition of inadequate response can change the composition of later-stage groups and therefore the policies identified as promising. Delayed effects of previous treatment must also be considered: a new intervention may act alongside a treatment whose benefit has not fully emerged. Methodology must account for repeated assignments, accumulated exposure, and treatment histories rather than analyze a later-stage subgroup as if it were a new simple trial.
This experimental framework connects naturally with MBC. Repeated measures provide the intermediate information needed to define decision occasions and determine which branch of an adaptive strategy applies. But measurement frequency is itself a candidate intervention component: more frequent assessments may enable timely action, or may increase workload without a corresponding improvement. A well-designed study can investigate separate effects of measurement, clinical response rules, and available treatment options by making each intervention component explicit.
Adaptive regimes should be interpreted as studied decision policies, not automatic instructions transferable without clinical judgment. A protocol that is effective in a specific randomized sample may operate differently when patients have more comorbidity, lower access, different preferences, or less consistent follow-up. Prescriptive models for dynamic care must establish not only predictive performance but also safety, transportability, and actual patient benefit across sequences.
Clinical Prediction Models, Calibration, and Net Benefit
Clinical prediction models can estimate a future symptom level, risk of discontinuation, probability of response, or expected outcome under a specified treatment. Their performance must be evaluated in terms of the precise prediction task. Discrimination asks how well the model ranks individuals by outcome; calibration asks how closely predicted values agree with observed frequencies or mean outcomes. A model may discriminate well while systematically exaggerating risks, making it dangerous when absolute predicted values determine treatment changes. Prescriptive models add the requirement that differences across treatments be estimated credibly Collins et al. (2024) Cohen & DeRubeis (2018).
Internal cross-validation and bootstrap methods can estimate optimism caused by fitting to the development sample. Independent external validation evaluates transport to a different sample, institution, geography, time period, or care system. The latter is especially demanding in clinical psychology, where treatment availability, diagnostic practices, and patient expectations change. Models trained on selective trial populations may be poorly calibrated for people with multiple diagnoses, disabilities, socioeconomic barriers, or uncommon clinical presentations. Good average calibration may conceal subgroup errors important to equity.
TRIPOD+AI, published in 2024, provides updated reporting guidance for prediction models developed or evaluated using regression or machine learning Collins et al. (2024). Transparent reporting should include predictors, outcomes, eligibility, source population, missing-data handling, model development, evaluation, and limitations. Publication of model coefficients or a reproducible computational implementation is additionally valuable where feasible. A black-box system cannot acquire scientific credibility solely from a high headline accuracy score or from the use of artificial intelligence as a technical label.
Clinical utility is a separate criterion from predictive accuracy. Vickers and Elkin (2006) introduced decision-curve analysis, which evaluates a model's net benefit across relevant decision thresholds relative to simple strategies, under stated assumptions about the balance of true and false positives Vickers & Elkin (2006). In mental health selection, the analogous policy question is whether acting on a forecast produces more desirable consequences than treating everyone according to an established pathway or leaving choices to usual care. The decision consequence matters: a false positive for urgent referral differs from an unnecessary treatment switch in cost and harm.
Robust evaluation therefore has a hierarchy: define the target decision; validate the observations; estimate model performance and its uncertainty; test external transport; assess decision consequences; and conduct a prospective impact evaluation. Updating model parameters without re-establishing clinical validity can change recommendations in unpredictable ways. A prediction model remains an investigational component of care until its specified use has a defensible evidentiary and governance basis.
PETRUSHKA: Prospective Testing of Antidepressant Decision Support
The PETRUSHKA randomized clinical trial, reported by Cipriani and colleagues in 2026, offers a contemporary prospective test of individualized selection rather than a retrospective accuracy demonstration Cipriani et al. (2026). The intervention was a web-based antidepressant decision-support system combining clinical and demographic predictors with patient preferences. Recruitment took place in Brazil, Canada, and the United Kingdom. Of 540 randomized patients, 520 were eligible for the analysis population, and 493 contributed to the primary analysis. The comparator was clinician-selected usual care, making the estimand the effect of deploying the complete decision-support pathway.
The study's primary outcome was antidepressant discontinuation for any cause by eight weeks. Discontinuation occurred in 41 of 241 participants (17%) in the intervention group and 69 of 252 (27%) in usual care, with an adjusted relative risk of 0.62 (95% confidence interval 0.44–0.88) Cipriani et al. (2026). This is a clinically relevant result because early discontinuation can reflect treatment tolerability, effectiveness, acceptability, or burdens. The outcome nevertheless represents treatment continuation, not by itself proof of reduced depression burden or long-term superiority.
At 24 weeks, among participants with observed questionnaire data, mean PHQ-9 scores were 7.1 in the intervention arm and 9.2 in usual care, with an adjusted mean difference of −1.92 (95% confidence interval −3.06 to −0.78). The GAD-7 analysis also favored the personalized support arm Cipriani et al. (2026). The investigators reported important limits: the trial was not double-blind, and substantial missing data affected later symptom comparisons. The observed late differences should therefore be interpreted with sensitivity to loss to follow-up rather than as unrestricted evidence of durable benefit for the original randomized population.
The system included patient preference as part of the decision, highlighting that personalization involves more than a biological or diagnostic risk score. Selecting a medication acceptable to the person can plausibly reduce discontinuation, even when predicted differential pharmacological response is modest. The randomized result supports the clinical utility of that overall combined approach under tested conditions. It does not identify which component—preference elicitation, predictor estimates, shared discussion, or workflow changes—caused the observed improvement.
PETRUSHKA advances the scientific field by providing a prospective randomized outcome contrast for a specific decision-support tool. Its interpretation should remain bounded by participating countries, eligibility criteria, available treatments, implementation fidelity, endpoint definition, and missing-data uncertainty. Independent testing, longer follow-up, adverse-effect assessment, cost-effectiveness analysis, and external calibration across underserved populations remain central questions. The trial is a reason to develop rigorous prospective precision-care research, not to presume that every automated recommendation system will have the same effect.
Implementation Science: Fidelity, Workflow, and Institutional Capacity
An intervention that succeeds in a controlled trial must still become an ordinary, sustainable form of care. Implementation requires choosing measures, assigning responsibility for completion and review, incorporating results into documentation, supporting shared interpretation, ensuring access to effective responses, and maintaining confidentiality. Every step consumes attention and institutional resources. The systematic review by Burr and colleagues (2026) examined interventions intended to support MBC implementation in adult mental health services, identifying organizational conditions, professional barriers, and the continuing need for standardized reporting Burr et al. (2026).
The distinction between adoption and fidelity is especially important. A clinic may report that it uses validated symptom questionnaires while rarely reviewing results with patients or acting on concerning trends. Another may regularly use measures but lose data when patients change clinicians or service settings. Implementation should be assessed by such proximal indicators as the proportion of eligible patients offered monitoring, valid completed observations, timeliness of feedback, documented interpretation, recorded decisions, and continuity across transitions. These are necessary process metrics; they are not substitutes for assessing patient outcomes and harms Fortney et al. (2017) Resnick (2025).
Clinician skepticism may reflect competing demands, uncertain interpretation, fear of reducing complex experience to scores, or experience with poorly integrated electronic workflows. Patient concerns may involve confidentiality, fatigue, distrust, digital access, or the meaning assigned to scores. Dey and colleagues (2025) synthesized qualitative and quantitative experience research from patients and providers, underscoring the importance of examining acceptability directly Dey et al. (2025). An implementation program that increases documentation burden without generating clear clinical decisions can undermine the very engagement it intends to improve.
Training should therefore include interpretation of longitudinal change, the meaning and limitations of each instrument, management of discordant reports, safe escalation pathways, and shared decision-making. Electronic record interfaces should display trajectories with observation dates and reasonable context while minimizing duplicate data entry. A dashboard is useful when the treatment team knows who receives an alert, what response is expected, and how decisions are communicated back to the patient. Standardized tools should fit the local service's capabilities and populations; implementation fidelity is inseparable from a feasible action pathway.
Economic and organizational evaluations should measure the labor, infrastructure, and opportunity costs of the monitoring process alongside clinical benefit. A pathway that provides earlier detection but requires substantial additional staffing may be worthwhile under particular conditions, yet it may displace other beneficial services when resources are scarce. Cost-effectiveness cannot be inferred from clinical efficacy alone. The same intervention's opportunity costs differ across primary care, specialist psychiatry, community services, and low-resource settings. Implementation science should test adaptation strategies while preserving the causal functions that produced benefit in trials.
Safety, Differential Diagnosis, and the Limits of Algorithmic Escalation
Clinical safety requires monitoring domains beyond a general symptom score. Severe deterioration, suicidality, emergent mania or psychosis, substance-related problems, medication adverse effects, and functional collapse may demand immediate assessment outside a routine questionnaire schedule. A low score on one scale does not establish the absence of a serious problem, while a high score does not by itself determine which diagnosis or intervention is appropriate. NICE guidance on adult depression instructs clinicians to ask directly about suicidal ideation and intent and to respond to concerning risk through appropriate professional pathways NICE (2022/2026).
Decision-support recommendations can create harm through delayed intervention, excessive switching, misclassification, inappropriate treatment intensity, or false reassurance. An automated rule that reacts to a single increase in symptoms without investigating context may misread an acute life event; a system that ignores worsening because it falls below a threshold may fail to identify a developing problem. Risk management should include clinician assessment, explicit override mechanisms, documentation of exceptions, periodic review of false alerts, and prospective evaluation of adverse outcomes. These design principles follow from the mismatch between probabilistic scores and consequential clinical decisions.
Treatment selection may also be limited by incomplete diagnostic assessment. Depressive symptoms occur in multiple psychiatric, neurological, medical, and situational conditions. Previous treatment response, concomitant illness, adverse effects, safety, and the possibility of bipolar-spectrum symptoms can affect what intervention is appropriate. NICE's guidance specifies a comprehensive clinical assessment and reassessment when treatment has not produced expected improvement NICE (2022/2026). An individualized algorithm trained only on unipolar depression trial data cannot reliably resolve a different diagnostic problem merely by comparing forecast symptom scores.
A safe monitoring program must make time and responsibility visible. If a patient reports a concerning change between appointments, it should be clear when data will be reviewed, who is accountable, and what channels are available for urgent contact. An algorithm that displays a reassuring score while no human reviewer is available can generate a misleading sense of oversight. Research should explicitly report response delays and safety incidents, including consequences of failed data transmission or incomplete follow-up.
Clinical validation should distinguish aggregate improvement from a favorable benefit–harm profile. A study reporting fewer treatment discontinuations may still require examination of adverse effects, quality of life, autonomy, and later relapse; a trial reporting symptom reduction may not adequately capture rare harms. Safety-sensitive care needs qualified professional assessment, locally applicable guidance, and procedures for exceptions that statistical performance summaries cannot replace. This monograph provides research analysis rather than patient-specific medical instruction.
Equity, Cultural Context, Accessibility, and Generalizability
Measurement-based care is implemented within unequal systems of access. Standardized self-report tools can make previously unrecognized distress visible, but their reliability and interpretation can change with language, culture, literacy, disability, institutional trust, and social context. The absence of a translated or accessible instrument may exclude a person from the monitoring program altogether. Comparative evaluation must therefore examine who is eligible, who completes measures, who receives results, and who benefits from subsequent clinical action. A program that improves average outcomes while widening differences in follow-up or treatment access has an incomplete account of success.
The geographic range of evidence matters. The Guo trial in Beijing and the Husain trial in Pakistan add important information beyond commonly studied Western service settings Guo et al. (2015) Husain et al. (2025). PETRUSHKA extended a decision-support trial across three countries Cipriani et al. (2026). Yet multiple settings alone do not prove transport to people with different diagnoses, care pathways, languages, socioeconomic conditions, or insurance arrangements. For every trial, the population, comparator, available treatment options, and service infrastructure determine the relevant scope of inference.
Clinical prediction may reproduce data-generation inequalities. If training samples contain fewer people from certain groups, calibrated probabilities and predicted treatment advantages may be less reliable for them. Differences in who remains engaged with care can further bias model outputs, since observed outcomes are conditional on follow-up and treatment access. Disaggregated evaluation should inspect missingness, calibration, error distributions, recommended treatment patterns, and actual outcome differences across relevant populations. Equity analysis must be integrated with scientific validity rather than appended as a purely rhetorical goal Collins et al. (2024).
Accessibility is both a measurement question and an implementation question. Sensory disabilities, cognitive difficulties, fluctuating energy, limited reading proficiency, assistive technology compatibility, and privacy in the home may all affect valid completion. Alternative administration modes can improve participation but may alter response patterns and need evaluation. A service should document mode of administration and accommodate disability without equating noncompletion with noncooperation. In culturally diverse settings, interpretation of distress may require contextual dialogue in addition to standardized categories.
Real-world individualization includes the feasibility of the chosen treatment. When clinicians have access to only a small set of medications, long waiting lists for psychotherapy, or unreliable follow-up infrastructure, the best statistically predicted option may not be attainable. A scientifically defensible selector should compare actually available alternatives and make constraints visible. Social and institutional conditions are therefore elements of the decision model, not noise to be removed from an otherwise idealized psychological process.
Digital Decision Support and the Artificial Era
Digital systems can combine repeated questionnaires, timestamps, treatment histories, alerts, and prediction models into a longitudinal decision interface. This offers practical possibilities: automated reminders, visualization of changing outcomes, consistent scoring, distributed access for the care team, and timely recognition of concerning trajectories. Some systems may support estimation of outcomes under multiple options and the communication of predicted tradeoffs. Their scientific status depends on validation of each component and on clinical evaluation of the integrated workflow. A technically functioning dashboard has no automatic claim to therapeutic effectiveness Collins et al. (2024) Burr et al. (2026).
The emergence of artificial systems capable of generating explanations or treatment suggestions changes the informational environment in which clinical decisions occur. A generated recommendation can appear fluent and personalized while drawing on unreliable sources, incorrectly inferred patient facts, miscalibrated predictions, or an incomplete record. The more clinically consequential the decision, the greater the need for documented provenance, validated input, uncertainty representation, and professional review. Prediction assistance and delegated clinical authority are different operational arrangements with different safety requirements. The Artificial Era invites systematic research on how nonbiological decision-support processes become coupled to human assessment and responsibility.
Artificially mediated longitudinal care also raises specific methodological questions. Which measured indicators are stable enough to support algorithmic adaptation? How does model updating alter patient trajectories and trial comparators? How do users interpret uncertainty when explanations are machine-generated? Can clinicians detect a fabricated citation, an unjustified threshold, or an extrapolation outside the training population? These are empirical questions for controlled experiments and observational implementation research. No philosophical declaration about artificial systems substitutes for such evidence of diagnostic, prognostic, or therapeutic utility.
The PETRUSHKA study demonstrates how a digital tool combining predictors and preferences can be evaluated prospectively Cipriani et al. (2026). Its reported benefit attaches to the tested decision-support package and context, not to every digital product or to large language models in general. Machine learning methods need the same scrutiny of external validation, calibration, prospective impact, and subgroup performance as conventional regression models, as recognized by TRIPOD+AI Collins et al. (2024). A model's sophistication is scientifically relevant only when it improves measurement, prediction, explanation, or outcomes under valid comparison.
The horizon From Homo to Artificial is pertinent here as a research context for the changing organization of knowledge and clinical decisions. Psychological science must distinguish patient experience, clinician authority, institutional governance, and artificial information-processing capabilities while investigating their interaction. A reproducible audit trail from measure through recommendation to human decision would make this relationship more accountable. Whether any specific human–artificial care arrangement improves outcomes remains a question for prospective evaluation.
Critical Discussion: Explanatory Conflicts and Research Priorities
The central empirical tension is between favorable trials of structured measurement-and-action pathways and less convincing results for generic routine feedback. Guo and colleagues observed substantially improved depression response and remission under a restricted, algorithm-guided comparison; Husain and colleagues found earlier response without significant differences in final response and remission proportions; PETRUSHKA demonstrated better early treatment continuation under personalized decision support Guo et al. (2015) Husain et al. (2025) Cipriani et al. (2026). The Cochrane review of feedback interventions found no clear pooled symptom advantage and rated its evidence low quality Kendrick et al. (2016). These results concern overlapping but nonidentical interventions, populations, comparators, and endpoints.
One explanatory hypothesis is that measurement has clinical effect only when paired with timely, feasible, and appropriate action. This is consistent with the Collect, Share, Act formulation, but it demands direct factorial, dismantling, or implementation research before being treated as a uniquely established mechanism Resnick (2025). Other plausible contributors include increased clinical contact, adherence support, greater patient agency, clinician expectancy, faster access to effective interventions, and improved care coordination. Distinguishing them requires explicit measurement of mediators and experimental designs that vary components without compromising safety.
A second conflict concerns personalized prediction. The PAI demonstrates a method for estimating differential treatment advantage, yet its retrospective analysis does not establish prospective clinical benefit DeRubeis et al. (2014). PETRUSHKA supplies an important randomized comparison of a specific decision-support system, but the trial's missing late outcome data and multifactorial intervention limit causal attribution to any particular predictor Cipriani et al. (2026). Model discrimination, treatment-effect estimation, patient preference elicitation, and actual improvement must remain separate evidentiary questions.
Future research should preregister clinically meaningful decision policies, specify estimands, and investigate which measure–action intervals are optimal for different conditions and modalities. SMART designs can compare adaptive sequences; pragmatic cluster trials can test service-level implementation; mixed-methods research can examine why patients and clinicians do or do not act on feedback Murphy (2005) Dey et al. (2025). External validation and transport analyses should assess model and measurement performance across settings, and decision-curve approaches can clarify the value of treating at different risk thresholds Vickers & Elkin (2006). Evaluations should preserve complete follow-up denominators, adverse effects, functioning, quality of life, costs, and patient-defined outcomes.
Four unanswered questions organize the strongest agenda. What combination of measurement, interpretation, and clinical action produces sustained benefit, and under what institutional conditions? When does baseline or longitudinal information reliably identify differential rather than general treatment prognosis? How can outcome thresholds, predicted advantages, and uncertainty be communicated without displacing patient preferences or clinical safety? How can digital and artificial decision systems be assessed for transportability, fairness, and accountable human oversight? Answering these questions requires trials of decisions, not solely tests of prediction accuracy.
Conclusion
Measurement-Based Care and Individualized Treatment Selection form a linked scientific field concerned with how clinical evidence is produced and used across the course of treatment. Standardized measures can disclose change that unaided impressions may miss, while structured reassessment can create opportunities to revise an ineffective or unacceptable plan. The clinical value emerges from a complete process in which valid observations are interpreted with the patient and incorporated into feasible, accountable decisions. The measurement instrument, decision rule, and care system are distinct objects of evaluation.
Historical research established concepts of reliable change, scalable patient-reported measures, collaborative follow-up, and multi-step treatment. Randomized evidence now includes trials favoring structured measurement-based treatment in defined depression populations, a prospective international test of preference-sensitive antidepressant decision support, and studies showing that more generic feedback may yield uncertain or small effects. These findings justify continued development of specific, evaluated pathways while preserving the distinction between favorable local results and general effectiveness.
Individualized selection requires evidence about alternatives. Prognostic models characterize expected outcome, while prescriptive models attempt to estimate differential treatment benefit; the latter must be externally validated and tested as decision policies. A patient-centered choice also weighs tolerability, functioning, clinical safety, affordability, access, and personal priorities. Sequential adaptation extends this logic into time, making intermediate measurements scientifically consequential because they can change the next treatment decision.
The central conclusion is methodological and clinical: accountable treatment individualization is an evidence-guided decision process under uncertainty. Its success is established by improved patient-important consequences under defensible comparison, not by the volume of collected data or the complexity of an algorithm. The continuing task of psychological science is to make measurement reliable, decisions transparent, participation meaningful, and benefits reproducible across diverse people and services.
References
- Bickman, L., Kelley, S. D., Breda, C., de Andrade, A. R., & Riemer, M. (2011). Effects of routine feedback to clinicians on mental health outcomes of youths: Results of a randomized trial. Psychiatric Services, 62(12), 1423–1429. https://doi.org/10.1176/appi.ps.002052011
- Burr, C., Hahn, S., Carr, A. M., & Schneeberger, A. R. (2026). Implementation of measurement-based care in mental health services: A systematic literature review. Psychiatric Services, 77(9), 855–870. https://doi.org/10.1176/appi.ps.20250065
- Cella, D., Riley, W., Stone, A., et al. (2010). The Patient-Reported Outcomes Measurement Information System (PROMIS) developed and tested its first wave of adult self-reported health outcome item banks: 2005–2008. Journal of Clinical Epidemiology, 63(11), 1179–1194. https://doi.org/10.1016/j.jclinepi.2010.04.011
- Cipriani, A., Fernandes, K. B. P., Mulsant, B. H., et al. (2026). A decision-support system to personalize antidepressant treatment in major depressive disorder: A randomized clinical trial. JAMA, 335(14), 1219–1231. https://doi.org/10.1001/jama.2026.1327
- Cohen, Z. D., & DeRubeis, R. J. (2018). Treatment selection in depression. Annual Review of Clinical Psychology, 14, 209–236. https://doi.org/10.1146/annurev-clinpsy-050817-084746
- Collins, G. S., Moons, K. G. M., Dhiman, P., et al. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, e078378. https://doi.org/10.1136/bmj-2023-078378
- DeRubeis, R. J., Cohen, Z. D., Forand, N. R., Fournier, J. C., Gelfand, L. A., & Lorenzo-Luaces, L. (2014). The personalized advantage index: Translating research on prediction into individualized treatment recommendations. A demonstration. PLOS ONE, 9(1), e83875. https://doi.org/10.1371/journal.pone.0083875
- Dey, A., Lewis, Z., Posel, J., Pan, R. Y., & Wang, K. (2025). Quantifying care, qualifying experiences: A systematic review of measurement-based care in psychiatry from patient and provider perspectives. BMJ Mental Health, 28, e301663. https://doi.org/10.1136/bmjment-2025-301663
- Fortney, J. C., Unützer, J., Wrenn, G., Pyne, J. M., Smith, G. R., Schoenbaum, M., & Harbin, H. T. (2017). A tipping point for measurement-based care. Psychiatric Services, 68(2), 179–188. https://doi.org/10.1176/appi.ps.201500439
- Guo, T., Xiang, Y.-T., Xiao, L., et al. (2015). Measurement-based care versus standard care for major depression: A randomized controlled trial with blind raters. American Journal of Psychiatry, 172(10), 1004–1013. https://doi.org/10.1176/appi.ajp.2015.14050652
- Husain, M. I., Nigah, Z., Ansari, S. U. H., et al. (2025). Measurement-based care to enhance antidepressant treatment outcomes in major depressive disorder: A randomized clinical trial. JAMA Network Open, 8(9), e2529427. https://doi.org/10.1001/jamanetworkopen.2025.29427
- Jacobson, N. S., & Truax, P. (1991). Clinical significance: A statistical approach to defining meaningful change in psychotherapy research. Journal of Consulting and Clinical Psychology, 59(1), 12–19. https://doi.org/10.1037/0022-006X.59.1.12
- Kendrick, T., El-Gohary, M., Stuart, B., et al. (2016). Routine use of patient reported outcome measures (PROMs) for improving treatment of common mental health disorders in adults. Cochrane Database of Systematic Reviews, 2016(7), CD011119. https://doi.org/10.1002/14651858.CD011119.pub2
- Kroenke, K., Spitzer, R. L., & Williams, J. B. W. (2001). The PHQ-9: Validity of a brief depression severity measure. Journal of General Internal Medicine, 16(9), 606–613. https://doi.org/10.1046/j.1525-1497.2001.016009606.x
- Löwe, B., Unützer, J., Callahan, C. M., Perkins, A. J., & Kroenke, K. (2004). Monitoring depression treatment outcomes with the Patient Health Questionnaire-9. Medical Care, 42(12), 1194–1201. https://doi.org/10.1097/00005650-200412000-00006
- Lutz, W., Schwartz, B., & Delgadillo, J. (2022). Measurement-based and data-informed psychological therapy. Annual Review of Clinical Psychology, 18, 71–98. https://doi.org/10.1146/annurev-clinpsy-071720-014821
- Murphy, S. A. (2005). An experimental design for the development of adaptive treatment strategies. Statistics in Medicine, 24(10), 1455–1481. https://doi.org/10.1002/sim.2022
- National Institute for Health and Care Excellence. (2022; surveillance review 2026). Depression in adults: Treatment and management (NICE Guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/Recommendations
- Resnick, S. G. (2025). Measurement-based care: Opportunities to improve global mental health care. World Psychiatry, 24(2), 193–194. https://doi.org/10.1002/wps.21303
- Rush, A. J., Trivedi, M. H., Wisniewski, S. R., et al. (2006). Acute and longer-term outcomes in depressed outpatients requiring one or several treatment steps: A STAR*D report. American Journal of Psychiatry, 163(11), 1905–1917. https://doi.org/10.1176/ajp.2006.163.11.1905
- Spitzer, R. L., Kroenke, K., Williams, J. B. W., & Löwe, B. (2006). A brief measure for assessing generalized anxiety disorder: The GAD-7. Archives of Internal Medicine, 166(10), 1092–1097. https://doi.org/10.1001/archinte.166.10.1092
- Unützer, J., Katon, W., Callahan, C. M., et al. (2002). Collaborative care management of late-life depression in the primary care setting: A randomized controlled trial. JAMA, 288(22), 2836–2845. https://doi.org/10.1001/jama.288.22.2836
- Vickers, A. J., & Elkin, E. B. (2006). Decision curve analysis: A novel method for evaluating prediction models. Medical Decision Making, 26(6), 565–574. https://doi.org/10.1177/0272989X06295361
- Zhu, M., Hong, R. H., Yang, T., et al. (2021). The efficacy of measurement-based care for depressive disorders: Systematic review and meta-analysis of randomized controlled trials. Journal of Clinical Psychiatry, 82(5), 21r14034. https://doi.org/10.4088/JCP.21r14034