Archive \ Volume.17 2026 Issue 2

Counterfactual Medication Planning before Treatment Is Changed

Constructing Alternative Treatment Trajectories

An alternative treatment trajectory should begin with a strategy definition rather than a medication name alone. Continuing therapy may mean continuing the same dose unless toxicity occurs, whereas switching may involve tapering, overlap, washout, intensified monitoring, or rescue treatment. Per-protocol causal analyses demonstrate why sustained strategies, adherence, treatment changes, censoring, and time-varying conditions must be defined when outcomes are attributed to a treatment course [13]. In medication planning, these elements should be specified prospectively within the conceptual comparison, even when the available evidence cannot estimate every component reliably.

The trajectory object is proposed as a linked set of clinically meaningful processes. It should include expected therapeutic response, adverse effects, drug–drug and drug–disease concerns, treatment burden, adherence feasibility, monitoring requirements, and outcomes important to the patient. Each process may follow a different time scale. Symptom benefit may emerge quickly, laboratory toxicity may develop later, withdrawal effects may occur during transition, and long-term complications may remain poorly represented in available data. A trajectory should therefore retain timing and uncertainty instead of combining all consequences immediately into one score.

Sequential policy-learning research illustrates how treatment decisions can be related to evolving patient states. The Artificial Intelligence Clinician study derived time-dependent treatment strategies from retrospective intensive-care data, demonstrating the technical possibility of estimating policies over a clinical trajectory [14]. That example does not establish that retrospectively derived policies are safe or appropriate for medication planning in other settings. It instead shows why the state, action, timing, reward definition, and observed clinician behavior must be examined before a learned strategy is interpreted.

Healthcare reinforcement-learning guidance identifies additional risks, including confounding by indication, incomplete state representation, inappropriate reward functions, unstable policy evaluation, and unsafe extrapolation [15]. These concerns apply even when the proposed framework does not use reinforcement learning. Any method that ranks longitudinal strategies may favor actions that appear beneficial because of historical treatment-selection patterns rather than causal effects. Consequently, trajectory construction should be accompanied by defensible comparators, off-policy or external evaluation where relevant, safety constraints, and an explicit abstention route.

Target-trial emulation provides a complementary design logic. It requires articulation of eligibility, treatment strategies, assignment procedures, follow-up, outcomes, causal contrasts, and analysis. Federated target-trial emulation further shows that treatment-effect estimation may be coordinated across distributed data while preserving local data control, although site heterogeneity and data compatibility remain important [16]. For counterfactual medication planning, the relevant principle is not that every decision requires a formal federated study. It is that the target comparison should be designed before analytic estimation and should not be reconstructed opportunistically from whichever data are most convenient.

Alternative trajectories may be assembled from multiple evidence sources, but their contributions must remain distinguishable. Randomized evidence may support average comparative effects; observational evidence may inform underrepresented populations or longer-term outcomes; pharmacologic knowledge may constrain implausible paths; and patient reports may reveal burdens absent from structured data. The framework does not propose that these sources be blended into an unqualified composite. Their designs, populations, outcomes, and assumptions should remain visible within the trajectory record.

The proposed construction process ends with a trajectory dossier for each feasible strategy. The dossier contains the intervention definition, outcome vector, time horizon, data sources, causal estimand, assumptions, uncertainty classes, known evidence gaps, and conditions that would invalidate or restrict the comparison. It should also state which outcomes have not been modeled. This negative specification is important because omitted harms or burdens can disappear from a visual comparison despite remaining clinically relevant.

A counterfactual trajectory is therefore an evidence-qualified representation of a possible future, not an observation of the future that would have occurred. Its role is to discipline deliberation by making alternatives, assumptions, and missing information explicit. It should not convert uncertainty into false precision or substitute model ranking for a medication decision.

Assumptions, Confounding, Uncertainty, and Nonidentifiability

Counterfactual medication planning depends on assumptions that cannot be confirmed solely by model performance. Exchangeability requires that relevant determinants of treatment selection and outcome are measured adequately. Consistency requires each treatment strategy to correspond to a sufficiently defined intervention. Positivity requires patients with comparable clinical states to have realistic opportunities to receive the alternatives being compared. Temporal ordering, exposure measurement, adherence, interference, and outcome ascertainment also affect whether an estimated contrast can be interpreted causally.

Unmeasured confounding is particularly consequential when treatment changes are prompted by deterioration, toxicity, clinician concern, or patient preference. Sensitivity analysis can estimate how strongly an omitted variable would need to relate to treatment and outcome to alter an estimated association. Such analysis makes dependence on assumptions more interpretable, but it does not demonstrate that unmeasured confounding is absent [17]. The proposed assumption ledger should therefore record both the analytical adjustment performed and the clinically plausible factors that remain unavailable or poorly measured.

Individualized treatment-effect estimates require uncertainty representations that respond to weakened causal assumptions. Robust conformal approaches illustrate how intervals for individual effects may widen when hidden confounding is permitted [18]. This principle supports a distinction between a narrow prediction interval and a well-supported treatment comparison. Precision conditional on a strong assumption should not be displayed as unconditional certainty.

Observational effect estimates may also contain systematic error not represented by conventional confidence intervals. Empirical calibration using control outcomes demonstrates that nominal intervals can substantially understate uncertainty when residual biases remain [19]. Although population-level calibration does not directly solve individual-effect uncertainty, it reinforces the need to treat sampling variation as only one component of the uncertainty surrounding medication trajectories.

The framework consequently proposes an uncertainty envelope with six distinguishable domains: random outcome variability, model uncertainty, measurement uncertainty, causal-identification uncertainty, transportability uncertainty, and preference uncertainty. These domains should not be collapsed into one confidence score. A model may be statistically stable while the treatment contrast remains unidentified; conversely, a causally defensible average contrast may remain uncertain for a particular patient because relevant effect modifiers or preferences are poorly known.

The identifiability and comparability gate is proposed to convert these limitations into explicit dispositions. A comparison may proceed when strategies are coherent and assumptions are sufficiently defensible; proceed conditionally when uncertainty can be communicated and managed; be restricted to a narrower population, outcome, or horizon; trigger collection of additional information; or be deferred. Where material confounding, absent overlap, inconsistent treatment versions, or missing outcomes prevent a defensible contrast, the framework requires abstention from trajectory ranking.

Nonidentifiability is not equivalent to indecision. It is a substantive finding that available evidence cannot distinguish the alternatives adequately for the proposed use. Sensitivity analysis can reveal dependence on assumptions, but it cannot convert an unsupported comparison into an identified causal effect.

Evaluation and Safe Decision Integration

Evaluation should begin by separating model performance from performance of the complete medication-planning process. A trajectory model may be reproducible and calibrated yet fail because relevant medicines are missing, outputs arrive too late, professionals misunderstand uncertainty, patients cannot express preferences, or responsibility for acting on the comparison is unclear. Early clinical evaluation of artificial-intelligence decision support therefore requires examination of intended users, workflow, safety, human factors, and clinical context rather than technical accuracy alone [20].

Medication-specific evaluation must also distinguish system activity from meaningful benefit. Electronic drug–drug interaction alerts may alter clinician behavior, but evidence of patient-important benefit is limited and alert burden can undermine use [21]. Counterfactual planning should consequently be evaluated for prescribing errors, omitted harms, delayed treatment, inappropriate continuation or switching, monitoring failures, workload, override patterns, and consequences for patients—not simply for the frequency with which its preferred trajectory is selected.

Responsible clinical decision support further requires transparent intended use, accountable human oversight, monitoring, and governance across the system lifecycle [22]. Technical validation should therefore be followed by causal evaluation, external and temporal validation, silent prospective operation, human-factors testing, medication-safety assessment, equity and transportability analysis, and controlled evaluation of decision and patient outcomes. Progression between stages should remain conditional rather than automatic.

Safe integration therefore requires staged clinical evaluation, patient-important medication-safety assessment, and lifecycle governance rather than model-performance reporting alone [20-22]. No universal readiness score is proposed. A serious failure in causal validity, safety, equity, comprehension, accountability, or monitoring should not be offset by strong performance in another domain.

During use, the output should be presented as a qualified comparison of alternatives, accompanied by assumptions, excluded outcomes, uncertainty, and reasons for any abstention. The interface should separate evidence description from the authorized medication decision. Passing technical evaluation would establish neither clinical utility nor permission to alter treatment; progression would remain conditional on medication-safety, workflow, equity, and governance evidence.

Clinical and Governance Implications

Counterfactual medication planning distributes work across clinical, technical, organizational, and patient roles. Responsible healthcare machine learning requires stakeholder involvement from problem formulation through implementation and monitoring [23]. In this framework, that principle means involving patients and medication professionals when defining outcomes and burdens, not only after a model has been developed.

Pharmacists are positioned to verify medication histories, identify treatment-version ambiguity, assess interactions and adverse-effect pathways, evaluate adherence and access, and design monitoring. Prescribers integrate the comparison with diagnosis, examination, comorbidity management, and the broader therapeutic plan. Patients contribute preferences, lived treatment burden, acceptable uncertainty, and goals that cannot be inferred reliably from structured records. Technical teams remain responsible for data provenance, model behavior, reproducibility, change control, and failure investigation.

Real-world implementation is not equivalent to installing software. Integration studies demonstrate the importance of workflow redesign, role definition, training, stakeholder alignment, and adaptation to local clinical practice [24]. An organization considering the proposed framework would therefore need named clinical, medication-safety, technical, and governance owners; defined escalation routes; documentation standards; and authority to restrict, suspend, or withdraw use.

Trust should be calibrated rather than maximized. Clinician trust in artificial intelligence depends on the user, task, system characteristics, evidence, and organizational context [25]. Excessive trust may produce automation bias, whereas indiscriminate distrust may prevent useful scrutiny of alternatives. Interfaces and training should support appropriate reliance by showing the basis and limitations of each trajectory, not by presenting confidence as certainty or explanation as proof.

The framework proposes that trajectory generation, assumption review, medication-safety interpretation, patient deliberation, treatment authorization, and monitoring be assigned explicitly. These assignments organize responsibility but do not redefine legal liability or professional scope. Organizations remain accountable for determining whether the system’s intended use is compatible with local policies, staffing, information systems, patient population, and capacity for continuing oversight.

Limitations

The framework cannot guarantee that explanations correspond to a model’s true reasoning or establish causal validity, fairness, safety, or clinical usefulness [26]. Explanations may support scrutiny but can also create false reassurance.

Equity cannot be inferred from aggregate accuracy. Healthcare algorithms may reproduce inequity when proxy outcomes encode unequal access or resource use rather than underlying need [27]. Counterfactual medication planning therefore requires examination of representation, treatment availability, measurement, proxy selection, subgroup performance, and unequal consequences of error.

Transportability is similarly conditional. Clinical and machine-learning findings may not generalize across populations, institutions, formularies, workflows, or treatment practices [28]. Local validation does not remove every limitation, but absence of local evidence restricts interpretation.

Explanation, equity, and transportability are independent decision boundaries, and apparent strength in one domain cannot compensate for failure in another [26-28]. The framework also cannot observe both treatment futures in the same patient, eliminate unmeasured confounding, ensure stable preferences, or represent every rare and delayed consequence. Its architecture and transition rules remain proposed and require empirical evaluation.

CONCLUSION

Medication planning before treatment is changed requires comparison of plausible futures, not merely prediction of risk under observed care. The proposed Counterfactual Medication-Planning Framework organizes this task around a defined decision time, a verified patient–medication state, feasible treatment strategies, explicit estimands, multidimensional alternative trajectories, an assumption and confounding ledger, an uncertainty envelope, an identifiability and comparability gate, human deliberation, and monitored decision dispositions.

The framework’s principal contribution is the separation of processes that are often conflated. Predicting an outcome is not equivalent to estimating a treatment contrast. Constructing a trajectory is not equivalent to identifying its causal effect. Producing an interpretable output is not equivalent to establishing safety or usefulness. Implementing a system is not equivalent to demonstrating patient benefit, and automated analysis does not confer professional authority to change medication.

The framework also treats abstention, restriction, deferment, and information acquisition as legitimate planning outcomes. This is important where data are incomplete, treatment alternatives lack overlap, adherence is uncertain, outcomes are omitted, patient preferences are unstable, or the comparison is materially nonidentifiable. The framework is intended to expose these boundaries rather than conceal them behind a single ranking.

Validation would require evidence across causal methodology, technical performance, medication safety, human factors, patient outcomes, equity, transportability, implementation, and lifecycle governance. Responsibility must remain explicit across pharmacists, prescribers, patients, technical teams, and accountable organizations. Until such validation is completed in defined settings and uses, the framework should be interpreted as an original conceptual architecture for research, evaluation, and disciplined deliberation—not as a clinical recommendation, universal standard, or deployment-ready system.

ACKNOWLEDGMENTS: None

CONFLICT OF INTEREST: None

FINANCIAL SUPPORT: None

ETHICS STATEMENT: None