Clinical AI and Precision Medicine: Philosophical Questions About Labels, Disease, and Evidence Authors: Antonis A. Armoundas Published: 2026 Introduction Precision medicine is often framed as a move from “average-patient” care toward prevention, diagnosis, and treatment tailored to individual variability. This aspiration has been articulated in major policy and scientific calls for a new evidence-generating health system and a new taxonomy of disease [1, 2]. Yet, the limiting factors are not only technical; in practice, clinicians repeatedly operationalize contested notions, such as disease, risk, diagnosis, and evidence, as if they were stable clinical elements. As clinical care is increasingly captured in electronic data systems, including electronic health records, imaging platforms, wearable devices, and large language model outputs, these data elements increasingly define the targets used to train and evaluate clinical AI models [2]. Clinicians frequently encounter this problem: the threshold that turns a continuous variable into a diagnosis is rarely purely objective; it is a decision made under constraints, including time, resources, risk tolerance, and downstream harms. While the question is often framed as a binary classification one, for example, hypertension or not, atrial fibrillation or not, in practice, however, the classification matters because it determines the next clinical action, such as, labelling, treatment, reassurance, or intensified surveillance. Artificial intelligence (AI) does not dissolve those choices, but it can make the trade-offs visible, especially when labels built for coding are re-used for care. This Comment focuses on three linked problems in clinical AI: medical classification, disease concepts, and evidence. These are not only technical problems, but philosophical ones, because they concern how medicine defines categories, draws thresholds, justifies action, and decides what counts as sufficient evidence. AI can be useful here not as a replacement for clinical judgement, but as a tool that makes these assumptions explicit enough to examine, evaluate, and revise. In this context, philosophical analysis means the practical examination of concepts and value judgements that clinical AI systems turn into operational choices: what counts as a disease label, where a diagnostic threshold is drawn, whether a proxy endpoint adequately represents clinical need, and what kind of evidence is sufficient to support action. The practical implication is a shift from label-centered care toward decision-centered precision medicine; by this we mean an approach that is both uncertainty-aware and time-varying: diagnosis is treated as a revisable hypothesis, classification is designed around its intended clinical use, and evidence is treated as transparent and auditable reasoning. This problem is increasingly relevant because health systems are embedding AI outputs into triage, documentation, risk prediction, and care pathways. At the same time, regulatory and guideline processes are still adapting to evolving models, dataset shift, and the reporting of model uncertainty. Clinical AI evaluation therefore requires fit-for-purpose labels, longitudinal assessment, and auditable reasoning, because errors in classification or proxy specification can affect safety, equity, and clinical trust. This Comment examines the fidelity of labels and proxy endpoints: whether the clinical label, terminology mapping, or proxy endpoint used to develop an AI model faithfully represents the same clinical concept across settings, workflows, and patient groups. Thus, while existing AI evaluation frameworks appropriately emphasize reporting, risk of bias, validation, calibration, and monitoring, those evaluations are incomplete if the target label or proxy does not represent the intended clinical concept. Medical Classification is Evidential Infrastructure Medical classifications and terminologies (such as the International Classification of Diseases (ICD) for reporting and billing, [3] and the Systematised Nomenclature of Medicine Clinical Terms (SNOMED CT) for clinical representation [4]) function as harmonization and interoperability tools [5, 6]: they standardize communication, reimbursement, and surveillance, but they also embed purpose-specific assumptions about what counts as the “same” disease [3, 4]. This matters because the same label is routinely used for population reporting, bedside decision support, and research cohort definition, even though these tasks demand different levels of granularity, validity, and explanatory scope [7]. Clinical AI can quantify the mismatch between a label and its intended use. Supervised models trained on diagnosis codes inherit the structure and limitations of those codes. When labels are optimized for surveillance or billing, models may learn disease as encoded rather than disease as it is clinically observed. Multimodal and unstructured data, including clinical narratives, laboratory results, imaging, waveforms, wearable signals, and longitudinal care patterns, can preserve phenotypic context, clinician uncertainty, temporal evolution, and treatment response that coded labels may omit. However, these data also reflect documentation practices, copied-forward text, differential access to evaluation, and clinician judgement. Multimodal AI can reduce label compression but does not remove the need to define the clinical concept, document data provenance, and specify intended use. A further complication is semantic drift across settings. The same clinical presentation may be encoded differently depending on local workflows (ongoing condition registries versus encounter diagnoses), incentive structures, and availability of confirmatory tests. Mapping across terminologies can mitigate this, for example, aligning ICD-coded diagnoses with SNOMED CT clinical concepts and mappings in the Unified Medical Language System (UMLS), [8] but mapping does not resolve conceptual disagreements. It can even conceal them by creating a false sense of equivalence between a billing label and a mechanistic phenotype. For developers, this is not semantics: the choice of labels and mapping decisions can affect an AI algorithm’s performance, portability, and downstream equity [9]. These concerns are not limited to abstract terminology definition; they appear in routine clinical informatics tasks such as EHR phenotyping, risk prediction, triage, population health management, and multimodal modelling. Table 1 gives concrete examples of how ICD codes, SNOMED CT concepts, terminology mappings, thresholds, proxy outcomes, and multimodal data can shape what AI models learn and how their outputs should be checked before implementation.Table 1Clinical informatics examples of classification problems in AI implementationClinical informatics exampleWhat can go wrongPractical check before implementationEHR phenotyping using ICD codesICD codes may identify patients for billing or reporting but may not capture diagnostic certainty, disease severity, mechanism, or treatment responseCheck whether the ICD-based label matches the intended clinical concept, and test label stability across sites and workflowsClinical representation using SNOMED CTSNOMED CT may provide richer clinical detail than ICD, but the same concept can still be used differently across institutions or workflowsCheck whether the SNOMED CT concept is used consistently across sites and whether it corresponds to the clinical phenotype needed for the AI modelMapping between ICD and SNOMED CTMapping can improve interoperability, but it may create a false sense that a billing code and a clinical phenotype are equivalentDocument whether each mapping is exact, broader, narrower, or approximate, and verify that mapped definitions perform similarly across settingsRisk prediction modelsA model may predict a coded outcome accurately but fail to support the clinical decision, especially if required inputs, such as laboratory or imaging results, are unavailable at the time the decision must be madeDetermine whether all required inputs are available at the intended decision point, and specify what should happen when key inputs are missing, delayed, or unavailableTriage algorithmsThresholds may change who is flagged as high risk, monitored more closely, or escalated to a higher level of careRe-run the model at several thresholds (sensitivity analyses) and report how each threshold changes escalation rates, false positives, false negatives, and subgroup performancePopulation health algorithmsCost, utilization, or coding frequency may be used as a proxy label for clinical need, but these may reflect access to care rather than disease burdenCompare the proxy endpoints with more direct markers of need, such as disease severity, missed care, complications, or clinician-reviewed needMultimodal AI using notes, imaging, waveforms, or wearablesMultimodal data sources may preserve context lost in codes, but they can also reflect documentation habits, missing tests, copied-forward notes, or unequal access to devicesCheck missingness, data provenance, device access, documentation quality, and performance across sites and subgroupsExamples showing how ICD codes, SNOMED CT concepts, terminology mappings, thresholds, proxy outcomes, and multimodal data can shape what AI models learn and how their outputs should be checked before implementation. EHR: electronic health record; ICD: International Classification of Diseases; SNOMED CT: Systematised Nomenclature of Medicine Clinical Terms; AI: artificial intelligence Table 1Clinical informatics examples of classification problems in AI implementationClinical informatics exampleWhat can go wrongPractical check before implementationEHR phenotyping using ICD codesICD codes may identify patients for billing or reporting but may not capture diagnostic certainty, disease severity, mechanism, or treatment responseCheck whether the ICD-based label matches the intended clinical concept, and test label stability across sites and workflowsClinical representation using SNOMED CTSNOMED CT may provide richer clinical detail than ICD, but the same concept can still be used differently across institutions or workflowsCheck whether the SNOMED CT concept is used consistently across sites and whether it corresponds to the clinical phenotype needed for the AI modelMapping between ICD and SNOMED CTMapping can improve interoperability, but it may create a false sense that a billing code and a clinical phenotype are equivalentDocument whether each mapping is exact, broader, narrower, or approximate, and verify that mapped definitions perform similarly across settingsRisk prediction modelsA model may predict a coded outcome accurately but fail to support the clinical decision, especially if required inputs, such as laboratory or imaging results, are unavailable at the time the decision must be madeDetermine whether all required inputs are available at the intended decision point, and specify what should happen when key inputs are missing, delayed, or unavailableTriage algorithmsThresholds may change who is flagged as high risk, monitored more closely, or escalated to a higher level of careRe-run the model at several thresholds (sensitivity analyses) and report how each threshold changes escalation rates, false positives, false negatives, and subgroup performancePopulation health algorithmsCost, utilization, or coding frequency may be used as a proxy label for clinical need, but these may reflect access to care rather than disease burdenCompare the proxy endpoints with more direct markers of need, such as disease severity, missed care, complications, or clinician-reviewed needMultimodal AI using notes, imaging, waveforms, or wearablesMultimodal data sources may preserve context lost in codes, but they can also reflect documentation habits, missing tests, copied-forward notes, or unequal access to devicesCheck missingness, data provenance, device access, documentation quality, and performance across sites and subgroupsExamples showing how ICD codes, SNOMED CT concepts, terminology mappings, thresholds, proxy outcomes, and multimodal data can shape what AI models learn and how their outputs should be checked before implementation. EHR: electronic health record; ICD: International Classification of Diseases; SNOMED CT: Systematised Nomenclature of Medicine Clinical Terms; AI: artificial intelligence Clinical informatics examples of classification problems in AI implementation Examples showing how ICD codes, SNOMED CT concepts, terminology mappings, thresholds, proxy outcomes, and multimodal data can shape what AI models learn and how their outputs should be checked before implementation. EHR: electronic health record; ICD: International Classification of Diseases; SNOMED CT: Systematised Nomenclature of Medicine Clinical Terms; AI: artificial intelligence Disease as a Continuum: Thresholds, Values, and What AI Can Clarify Philosophers of medicine have long debated whether disease can be defined purely biologically (naturalism) or whether it necessarily includes evaluative components. Boorse’s naturalist account centers disease on dysfunction relative to species-typical functioning, whereas Wakefield’s “harmful dysfunction” analysis treats disorder as hybrid: dysfunction plus harm [10, 11]. These debates surface in daily clinical work because many diagnoses that manifest themselves as categorical labels are derived from continuous, quantitative measures, such as blood pressure, glucose, ejection fraction, bone density, and symptom scales. However, thresholds are unavoidable because clinical action is discrete. The philosophical and clinical question is whether a threshold supports clinical decisions or masks clinically important heterogeneity. Jutel’s history of overweight and obesity illustrates how boundaries can shift over time, with downstream clinical and social consequences [7]. Three evaluation strategies make threshold effects measurable. First, threshold sensitivity analyses can show how changing a cut-off affects diagnosis, treatment eligibility, false positives, false negatives, and subgroup performance. Second, multimodal and longitudinal models can identify heterogeneity within a labelled condition, for example, patients who share a diagnosis code but differ in mechanism, trajectory, or likely treatment response. Third, decision-analytic methods can evaluate whether an alternative threshold or representation improves clinically relevant choices, such as when to monitor, treat, escalate therapy, or allocate preventive resources. In this sense, AI does not remove the need for judgement; it makes the consequences of classification-based judgement visible. Beyond Classification: Continuous, Multidimensional Disease Representations And Continuous Disease Assessment If disease is multidimensional and continuous, categorical diagnoses may be insufficient representations for precision medicine. Petersen and Ursin argue that forcing complex states into diagnostic classes creates an “information bottleneck” that discards clinically meaningful variation and can propagate historical and social biases when treated as ground truth for model development [12]. They propose “continuous disease assessment”: modelling severity, progression, and treatment response as evolving estimates in physiological space, with a diagnosis being a revisable hypothesis rather than an endpoint. Longitudinal modelling, multimodal representation learning, and time-series prediction can estimate latent states and trajectories from heterogeneous data streams, including laboratory results, imaging, wearable signals, and clinical text. The key philosophical move is to treat disease as state estimation under uncertainty, not merely an association with a class of codes, which is closer to how clinicians update their beliefs over time and is more aligned with the aims of precision medicine. Continuous assessment is most useful when linked to action. A patient-specific trajectory estimate can inform the monitoring interval, preventive threshold, or timing of treatment escalation, which are usually implicit in clinical reasoning [13]. Displaying the trajectory with uncertainty also specifies the evidence that would support continuation, escalation, or revision of the clinical plan. Conceptual Engineering: When AI Motivates Revising Disease Concepts Precision medicine requires disease concepts that support explanation and intervention. Network medicine provides one basis for this shift by treating disease phenotypes as products of interacting biological networks and by defining disease modules and pathway-level organization [14, 15]. AI can operationalize these concepts by identifying subgroups with shared trajectories or treatment responses, by learning representations that capture comorbidity structure, and by evaluating whether alternative classifications improve prediction or clinical decision-making. AI should therefore be used to evaluate purpose-specific disease representations, not to define a single universal disease taxonomy. Different representations may be appropriate for surveillance, reimbursement, mechanism discovery, treatment selection, prognosis, or preventive resource allocation. Evidence is Itself a Classification Problem and AI May Clarify Disagreement and Uncertainty Evidence-based medicine aims to ground clinical decisions in the best available evidence, but evidence requires appraisal and interpretation. It is produced and ranked through appraisal methods and tools, bias frameworks, [16, 17] and judgements about relevance and applicability [18]. Empirical work shows that different appraisal tools applied to the same studies can yield divergent judgements, and even within the same tool, agreement can be limited [19]. Contemporary frameworks seek to standardize these judgements. GRADE (Grading of Recommendations, Assessment, Development and Evaluations) provides a structured approach for rating certainty of evidence and strength of recommendations [18]. RoB 2 (the revised Cochrane Risk of Bias tool for randomized trials) operationalizes bias assessment across domains such as randomization, deviations from intended interventions, missing data, outcome measurement, and selective reporting [20]. PROBAST (Prediction model Risk Of Bias ASsessment Tool) provides analogous structure for prediction model studies, focusing on risk of bias and applicability, [21] and PROBAST + AI is especially relevant when AI prediction models are used for enrichment, prescreening, endpoint support, or risk forecasting [22]. AI tools can support evidence appraisal by extracting study features, enabling sensitivity analyses, and making uncertainty explicit. They can also embed appraisal rules in workflows that are difficult to scrutinize. The reproducibility of the model’s output should be measured when LLMs are used for evidence screening, summarization, or quality appraisal; the stability of the output may vary with its version, the structure of the prompt, the sources of evidence, the model settings, the date of use, the inference environment, and across repeated runs. LLM-assisted appraisal should therefore report these conditions, measure the model’s repeated-run stability, and include human adjudication when outputs inform evidence judgments [23]. The strength of evidence for clinical AI remains uneven. Many systems are supported primarily by retrospective or observational validation, with limited prospective evaluation of how predictions change management or improve outcomes. This concern is illustrated by reviews in perioperative [24] and pediatric anesthesia [25] as well as in cardiology [17], which have identified limitations related to risk of bias, limited external validation, [26] and uncertain generalizability as barriers to clinical integration of AI algorithms. A GRADE-style perspective is helpful here: confidence should be reduced when studies have high risk of bias, when external validity is uncertain, or when outcomes are surrogates rather than patient-important endpoints [18]. AI tools may support the evaluation and accountability of evidence, but they do not eliminate the need for explicit intended-use statements, prespecified evaluation plans, and, where feasible, prospective assessment of clinical impact. Diagnosis, causality, and endpoints Diagnostic reasoning is not simply label assignment. Stanley and Hanna describe it as an abductive hypothesis generation and refutation: clinicians propose explanations, seek discriminating evidence, revise as data accumulate, and treat diagnoses as provisional and corrigible [27]. Because diagnosis is often revised as new evidence emerges, clinical AI should report uncertainty, support updating with new data, and align predictions with the patient’s course over time, rather than provide only one-time diagnostic accuracy. Precision medicine requires more than prediction; it requires actionable understanding about which interventions will help which patients [28]. Philosophically, this is a causal requirement. Russo and Williamson argue that causal claims in health science typically rely on both mechanistic evidence and difference-making (probabilistic) evidence, and that neither alone suffices [28]. Causal inference frameworks (for example, counterfactual reasoning as developed in epidemiology) formalize the distinction between association and intervention effects and clarify the assumptions needed to translate observational data into causal claims [29]. However, a central vulnerability arises when the AI model endpoints are proxies. Obermeyer and colleagues showed that using future healthcare costs as a proxy for health need can produce substantial racial bias because costs reflect differential access and spending rather than morbidity [30]. Therefore, a model can be accurate for the measured endpoint while still misclassifying clinical need if the endpoint, data source, and subgroup performance are not examined directly. Endpoint specification should therefore be treated as a primary methodological step. Developers should define the clinical concept of interest, justify any proxy endpoint, and evaluate whether the proxy behaves similarly across subgroups and settings. When differences reflect structural inequities, such as differential access to digital tools or care utilization, [31] statistical adjustment alone is insufficient. Measurement and deployment should be redesigned so that model outputs reflect clinical need rather than access or utilization [31]. What this Means for Precision Medicine Taken together, these arguments point toward decision-centered precision medicine. First, clinical work shifts from label correctness to trajectory estimation and action under uncertainty. Second, classification becomes explicitly designed for purpose, such as surveillance, reimbursement, decision support, or treatment selection, with mappings between representations rather than a single universal taxonomy. Third, evidence becomes an auditable process in which assumptions, endpoint definitions, and bias controls must be visible if AI models are to earn clinical trust. This implementation approach can be operationalized by model developers as well as clinicians and health systems. For model developers, the first step is to define the clinical decision the model is meant to support: diagnosis, triage, monitoring, treatment escalation, prevention, or resource allocation. The second is to define the clinical concept of interest and state whether the model uses a direct measure or a proxy label. If a proxy label is used, such as healthcare cost, utilization, coding frequency, or clinician documentation, developers should explain why it is appropriate and evaluate whether it behaves similarly across sites and patient groups. Third, developers should evaluate whether the model’s labels, inputs, and outputs remain stable across settings, including different electronic health record systems, coding practices, workflows, and patient populations. Fourth, evaluation should include decision-relevant outcomes, not label accuracy alone: whether the model changes care, improves calibration of risk, reduces missed or delayed diagnosis, supports safer treatment decisions, or improves equity (Table 2).Table 2Practical questions for implementing clinical AI in precision medicineStep (User)Actionable question (s)What is requiredDefine the decision (Developer)What clinical decision is the model meant to support?• State the intended use clearly: diagnosis, triage, monitoring, treatment escalation, prevention, or resource allocationVerify label (endpoint)Appropriateness (Developer)Does the label match the intended clinical concept?• Define the clinical concept, endpoint, and any proxy label used• Explain how the label was defined and whether it reflects the clinical state of interestJustify proxy labels (Developer)Is the endpoint direct, or is there a proxy label being used? Does the proxy label behave the same across groups?• Justify proxy labels such as cost, utilization, coding frequency, or documentation• Test performance across sites, patient subgroups, workflows, and coding practicesAssess coding drift (Developer)Does coding vary across sites or workflows?• Evaluate whether labels are stable across electronic health record systems, institutions, and clinical settingsEvaluate subgroup validity (Clinician or Clinical System)Does the model perform similarly across groups?• Require external/local validation in a comparable population and workflow• Report calibration, errors, and usefulness across relevant patient subgroupsExamine thresholds (Clinician or Clinical System)Do thresholds change who is diagnosed or treated? Can clinicians use the uncertainty?• Perform threshold sensitivity analyses and report effects on treatment eligibility, escalation rates, false positives, false negatives, and subgroup performanceEvaluate decisions (Clinician or Clinical System)Does the model improve care? Are model outputs reproducible?• Measure decision-relevant outcomes, including changes in management, safety, equity, workflow burden, and patient-important outcomes. For tools that use LLMs, report model version, prompt structure, source documents, repeated-run stability, and human reviewMonitor after deployment (Clinician or Clinical System)What happens after deployment? Does performance remain safe over time?• Monitor calibration, errors, workflow effects, subgroup performance, and patient-important outcomes over timeKey questions that model developers, clinicians, and health systems may use before and after implementation to assess whether AI tools are aligned with the intended clinical decision, evaluated across settings and patient groups, and monitored for safety, equity, and usefulness over time Table 2Practical questions for implementing clinical AI in precision medicineStep (User)Actionable question (s)What is requiredDefine the decision (Developer)What clinical decision is the model meant to support?• State the intended use clearly: diagnosis, triage, monitoring, treatment escalation, prevention, or resource allocationVerify label (endpoint)Appropriateness (Developer)Does the label match the intended clinical concept?• Define the clinical concept, endpoint, and any proxy label used• Explain how the label was defined and whether it reflects the clinical state of interestJustify proxy labels (Developer)Is the endpoint direct, or is there a proxy label being used? Does the proxy label behave the same across groups?• Justify proxy labels such as cost, utilization, coding frequency, or documentation• Test performance across sites, patient subgroups, workflows, and coding practicesAssess coding drift (Developer)Does coding vary across sites or workflows?• Evaluate whether labels are stable across electronic health record systems, institutions, and clinical settingsEvaluate subgroup validity (Clinician or Clinical System)Does the model perform similarly across groups?• Require external/local validation in a comparable population and workflow• Report calibration, errors, and usefulness across relevant patient subgroupsExamine thresholds (Clinician or Clinical System)Do thresholds change who is diagnosed or treated? Can clinicians use the uncertainty?• Perform threshold sensitivity analyses and report effects on treatment eligibility, escalation rates, false positives, false negatives, and subgroup performanceEvaluate decisions (Clinician or Clinical System)Does the model improve care? Are model outputs reproducible?• Measure decision-relevant outcomes, including changes in management, safety, equity, workflow burden, and patient-important outcomes. For tools that use LLMs, report model version, prompt structure, source documents, repeated-run stability, and human reviewMonitor after deployment (Clinician or Clinical System)What happens after deployment? Does performance remain safe over time?• Monitor calibration, errors, workflow effects, subgroup performance, and patient-important outcomes over timeKey questions that model developers, clinicians, and health systems may use before and after implementation to assess whether AI tools are aligned with the intended clinical decision, evaluated across settings and patient groups, and monitored for safety, equity, and usefulness over time Practical questions for implementing clinical AI in precision medicine Verify label (endpoint) Appropriateness (Developer) • Define the clinical concept, endpoint, and any proxy label used • Explain how the label was defined and whether it reflects the clinical state of interest • Justify proxy labels such as cost, utilization, coding frequency, or documentation • Test performance across sites, patient subgroups, workflows, and coding practices • Require external/local validation in a comparable population and workflow • Report calibration, errors, and usefulness across relevant patient subgroups Key questions that model developers, clinicians, and health systems may use before and after implementation to assess whether AI tools are aligned with the intended clinical decision, evaluated across settings and patient groups, and monitored for safety, equity, and usefulness over time For clinicians and health systems that select AI tools, the same questions can guide implementation. Before adoption, they should ask what decision the tool is designed to support, what label or endpoint was used during development, whether the tool was externally validated in a population similar to their own, whether subgroup performance was reported, and whether uncertainty is shown in a form clinicians can use. For tools that use LLMs, they should also ask whether the model version, prompt structure, source document set, repeated-run stability, and human review process were reported [23]. After deployment, health systems should monitor calibration, errors, subgroup performance, workflow effects, and patient-important outcomes over time [23]. In this view, implementation is not the final step after model development; it is part of the evidence process itself (Table 2). This approach complements precision prevention efforts that emphasize which patients should receive preventive resources or intensified follow-up, which interventions should be selected, and how equity should be protected. Conclusion AI can contribute to questions about medical classification, disease, and evidence by turning conceptual ambiguities into testable and governable design choices. It can quantify what is lost when physiology is compressed into codes, support continuous state and trajectory estimation, and make evidence appraisal more transparent. For precision medicine, clinical AI should be developed and evaluated using explicit clinical concepts, auditable evidence pipelines, and purpose-specific representations of disease (Fig. 1).Fig. 1AI as an instrument for philosophical inquiry in medicine and a driver of decision-centered precision care AI as an instrument for philosophical inquiry in medicine and a driver of decision-centered precision care Key Messages • Medical labels and terminologies are purpose-built infrastructure; AI trained on them learns disease-as-encoded and can quantify clinically important information loss.• Many conditions behave as continuous, multidimensional states; AI enables longitudinal state and trajectory estimation that better supports precision decisions than discrete classes of disease.• Evidence appraisal is itself a classification problem; AI can improve transparency and sensitivity analyses, but appraisal rules and endpoints must remain explicit and contestable.• Precision medicine should evaluate AI model performance against decision-relevant outcomes and equity-sensitive endpoints, not label accuracy alone. • Medical labels and terminologies are purpose-built infrastructure; AI trained on them learns disease-as-encoded and can quantify clinically important information loss. • Medical labels and terminologies are purpose-built infrastructure; AI trained on them learns disease-as-encoded and can quantify clinically important information loss. • Many conditions behave as continuous, multidimensional states; AI enables longitudinal state and trajectory estimation that better supports precision decisions than discrete classes of disease. • Many conditions behave as continuous, multidimensional states; AI enables longitudinal state and trajectory estimation that better supports precision decisions than discrete classes of disease. • Evidence appraisal is itself a classification problem; AI can improve transparency and sensitivity analyses, but appraisal rules and endpoints must remain explicit and contestable. • Evidence appraisal is itself a classification problem; AI can improve transparency and sensitivity analyses, but appraisal rules and endpoints must remain explicit and contestable. • Precision medicine should evaluate AI model performance against decision-relevant outcomes and equity-sensitive endpoints, not label accuracy alone. • Precision medicine should evaluate AI model performance against decision-relevant outcomes and equity-sensitive endpoints, not label accuracy alone.