Classification Without Reduction: The Epistemology of Clinical Taxonomy Examined Through TCM Constitution Theory

ABSTRACT

Rationale, Aims and Objectives

Clinical classifications operate at different explanatory levels—from molecular subtypes to phenotypic syndromes—yet the epistemological standing of a classification is often treated as though it depends on which level it occupies. This conflation of explanatory level with epistemological standing is unfounded. This essay identifies the minimal epistemological conditions that any clinical classification must satisfy and examines whether these conditions depend on explanatory level.

Methods

Philosophical analysis drawing on natural kinds theory (Boyd's homoeostatic property cluster theory), operationalism (Bridgman) and interventionist causation (Woodward). Traditional Chinese Medicine (TCM) constitution theory, developed by Wang Qi over four decades, is used as a developing top‐down case study whose three core tenets—classifiable, measurable and modifiable—map onto the three philosophical frameworks.

Results

Three conditions were identified as necessary for a clinical classification to be scientifically tractable: ontologically grounded categories, operationalized measurement and testable causal propositions. These conditions are level‐independent: they can be met by bottom‐up molecular taxonomies and top‐down phenotypic classifications alike. The TCM constitution case demonstrates uneven progress across the three conditions—operationalization is relatively advanced, ontological grounding has partial empirical support and causal validation remains nascent—illustrating the framework's utility as a diagnostic tool for assessing research programme maturity.

Conclusion

The epistemological standing of a clinical classification depends not on the explanatory level at which it operates, but on whether it satisfies the three identified conditions. This analysis is consistent with explanatory pluralism: molecular investigation and phenotypic classification address different questions along different axes, and evidence‐based medicine's evaluative framework applies to both without prejudice.

Plain Language Summary

In medicine, patients are grouped into categories—disease types, risk groups or constitutional types—to guide diagnosis and treatment. Some of these categories are defined by molecular tests (such as genetic subtypes of cancer), while others are defined by observable symptoms and characteristics (such as psychiatric diagnoses or body constitution types in Chinese medicine). There is a widespread assumption that molecular categories are more scientifically valid than categories based on observable features. This essay argues that this assumption is mistaken. It identifies three requirements that any clinical classification must meet to be scientifically useful: the categories must reflect real patterns in patients (not just observer convenience), they must be measurable in a standardized way and it must be possible to test whether acting on the classification improves health outcomes. These requirements apply equally regardless of whether the classification is molecular or observational. The essay uses Chinese medicine's constitution classification system as an example of a developing observational system, showing where it has made progress and where important work remains. The analysis suggests that clinical classifications should be judged by whether they meet these three requirements, not by how close they are to molecular biology.

1. Introduction

Clinical classifications operate at different explanatory levels. Some begin at the molecular level and work upward: breast cancer's molecular subtypes—Luminal A, Luminal B, HER2‐positive, triple‐negative—were defined through gene expression profiling and subsequently validated against clinical outcomes [1]. level and work downward: psychiatric diagnoses were defined by symptom clusters long before any molecular correlates were sought, and many remain defined at that level today [2, 3]. Both directions have produced clinically useful classifications. Both have also produced classifications that remain contested.

Yet there is a pervasive assumption—rarely stated explicitly, but evident in how classifications are evaluated—that a classification's epistemological standing depends on the explanatory level at which it operates. Molecular subtypes are treated as more ‘real’ or more ‘scientific’ than phenomenological categories, as though proximity to molecular mechanisms were itself a mark of epistemic rigour. This assumption conflates two distinct questions: the question of explanatory level (at which level of biological organization does the classification operate?) and the question of epistemological standing (does the classification satisfy the conditions required for scientific tractability?).

This essay argues that these two questions are orthogonal. To make this argument, it identifies the minimal epistemological conditions that any clinical classification must satisfy—ontologically grounded categories, operationalized measurement and testable causal propositions—and shows that these conditions are level‐independent. A phenotypic classification that satisfies all three has the same epistemological standing as a molecular one that does so, and a molecular classification that fails to satisfy them is no more scientifically tractable for being molecular.

The three‐stage structure—classification, operationalization, causal testability—is not new. It is implicit in clinical research practice and can be recognized in the logic of evidence‐based medicine (EBM). What has been lacking is an explicit philosophical analysis that grounds each stage in established epistemological frameworks (natural kinds theory, operationalism, interventionist causation) and that identifies the conditions as level‐independent. The present essay undertakes that analysis.

Robins and Guze [4] addressed a related question in 1970, proposing five phases for establishing diagnostic validity in psychiatry: clinical description, laboratory study, exclusion of other disorders, follow‐up study and family study. Their framework has been enormously influential and remains the standard reference for diagnostic validation [5]. However, it operates at the methodological level—specifying what kinds of evidence to gather—rather than at the epistemological level—identifying what philosophical conditions a classification must satisfy to be scientifically tractable. The present analysis complements Robins and Guze by asking a prior question: what are the epistemological conditions that make diagnostic validation possible in the first place?

To examine these conditions, it is useful to have a case in which the three stages can be observed in explicit, temporally separated form. In many areas of clinical science, the stages are compressed—classification, measurement and causal validation develop in close succession, making it difficult to distinguish them as independent requirements. A case in which the stages are temporally extended and independently documented allows each to be examined on its own terms.

TCM constitution theory, developed by Wang Qi beginning in 1978 [6], provides such a case. This theory classifies individual constitution into nine types based on phenotypic characteristics and has undergone four decades of development—from initial taxonomy through standardized measurement to population‐level survey and ongoing causal investigation. The interest of this case lies not in its medical content per se, but in its structural properties: it is a top‐down classification system whose conceptual framework developed independently within TCM's intellectual lineage, while employing shared scientific methodology (psychometrics, epidemiological survey design, metabolomic profiling, statistical inference) to formalize and test that framework. Crucially, the research programme has reached different stages of maturity with respect to each of the three epistemological conditions—operationalization is relatively advanced, ontological grounding has partial empirical support, and causal testability remains largely at the hypothesis stage. This unevenness makes the case analytically informative: it allows the three conditions to be examined not only as abstract requirements but as a diagnostic tool for assessing where a developing research programme stands and what it still needs to accomplish.

The analysis is organized around the theory's three core tenets—classifiable, measurable and modifiable—examining how each maps onto a well‐established philosophical framework.

2. Background: TCM Constitution Theory in Brief

Wang Qi proposed that individual constitution—the relatively stable physiological and psychological characteristics that shape a person's susceptibility to disease and response to treatment—can be classified into nine types: one balanced type (Pinghe) and eight biased types, each defined by a characteristic cluster of morphological, physiological and psychological features [6, 7]. The classification was derived from extensive clinical observation within the conceptual framework of Chinese medicine, drawing on classical sources while introducing a systematic typological structure absent from earlier writings.

To operationalize this classification, the Constitution in Chinese Medicine Questionnaire (CCMQ) was developed—a 60‐item self‐report instrument subjected to psychometric validation procedures including tests of internal consistency, test−retest reliability and criterion validity [8]. Independent validation in Hong Kong Chinese (n = 1084) confirmed the nine‐factor structure through confirmatory factor analysis and demonstrated satisfactory reliability and construct validity [9], and subsequent cross‐cultural adaptations (Japanese, Korean, Vietnamese versions) have provided further evidence of structural replicability. The CCMQ enabled large‐scale epidemiological investigation: a 2007 national survey across nine provinces (n = 21,948) established population‐level distributions of constitution types across demographic and geographic strata [10].

Biological investigation has sought molecular correlates. Metabolomic and proteomic studies have reported that constitution types exhibit distinct molecular profiles, providing preliminary evidence that the phenotypically defined categories correspond to detectable differences at the molecular level [11]. More recently, digitization efforts have begun to expand the observational base to include tongue surface imaging, pulse waveform analysis and infra‐red thermography, while knowledge ontology work has mapped constitution‐related concepts to standard biomedical terminologies including the Human Phenotype Ontology (HPO), Disease Ontology (DO) and Gene Ontology (GO) [6]. These extensions remain at varying stages of development; their significance for the present analysis lies not in their current maturity but in the direction they represent—an expansion of the measurement frontier at the phenotypic level.

The theoretical backbone of this research programme is captured in three tenets: constitutions are classifiable (可分, kě fēn), measurable (可测, kě cè) and modifiable (可调, kě tiáo) [6]. These three tenets structure the analysis that follows—not because they are uniquely Chinese concepts, but because they articulate, with unusual explicitness, three epistemological problems that any clinical classification system must confront. The following sections examine how each tenet maps onto an independently established philosophical framework, and what this mapping reveals about the structure of clinical knowledge maturation.

3. Classifiable: The Ontology of Clinical Categories

The act of classifying patients into constitution types carries an ontological commitment: that the categories reflect structure in the phenomena themselves, not merely the convenience of the observer. In philosophy of science, this is the problem of natural kinds—whether a classification carves nature at its joints or imposes boundaries on a continuous reality [12].

The classical theory of natural kinds, associated with Kripke and Putnam, requires that members of a natural kind share a common microstructural essence—water is H₂O, gold has atomic number 79 [13, 14]. By this criterion, few clinical categories qualify as natural kinds. Most diseases, and certainly constitutional types, are not defined by a single molecular essence. Applied strictly, the essentialist standard would render most of clinical medicine's categories ontologically suspect—an unhelpful conclusion.

Boyd's homoeostatic property cluster (HPC) theory offers a more applicable framework [15]. On this account, a natural kind need not be defined by a single essence. It may instead be constituted by a cluster of properties that tend to co‐occur, where the co‐occurrence is maintained by some underlying homoeostatic mechanism. The mechanism need not be fully elucidated; what matters is that the clustering is stable and non‐accidental.

TCM constitution types fit this description in form, though whether they satisfy it in substance remains an open empirical question. Each type is defined by a cluster of phenotypic features—morphological, physiological and psychological—that tend to co‐occur across individuals. The ‘Yangxu’ (yang‐deficient) type, for instance, is characterized by cold intolerance, a slow pulse and a pale complexion—features that cluster together with sufficient regularity to be clinically recognizable. The co‐occurrence is hypothesized to be maintained by underlying physiological mechanisms, and metabolomic evidence showing distinct molecular profiles across types provides partial support for this hypothesis [11]. However, whether the clustering is statistically robust—whether, for example, latent class analysis or mixture modelling on CCMQ data would recover the nine‐type structure rather than a different number of classes—has not been rigorously tested through independent studies. The HPC status of constitution types is therefore best understood as a working hypothesis that has been posed in a form amenable to empirical investigation, not as an established conclusion.

The philosophical point is that HPC theory grants ontological standing to categories defined at the phenotypic level without requiring molecular specification. This is not a concession or a temporary expedient. It is a principled ontological position, and it applies equally to psychiatric diagnoses defined by symptom clusters, to metabolic syndrome defined by co‐occurring metabolic abnormalities and to constitutional types defined by co‐occurring phenotypic features. Importantly, the HPC framework also exposes vulnerabilities: metabolic syndrome has been criticized on precisely these grounds. Kahn et al. [16] argued that the co‐occurring abnormalities may represent distinct pathological processes with different etiologies rather than a single coherent entity, and that the syndrome's clinical utility does not establish its ontological coherence. This critique applies with equal force to constitution types: clinical recognizability is not the same as ontological robustness. In each case, the ontological question—whether the category reflects real structure—is an empirical question that can be posed and investigated at the phenotypic level. Molecular investigation may enrich the answer, but it is not a prerequisite for posing the question.

4. Measurable: The Epistemology of Operationalization

A classification with ontological standing is not yet scientifically tractable. To enter the domain of systematic enquiry, the categories must be operationalized—translated into measurement procedures that are standardized, reliable and valid. Bridgman's operationalism holds that a scientific concept is defined by the operations used to measure it [17]. Whatever ‘Yangxu constitution’ means in theoretical terms, its scientific meaning is constituted by the procedures through which it is identified and scored.

The CCMQ achieves this translation. Its development followed the standard psychometric pipeline: item generation from clinical theory, expert review, pilot testing, factor analysis and validation against clinical criteria. Independent validation in a Hong Kong population (n = 1084) confirmed the nine‐factor structure through confirmatory factor analysis and demonstrated satisfactory reliability (Cronbach's ⍺ > 0.6, ICC > 0.6) and construct validity against the SF‐12v2 [9]. Cross‐cultural adaptations in Japanese, Korean and Vietnamese populations have provided further evidence of structural replicability. The result is a measurement instrument whose properties are assessable by the same standards applied to any clinical questionnaire, from the PHQ‐9 for depression to the IPSS for lower urinary tract symptoms. Structurally, this is the same move that the DSM‐III made for psychiatric diagnosis in 1980: replacing clinician gestalt with operationalized criteria, thereby making the classification amenable to epidemiological investigation, inter‐rater comparison and outcome research [2].

That said, the psychometric evidence should be assessed with appropriate caution. The original validation [8] was described as ‘preliminary’, and while the Hong Kong CFA confirmed the hypothesized structure, sensitivity and specificity in predicting constitution types diagnosed by practitioners were only fair (42.7%–82.7%) [9]. Whether the nine‐type structure represents the optimal number of classes—as opposed to a larger or smaller number—has not been tested through data‐driven approaches such as latent class analysis on independent samples. These are not disqualifying weaknesses; they are the kinds of open questions that characterize a developing measurement programme. The DSM's own operationalized criteria have been subject to decades of revision and remain contested on grounds of construct validity and comorbidity [2, 3]. The relevant point is that the CCMQ has entered the domain of systematic psychometric scrutiny—its properties can be debated, tested and improved using shared methodological standards.

But operationalization through a single modality has an inherent epistemic limitation. A self‐report questionnaire constitutes what philosophers of science call underdetermination: the same score can correspond to different underlying physiological states. This limitation is not specific to the CCMQ—it applies to any single‐instrument assessment. The PHQ‐9 score of 15 may reflect major depression, adjustment disorder or medical illness with somatic symptom overlap. A single measurement channel cannot resolve this ambiguity.

The response to underdetermination is triangulation—the convergence of independent measurement modalities on the same construct. Campbell and Fiske's multitrait‐multimethod framework provides the rationale: when measurements obtained through independent methods converge on the same classification structure, this constitutes evidence that the classification reflects something beyond any single instrument's artifact [18]. Denzin's triangulation principle generalizes this logic: cross‐validation across independent data sources is a fundamental epistemic strategy for strengthening inference [19].

In the case of constitution theory, convergent evidence comes from multiple independent sources: self‐report questionnaire data, metabolomic profiles and proteomic arrays point to the same nine‐type classification structure. This convergence is epistemically significant—it is the same logic by which construct validity is established in any field of measurement science.

Wang Qi's recent digitization strategy extends this logic further. The integration of tongue surface imaging, pulse waveform analysis and infra‐red thermography into constitution assessment is not merely a technological upgrade. It represents a deliberate expansion of the observation space at the phenotypic level—increasing the dimensionality and resolution of observation without descending to a lower explanatory level. This is analogous to the expansion of cardiac assessment from auscultation to electrocardiography to echocardiography to cardiac MRI: each modality captures different dimensions of the same organ system, and the convergence or divergence of their findings has diagnostic significance that no single modality can provide.

5. Modifiable: From Correlation to Causal Intervention

Classification and measurement establish that constitution types can be defined and reliably identified. But clinical utility demands a stronger claim: that constitution is causally relevant to health outcomes, and that modifying constitutional status changes those outcomes. This is the transition from descriptive taxonomy to causal knowledge—and it is here that the epistemological analysis connects most directly to questions about the nature of clinical evidence.

5.1. The Interventionist Framework

Woodward's interventionist account of causation provides the relevant philosophical framework: X causes Y if and only if there exists an intervention on X that, under appropriate conditions, changes Y [20]. This framework is explicitly mechanism‐agnostic. It does not require specification of the pathway linking X to Y. It requires only that the intervention‐outcome relationship be demonstrable—that a well‐defined change in the putative cause produces a detectable change in the outcome.

Wang Qi's tenet that constitutions are modifiable (体质可调) is, in this framework, a causal hypothesis: intervening on constitutional status changes health outcomes. The hypothesis decomposes into two layers, each with distinct evidential requirements.

The first is a predictive claim: a biased constitution predicts specific health outcomes. This is testable through longitudinal cohort designs, survival analysis and state‐transition models. It asks whether constitutional classification carries prognostic information—whether knowing a person's constitution type improves prediction of their future health trajectory beyond what is available from conventional risk factors.

The second is an interventional claim: modifying the constitution changes outcome risk. This is the stronger claim, requiring evidence that the constitution‐outcome relationship is causal rather than merely correlational. The appropriate evidential tools—randomized intervention trials, or quasi‐experimental causal inference methods such as instrumental variables, Mendelian randomization and target trial emulation—are part of EBM's standard inferential toolkit. None of them requires a molecular mechanism as a prerequisite.

5.2. Mechanism‐Agnostic Causation in Clinical Practice

The mechanism‐agnostic character of causal inference is not a novelty introduced by this analysis. It is the operating logic of much of clinical medicine. Aspirin's antiplatelet effect was established through clinical trials and used to prevent cardiovascular events decades before the COX‐1 inhibition mechanism was characterized. Lithium's efficacy in bipolar disorder was demonstrated well before its molecular targets were identified. In each case, the causal claim rests on interventionist evidence: the intervention was performed, the outcome changed, and confounding was controlled. The mechanism is a separate question—valuable when available, but not a gatekeeper for the causal inference itself.

The same logic applies in principle to constitution‐directed interventions. If a well‐designed trial demonstrates that an intervention targeting Yangxu constitution reduces the incidence of a specific outcome—say, recurrent respiratory infection—the causal claim would stand on the strength of the trial design, not on whether the molecular pathway from ‘Yangxu constitution’ to ‘immune susceptibility’ has been elucidated.

However, it is important to state clearly where the constitution research programme currently stands with respect to this causal agenda. Observational evidence has begun to accumulate: a recent retrospective cohort study (n > 17,000 elderly individuals) reported that Qi‐deficiency constitution was associated with the highest mortality risk (HR = 1.83) after controlling for conventional risk factors [21]. This represents evidence at the predictive layer—constitution classification appears to carry prognostic information. But evidence at the interventional layer—demonstrating that modifying constitutional status causally changes health outcomes through well‐designed randomized trials—remains sparse. This gap is not a minor omission in an otherwise complete evidence base; it is the central unfinished task of the research programme. The three‐stage framework developed in this essay is useful precisely because it can identify this gap with precision: the programme has made substantial progress on operationalization, has partial evidence for ontological grounding and has barely begun the work of causal validation.

5.3. The Structure of the Causal Challenge

This is not to suggest that the causal agenda for constitution theory is simple. It is not. The challenges are considerable, but they are challenges of study design, not of epistemic legitimacy.

Constitution is not a single variable that can be straightforwardly randomized. It is a multi‐dimensional phenotypic configuration. Interventions targeting constitution are complex—typically involving combinations of dietary, behavioural and pharmacological components—and their effects may unfold over long time horizons. These features make constitution‐directed research methodologically demanding.

But these are familiar challenges in contemporary clinical science, not unique obstacles. The MRC framework for evaluating complex interventions explicitly addresses multi‐component, context‐dependent interventions whose effects cannot be reduced to single causal chains [22]. Heterogeneity of treatment effect—the recognition that average treatment effects may obscure clinically important variation in individual responses—is now a central concern of precision medicine [23]. The evolution of EBM's own methodological toolkit, from simple two‐arm trials to adaptive designs, factorial experiments and n‐of‐1 trials, reflects a growing recognition that clinical causation is often non‐linear, multi‐dimensional and individual‐level.

This observation points towards a broader reflection. The kind of clinical knowledge that constitution theory seeks to produce—individualized, pattern‐based, multi‐dimensional—is the kind that EBM's evolved methodological framework is increasingly designed to evaluate. The early toolkit of EBM—the simple RCT, the average treatment effect, the single‐outcome hypothesis test—was optimized for linear, single‐factor causal questions. As EBM's methods have matured to accommodate complexity, heterogeneity and individual‐level variation, the methodological tools needed to test constitution‐directed hypotheses have become available in principle. Whether they will be applied with sufficient rigour to produce convincing evidence remains to be seen. The point here is structural: the methodological distance between what constitution theory asks and what EBM's methods can address has narrowed—not because constitution theory has moved towards EBM, but because both are responding to the same recognition that clinical causation is often non‐linear, multi‐dimensional and individual‐level.

6. Discussion

6.1. Convergence and Level‐Independence

Two knowledge systems, starting from different conceptual origins but employing shared methodological instruments, have arrived at the same epistemological stage‐structure: classification, operationalization, causal testability. Constitution theory moved from phenomenological classification to standardized questionnaire to epidemiological survey to—in its current phase—the beginnings of causal investigation. EBM codified the same logic: define, measure, test causally. This parallel should not be overstated: at the most generic level, any systematic research programme will involve defining categories, measuring them and testing causal claims, and the convergence may reflect nothing more than the basic structure of empirical enquiry. The more specific claim is that both trajectories encounter the same three epistemological conditions as distinct, independently satisfiable requirements—and that a research programme's maturity can be assessed by how far it has progressed along each. Constitution theory has made uneven progress: operationalization is relatively advanced, ontological grounding has partial support, and causal validation is nascent. This unevenness is itself informative—it demonstrates that the three conditions are not merely aspects of a single process but separable stages that a research programme may traverse at different rates.

But the convergence reveals something more specific than universality. Constitution theory is a top‐down system: it begins with phenotypic patterns and subsequently seeks molecular correlates. Molecular oncology, by contrast, is bottom‐up: breast cancer's molecular subtypes were identified through gene expression profiling and then validated against clinical outcomes. Yet both trajectories must satisfy the same three epistemological conditions. The molecular subtypes required ontological grounding (do these expression clusters correspond to biologically distinct disease entities?), operationalized measurement (can the subtypes be reliably identified in clinical practice?) and causal testability (does subtype‐directed therapy change outcomes?). The fact that the molecular subtypes entered the sequence at a different point—beginning with molecular characterization rather than arriving at it—does not exempt them from the sequence itself. And molecular classification has its own record of difficulty: sepsis has been repeatedly reclassified—from SIRS criteria to Sepsis‐2 to Sepsis‐3—in large part because earlier operationalizations proved insufficiently specific, a failure at the second epistemological condition despite the availability of molecular and pathophysiological detail.

This is the central observation: the three epistemological conditions are level‐independent. They specify what a classification must achieve to be scientifically tractable, not the explanatory level at which it must operate. A phenotypic classification that satisfies all three has the same epistemological standing as a molecular one that does so—and, equally, a phenotypic classification that has satisfied only one or two is no less scientifically tractable than a molecular classification in the same position. This is not a novel claim. It is the logic by which John Snow inferred waterborne cholera transmission through spatial mapping, by which Semmelweis established the causal role of hand hygiene through comparative mortality data, and by which clinical pharmacology routinely demonstrates drug efficacy through controlled trials in advance of complete mechanistic accounts. In each case, rigorous clinical knowledge was produced without molecular specification. The mechanisms were elucidated later, or not at all.

Any clinical classification—in psychiatry, in functional somatic syndromes, in the emerging frameworks of person‐centred typology—can be evaluated by these three criteria. Psychiatric nosology is instructive here not as a normalizing example but as a diagnostic one: the DSM's operationalized criteria have been criticized for poor construct validity, high comorbidity rates and weak predictive power [2, 3], which, in the terms of this analysis, represents difficulty at the first condition (ontological grounding) despite relative success at the second (operationalization). The three‐stage framework can diagnose where a classification system's weaknesses lie, regardless of the tradition from which it originates. The question worth asking is not at which explanatory level a classification operates, but whether it has been defined with ontological rigour, measured reliably and tested causally—and, for a developing research programme, which of these conditions have been addressed and which remain open.

6.2. The Place of Molecular Reduction

If the three epistemological conditions are level‐independent, a natural question follows: what role does molecular reduction play? The answer suggested by this analysis is that molecular investigation is a direction of enquiry, not an epistemological prerequisite.

This is an observation, not a critique. Molecular characterization has transformed clinical practice in oncological subtyping, pharmacogenomics and therapeutic target identification. These contributions are substantial. But they are contributions along a particular axis—the axis of mechanistic explanation—which is distinct from the axis along which the three epistemological conditions operate. Molecular investigation asks through what mechanism a phenomenon operates. Phenotypic classification asks what patterns characterize different patient groups and how those patterns relate to health outcomes. These are orthogonal inquiries, each with its own criteria for success.

The philosophical framework for this position is explanatory pluralism, developed by Cartwright [24], Dupré [25] and Wimsatt [26]: scientific phenomena can be legitimately investigated at multiple levels of organization, each with its own regularities and evidential standards. On this view, phenotypic classification and molecular investigation are not competing for the same explanatory territory. They address different questions, and progress along one axis does not substitute for progress along the other.

The practical relevance of this distinction is visible in clinical areas where molecular investigation has not resolved the classificatory challenges. Psychiatric nosology is a case in point: despite decades of intensive molecular and genetic research, phenomenological categories remain the working units of clinical practice [2, 3]. But the lesson here is not simply that phenotypic classification persists—it is that the persistence itself exposes specific epistemological gaps. The DSM categories have been operationalized (condition two is largely met), but their ontological grounding remains contested (condition one is weak), and the causal link between diagnostic category and treatment response is often modest (condition three is partially met at best). The RDoC initiative [3] represents an attempt to address condition one through a different ontological strategy—dimensional constructs anchored in neurobiology—but it has not yet produced classifications that satisfy condition two (clinical operationalization) with comparable success. These are not failures of molecular science or phenotypic classification per se; they are indications that satisfying the three conditions is difficult at any explanatory level, and that the framework developed here can diagnose the specific nature of the difficulty.

Comorbidity patterns—the stable co‐occurrence of conditions across organ systems—represent another domain where molecular pathway‐level explanations have not yet provided the integrative framework that clinicians need. These cases suggest that some clinical questions are most naturally posed and investigated at the phenotypic level, at least as a starting point.

EBM's own inferential logic is consistent with this pluralist position. The evidence hierarchy places systematic reviews and RCTs—mechanism‐agnostic instruments—above mechanistic reasoning [27]. Fisher's hypothesis testing framework evaluates the rigour of testing, not the theoretical pedigree of what is tested [28]. The widely recognized principle that the evaluation of a hypothesis is independent of its origin—however contested its strong philosophical formulations may be—reflects the operational logic of clinical evidence appraisal. EBM's evaluative framework was designed to assess evidential rigour, not explanatory level—and it applies to top‐down and bottom‐up classifications without prejudice.

7. Conclusion

This essay has identified three epistemological conditions—ontologically grounded categories, operationalized measurement and testable causal propositions—as the minimal requirements for a clinical classification to be scientifically tractable, and has argued that these conditions are independent of the explanatory level at which the classification operates. The analysis used TCM constitution theory as a developing, top‐down case to demonstrate that all three conditions can be posed and pursued at the phenotypic level, and that the three‐stage framework serves not only as a set of requirements but as a diagnostic tool for assessing a research programme's current maturity: constitution theory has made substantial progress on operationalization, has partial evidence for ontological grounding and has barely begun causal validation.

The parallel between this independently conceived research programme and EBM's core epistemic logic supports the claim that the three‐stage structure reflects the demands of clinical enquiry itself, though the parallel should not be overstated—at the most generic level, it may reflect the structure of empirical investigation more broadly. Molecular reduction contributes along a distinct axis—mechanistic explanation—but is neither a prerequisite for nor a substitute for the three conditions identified here. Recognizing the orthogonality of explanatory level and epistemological standing has practical consequences: it suggests that clinical classifications should be evaluated not by their proximity to molecular mechanisms, but by whether they have been defined with ontological rigour, measured reliably and tested causally—and, for developing research programmes, by which of these conditions have been addressed and which remain open.

Author Contributions

Sikai Ge: conceptualization, investigation, writing – original draft, writing – review and editing. Xing Zhou: investigation, resources, writing – review and editing. Xuejun Shang: supervision, writing – review and editing.

Funding

The authors have nothing to report.

Ethics Statement

This is a philosophical essay and does not involve human subjects, animal experiment or clinical data. No ethical approval was required.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

Data sharing is not applicable to this article as no data sets were generated or analysed during the current study.