The starting point of the argument is that the labels on which AI models are trained are not neutral descriptions of reality. Medical terminologies such as the International Classification of Diseases, used for reporting and billing, and SNOMED CT, used for clinical representation, were built as harmonization and interoperability tools. They standardize communication, reimbursement, and surveillance, but they also embed purpose-specific assumptions about what counts as the same disease. The same diagnostic label is routinely used for population reporting, bedside decision support, and research cohort definition, even though these tasks demand very different levels of granularity, validity, and explanatory scope. When supervised models are trained on diagnosis codes, the commentary warns, they inherit the structure and limitations of those codes and may end up learning disease as it is encoded rather than disease as it is clinically observed.

This mismatch is not merely theoretical. A clinician deciding whether a patient’s blood pressure crosses the threshold for hypertension is making a judgment shaped by time constraints, available resources, risk tolerance, and the downstream harms of labeling someone with a chronic condition. The threshold that turns a continuous variable into a diagnosis is rarely purely objective; it is a decision made under constraints. Classification matters because it determines the next clinical action, whether that is treatment, reassurance, or intensified surveillance. AI does not dissolve those choices, but it can make the trade-offs visible, especially when labels originally designed for coding are quietly reused for care. Models optimized for surveillance or billing may consequently misclassify clinical need even while achieving impressive accuracy statistics.

The commentary draws on a long-running debate in the philosophy of medicine about whether disease can be defined purely biologically or whether it necessarily includes evaluative components. The naturalist account associated with philosopher Christopher Boorse centers disease on dysfunction relative to species-typical functioning, whereas Jerome Wakefield’s influential harmful dysfunction analysis treats disorder as a hybrid of biological dysfunction and judged harm. These abstract debates surface in daily practice because many diagnoses that appear as categorical labels are actually derived from continuous quantitative measures: blood pressure, glucose, ejection fraction, bone density, symptom scales. The history of overweight and obesity illustrates how such boundaries can shift over time, with substantial clinical and social consequences for the people reclassified on either side of the line.

Because clinical action is discrete while physiology is continuous, thresholds are unavoidable. The practical question, the author argues, is whether a given threshold supports clinical decisions or masks clinically important heterogeneity. Three evaluation strategies can make threshold effects measurable. Threshold sensitivity analyses reveal how changing a cut-off alters diagnosis rates, treatment eligibility, false positives, false negatives, and subgroup performance. Multimodal and longitudinal models can identify heterogeneity within a labeled condition, distinguishing patients who share a diagnosis code but differ in underlying mechanism, disease trajectory, or likely treatment response. Decision-analytic methods can then test whether an alternative threshold or representation genuinely improves clinically relevant choices, such as when to monitor, treat, escalate therapy, or allocate preventive resources. In this framing, AI does not replace clinical judgment; it exposes the consequences of classification-based judgment to scrutiny.

Taking continuity seriously leads to a more radical proposal. Petersen and Ursin, cited in the commentary, 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. They advocate continuous disease assessment: modeling severity, progression, and treatment response as evolving estimates in physiological space, with a diagnosis treated as a revisable hypothesis rather than an endpoint. Longitudinal modeling, 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 rather than association with a class of codes, which is closer to how experienced clinicians actually update their beliefs over time. Network medicine reinforces this shift, treating disease phenotypes as products of interacting biological networks rather than isolated categories.

Evidence itself, the commentary insists, is a classification problem. Evidence-based medicine ranks studies through appraisal tools such as GRADE for rating certainty of evidence, RoB 2 for assessing risk of bias in randomized trials, and PROBAST for prediction model studies, with the recent PROBAST+AI extension addressing machine learning prediction tools. Yet empirical work shows that different appraisal tools applied to the same studies can yield divergent judgments, and even within a single tool, agreement between assessors can be limited. AI tools can support evidence appraisal by extracting study features, enabling sensitivity analyses, and making uncertainty explicit, but they can also embed appraisal rules in workflows that are difficult to scrutinize. When large language models are used for evidence screening, summarization, or quality appraisal, their outputs may vary with model version, prompt structure, evidence sources, settings, and repeated runs, demanding stability measurement and human adjudication before informing evidence judgments.

The stakes become vivid in the case of proxy endpoints. A landmark study by Obermeyer and colleagues showed that an algorithm using future healthcare costs as a proxy for health need produced substantial racial bias, because costs reflect differential access and spending rather than morbidity. A model can be accurate for its measured endpoint while still misclassifying clinical need if the endpoint, data source, and subgroup performance are never examined directly. The commentary argues that endpoint specification must be treated as a primary methodological step: developers should define the clinical concept of interest, justify any proxy, and evaluate whether the proxy behaves similarly across subgroups and settings. When differences reflect structural inequities such as unequal access to digital tools or care, statistical adjustment alone is insufficient; measurement and deployment must be redesigned so that model outputs reflect clinical need rather than access or utilization. Related systematic reviews in perioperative medicine, pediatric anesthesia, and cardiology have found high risk of bias, limited external validation, and uncertain generalizability as persistent barriers to clinical integration.

Diagnostic reasoning, on this account, is not label assignment but abductive hypothesis generation and refutation: clinicians propose explanations, seek discriminating evidence, revise as data accumulate, and treat diagnoses as provisional. Clinical AI should mirror this structure by reporting uncertainty, supporting updating as new data arrive, and aligning predictions with the patient’s course over time rather than offering one-time diagnostic accuracy. Precision medicine also demands causal understanding of which interventions will help which patients, a requirement that philosophers Francesco Russo and John Williamson argue depends on combining mechanistic evidence with difference-making probabilistic evidence, neither of which suffices alone. Causal inference frameworks formalize the distinction between association and intervention effects, clarifying the assumptions needed to translate observational data into causal claims.

The practical payoff is what the commentary calls decision-centered precision medicine. Developers should first define the clinical decision a model supports, then define the clinical concept of interest and disclose whether a direct measure or a proxy label is used, evaluate stability across different electronic health record systems and patient populations, and finally assess decision-relevant outcomes rather than label accuracy alone, including effects on calibration, missed diagnoses, safer treatment, and equity. Health systems selecting AI tools should ask what decision the tool supports, what endpoint was used, whether external validation covered a population like theirs, whether subgroup performance was reported, and whether uncertainty is displayed in a usable form. After deployment, calibration, errors, subgroup performance, and patient-important outcomes should be monitored continuously. In this vision, implementation is not the final step after development; it is part of the evidence process itself. By turning conceptual ambiguities into testable, governable design choices, AI may ultimately sharpen medicine’s understanding of its own most basic categories.

Subject of Research: Philosophical foundations of clinical artificial intelligence and precision medicine, including disease classification, diagnostic thresholds, and evidence appraisal

Article Title: Clinical AI and Precision Medicine: Philosophical Questions About Labels, Disease, and Evidence

Article References: Armoundas, A. A. (2026). Clinical AI and Precision Medicine: Philosophical Questions About Labels, Disease, and Evidence. Journal of Medical Systems, 50(1), Article 130. https://doi.org/10.1007/s10916-026-02457-3

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02457-3

Keywords: clinical AI, precision medicine, disease classification, diagnostic thresholds, medical labels, proxy endpoints, evidence-based medicine, algorithmic bias, large language models, continuous disease assessment, electronic health records, health equity

The study, published in Nature Genetics, describes a clinically annotated, population-scale single-cell atlas built from bone marrow samples of 341 patients, spanning the full continuum of the disease from precursor conditions through newly diagnosed myeloma to relapsed and refractory disease after multiple lines of therapy. Using single-cell RNA sequencing enriched for plasma cells and the surrounding CD45-positive immune compartment, the researchers captured the transcriptomes of tens of thousands of individual cells, allowing them to dissect the malignant compartment cell by cell rather than averaging signals across bulk tumor tissue, which has historically masked the very heterogeneity that drives treatment failure.

Technically, the effort was formidable. Samples were processed using the MARS-seq platform and an updated version, MARS-seq2.0, with rigorous quality control on mitochondrial content, unique molecular identifier counts and detected genes. Cells were annotated through a computational pipeline combining scVI latent-space integration, uniform manifold approximation and projection embeddings, and inferCNV-based inference of copy number alterations to distinguish malignant plasma cells from their normal counterparts. The team then applied non-negative matrix factorization to decompose malignant gene expression into recurrent transcriptional programs, validated for stability through hundreds of bootstrapped iterations. This dual-layer analytical strategy allowed the investigators to separate two largely independent axes of tumor biology: what kind of myeloma a patient has, and how fast that myeloma is growing.

The first axis yielded five recurrent malignant transcriptional archetypes, designated MM1 through MM5, each corresponding to a stable pattern of gene expression anchored in distinct biological pathways. These archetypes align with known myeloma biology, including immunoglobulin heavy chain translocations such as t(11;14) with its cyclin D and BCL-2 dependencies, t(4;14) with NSD2 dysregulation, MAF and MAFB associated programs, and features reflecting unfolded protein response burden and bone marrow niche interactions. Crucially, the archetypes were not merely descriptive. They correlated with genomic features, therapeutic sensitivity patterns and clinical outcomes, and the team demonstrated that the classification could be ported to independent bulk RNA datasets, including the Blueprint cohort and the Multiple Myeloma Research Foundation’s CoMMpass cohort of treatment-naive patients, confirming that the single-cell-defined signatures retain prognostic power even when measured on standard clinical platforms.

The second axis, orthogonal to the archetypes, is a proliferative program. By scoring single-cell proliferation signatures and characterizing plasmablastic cells, the rapidly dividing precursors of antibody-secreting plasma cells, the researchers quantified the fraction of malignant cells actively cycling in each patient’s marrow. Proliferation has long been recognized as a poor prognostic marker in myeloma, measured historically by crude methods such as plasma cell labeling indices. The new work refines this concept at single-cell resolution, showing that the proportion of proliferating malignant plasma cells stratifies patients within every archetype, revealing intra-archetypal heterogeneity that earlier bulk approaches could not detect. In relapsed and refractory patients, higher proliferative fractions predicted shorter progression-free survival, and the effect persisted in multivariate Cox regression models adjusting for cytogenetic risk, age and prior treatment lines.

Combining the two axes produced an improved risk stratifier that outperformed existing molecular subtyping schemes. Patients could be placed into joint archetype-proliferation subgroups with meaningfully distinct progression-free and overall survival, and the framework added prognostic information beyond standard clinical variables including high-risk cytogenetics and chromosome 1p deletion. The validation in CoMMpass, one of the largest longitudinally followed myeloma cohorts in the world, demonstrated robustness and portability across sequencing platforms, an essential prerequisite for clinical translation. In principle, a myeloma patient’s tumor could one day be assigned to an archetype and proliferation state from a routine biopsy, guiding intensity of upfront therapy and informing decisions about transplantation, novel agents or early escalation.

Perhaps the most clinically electrifying result, however, came from the atlas’s use as a target-discovery engine. The team built a computational pipeline that ranked every protein-coding gene by a composite score integrating malignant enrichment, specificity for malignant plasma cells relative to normal plasma cells, and restriction across healthy tissues, the latter being critical to minimize off-tumor toxicity for any future immunotherapy. This screen surfaced FCRL2, an Fc receptor-like molecule with established roles in B cell biology, as a surface target expressed by malignant plasma cells but largely restricted to the B cell lineage elsewhere in the body. The atlas approach meant the researchers could verify not just that myeloma cells express FCRL2, but that expression is preserved across archetypes and proliferation states, addressing the antigen escape problem that plagues current myeloma immunotherapies.

The translational proof followed swiftly. The researchers engineered chimeric antigen receptor T cells directed against FCRL2 and tested them against myeloma cell lines with varying levels of target expression. In vitro, FCRL2-redirected CAR-T cells killed antigen-positive myeloma cells in an antigen-specific manner, with luciferase-based co-culture assays showing progressive suppression of tumor cell growth compared to non-transduced controls, and detailed immunophenotyping confirming proper CAR expression and memory-phenotype differentiation of the engineered cells. In mouse models, FCRL2-targeted CAR-T cells conferred a significant survival benefit. Given that existing myeloma immunotherapies targeting BCMA and GPRC5D eventually fail through antigen loss and relapse, a third lineage-restricted target backed by a genome-wide, single-cell-verified prioritization pipeline offers a credible path toward combination or sequential immunotherapy strategies.

For patients, the near-term significance is prognostic rather than therapeutic: an archetype and proliferation score could refine risk assessment well before relapse, when treatment decisions matter most. For the field, the study establishes a template for how population-scale single-cell atlases can move beyond description into actionable classification and target nomination. All of the underlying data, including the full scRNA-seq dataset deposited in the Gene Expression Omnibus and the analysis code released openly by the Amit lab, are publicly available, ensuring that other groups can interrogate, extend and challenge the framework. As single-cell sequencing costs fall and clinical grade assays mature, the line between research atlases and routine diagnostics grows thinner, and this myeloma atlas may be remembered as a turning point where that line was crossed for a historically intractable cancer.

Subject of Research: Single-cell transcriptomic atlas of multiple myeloma defining malignant archetypes, proliferative states, and the immunotherapy target FCRL2

Article Title: A single-cell atlas of multiple myeloma defines malignant archetypes and proliferative states

Article References: Zada, M., Kurilovich, A., Shapira, N., Wang, S.-Y., Sharet-Eshed, R., Kfir-Erenfeld, S., Schlossberg, M., Zorde, E., Asherie, N., Gur, C., Chalan, P., Shalita, R., Ben Yehuda, M., Zwicky, P., von Locquenghien, M., Ingelfinger, F., Mazuz, K., David, E., Gurevich-Shapiro, A., … Amit, I. (2026). A single-cell atlas of multiple myeloma defines malignant archetypes and proliferative states. Nature Genetics, 58(9), 2254-2269. https://doi.org/10.1038/s41588-026-02725-5

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02725-5

Keywords: multiple myeloma, single-cell RNA sequencing, transcriptional archetypes, proliferation, FCRL2, CAR-T cell therapy, risk stratification, plasma cells, Nature Genetics, precision medicine, CoMMpass cohort, target discovery