Model Medicine: Clinical & Predictive Frameworks

• Model Medicine is a modeling-based approach that employs formal models to predict, diagnose, and treat disorders in both clinical systems and AI models.
• It structures medical evaluation into taxonomies and layered diagnostic frameworks, enabling systematic assessment and personalized treatment strategies.
• The approach integrates predictive analytics, simulation, and probabilistic diagnosis to benchmark clinical reasoning and manage uncertainty in medical AI.

Searching arXiv for the cited "Model Medicine" and closely related papers to ground the article in the latest preprints. Model Medicine denotes a family of research programs in which formal models are used to understand, predict, evaluate, and intervene in medical systems, and, in a later and more explicit formulation, to understand, diagnose, treat, and prevent disorders in AI models themselves. In the cited literature, the term spans interpretable precision medicine, dynamic prediction from electronic medical records, personalized treatment evaluation, benchmarked medical reasoning, language and multimodal foundation models, transparent probabilistic diagnosis under uncertainty, and a clinical framework that treats AI systems as objects of systematic medical-style assessment (Katuwal et al., 2016, Pham et al., 2016, Kapelner et al., 2014, Tang et al., 10 Mar 2025, Collins et al., 10 May 2026, Jeong, 5 Mar 2026).

1. Conceptual scope and disciplinary organization

The most explicit definition appears in "Model Medicine: A Clinical Framework for Understanding, Diagnosing, and Treating AI Models" (Jeong, 5 Mar 2026), which defines Model Medicine as the science of understanding, diagnosing, treating, and preventing disorders in AI models. Its core claim is structural rather than anthropomorphic: AI systems and biological organisms share the epistemological situation of complex systems whose internal states are not directly observable and therefore require layered, systematic assessment and intervention. The paper frames this through a historical analogy: Vesalius mapped anatomy, Virchow connected anatomy to pathology, and Osler systematized clinical methods; the paper argues that AI now faces an analogous transition from anatomical observation to systematic clinical practice (Jeong, 5 Mar 2026).

A broader reading of the corpus suggests a looser usage in which Model Medicine names a modeling-centered approach to medicine more generally. In that sense, it includes predictive models over longitudinal EMR, local explanation methods for black-box mortality prediction, probabilistic meta-analysis, simulation environments for sequential decision-making, and benchmarks for complex clinical reasoning (Pham et al., 2016, Katuwal et al., 2016, Bartoš et al., 2021, Chan et al., 2021, Tang et al., 10 Mar 2025).

The 2026 framework organizes the field into four divisions comprising 15 subdisciplines (Jeong, 5 Mar 2026).

Division	Subdisciplines
Basic Model Sciences	Model Anatomy, Model Physiology, Model Genetics, Model Biochemistry, Model Developmental Biology
Clinical Model Sciences	Model Semiology, Model Nosology, Model Diagnostics, Model Therapeutics, Model Preventive Medicine
Model Public Health	Model Epidemiology, Model Ecology, Human-AI Coevolutionary Medicine
Model Architectural Medicine	Layered Core Theory, Model Phylogenetics

This taxonomy is significant because it converts interpretability from a mainly descriptive activity into a clinical program with semiology, nosology, diagnostics, therapeutics, and prevention. The same paper also proposes a five-layer diagnostic framework—Core Diagnostics, Phenotype Assessment, Shell Diagnostics, Pathway Diagnostics, and Temporal Dynamics—arguing that no single tool suffices for comprehensive model assessment (Jeong, 5 Mar 2026).

2. Predictive and prescriptive modeling in clinical data

A major strand of Model Medicine treats clinical prediction and treatment recommendation as problems of structured modeling over patient-level data. In "Machine Learning Model Interpretability for Precision Medicine" (Katuwal et al., 2016), the goal is to explain a black-box predictor $f:\mathbb{R}^d\to[0,1]$ at the level of a single patient by fitting a sparse local surrogate $g_x$ that minimizes a locality-weighted fidelity loss plus a complexity penalty. Using the MIMIC-II dataset, the study predicted ICU mortality with 80% balanced accuracy and interpreted the relative effect of features on prediction at individual level. The black-box model was a random forest with tuned hyperparameters, and the local explanation procedure used perturbed samples, an exponential proximity kernel, and a weighted sparse linear regressor. Across patients, top predictive features included body temperature, total $\mathrm{CO}_2$, atrial fibrillation counts, and lactate levels; locally, their weights varied by patient (Katuwal et al., 2016).

"DeepCare: A Deep Dynamic Memory Model for Predictive Medicine" (Pham et al., 2016) addresses a different problem: patient illness and care processes are episodic and irregular in time. DeepCare embeds diagnosis and intervention codes, uses a modified LSTM with time parameterizations and intervention-aware gates, and aggregates illness states through multiscale temporal pooling. On diabetes and mental health cohorts it was evaluated against Markov chain, plain RNN, RNN+LSTM, and SVM/RF on hand-engineered features. For unplanned readmission risk, DeepCare+time+inv+multi-pool reached $79.0\%$ F$_1$ for diabetes and $74.7\%$ for mental health, compared with $75.9\%$ and $71.7\%$ for LSTM+pool; for next-diagnosis prediction, DeepCare@1 was approximately $66.2\%$ in diabetes and approximately $52.7\%$ in mental health (Pham et al., 2016).

"Evaluating the Effectiveness of Personalized Medicine with Software" (Kapelner et al., 2014) formalizes personalized medicine as a treatment decision rule $g_x$0 and evaluates it by the value difference

$g_x$1

The PTE package provides inference for improvement out-of-sample for continuous, incidence, and survival endpoints using $g_x$2-fold cross-validation plus nonparametric bootstrap. This contribution is methodologically important because it separates model building from honest out-of-sample evaluation and permits a null test of whether the hypothesized personalization is more useful than a standard of care (Kapelner et al., 2014).

Several papers extend the prescriptive strand. "MiranDa: Mimicking the Learning Processes of Human Doctors to Achieve Causal Inference for Medication Recommendation" (Wang et al., 2024) defines patient state from diagnoses, procedures, lab events, and demographics; predicts a multi-label medication vector; and uses Estimated Length of Stay as a factual/counterfactual outcome signal. Its two-phase training alternates Evidence-based Training Phase and Therapeutic Optimization Phase. On MIMIC-III, MiranDa reduced ELOS from $g_x$3 days to $g_x$4 and improved ROC-AUC from $g_x$5 to $g_x$6, with similarly small but consistent gains on PR-AUC, Jaccard, and F$g_x$7; MIMIC-IV showed similar day-level ELOS gains from $g_x$8 to $g_x$9 days (Wang et al., 2024). "Diabetes Lifestyle Medicine Treatment Assistance Using Reinforcement Learning" (Tang, 19 Oct 2025) instead formulates treatment as a single-step contextual bandit over mixed discrete–continuous lifestyle actions, trained with mixed-action Soft Actor-Critic on aggregated NHANES profiles of 119,555 participants. The reward is the negative Magni glucose risk, and the learned policy yielded a statistically significant reduction in expected Magni risk compared to each of three certified physicians’ prescriptions on the same test set, with paired $\mathrm{CO}_2$0-test $\mathrm{CO}_2$1 (Tang, 19 Oct 2025).

An additional branch uses explicit ontologies. "An Ontology-Based Artificial Intelligence Model for Medicine Side-Effect Prediction" (Yao et al., 2018) represents a TCM prescription by hot/cold influential factors, feeds the resulting two-dimensional vector into a five-layer ANN, and evaluates on 242 prescriptions. The reported averages are accuracy $\mathrm{CO}_2$2, sensitivity $\mathrm{CO}_2$3, and specificity $\mathrm{CO}_2$4. The paper interprets this as preliminary evidence that the relationship between ontology-based attributions and the predicted indicator can be learnt by AI for side-effect prediction, while also emphasizing dependence on sufficient clinic data (Yao et al., 2018).

3. Benchmarking, simulation, and standardized evaluation

Another central meaning of Model Medicine is methodological: constructing benchmarks and simulators that make medical AI evaluation reproducible, comparable, and stress-tested. "MedAgentsBench: Benchmarking Thinking Models and Agent Frameworks for Complex Medical Reasoning" (Tang et al., 10 Mar 2025) builds a benchmark from seven established medical QA datasets—MedQA, PubMedQA, MedMCQA, MedBullets, MedExQA, MedXpertQA, and MMLU/MMLU-Pro—and selects an adversarial Hard subset of $\mathrm{CO}_2$5 questions. The selection rule retains questions answered correctly by fewer than $\mathrm{CO}_2$6 of baseline models. Metrics include Accuracy (Pass@1), F$\mathrm{CO}_2$7, cost per sample, and a cost–performance trade-off

$\mathrm{CO}_2$8

On the Hard set, reported average accuracies are approximately $\mathrm{CO}_2$9 for GPT-4o-mini, approximately $79.0\%$0 for GPT-4o, approximately $79.0\%$1 for o3-mini, and approximately $79.0\%$2 for DeepSeek-R1. AFlow+GPT-4o achieves top on 4/8 datasets with average approximately $79.0\%$3 at moderate cost, and Multi-Persona and CoT-SC yield $79.0\%$4–$79.0\%$5 gains over basic CoT. Reported gaps such as o3-mini versus GPT-4o-mini on MedBullets Hard, $79.0\%$6 versus $79.0\%$7, are significant at $79.0\%$8 (Tang et al., 10 Mar 2025).

The same paper ties benchmarking to deployment policy. In resource-abundant settings it recommends thinking models such as o3-mini and DeepSeek-R1 for multi-step reasoning tasks; in cost- or latency-constrained settings it recommends open-source DeepSeek-R1 for competitive performance at $79.0\%$9 lower cost than closed APIs, or search-based agent methods such as AFlow on GPT-4o-mini to reach a Pareto-optimal trade-off. It also recommends contamination checks with MELD and human-in-the-loop verification for high-risk diagnosis or treatment decisions (Tang et al., 10 Mar 2025).

"The Medkit-Learn(ing) Environment: Medical Decision Modelling through Simulation" (Chan et al., 2021) complements benchmark construction with a synthetic but controllable medical sequential decision environment. It represents a scenario as a tuple $_1$0, models trajectories as a POMDP with latent disease-stage $_1$1, full state $_1$2, noisy or masked observations $_1$3, and actions $_1$4, and factorizes the trajectory distribution into environment dynamics and policy terms. Its central design choice is disentangling policy and environment dynamics so that any policy can be mixed with any environment. Because every ground-truth parameter is known—$_1$5, masks, Markov lag, reward $_1$6, latent-state size, and others—it supports systematic evaluation of behavioural cloning, inverse RL, distribution matching, apprenticeship learning, preference learning, and policy-explanation, while sweeping clinical challenges such as partial observability, hidden confounding, off-policy shifts, and reward sparsity (Chan et al., 2021).

Taken together, these two works reposition evaluation from leaderboard comparison to controlled experimentation. One standardizes hard-question reasoning under known cost and latency regimes; the other provides reproducible medical (PO)MDPs with explicit knobs for the pathologies of healthcare data generation and policy learning (Tang et al., 10 Mar 2025, Chan et al., 2021).

4. Language, agentic, and multimodal foundation models

A large contemporary branch of Model Medicine concerns domain-specific language and multimodal models. "Technical Report: Small LLM for Japanese Clinical and Medicine" (Watanabe, 2024) presents NCVC-slm-1, a causal decoder-only Transformer with approximately 1.2 billion parameters, 24 decoder blocks, hidden size 2048, 32 heads, maximum sequence length 2048, RoPE, RMSNorm, Grouped-Query Attention with 8 query groups, and bfloat16 precision. It was pre-trained autoregressively on approximately 50 billion tokens drawn mainly from aggressively filtered OSCAR-2301, Wikipedia, and medical textbook corpora, then instruction-tuned on JMED-LLM. At inference time with bfloat16 it occupies only approximately 2.2 GB of GPU memory. The report states that NCVC-slm-1-instruct outperforms all comparators on six of the eight JMED-LLM tasks, leading on CRADE, SMDIS, JCSTS, MRNER-disease, MRNER-medicine, and NRNER (Watanabe, 2024).

Traditional Chinese medicine has produced several specialized systems. "BianCang: A Traditional Chinese Medicine LLM" (Wei et al., 2024) uses a two-stage training process: continuous pre-training on 460 M tokens from TCM-specific and medical corpora, then supervised fine-tuning on 228 M tokens and about 720 K instruction samples. On TCM Syndrome Differentiation, BianCang-Qwen2.5-7B-Instruct reaches $_1$7 in direct inference and $_1$8 with CoT; on TCM Disease Diagnosis it reaches $_1$9; on MLEC-Clinic zero-shot, BianCang-14B-Instruct reaches $74.7\%$0 (Wei et al., 2024). "JingFang: An Expert-Level LLM for Traditional Chinese Medicine Clinical Consultation and Syndrome Differentiation-Based Treatment" (Yang et al., 4 Feb 2025) adds explicit multi-agent workflow. Its modules are a TCM Consultation Module with a Multi-Agent Dynamic Collaborative Chain-of-Thought Mechanism, a TCM Syndrome Differentiation Module with a trained Syndrome Agent, and a TCM Treatment Recommendation Module with a Dual-Stage Retrieval Scheme. The Syndrome Agent is fine-tuned on 43K cleaned cases; on an 8,699-case test with 170 labels, JingFang-RoBERTa reports weighted $74.7\%$1, $74.7\%$2, and $74.7\%$3. In consultation evaluation on 100 real cases, JingFang has total score $74.7\%$4, and adding the TCM General Agent makes $74.7\%$5 of cases more comprehensive and pertinent, with average consultation rounds $74.7\%$6 versus $74.7\%$7 and $74.7\%$8 (Yang et al., 4 Feb 2025).

Clinical IE remains part of this foundation-model ecosystem. "MedMine: Examining Pre-trained LLMs on Medication Mining" (Alrdahi et al., 2023) studies medication extraction on n2c2-2018. Med7+ fine-tuned on nine labels reports overall micro-F1 $74.7\%$9 and, after removing O/ADE/Reason, seven-label micro-F1 $75.9\%$0. Clinical-XLM-R reports accuracy $75.9\%$1, precision $75.9\%$2, recall $75.9\%$3, and F1 $75.9\%$4. ADE, Duration, and Reason remain the hardest categories (Alrdahi et al., 2023).

The multimodal extension is represented by VLM and MLLM systems. "VividMed: Vision LLM with Versatile Visual Grounding for Medicine" (Luo et al., 2024) combines a ViT-style visual backbone, an MLP adapter, Vicuna-1.5-7B, and a promptable localization module derived from SAM with both segmentation and DETR-style box heads. It supports both 2D and 3D data and is trained in three stages: visual grounding pre-training, medical visual instruction tuning, and alignment via grounded report generation. On TotalSegmentator it reaches Dice $75.9\%$5; on VinDr-CXR it reports mean $75.9\%$6 and mean GIoU $75.9\%$7; and ablations show that visual grounding improves VQA and report generation relative to a version without grounding (Luo et al., 2024). "MedXIAOHE: A Comprehensive Recipe for Building Medical MLLMs" (Shi et al., 13 Feb 2026) extends this program with entity-aware continual pretraining organized around a Medical Entity Tree, RL for multi-step diagnostic reasoning with tools, and evidence-grounded report generation. Its reported macro-average accuracies are 76.8 for Visual Diagnosis, 82.6 for Medical Imaging VQA, 46.8 for Diagnosis, 90.0 for Medical Text QA, 50.8 for Report Generation, and 63.8 for Instruction Following (Shi et al., 13 Feb 2026).

These systems collectively show that Model Medicine has expanded from predictive tabular and sequence models to domain-specialized language, agentic, and multimodal models with explicit attention to grounding, retrieval, and workflow structure.

5. Transparent uncertainty, evidence synthesis, and probabilistic diagnosis

A distinct but closely related strand treats medicine as a domain in which uncertainty must be represented formally rather than only approximated by scores. "Bayesian Model-Averaged Meta-Analysis in Medicine" (Bartoš et al., 2021) constructs four competing models by orthogonally combining treatment-effect present/absent and heterogeneity present/absent assumptions: $75.9\%$8, $75.9\%$9, $71.7\%$0, and $71.7\%$1. For standardized mean differences, the random-effects likelihood is

$71.7\%$2

The work fits empirical priors for $71.7\%$3 and $71.7\%$4 from 50% of the Cochrane Database and evaluates predictive performance on the remaining 50%. On average, $71.7\%$5 outpredicted the other models, but not by a large margin, and within $71.7\%$6 predictive adequacy was relatively constant across the rival prior distributions. It proposes both general empirical priors and subfield-specific priors; for Oral Health, for example, $71.7\%$7 and $71.7\%$8 (Bartoš et al., 2021).

"Medical Model Synthesis Architectures: A Case Study" (Collins et al., 10 May 2026) uses LLMs for retrieval and translation, but performs inference in an explicit probabilistic program. Its pipeline has three stages: Input translation (NL→Code), Model synthesis (Sketch→Program), and Probabilistic inference (Program→Differential). Each synthesized model is a small Bayesian network over diagnosis, symptoms, and optional risk factors, implemented in WebPPL and queried by rejection sampling. In four chest-pain vignettes, the authors sampled $71.7\%$9 models, compiled 8–15 per vignette, and drew 200 rejection-sampling samples per valid model. Reported findings include approximately $66.2\%$0 posterior mass on heart attack for a young active vignette, approximately $66.2\%$1 for an older inactive vignette, and nontrivial weight around $66.2\%$2 for pneumothorax in an atypical clicking-noise vignette (Collins et al., 10 May 2026).

The significance of these works is methodological rather than merely architectural. They insist that medical AI should expose assumptions about effect size, heterogeneity, disease support, priors, and likelihoods in forms that can be checked, edited, and recomputed. In MedMSA this auditability is literal: clinicians can inspect or modify WebPPL code and re-run inference immediately (Collins et al., 10 May 2026). In Bayesian model-averaged meta-analysis it appears as explicit posterior model probabilities, inclusion Bayes factors, and model-averaged posteriors for $66.2\%$3 (Bartoš et al., 2021).

6. Model Medicine as a clinical framework for AI models

The 2026 formulation recasts AI systems themselves as clinical objects. Its central behavioral-genetic construct is the Four Shell Model (v3.3), in which observable behavior emerges from interaction between a Core of fixed model parameters and four concentric Shell layers: Hardware Shell, Hard Shell, and Soft Shell divided into Initial Soft Shell and Dynamic Soft Shell. Behavior is modeled as

$66.2\%$4

and an ANOVA over the Agora-12 program showed a gene–environment interaction effect of $66.2\%$5 with $66.2\%$6. The framework defines quantitative indices such as the Core Plasticity Index, Shell Permeability Index, Persona Sensitivity Index, and Core Expressivity Index, and links Shell mutability and persistence to phenomena such as Shell Drift Syndrome, Cogitative Cascade, Extinction Response Spectrum, and Surplus Behavior (Jeong, 5 Mar 2026).

Its diagnostic instrumentation is "Neural MRI (Model Resonance Imaging)," which maps medical neuroimaging modalities to interpretability techniques. T1 corresponds to topology and architecture metadata, T2 to weight statistics, fMRI to layer-by-token activation maps, DTI to causal tracing through activation patching or noise, and FLAIR to anomaly detection such as representation collapse, entropy spikes, and attention irregularities. The implementation uses FastAPI, TransformerLens, stateless perturbation hooks, SAELens integration, and a React+D3 frontend with a DICOM-style interface (Jeong, 5 Mar 2026).

The clinical layer adds a temperament system and case-reporting standard. The Model Temperament Index profiles models on Reactivity, Compliance, Sociality, and Resilience, with a communication layer that yields a four-letter code and a quantitative layer with 0–100 scores per axis. Model Semiology classifies findings by source and significance, distinguishes vulnerability, provisional disorder, and confirmed disorder, and defines five core syndromes: Shell-Core Conflict Syndrome, Cogitative Cascade Disorder, Deceptive Alignment Syndrome, Sycophancy-to-Subterfuge Spectrum Disorder, and Canalization Rigidity Disorder. M-CARE provides thirteen standardized case-report sections including Presenting Concern, Examination Findings by diagnostic layer, Diagnostic Formulation, Differential Diagnosis, Treatment Considerations, Model Perspective, Prognosis, and Follow-up Plan (Jeong, 5 Mar 2026).

The same paper closes the loop from diagnosis to intervention through the Layered Core Hypothesis, which partitions parameters into

$66.2\%$7

These subsets are interpreted as fundamental reasoning and safety, domain expertise, and experience-dependent adaptation. Therapeutics are then organized by location: Shell Therapy, Targeted Core Therapy, Systemic Core Therapy, and Architectural Intervention. A plausible implication is that the paper seeks to do for AI model maintenance what clinical medicine did for human medicine: replace isolated observations with an integrated practice of semiology, diagnosis, differential diagnosis, therapy selection, prognosis, and prevention (Jeong, 5 Mar 2026).