AI Multimorbidity Score Outperforms Classic Risk Indices Across Nine Global Cohorts

As populations around the world age and chronic diseases accumulate within individual patients, clinicians are increasingly forced to make decisions for people who carry several conditions at once. Multimorbidity, the coexistence of two or more chronic disorders in the same person, has become one of the most consequential challenges in modern medicine, shaping clinical decision-making, the allocation of healthcare resources, and long-term prognosis. Yet the tools used to quantify this burden remain surprisingly crude. A new study from researchers at Guangzhou Medical University, published in the journal Cyborg and Bionic Systems, argues that the next generation of risk stratification may come not from longer disease checklists but from large language models capable of reasoning across the whole patient record.

Conventional comorbidity indices such as the Charlson Comorbidity Index, the Elixhauser score, the Functional Comorbidity Index, the Hospital Frailty Risk Score, and the Multimorbidity Weighted Index are widely used in hospitals and research settings. These instruments typically depend on fixed disease lists and fixed weights derived from specific populations. That design choice creates a persistent problem: when such scores are transported to a different country, a different healthcare system, or a different age structure, their calibration and predictive performance can deteriorate. A weighting scheme tuned to older European patients may misjudge risk in younger Asian cohorts, limiting genuinely transportable risk stratification across heterogeneous populations.

Large language models offer a fundamentally different computational approach. Rather than counting diagnoses, an LLM can integrate heterogeneous health information, including medical history, laboratory measurements, lifestyle factors, functional status, and social context, and approximate the cross-system reasoning that experienced clinicians perform intuitively. The research team, led by Jinghao Liang of Guangzhou Medical University, set out to test whether this capability could be converted into a stable, interpretable multimorbidity phenotype score. However, the authors were careful to acknowledge the caveats that accompany medical artificial intelligence. Medical LLMs face concerns regarding reliability, bias, and cross-population generalizability, which makes large-scale external validation essential before any clinical deployment.

The resulting framework, called the LLM-based multimorbidity score, or LLM-MMS, was developed and validated for all-cause mortality risk stratification using nine international longitudinal cohorts: the UK Biobank, NHANES, SHARE, HRS, CHARLS, CLHLS, ELSA, KLoSA, and MHAS, together comprising 770,439 participants. The pipeline begins with harmonization. Demographic characteristics, lifestyle factors, physical examinations, biomarkers, and medical histories from all cohorts were mapped through a unified data dictionary and converted into standardized English natural-language clinical reports, with explicit hallucination control during the textification step. This conversion of structured data into narrative form is what allows a language model to read a patient profile the way a physician would.

At the core of the system sits DeepSeek-V3, used as the reasoning engine with a fixed prompt template and chain-of-thought prompting. Notably, the model was not fine-tuned on any clinical data. Through prompt-based learning alone, it generated seventeen health phenotypes for each participant, including an overall biological age, multiple organ-specific biological ages, a frailty age, and multidimensional health gradings covering psychological health, diet, behaviors, medical history, income, and family history. These generated phenotypes were then entered into standard Cox proportional hazards models for mortality prediction. The score was developed in the UK Biobank and externally validated in the other eight cohorts, with head-to-head comparisons against the Charlson, Elixhauser, FCI, HFRS, and MWI indices using the C-index, calibration plots, Brier scores, and decision-curve analysis.

The performance gap was striking. In the UK Biobank, LLM-MMS reached a C-index of 0.81, substantially higher than the best conventional comorbidity index at 0.61. Across the eight external validation cohorts, C-indices ranged from 0.67 to 0.86, exceeding even recalibrated conventional indices by 0.12 to 0.21. Calibration remained stable, with observed-to-expected ratios generally falling between 0.94 and 1.04, meaning the model’s predicted risks closely matched actual mortality rates. Stratified analyses showed relatively consistent discrimination across sex and age groups, with particularly clear advantages among adults younger than sixty years, a population in which traditional comorbidity scores often exhibit floor effects because younger patients accumulate few listed diagnoses even when their overall health trajectory is concerning.

Two further tests addressed the most common objections to LLM-based medicine. First, LLM-MMS outperformed survival models trained directly on the same raw structured variables in the UK Biobank, suggesting that its advantage was not simply a consequence of richer input information but of the model’s ability to transform multidimensional signals into a more holistic phenotype representation without fixed weighting schemes. Second, portability and reproducibility were tested by swapping in alternative LLM backbones and repeating inference runs; performance varied only modestly across different base models and remained highly reproducible, indicating that the framework does not depend on one proprietary engine or on stochastic quirks of individual inference passes.

The biological plausibility of the score was examined through proteomic profiling, model-generated reasoning summaries, and input-masking analyses. High-risk individuals carried molecular signatures related to inflammation, tissue injury, and metabolic stress, precisely the processes one would expect to underlie elevated mortality. Input-masking experiments, in which individual variables were systematically hidden from the model, identified smoking, inflammatory markers, systolic blood pressure, body mass index, and HbA1c as the major contributors to the final risk score. These contributors align with well-established epidemiological risk factors, supporting the interpretation that the language model had learned clinically meaningful patterns rather than exploiting spurious correlations hidden in the harmonized data.

The significance of the work lies in moving beyond the conventional paradigm of disease counting plus fixed weights. By integrating cross-organ and multidimensional health information into an interpretable phenotype, LLM-MMS maintained more stable discrimination and calibration across countries, age groups, and sexes without cohort-specific retraining, suggesting that LLM-derived integrated health representations may offer genuinely improved transportability across heterogeneous populations. The authors are candid about the limitations. LLM-MMS functions more like a broad prognostic health phenotype than a purely diagnosis-based multimorbidity score, and the inclusion of socioeconomic and other non-diagnostic domains raises legitimate questions about fairness and appropriate use. The framework also remains anchored to its UK Biobank development cohort, to English-language textification, and to potential biases inherent in large language models themselves.

Looking forward, the team outlines a roadmap toward safe clinical integration: prospective clinical validation, testing in non-English environments, continuous calibration, fairness audits, and the establishment of model version management and post-deployment monitoring mechanisms. The goal, as Jinghao Liang describes it, is to promote the safe use of such systems for human-machine collaborative risk stratification, in which the AI-generated phenotype serves as a decision-support layer alongside clinician judgment rather than a replacement for it. The paper, Bio-Inspired Artificial Intelligence Multimorbidity Score for Human-Machine Clinical Risk Stratification, was published in Cyborg and Bionic Systems on September 30, 2026, and the work was supported by the Natural Science Foundation of Guangdong Province, the National Natural Science Foundation of China, and the R&D Program of Guangzhou National Laboratory.

Subject of Research: A large language model-based multimorbidity score for all-cause mortality risk stratification across international cohorts

Article Title: Bio-inspired artificial intelligence multimorbidity score for human–machine clinical risk stratification

Article References: Bio-inspired artificial intelligence multimorbidity score for human–machine clinical risk stratification. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, multimorbidity, risk stratification, all-cause mortality, UK Biobank, biological age, comorbidity indices, DeepSeek-V3, clinical decision support, artificial intelligence, proteomics, population aging