An LLM turned 17 health phenotypes into a mortality score that travelled across nine cohorts

A conventional comorbidity index usually begins with a fixed list: particular diagnoses receive particular weights, and the total becomes a risk score. LLM-MMS takes a different route. It turns harmonised demographics, lifestyle, examinations, biomarkers and medical history into a standardised English report, asks a large language model for 17 predefined health phenotypes, and then feeds those outputs into a survival model.[1]

The framework was developed and tested on 770,439 adults from nine international longitudinal cohorts. The language model was not fine-tuned for each cohort. The researchers instead held the prompt and output schema steady, developed the survival model in UK Biobank and evaluated transportability across eight external populations.[1][2]

The model’s job was representation

The 17 outputs included biological ages for major organ systems, frailty age, psychological and dietary health grades, behavioural health, medical and family history, income grading and an overall biological age. Those values were then combined through fixed Cox coefficients to estimate all-cause mortality risk. This architecture separates two jobs: the language model compresses a complex narrative into structured phenotypes, while the survival model performs the risk calculation.[1]

Why fixed disease lists can fail when populations change

Traditional indices are useful because they are simple and interpretable. Their weakness is that a disease list and its weights may encode one clinical era, one population or one purpose. The same diagnosis can carry different implications depending on age, function, biomarkers, behaviour and the conditions surrounding it. LLM-MMS attempts to represent that broader context before risk is calculated.

Across the external cohorts, observed-to-expected ratios were close to one, ranging from 0.94 to 1.04, and Brier scores were lower than those of the conventional indices. Performance remained comparatively stable across sex and age groups, including adults younger than 60, where some disease-counting tools have a floor problem because severe diagnoses are less common.[1]

Interpretability is broader than a fluent rationale

The study did not rely only on model-written explanations. It masked prespecified inputs—such as smoking, inflammatory markers, blood pressure, body mass index and kidney dysfunction—and measured how the 17 outputs and final risk score changed. The direction of those changes was clinically coherent, and proteomic analysis linked higher predicted burden to inflammatory, metabolic and tissue-damage pathways.[1]

That does not prove the language model discovered a causal biological mechanism. It shows that the score’s internal response and external molecular associations were not obviously arbitrary. A convincing clinical interpretation still requires checking whether the score changes useful decisions, whether errors cluster in particular groups and whether model updates alter the mapping.

The most interesting contribution is architectural. LLM-MMS uses language not as a substitute for statistics but as a representation layer between heterogeneous health records and a conventional survival model. Its success suggests that the hard part of transportability may be learning a richer common description of health—not merely fitting another set of weights to the same disease list.

Report an error or suggest a correction