A fixed comorbidity index works like a scorecard: count selected diseases, apply predetermined weights and add the result. That is useful when the population resembles the one used to build the index. It can fail when age, disease patterns and available measurements change. Researchers have tested a different approach across 770,439 adults in nine international cohorts: convert harmonised health data into standardised narratives, then ask a language model to derive a multidimensional multimorbidity phenotype.[1][2]
The system, called LLM-MMS, was evaluated for all-cause mortality. It was not fine-tuned on each cohort. Across eight external cohorts, its C-index ranged from 0.67 to 0.86 and exceeded recalibrated conventional comorbidity indices by 0.12 to 0.21. Observed-to-expected ratios remained close to 1, indicating that predicted average risk was broadly aligned with observed outcomes in the reported tests.[1][2]
The narrative is a representation layer
The input begins as structured information: demographics, lifestyle, examinations, biomarkers and medical history. A controlled mapping converts those fields into standardised English clinical reports, with safeguards intended to prevent the model from inventing facts. The language model then produces features such as overall biological age, organ-system ages, frailty age and health gradings. Those outputs feed the final survival model.[1][3]
Three different ways to turn multimorbidity into risk
| Approach | What it represents | Main weakness |
|---|---|---|
| Traditional comorbidity index | A fixed list of diagnoses with fixed weights | Weights and disease lists can travel poorly across populations.Source 1 |
| Conventional machine learning on raw fields | Patterns in the original structured variables | Can capture nonlinearity but may not create an interpretable cross-domain phenotype.Source 1 |
| LLM-MMS | A standardised narrative transformed into biological-age and health-domain features | Depends on prompt, model version, language mapping and governance of generated representations.Source 1Source 3 |
How to read the performance numbers
Reported validation results
| Measure | Value | Context |
|---|---|---|
| Participants | 770,439 adults | Nine longitudinal cohorts spanning several countries and age profiles.Source 1Source 2 |
| External discrimination | C-index 0.67–0.86 | A higher value means better ordering of people who experience the outcome earlier versus later.Source 1Source 2 |
| Gain over conventional indices | 0.12–0.21 C-index points | Reported after conventional comparators were recalibrated in each cohort.Source 1 |
| Calibration | Observed-to-expected ratios near 1.00 | Average predicted and observed mortality were close in the reported validation.Source 1Source 2 |
| Fine-tuning | None for the cohort narratives | The method used prompt-based inference rather than model retraining for each population.Source 1Source 3 |
A C-index is a ranking measure. If two people can be ordered by who experiences death earlier during follow-up, the model receives credit when its risk ordering agrees. A value of 0.81 does not mean an individual has an 81% chance of receiving a correct forecast. Calibration answers another question: whether predicted probabilities match the event rate. A useful risk model needs both ranking and calibration, then prospective evidence that acting on it improves care.[1]
The model is broader than a diagnosis count
LLM-MMS incorporates behaviour, physiology, laboratory values and functional information alongside diagnoses. That breadth may explain why it predicts mortality better than a fixed disease count. It also changes what the score is. It is closer to a general prognostic health phenotype than a pure measure of multimorbidity, so it may not be suitable for every task that currently uses a comorbidity index, such as administrative risk adjustment.[1]
The scale and cross-cohort performance make the study important. The deployment question is harder. The work is anchored in standardised English narratives and a particular model pipeline; real health systems contain missing data, changing coding practices, multilingual records and decisions with unequal consequences. LLM-MMS shows that language models can be used as representation engines for population risk. It does not yet show that a hospital should use the score to make an individual treatment decision.[1][2]
Sources & further reading
- Bio-Inspired Artificial Intelligence Multimorbidity Score for Human–Machine Clinical Risk Stratification Cyborg and Bionic Systems · 30 Sep 2026 · Accessed 16 Oct 2026
- Bio-Inspired Artificial Intelligence Multimorbidity Score for Human-Machine Clinical Risk Stratification PubMed · 30 Sep 2026 · Accessed 16 Oct 2026
- Bio-inspired artificial intelligence multimorbidity score for human–machine clinical risk stratification EurekAlert · 12 Oct 2026 · Accessed 16 Oct 2026