A new mortality-risk system used a large language model in an unusual role. It did not read doctors' notes and directly announce who would die. Researchers converted structured health data into standardised English narratives, asked DeepSeek-V3 to generate 17 interpretable health phenotypes, and then placed those outputs into conventional Cox survival models.[1]
Across 770,439 adults in nine international longitudinal cohorts, the resulting LLM-MMS score reported C-indices from 0.67 to 0.86 in eight external cohorts and gains of 0.12 to 0.21 over recalibrated conventional comorbidity indices. In the UK Biobank development cohort, its C-index was 0.81, compared with 0.61 for the strongest traditional index in that analysis.[1][2]
The language model generated features; the survival model generated risk
The LLM-MMS pipeline
| Stage | Input | Output |
|---|---|---|
| 1. Harmonisation | Demographics, lifestyle, examinations, biomarkers and medical history from nine cohorts | A common data dictionary.Source 1 |
| 2. Textification | Structured variables mapped into controlled English health reports | Standardised narratives with rules intended to limit hallucinated facts.Source 1 |
| 3. LLM inference | Fixed prompt and DeepSeek-V3-0324; no model fine-tuning | Seventeen phenotypes, including overall and organ-system biological ages, frailty age and health grades.Source 1 |
| 4. Survival modelling | The 17 generated variables | Cox proportional-hazards prediction of all-cause mortality.Source 1 |
| 5. Validation | UK Biobank development plus eight external cohorts | Discrimination, calibration, Brier score, subgroup stability and decision-curve results.Source 1 |
That architecture matters when interpreting the result. The gain is not evidence that free-form language alone predicts mortality. It is evidence that an LLM-generated representation of many health domains carried information that fixed disease lists and their inherited weights missed.[1]
The comparison was broader than disease counting
Selected UK Biobank discrimination results
| Approach | C-index | What it used |
|---|---|---|
| Best conventional comorbidity index in the reported comparison | 0.61 | A fixed disease/weight framework.Source 1 |
| Random survival forest on raw structured variables | 0.725 | Non-linear survival modelling of the original variables.Source 1 |
| Elastic-net Cox model on raw variables | 0.771 | Regularised linear survival model.Source 1 |
| XGBoost survival model on raw variables | 0.778 | Tree-based modelling of the original variables.Source 1 |
| LLM-MMS | 0.810 | Seventeen LLM-derived phenotypes entered into a Cox model.Source 1 |
The raw-variable comparisons are important because they test a simpler explanation: perhaps LLM-MMS wins only because it receives richer inputs than the traditional scores. In the UK Biobank, it also exceeded three survival models trained directly on those structured variables. That supports an incremental representation effect, although it does not reveal which portion comes from the prompt, the language model, the 17-dimensional bottleneck or the downstream Cox model.[1]
Portability was the main test
External-validation results
| Measure | Value | Context |
|---|---|---|
| External cohorts | 8 | The system was developed in UK Biobank and evaluated across cohorts in several countries and age ranges.Source 1 |
| External C-index range | 0.67–0.86 | Discrimination varied by cohort but exceeded the recalibrated conventional comparators in each.Source 2 |
| Absolute gain over conventional indices | 0.12–0.21 | The comparator was the strongest recalibrated conventional index within each cohort.Source 2 |
| Observed-to-expected ratio | Generally 0.94–1.04 | Reported calibration was close to one across the external cohorts.Source 1 |
| Repeated model runs | Five | Reported metric variation was small under repeated inference with the fixed base model.Source 1 |
Traditional indices showed a familiar deployment failure: a weight fitted to one population can rank or calibrate risk poorly in another. Recalibration can repair average predicted risk without restoring the ordering of individuals. LLM-MMS was designed to create a more portable health representation before the survival model is fitted.[1]
Interpretability is useful, but not the same as clinical evidence
The generated phenotypes included organ-specific ages, frailty age and grades covering behaviour, diet, psychological health, medical history and social context. Input-masking analyses produced clinically coherent shifts, and high-risk groups showed proteomic patterns related to inflammation, metabolism and tissue damage. Those analyses make the score less opaque than one unexplained scalar.[1]
But a model-generated rationale is an audit trail of the model's representation, not proof that its biological explanation is correct. Chain-of-thought output can be unstable, incomplete or influenced by the prompt. A clinician should not treat a generated 'biological age' as a measured organ state unless it has its own validated meaning and uncertainty.[1]
Evidence required before clinical use
| Risk | Why the paper does not close it | Necessary next test |
|---|---|---|
| Prospective impact | The study is retrospective and predicts an outcome already recorded in cohorts. | Test whether using the score changes decisions and improves outcomes without harmful over-treatment.Source 1 |
| Model-version drift | The core engine is a named 2026 DeepSeek release. | Lock versions, revalidate after updates and preserve reproducible inference records.Source 1 |
| Language and geography | All cohorts were converted to standardised English narratives. | Validate other languages and health systems without assuming English textification is neutral.Source 1 |
| Fairness and subgroup calibration | Sex and age analyses were encouraging but do not cover every protected or clinical group. | Pre-specify subgroup audits, local recalibration and monitoring after deployment.Source 1 |
| Actionability | Mortality prediction does not say which intervention will help. | Connect risk components to evidence-based clinical actions and test them separately.Source 1 |
LLM-MMS is compelling because it reframes multimorbidity as a whole-system state rather than a disease count. Its nine-cohort validation is unusually broad for an LLM-derived clinical score. The responsible next step is not to ask the model for bedside verdicts. It is to test whether this representation remains calibrated, reproducible and clinically useful when real decisions depend on it.[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: PubMed record US National Library of Medicine · 30 Sep 2026 · Accessed 16 Oct 2026