Summary
A medRxiv preprint describes a method that distils multimorbidity reasoning from large language models into smaller models using synthetic cohorts shaped like UK Biobank data. The resulting models achieved Spearman correlations of 0.75–0.89 and produced scores associated with improved survival prediction in a UK Biobank analysis.
Researchers from Case Western Reserve University have described a privacy-aware way to use large language models (LLMs) for multimorbidity scoring, a task that estimates the burden of multiple health conditions in an individual. The medRxiv preprint transfers reasoning from larger LLMs into smaller models using synthetic cohorts designed to preserve the statistical distributions of UK Biobank data rather than exposing real patient records to the teacher models.
The resulting student models, called CoLLMs, showed Spearman correlations of 0.75–0.89 with the knowledge being transferred. When their scores were applied to real UK Biobank data, the researchers reported better survival prediction, with a concordance index (C-index) of up to 0.91, and SNP heritability of approximately 0.05.
The work is a preprint and presents a computational evaluation using biobank data, rather than evidence from a prospective clinical trial or patient treatment study.
From synthetic cohorts to UK Biobank analysis
Multimorbidity scoring is commonly used to represent the combined effect of several conditions rather than examining diseases one at a time. Traditional scoring systems generally rely on predefined rules. The researchers argue that LLMs can encode richer clinical knowledge, but that directly evaluating them on large biobank datasets creates privacy challenges because those datasets contain sensitive health information.
Their framework uses two model roles. Larger LLMs act as “teacher” models, supplying multimorbidity reasoning. Smaller student models—the CoLLMs—learn from that reasoning through synthetic cohorts whose distributions are designed to resemble those in the UK Biobank. According to the authors, this allows the knowledge-transfer process to avoid exposing real patient data during distillation.
The CoLLMs were then used to generate multimorbidity scores for real UK Biobank data. The study was conducted under the ethical approval covering the UK Biobank resource from the North West-Haydock Research Ethics Committee of the UK Health Research Authority, with informed consent from UK Biobank participants.
What the evaluation found
The reported Spearman correlation range of 0.75–0.89 measures how closely the distilled models preserved the ranking or pattern of the teacher models’ outputs. The authors describe this as high-fidelity knowledge transfer. An independent LLM-as-a-Judge assessment also found the distilled knowledge clinically significant, while showing substantial differences among the teacher models.
The researchers next tested whether the generated multimorbidity scores were useful for survival prediction in the UK Biobank data. They report a C-index as high as 0.91. The C-index is a ranking measure: it assesses how consistently a model assigns higher predicted risk to people who experience an event earlier than people who experience it later. The abstract describes the CoLLM-derived scores as improving survival prediction, but does not provide the detailed comparator or cohort-size information needed to interpret the headline value fully.
The scores also showed SNP heritability of approximately 0.05. In this context, SNP heritability refers to the proportion of variation in the score associated with measured genetic variants under the study’s statistical model. It describes a property of the score and its data, not a direct estimate of how strongly genetics causes multimorbidity in an individual.
The main significance of the approach is methodological. It offers a way to transfer some of the information encoded by large clinical language models into smaller models while using synthetic data during the sensitive distillation stage. That could make large-scale health-data modelling easier to conduct under privacy constraints. The preprint’s results support further evaluation of the framework, while its evidence remains at the level of model assessment and secondary analysis of UK Biobank data.