Towards Maps of Disease Progression: Biomedical Large Language Model Latent Spaces For Representing Disease Phenotypes And Pseudotime

In this study, we propose a scientific framework to detect capability among biomedical large language models (LLMs) for organizing expressions of comorbid disease and temporal progression. We hypothesize that biomedical LLMs pretrained on next-token prediction produce latent spaces that implicitly capture "disease states" and disease progression, i.e., the transitions over disease states over time. We describe how foundation models may capture and transfer knowledge from explicit pretraining tasks to specific clinical applications. A scoring function based on Kullback-Leibler divergence was developed to measure "surprise" in seeing specialization when subsetting admissions along 13 biomedical LLM latent spaces. By detecting implicit ordering of longitudinal data, we aim to understand how these models self-organize clinical information and support tasks such as phenotypic classification and mortality prediction. We test our hypothesis along a case study for obstructive sleep apnea (OSA) in the publicly available MIMIC-IV dataset, finding ordering of phenotypic clusters and temporality within latent spaces. Our quantitative findings suggest that increased compute, conformance with compute-optimal training, and widening contexts promote better implicit ordering of clinical admissions by disease states, explaining 60.3% of the variance in our proposed implicit task. Preliminary qualitative findings suggest LLMs latent spaces trace patient trajectories through different phenotypic clusters, terminating at end-of-life phenotypes. This approach highlights the potential of biomedical LLMs in modeling disease progression, identifying new patterns in disease pathways and interventions, and evaluating clinical hypotheses related to drivers of severe illness. We underscore the need for larger, high-resolution longitudinal datasets to further validate and enhance understanding of the utility of LLMs in modeling patient trajectories along clinical text and advancing precision medicine. Key PointsO_ST_ABSQuestionC_ST_ABSDo LLMs sensibly organize cilnical data with respect to applications in precision medicine? FindingsBiomedically-trained LLMs show increasing potential in promoting the organization of patient data to reflect disease progression. In a subcohort of OSA patients, maps derived from LLMs latent representations reveal traceable disease trajectories. MeaningMaps of disease progression offer an explanation to the utility of LLMs in precision medicine. Following current pretraining conventions in foundation modeling, scientific inquiry into these maps may help anticipate progress in applications of LLMs for healthcare.