Skip to content

Multimodal Disease Risk Prediction

Summary

Multimodal Disease Risk Prediction represents a shift in clinical informatics away from evaluating diseases as isolated endpoints. By modeling joint physiological systems, electronic health records (EHR), genetics, lifestyle, and environmental factors across continuous time scales, this approach captures shared disease susceptibilities. Implementations include RisQ, which scales input modalities and disease targets to improve risk prediction, and Delphi-2M, which extends the paradigm to generate synthetic health trajectories.

Key Principles

Multimodal disease risk prediction is governed by several core properties: - Shared Susceptibility Structure: Human diseases frequently co-occur due to overlapping genetic, lifestyle, and environmental risk factors. Modeling these diseases jointly allows the model to learn a shared representation of health that captures cross-disease patterns. - Multimodal Integration: Combining diverse data types—such as demographics, blood and urine biomarkers, physical measurements, self-reported lifestyle surveys, diagnostic histories, and polygenic risk scores—provides a more complete picture of an individual's health trajectory. - Continuous Temporal Modeling: Disease risk changes dynamically over time. Incorporating continuous time representations (e.g., via sinusoidal encodings or waiting time heads) allows predictions to be made across arbitrary horizons. - Zero-Shot Transfer: Generalization to unseen disease chapters is enabled by using natural language embeddings of disease targets, allowing a single model to predict risk for conditions it was not explicitly trained to classify. - Generative Trajectory Sampling: Autoregressive models can iteratively sample the identity and continuous time to the next clinical event, enabling long-term disease burden projection and privacy-preserving synthetic data generation.

Clinical and Genetic Grounding

  • Risk Group Profiling: Unsupervised clustering of multimodal disease-risk representations reveals distinct, reproducible patient trajectories that align with specific physiological states (e.g., cardiometabolic, tobacco-associated, and respiratory-enriched profiles).
  • Genetic Associations: The learned risk structure shows a strong biological grounding. Exome-wide gene-level burden analyses link the risk profiles to rare loss-of-function variants in key physiological genes (such as OMA1, HBB, LDLR, and CASR).
  • Comorbidity Clustering and Attribution: Applying explainable AI methods (e.g., SHAP) to multi-disease models reveals distinct disease clusters and distinguishes acute clinical events (which drive short-term risk) from chronic conditions (which exert a long-term, multi-year influence).

Key Implementations

  • RisQ: A discriminative framework using a transformer encoder and cross-attention decoder to predict risks for arbitrary diseases and horizons using natural-language queries.
  • Delphi-2M: An autoregressive generative transformer (GPT-2 based) that treats health histories as sequences of disease events and predicts both the identity and continuous timing of future diagnoses.

Citations