New method predicts cancer risk from health records
GraphBEAT uses temporal graphs to identify risks for pancreatic, liver, and intestinal cancers.
GraphBEAT is a temporal graph representation learning framework. It helps with early risk stratification for gastrointestinal cancers.
The system uses longitudinal electronic health records (EHRs). These records contain diagnostic and laboratory information about patient health.
The framework groups clinical events into time buckets. It transforms observed features into patient-specific, temporally directed graphs.
Each node encodes feature identity, observed value, and temporal position. A context-aware, time-decay message-passing mechanism models transitions between successive buckets.
Multi-head attention pooling generates patient-level representations. This method was tested on multi-site longitudinal diagnosis codes and laboratory tests.
GraphBEAT predicts pancreatic, liver, and intestinal cancer at 12-, 24-, and 36-month diagnostic lead times. It requires over 250x fewer floating-point operations per patient than Transformer and long short-term memory (LSTM) baselines.
Key facts
- GraphBEAT predicts pancreatic, liver, and intestinal cancer at 12-, 24-, and 36-month lead times.
- The method uses over 20x fewer parameters than Transformer and long short-term memory (LSTM) baselines.
- GraphBEAT requires over 250x fewer floating-point operations per patient than Transformer and long short-term memory (LSTM) baselines.