mims-harvard/Raindrop

Combines graph neural networks with temporal self-attention to model inter-sensor dependencies in irregularly sampled multivariate time series, using neural message passing where each sample is represented as a graph with sensors as nodes. The architecture generates observation embeddings that aggregate into sensor-level representations through attention mechanisms, enabling robust handling of missing sensors and variable sampling rates. Evaluated on clinical datasets (PhysioNet P19, P12) and activity monitoring (PAMAP2) with specialized settings including leave-sensor-out robustness testing.

219 stars. No commits in the last 6 months.