Reproduce “Forecasting adverse surgical events using self-supervised transfer learning for physiological signals” on OpenResearch

About this paper

Hugh Chen, Scott Lundberg, Gabe Erion, Jerry H. Kim, Su-In Lee

A framework called PHASE (PHysiologicAl Signal Embeddings) leverages self-supervised transfer learning on physiological signals to forecast adverse surgical events. This method significantly improves predictive accuracy for outcomes like hypoxemia and hypotension while enabling robust model transferability across different hospital datasets, potentially forecasting an additional million hypoxemic events annually in the US.

Key insights

• Self-supervised learning tasks, particularly predicting future signal values or minimums, extract more robust and generalizable features from physiological time series than traditional methods or simple autoencoders.
• Pre-trained embedding models can be effectively transferred across distinct hospital datasets, maintaining high predictive performance and significantly reducing computational costs for target institutions while addressing data privacy concerns.
• Fine-tuning transferred embedding models on target hospital data further enhances predictive performance, achieving the best results by combining general pre-trained knowledge with domain-specific adaptation.
• The per-signal embedding design facilitates model interpretability using Shapley values, aligning important features with clinical knowledge and building trust in AI-driven predictions.

Summary by alphaXiv.