Forecasting adverse surgical events using self-supervised transfer learning for physiological signals

Hugh Chen, Scott M. Lundberg, Gabriel G. Erion, Jerry H. Kim, Su-In LeeNpj Digit. Medicine 2021 · DOI

TL;DR

This paper introduces PHASE, a novel self-supervised transfer learning framework that embeds physiological signals from electronic health records to predict adverse surgical events with improved accuracy and reduced computational cost. PHASE offers a way for institutions with limited resources to leverage powerful predictive models without sharing sensitive patient data.

What Problem Does This Paper Address?

Millions of surgical procedures occur annually, and a significant portion result in adverse events, many of which are preventable. Accurate prediction of these events could significantly improve patient outcomes. However, training robust predictive models is challenging due to data scarcity, especially in smaller hospitals lacking computational resources and large, privacy-protected datasets. Traditional methods often rely on handcrafted features or deep learning models trained on raw data, which can be suboptimal. Transfer learning, a technique proven effective in domains like medical imaging and natural language processing, offers a promising avenue for physiological signals, but its application in this area, particularly with self-supervised learning and explainability, remains underexplored.

Methods

The core of this work is the PHASE (PHysiologicAl Signal Embeddings) framework. PHASE employs a two-stage approach:

  1. Upstream Embedding Model: This stage uses Long Short-Term Memory (LSTM) networks to learn generalizable features (embeddings) from raw physiological signals. Crucially, these LSTMs are trained using self-supervised learning. Instead of predicting the adverse event directly, the LSTMs are trained on tasks derived from the physiological signals themselves, such as predicting future signal values ("next") or the minimum future signal value ("min"). This self-supervision allows the models to learn robust representations without direct supervision on the outcome of interest.
  2. Downstream Prediction Model: The extracted embeddings are then concatenated with static patient data (e.g., age, weight) and fed into a downstream model, typically a Gradient Boosted Tree (GBT), to predict the adverse surgical outcome.

A key innovation is the transfer learning capability. PHASE allows pre-trained upstream embedding models from a "source" hospital to be used to embed signals from a "target" hospital, even if the target hospital has fewer resources or different data characteristics. This "transferred embedding" approach significantly reduces computational burden and data requirements for the target institution. The authors also explore fine-tuning these transferred models on the target data for even better performance and faster convergence.

Datasets

The study utilized three distinct datasets:

  • Two Operating Room (OR) datasets: Collected from the Anesthesia Information Management System (AIMS) at two different medical centers (denoted OR₀ and OR₁). These datasets contain minute-by-minute physiological signals and static patient information.
  • One Intensive Care Unit (ICU) dataset: A subset of the publicly available Medical Information Mart for Intensive Care (MIMIC-III) dataset (denoted ICUM). For this dataset, due to significant missingness in many signals, only the minute-by-minute SAO₂ (blood oxygen saturation) signal was consistently used.

These datasets were used to forecast six adverse surgical outcomes: hypoxemia, hypocapnia, hypotension, hypertension, phenylephrine administration, and epinephrine administration.

Key Insights and Ideas

  • Self-supervised learning is powerful for physiological signals: The "next" and "min" self-supervised tasks for LSTMs proved to be highly effective in learning generalizable embeddings, outperforming random initialization, autoencoders, and even traditional handcrafted features.
  • Transfer learning without data sharing is feasible and effective: PHASE enables institutions to leverage pre-trained models from other hospitals, addressing data scarcity and privacy concerns. This is a significant step towards democratizing advanced predictive analytics in healthcare.
  • Per-signal embedding is advantageous: Training separate LSTMs for each physiological signal, rather than a single model for all signals, proved more robust, especially in heterogeneous settings where target hospitals might have different sets of available signals.
  • Fine-tuning accelerates convergence and improves performance: While transferred embeddings offer computational savings, fine-tuning these models on target data leads to faster convergence and often superior predictive accuracy.
  • Explainability is crucial for clinical adoption: The authors demonstrate that PHASE models can be interpreted using local feature attribution methods, validating that the learned patterns align with clinical knowledge and building trust in the predictions.

Key Results

  • PHASE significantly outperformed conventional approaches (e.g., GBTs with raw or handcrafted features, LSTMs on raw data) in predicting six adverse surgical outcomes across the three datasets. For hypoxemia, PHASE achieved an approximate 5% improvement in average precision over the state-of-the-art.
  • In a transfer learning setting, PHASE achieved significantly higher prediction accuracy at lower computational cost compared to traditional methods.
  • Fine-tuned PHASE models showed faster convergence and improved performance compared to both standard and transferred embedding models.
  • Local feature attribution methods confirmed that PHASE models learn clinically relevant patterns, identifying important physiological variables for each outcome that align with existing medical literature.

Limitations and Future Work

  • Privacy concerns with model inversion: While sharing models is more private than sharing data, the authors acknowledge the theoretical possibility of model inversion attacks to infer training data. They mention efforts with differential privacy but found it challenging to train effective models.
  • Sampling frequency: The current models are designed for minute-by-minute physiological signals. Adapting them to accommodate varying or irregular sampling frequencies would require further research.
  • Threshold selection: The definition of adverse events often relies on specific thresholds (e.g., for hypoxemia). The paper notes that identifying optimal thresholds is a separate research direction.
  • Heterogeneous features: While per-signal embedding helps, further work on applying PHASE to scenarios with significant differences in available features between source and target hospitals is suggested.
  • Per-user embeddings: The authors suggest exploring per-user embeddings or incorporating user IDs/demographics to capture more personalized patterns, as aggregating embeddings across time might lose crucial information.

Summary updated

BibTeX
@article{chen2021forecasting,
  author = {Hugh Chen and Scott M. Lundberg and Gabriel G. Erion and Jerry H. Kim and Su-In Lee},
  title = {{Forecasting adverse surgical events using self-supervised transfer learning for physiological signals}},
  journal = {Npj Digit. Medicine},
  year = {2021},
  doi = {10.1038/s41746-021-00536-y},
}
Figure