Modeling Dynamic Intraoperative Physiological Signals via Spatio-temporal Graphs for Predicting Postoperative Adverse Outcomes

Co-Supervised by: Dr. Linlin Jia and Dr. Xiao Ning (Lecturer at the China Pharmaceutical University)

If you are interested in this topic or have further questions, do not hesitate to contact linlin.jia@unibe.ch.

Background / Context

Perioperative adverse outcomes, such as postoperative complications, organ dysfunction, and mortality, remain a significant concern in surgical care. Traditional risk assessment often relies on static preoperative indicators and single-point intraoperative measurements, which may fail to capture the complex dynamics of physiological responses during surgery. Continuous intraoperative monitoring provides rich time-series data, including blood pressure, heart rate, and oxygen saturation, offering an opportunity to investigate how dynamic signal patterns relate to postoperative outcomes, such as AKI and AKD. Understanding these relationships may improve risk stratification and guide real-time intraoperative management.

In this project, we aim to bridge this gap by modeling intraoperative physiological signals as spatio-temporal graphs, where each node represents a patient, and edges capture the inter-patient and inter-signal relations. Using spatio-temporal graph neural networks (ST-GNNs) and Transformer-based temporal models, we will explore how dynamic signal patterns can predict postoperative adverse outcomes. Meanwhile, we will examine the dynamic correlations among patients, uncovering interpretable temporal dependencies that can inform real-time intraoperative risk prediction and personalized intervention strategies.

Research Question(s) / Goals

In this project, we would like to answer the following research questions:

  • Q1: Can spatio-temporal representations of intraoperative physiological signals improve the prediction of postoperative adverse outcomes compared to conventional static or sequence-based models?
  • Q2: Which dynamic or relational features among physiological signals are most indicative of postoperative risk?
  • Q3: How can ST-GNNs and Transformer-based temporal attention models be adapted for multi-signal physiological data with varying lengths and sampling rates?

To answer these questions, we define the following goals

  • G1: Extract and represent intraoperative physiological signals in a unified spatio-temporal graph format.
  • G2: Develop and evaluate ST-GNNs and Transformer-based models for predicting postoperative adverse outcomes.
  • G3: Compare the proposed models against static or sequence-based baselines
  • G4: Analyze and visualize learned temporal dependencies for clinical interpretability.

Approach / Methods

  • Data Source: Utilize the INSPIRE perioperative database, which contains comprehensive intraoperative physiological signals and postoperative outcome data.
  • Data Processing: Extract signal characteristics, including mean, variability, extreme values, cumulative deviation, and temporal trends of blood pressure, heart rate, oxygen saturation, etc.
  • Modeling Approaches: Model temporal and relational dependencies between multiple intraoperative signals and postoperative outcomes using ST-GNNs and transformer-based models.
  • Evaluation: Predict postoperative AKI/AKD or other adverse outcomes, interpret model attention or graph weights to highlight critical periods or signals, and group patients based on learnable dynamic patterns tracing back to intraoperative signals and to identify common trajectories.

Expected Contributions / Outcomes

  • A novel spatio-temporal graph framework for modeling intraoperative physiological dynamics.
  • Benchmarking of ST-GNNs and Transformer models against traditional approaches for postoperative risk prediction
  • Quantitative description of which features (e.g., duration of hypotension, heart rate variability) most strongly correlate with postoperative events. Identification of intraoperative signal dynamic patterns associated with higher risk of postoperative adverse outcomes
  • Open and reproducible codebase, including preprocessing pipelines and model implementations
  • A thesis and potentially a research paper submission

Required Skills / Prerequisites

  • Solid knowledge of statistics and machine learning, with an understanding of deep learning, learning on graphs, and time-series analysis being a plus.
  • Strong programming skills, preferably in Python and PyTorch.
  • Interest in biomedical data analysis and healthcare applications.
  • Proficiency in English for effective communication and presentations at the research level.

Possible Extensions

  • Build a real-time intraoperative risk prediction system integrating streaming data.
  • Explore graph-transformer hybrids for improved long-term temporal dependency modeling.
  • Build a foundation for developing risk stratification tools that integrate dynamic intraoperative monitoring into clinical decision support systems.

Further Reading / Starting Literature

  • Chen, L., Hong, L., Ma, A., Chen, Y., Xiao, Y., Jiang, F., … & Zhou, J. (2022). Intraoperative venous congestion rather than hypotension is associated with acute adverse kidney events after cardiac surgery: a retrospective cohort study. British Journal of Anaesthesia, 128(5), 785-795.
  • Yu, B., Yin, H., & Zhu, Z. (2017). Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875.
  • Lim, L., Lee, H., Jung, C. W., Sim, D., Borrat, X., Pollard, T. J., … & Lee, H. C. (2024). INSPIRE, a publicly available research dataset for perioperative medicine. Scientific Data, 11(1), 655.