Details

Intraoperative Mean Arterial Pressure Time Series Forecasting Using Deep Learning Methods

Year: 2026

Term: Winter

Student Name: Josiah Mak

Supervisor: Matthew Holden

Abstract: Intraoperative hemodynamic instability, such as intraoperative hypotension and hypertension, is associated with adverse postoperative outcomes, making accurate real-time mean arterial pressure (MAP) forecasting important for improving patient safety. This study pro- poses a machine learning framework for multi-horizon MAP forecasting using intraoperative physiological time-series data. A total of 570 non-cardiac surgical cases were used, incorporating vitals such as arterial pressure, heart rate, electroencephalogram (EEG), and cardiac output, along with auxiliary patient features. The proposed method is based on TSMixer architecture, a multilayer perceptron model designed to capture temporal and cross-feature relationships through its mixing layers. This study also explores a novel approach to TSMixer that includes recon- structing the input instead of just forecasting the output. Variants of TSMixer, such as including joint training and auxiliary patient features, are evaluated against baseline models such as Long Short-Term Memory (LSTM) and Temporal Fusion Transformer (TFT). Results show that TSMixer-based models consistently outperform baseline methods across 1-, 3-, and 5-minute forecasting horizons. The base TSMixer model achieved up to 3.0% improvement over LSTM and over 54.8% improvement over TFT at the 1-minute horizon. These findings suggest that structured MLP-based models are more effective for intraoperative MAP forecasting.