Towards personalised early prediction of Intra-Operative Hypotension following anesthesia using Deep Learning and phenotypic heterogeneity

Abstract

Intra-Operative Hypotension (IOH) is a haemodynamic abnormality that is commonly observed in operating theatres following general anesthesia with associated life-threatening post-operative complications. Here, we apply Deep Learning (DL) and more specifically Long Short Term Memory (LSTM) models across different patient groups for the classification of IOH events five minutes before onset, using Electronic Health Records (EHR) and time-series intra-operative data of 604 patients that had undergone colorectal surgery. K-means clustering was used to group patients based on pre-clinical data. As part of a sensitivity analysis, the model was also trained on patients clustered according to Mean arterial Blood Pressure (MBP) time-series trends at the start of the operation using K-means with Dynamic Time Warping. The baseline LSTM model trained on all patients yielded an Area Under the Curve (AUC) value of 0.83. In contrast, training the model on smaller sized clusters (grouped by EHR) improved the AUC value (0.85). Similarly, the AUC was increased by 4.8% (0.87) when training the model on clusters grouped by MBP. The encouraging results of the baseline model demonstrate the applicability of the approach in a clinical setting. Furthermore, the increased predictive performance of the model after being trained using a clustering approach first, paves the way for a more personalised patient stratification approach to IOH prediction using clinical data.