The integrated multiple event representation framework (IMERF): a case study on critically-ill patients

This study introduces the Integrated Multiple Event Representation Framework (IMERF), a novel methodological approach for developing risk prediction models for multiple clinical events. Using a two-stage process involving multi-task learning and dimensionality reduction, IMERF creates a visual representation of predicted event risks and identifies clusters based on overlapping risks. The proposed framework is showcased through a case study modelling nine adverse events in critically ill patients admitted to intensive care units (ICUs). Stage 1 was implemented using convolutional neural networks, which displayed superior performance to logistic regression and random forest algorithms. The generative topographic mapping (GTM) algorithm was implemented in stage 2 for data visualisation and clustering. It revealed clear patterns of adverse event risk clusters. GTM in combination with class activation maps was also employed to trace input factors influencing cluster membership, highlighting distinct risk profiles among patients. Macro-clusters representing distinctive combinations of adverse event risk levels were also identified by performing a hierarchical clustering on the GTM results. In conclusion, IMERF could represent a significant advancement in multiple event risk modelling by enabling simultaneous prediction and characterisation of overlapping events and providing an interpretable framework for understanding their complex patterns. Its application in ICUs underscores its potential for broader clinical use, including modelling clusters of conditions or multiple instances of events.