Enriching Predictive Process Monitoring with Discovery and Validation Analysis of Process-Oriented Features: An Exploration in Adverse Patient Outcomes Prediction

Abstract

Static features from Electronic Health Records (EHRs) are widely used for predicting adverse patient outcomes. However, these features often ignore the depth of data in healthcare processes. EHRs containing treatment data offer a unique opportunity to integrate process-oriented features into predictive models. Event logs in EHRs are prone to anomalies like incorrect activity labels, necessitating preprocessing for accurate predictions. Our methodology begins with a specialised event log preprocessing approach. This process refines raw event logs by identifying and correcting problematic activity labels and preparing them for further analysis. Subsequently, we introduce a Predictive Process Monitoring (PPM) framework comprising three key components. Each component enhances static and process-oriented features from the refined event logs and patient data. The first component focuses on identifying process-oriented features, establishing a basis for in-depth analyses of intricate healthcare processes. The second component employs revised statistical methods to validate the relevance and consistency of these dynamic features for patient outcome prediction. This validation enhances model interpretability and readiness for practical use. Using a customised Deep Learning architecture, the final component efficiently combines dynamic features with traditional static ones, augmenting the PPM framework’s predictive accuracy and practicality. We demonstrate our framework's effectiveness using cardiovascular patient EHRs, showing its superior predictive accuracy over traditional methods reliant solely on static features. Our approach's precision and timely predictions highlight the significance of including process-oriented features in adverse patient outcome predictions. This work paves the way for future use of EHR process information to improve healthcare services.