Conformance-Aware Predictive Process Monitoring for Early Detection of Sepsis Deterioration Using Incomplete Care Pathways

TL;DR

This paper introduces a new framework that uses process mining and machine learning to detect sepsis deterioration early by analyzing deviations in care pathways.

Contribution

The novel CAPPM framework integrates process mining with predictive modeling to detect sepsis deterioration using incomplete care pathways.

Findings

Incorporating conformance and pathway-based features improved predictive performance over traditional models.

Adaptive Boosting and Gradient Boosting achieved AUROC values of 0.744 and 0.731, respectively.

Early deviations in care pathways provide meaningful signals for predicting sepsis deterioration.

Abstract

Background/Objectives: Sepsis is a leading cause of morbidity and mortality due to its rapid progression and variability in care delivery. While existing predictive models estimate sepsis risk using clinical variables, they typically rely on static attributes and overlook temporal, behavioral, and process-related characteristics of care pathways. In particular, deviations from recommended protocols and process inefficiencies are rarely incorporated into early deterioration prediction. This study proposes a Conformance-Aware Predictive Process Monitoring (CAPPM) framework to enable early detection of sepsis deterioration using incomplete care pathways. Methods: The proposed framework integrates process mining with predictive modeling. Using the publicly available Sepsis Cases Event Log, we first discovered the reference care pathway and generated prefix-level representations of ongoing cases. Temporal and behavioral features were engineered alongside alignment-based and declarative conformance metrics to quantify pathway deviations. These features were used to train and evaluate multiple supervised learning models, including Adaptive Boosting and Gradient Boosting. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUROC). Results: Incorporating conformance and pathway-based features improved predictive performance compared to models relying solely on traditional attributes. Adaptive Boosting and Gradient Boosting achieved the strongest results, with AUROC values of 0.744 and 0.731, respectively, demonstrating enhanced early detection ability. Conclusions: The findings indicate that early deviations in care pathways and temporal progression patterns provide meaningful predictive signals for sepsis deterioration. Integrating process mining with machine learning offers a promising approach for time-critical clinical decision support and proactive intervention.

Figures