Machine Learning–Driven Identification of Determinants Affecting Clinical Outcomes in Outpatient Surgery

Abstract

The increasing use of Artificial Intelligence (AI) and Machine Learning (ML) techniques in healthcare delivery organizations has opened up many possibilities to predict patients' outcomes and to optimize throughput in those organizations. The current paper introduces a simulation modeling framework to predict the outcomes of patients in outpatient surgery settings using the power of ML algorithms with a special focus on ensemble approaches to Random Forest models. Based on the ideas from Process Mining (PM) and predictive analytics in perioperative healthcare systems, the current paper proposes the process-aware modeling framework that consists of four steps: 1) Creation of synthetic data sets based on empirical probability distributions which are used to mimic the real-world patient flows, 2) Feature engineering and process-aware variables creation (including pre-operative risk factors – ASA physical status, intra-operative parameters and temporal metrics obtained from PM), 3) prediction of Length of Stay (LOS) and complication occurrence using the Random Forest Regressors and Random Forest Classifiers, and 4) system-level operational performance assessment based on the Key Performance Indicators (KPIs) which evaluate the effectiveness and efficiency of the workflow and bed occupancy. Applying the proposed modeling framework to synthetic patient flows results in identification of three most important factors affecting the post-surgical recovery: Surgical duration time, pre-operative physical status of a patient and process-aware waiting times. The predictive model shows high predictive performance: R²=0.81 and MAE=0.29 days for length-of-stay estimation and strong discriminative power for complication prediction.