Predicting Oncology Readmissions: A Machine Learning Approach Using MIMIC-IV-ED Database

Abstract

This study investigates the artificial intelligence (AI) impact on hospital admission classification, specifically focusing on the identification of patients discharged from the emergency room who required readmission within seven days. The research aims to scrutinize potential inconsistencies in human classification within medical triage records and understand the disparities between human decisions and AI predictions. Class balancing techniques like Class Weights and K-Fold Cross-Validation (AI model) and SMOTE, Random Oversampling (The Random Forest) were employed to address imbalanced data, improving the model's ability to generalize across different patient groups. Following the delimitation of cancer cases, the dataset was refined to 7,525 rows, involving 1,945 unique patients. After machine learning training for screening and hospital admission, 80 cases (4.11%) of inconsistent data were observed in triage records, highlighting the critical importance of accurate classification in preventing unnecessary hospital readmissions. The study analyzed key features influencing predictions using the Random Forest algorithm, providing insights to enhance patient management and reduce mortality. The Random Forest classifier exhibited a median accuracy of 72.79% (Mean Precision: 0.76%, Mean Recall: 0.84%, AUC Score: 0.76% and Mean F1-Score: 0.80%), while the neural network showed a Mean Accuracy of 70.3% (Mean Precision: 0.80, Mean Recall: 0.71, AUC Score: 0.75 and Mean F1-Score: 0.75), indicating consistent performance. Feature importance analysis emphasized vital signs, showcasing AI's capability to identify inconsistencies and the need for evolution in triage practices.