An Explainable AI Framework for Predictive Healthcare Intelligence, Clinical Risk Assessment, and Data-Driven Medical Decision Support

Abstract

In patients with diabetes, it is still a significant issue for healthcare systems that the patients get admitted again to the hospital, and it is a sign of clinical complexity, repeated healthcare utilization and also of the burden of resources. This work created an XAI model for predictive healthcare intelligence, clinical risk assessment and data-driven medical decision support from hospital encounter data of patients with diabetes. 99,343 records were analyzed, excluding those involving death or hospice care. The readmission to hospital within 30 days was considered as the primary prediction outcome. Logistic regression and XGBoost models were tested with patient level train-test split, and the performance was assessed with respect to ROC-AUC, precision-recall AUC, accuracy, precision, sensitivity and F1-score. The feature importance of permutation was used to explain the model. XGBoost showed slightly better discrimination compared to logistic regression, with the ROC-AUC scores 0.643 and 0.638, respectively. The five most influential predictors of early readmission were previous inpatient utilization, primary diagnosis group, hospital length of stay, age, and number of diagnoses. Prior inpatient use was significantly associated with risk of readmission, with 5 or more inpatient visits resulting in a 37.13% risk of readmission. The results show that the explainable machine learning approach can be beneficial in transparent clinical risk stratification not only in terms of predictive performance, but also by providing interpretable information about contributing factors. The moderate discrimination, however, suggests that these models should be used in conjunction with, not instead of, clinical judgment.