Forecasting Patient Early Readmission from Irish Hospital Discharge Records Using Conventional Machine Learning...

Area of Science:

• Healthcare Informatics
• Clinical Risk Management
• Machine Learning in Medicine

Background:

• Patient readmission poses significant challenges in healthcare, impacting costs and patient outcomes.
• Accurate prediction of readmissions is crucial for effective risk management and intervention strategies.

Purpose of the Study:

• To compare conventional and deep learning models for predicting patient readmission.
• To identify key clinical and demographic features associated with 30-day readmission risk.
• To apply explainable AI techniques for model interpretability.

Main Methods:

• Evaluation of machine learning models using multimodal electronic discharge records.
• Addressing data imbalance and data type variety to enhance algorithm performance.
• Utilizing SHapley Additive Explanations (SHAP) for feature and diagnosis code interpretation.

Main Results:

• Achieved an improvement in Area Under the Receiver Operating Characteristic Curve (AUROC) from 0.628 to 0.7.
• Identified cancer, COPD, and social factors as significant predictors of 30-day readmission.
• Determined that bacterial carrier status had minimal impact due to low frequency.

Conclusions:

• Routinely collected hospital data can be effectively used for patient readmission forecasting.
• Conventional machine learning combined with explainable AI provides insights into readmission risk factors.