AI-RiskX: An Explainable Deep Learning Approach for Identifying At-Risk Patients During Pandemics

Area of Science:

• Public Health
• Artificial Intelligence
• Computational Biology

Background:

• Pandemics strain healthcare systems, necessitating accurate identification of high-risk individuals.
• Existing AI models lack interpretability and fail to account for diverse patient vulnerabilities.

Purpose of the Study:

• To develop an explainable deep learning model (AI-RiskX) for identifying at-risk patients during pandemics.
• To enhance timely intervention and resource allocation for COVID-19 and related infections.

Main Methods:

• Integrated five public health datasets (asthma, diabetes, heart, kidney, thyroid).
• Utilized Synthetic Minority Over-sampling Technique (SMOTE) for class balancing.
• Employed a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model.
• Incorporated SHAP for model interpretability and a rule-based module for stratification.

Main Results:

• Achieved 98.78% accuracy in classifying at-risk patients.
• Provided both individual and population-level interpretability.
• Successfully stratified patients by age and pregnancy status.

Conclusions:

• AI-RiskX offers a scalable and interpretable solution for equitable patient classification.
• The model supports critical decision-making in public health emergencies.
• Addresses limitations of previous AI models by integrating diverse data and prioritizing explainability.