Treatment effect prediction with adversarial deep learning using electronic health records

Area of Science:

• Medical Informatics
• Machine Learning
• Clinical Decision Support

Background:

• Treatment effect prediction (TEP) is crucial for personalized disease management.
• Electronic Health Records (EHRs) offer a rich data source for clinical applications like TEP.

Purpose of the Study:

• To develop and evaluate an adversarial deep treatment effect prediction (ADTEP) model using heterogeneous EHR data.
• To enhance the accuracy of predicting treatment effects by leveraging patient characteristics and treatment information.

Main Methods:

• Developed an adversarial deep learning model employing auto-encoders to learn patient and treatment features from EHR data.
• Utilized adversarial learning to enhance feature discriminative power by decoding correlations between patient characteristics and treatments.
• Integrated a logistic regression layer for the final treatment effect prediction.

Main Results:

• The ADTEP model achieved superior performance on acute coronary syndrome (ACS) and heart failure (HF) datasets.
• Demonstrated significant performance gains (up to 6.3% AUC improvement) over benchmark models like DTEP, LR, and SVM on the ACS dataset.
• Outperformed all evaluated benchmarks in the heart failure case study.

Conclusions:

• The proposed ADTEP model effectively utilizes EHR data and adversarial learning for robust treatment effect prediction.
• The model's ability to extract discriminative representations from EHR data leads to improved TEP.
• Experimental results confirm the superiority of the proposed method over state-of-the-art approaches.