Automated Fusion of Multimodal Electronic Health Records for Better Medical Predictions

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Electronic Health Record (EHR) systems generate vast medical data, presenting opportunities for deep learning.
• Real-world EHR data's complexity and multi-modality challenge current deep learning model design.
• Existing methods often rely on manual, intuition-based architectures, leading to suboptimal performance.

Purpose of the Study:

• To address the multi-modality challenge in EHR data by automating model architecture design.
• To propose a novel neural architecture search (NAS) framework, AutoFM, for optimal EHR data encoding and fusion.
• To enhance the performance of deep learning models in mining complex EHR data.

Main Methods:

• Developed AutoFM, a neural architecture search (NAS) framework.
• The framework automatically searches for optimal architectures for diverse input modalities and fusion strategies.
• Conducted experiments on real-world multi-modal EHR data for prediction tasks.

Main Results:

• AutoFM achieved significant performance improvements over state-of-the-art methods.
• The framework effectively discovered meaningful and optimized network architectures.
• Demonstrated superior performance in mining multi-modal EHR data.

Conclusions:

• Automated model design using NAS frameworks like AutoFM is effective for EHR data.
• AutoFM offers a powerful solution for overcoming the multi-modality challenge in EHR analysis.
• The proposed framework holds potential for advancing deep learning applications in healthcare.