Encoding Health Records into Pathway Representations for Deep Learning.

Area of Science:

• Health Informatics
• Machine Learning
• Computational Biology

Background:

• Deep learning models increasingly utilize patient health records for comprehensive data analysis.
• Developing effective patient representations is crucial for machine learning tasks in healthcare.

Purpose of the Study:

• To propose a reproducible method for generating patient pathways from health records.
• To transform these pathways into an image-like structure suitable for deep learning.
• To train and evaluate a convolutional neural network (CNN) using these pathways.

Main Methods:

• Developed a reproducible approach to generate patient pathways from electronic health records.
• Transformed patient pathways into a machine-processable, image-like format.
• Generated over a million pathways from FAIR synthetic health records.
• Trained a convolutional neural network (CNN) on the generated pathways.
• Assessed the impact of training dataset size on autoencoder performance.

Main Results:

• The CNN achieved prediction task accuracy comparable to or better than autoencoders trained on the same data.
• The CNN required significantly fewer computational resources for training compared to autoencoders.
• Initial experiments demonstrated the effectiveness of the pathway-based deep learning approach.

Conclusions:

• The proposed method offers an efficient and effective way to create deep learning-ready patient representations from health records.
• This approach holds promise for improving machine learning applications in healthcare by leveraging structured patient pathways.
• The open-source availability of the code facilitates reproducibility and further research.