Autoencoder-Based Representation Learning for Similar Patients Retrieval From Electronic Health Records: Comparative...

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Computational biology

Background:

• Electronic health record (EHR) data modeling is challenging due to high dimensionality, mixed features, noise, bias, and sparsity.
• Patient representation learning using autoencoders (AEs) offers a promising approach to address these EHR data challenges.
• Understanding the impact of different AE designs and distance measures on retrieving similar patient cohorts is crucial.

Purpose of the Study:

• To evaluate the performance of five common autoencoder (AE) variants in retrieving similar patients.
• To investigate the influence of various distance measures and hyperparameter configurations on AE model performance.
• To assess the effectiveness of AE-based patient similarity estimation for clinical outcome prediction.

Main Methods:

• Tested five AE variants (vanilla, denoising, contractive, sparse, robust) on two real-world EHR datasets.
• Applied k-nearest neighbors (k-NN) with Euclidean and Mahalanobis distances to AE-produced latent representations.
• Evaluated model performance on predicting acute kidney injury onset and 1-year postdischarge mortality.

Main Results:

• Denoising autoencoders significantly outperformed other AE variants when paired with Euclidean distance (P<.001).
• Learning rates were identified as a critical hyperparameter influencing AE model performance.
• Mahalanobis distance-based k-NN frequently outperformed Euclidean distance-based k-NN on latent representations, though direct application to raw data was data-dependent.

Conclusions:

• This study provides a comprehensive analysis of AE variants for patient similarity retrieval.
• Findings highlight the importance of AE design and hyperparameter tuning for effective patient representation learning.
• The results lay the groundwork for developing advanced AE-based methods for personalized medicine and improved patient care.