JMIR Publications

ABSTRACT

Background:

By analyzing electronic health record (EHR) snapshots of similar patients, physicians can proactively predict disease onset, customize treatment plans, and anticipate patient-specific trajectories. However, EHR data are inherently challenging to model due to high dimensionality, mixed feature types, noise, bias, and sparsity. Patient representation learning with autoencoders (AE) presents promising opportunities. A critical question remains: how do different AE designs and distance measures impact the quality of retrieved similar patient cohorts?

Objective:

This study aims to evaluate the performance of five common AE variants—vanilla AE (AE), denoising AE (DAE), contractive AE (CAE), sparse AE (SAE), and robust AE (RAE)—in retrieving similar patients. Additionally, it investigates the impact of different distance measures and hyperparameter configurations on model performance.

Methods:

We tested the five AE variants on two real-world datasets from the University of Kansas Medical Center (KUMC, n = 13,752) and the Medical College of Wisconsin (MCW, n = 9,568) across 168 different hyperparameter configurations. Euclidean distance-based k-nearest neighbors (k-NN) and Mahalanobis distance-based k-NN were then applied on the latent representations to retrieve similar patients. Two prediction targets were evaluated: Acute Kidney Injury (AKI) onset and post-discharge 1-year mortality, with F1 score as the evaluation metric.

Results:

Our findings demonstrate that (1) DAE outperformed other AE variants when paired with Euclidean distance (P<.001), followed by vanilla AE and CAE; (2) learning rates significantly influenced the performance of AE variants; and (3) Mahalanobis distance-based k-NN frequently outperformed Euclidean distance-based k-NN when applied to latent representations. However, whether AE models yield better performance by transforming raw data into latent representations, compared to applying Mahalanobis distance-based k-NN directly to raw data, was observed to be data-dependent.

Conclusions:

This study provides a comprehensive analysis of the performance of different AE variants in retrieving similar patients and evaluates the impact of various hyperparameter configurations on model performance. It lays the groundwork for future development of AE-based patient similarity estimation and personalized medicine models.