[Paper Review] Disease Prediction from Electronic Health Records Using Generative Adversarial Networks.

Uiwon Hwang, Sung‐Woon Choi+1 more|arXiv (Cornell University)|Nov 11, 2017

Machine Learning in Healthcare6 references12 citations

TL;DR

This study proposes a novel framework combining stacked autoencoders for missing value imputation and auxiliary classifier GANs (AC-GANs) for disease prediction in electronic health records (EHRs). It achieves a 98.05% prediction accuracy, significantly outperforming conventional algorithms in handling missing data and class imbalance.

ABSTRACT

Electronic health records (EHRs) have contributed to the computerization of patient records so that it can be used not only for efficient and systematic medical services, but also for research on data science. In this paper, we compared disease prediction performance of generative adversarial networks (GANs) and conventional learning algorithms in combination with missing value prediction methods. As a result, the highest accuracy of 98.05% was obtained using stacked autoencoder as the missing value prediction method and auxiliary classifier GANs (AC-GANs) as the disease predicting method. Results show that the combination of stacked autoencoder and AC-GANs performs significantly greater than existing algorithms at the problem of disease prediction in which missing values and class imbalance exist.

Motivation & Objective

• To address the challenges of missing values and class imbalance in electronic health records (EHRs) for disease prediction.
• To evaluate the performance of generative adversarial networks (GANs) against conventional machine learning algorithms in EHR-based disease prediction.
• To identify the optimal combination of missing value imputation and deep learning models for maximizing prediction accuracy.

Proposed method

• Employed stacked autoencoders to impute missing values in EHRs by learning hierarchical representations of the input data.
• Utilized auxiliary classifier GANs (AC-GANs) to generate realistic synthetic EHR samples while simultaneously classifying disease outcomes.
• Trained the AC-GANs with a conditional loss to improve disease prediction by leveraging class labels during the adversarial training process.
• Compared the performance of AC-GANs with stacked autoencoders against conventional learning algorithms using the same imputation and prediction pipeline.
• Applied feature engineering and normalization to preprocess EHR data before feeding it into the deep learning models.
• Evaluated model performance using standard classification metrics, with accuracy as the primary outcome.

Experimental results

Research questions

• RQ1 Can AC-GANs with stacked autoencoder-based imputation outperform conventional machine learning models in disease prediction from EHRs with missing data?
• RQ2 How does the integration of stacked autoencoders and AC-GANs affect prediction accuracy in the presence of class imbalance?
• RQ3 What is the optimal combination of missing value imputation and deep learning model for disease prediction in real-world EHR datasets?

Key findings

• The combination of stacked autoencoders for missing value imputation and AC-GANs for disease prediction achieved the highest accuracy of 98.05%.
• AC-GANs with stacked autoencoder imputation significantly outperformed conventional learning algorithms in disease prediction tasks.
• The proposed method demonstrated robust performance in handling both missing data and class imbalance, common challenges in EHR datasets.
• The results indicate that generative modeling via GANs, when combined with effective imputation, enhances predictive performance beyond discriminative models.