The Quality Application of Deep Learning in Clinical Outcome Predictions Using Electronic Health Record Data: A Systematic Review

Abstract

Introduction: Electronic Health Record (EHR) is a significant source of medical data that can be used to develop predictive modelling with therapeutically useful outcomes. Predictive modelling using EHR data has been increasingly utilized in healthcare, achieving outstanding performance and improving healthcare outcomes.

Objectives: The main goal of this review study is to examine different deep learning approaches and techniques used to EHR data processing.

Methods: To find possibly pertinent articles that have used deep learning on EHR data, the PubMed database was searched. Using EHR data, we assessed and summarized deep learning performance in a number of clinical applications that focus on making specific predictions about clinical outcomes, and we compared the outcomes with those of conventional machine learning models.

Results: For this study, a total of 57 papers were chosen. There have been five identified clinical outcome predictions: illness (n=33), intervention (n=6), mortality (n=5), Hospital readmission (n=7), and duration of stay (n=1). The majority of research (39 out of 57) used structured EHR data. RNNs were used as deep learning models the most frequently (LSTM: 17 studies, GRU: 6 research). The analysis shows that deep learning models have excelled when applied to a variety of clinical outcome predictions. While deep learning's application to EHR data has advanced rapidly, it's crucial that these models remain reliable, offering critical insights to assist clinicians in making informed decision.

Conclusions: The findings demonstrate that deep learning can outperform classic machine learning techniques since it has the advantage of utilizing extensive and sophisticated datasets, such as longitudinal data seen in EHR. We think that deep learning will keep expanding because it has been quite successful in enhancing healthcare outcomes utilizing EHR data.