JMIR Publications

ABSTRACT

Background:

Clinical notes contain contextualized information beyond structured data related to patients’ past and current health status. This study aimed to design a multimodal deep learning approach for improving the outcome evaluation precision of heart failure (HF) using admission clinical notes and easily collected tabular data.

Objective:

We aim to develop multimodal deep learning (DL) model to improve the precision prognostic evaluation of hospital outcome for heart failure leveraging admission notes and clinical tabular data.

Methods:

Data for the development and validation of the multimodal model were retrospectively derived from three open-access USA databases including the Medical Information Mart for Intensive Care Database v1.4 (MIMIC-III, CareVue) and MIMIC-IV v1.0 collected from a teaching hospital covering 2001 to 2019; the eICU Collaborative Research Database v1.2 (eICU-CRD) collected from 208 hospitals covering 2014 to 2015, respectively. The study cohorts consisted of all critical illness HF patients. The clinical notes including chief complaint, history of present illness, physical examination, medical history and admission medication, and clinical variables recorded in electronic health records were analyzed. We developed a deep learning mortality prediction model for in-hospital, which was completely internally, prospectively and externally evaluated. The Integrated Gradients (IG) and SHapley Additive exPlanations (SHAP) methods were used to analyze the importance of risk factors.

Results:

The study contained 9,989 (development set, 16.4%), 2,497 (internal validation set, 14.1%), 1,896 (prospective validation set, 18.3%) and 7,432 (external validation set, 15.0%) patients. The area under the receiver operating characteristic curves (AUCs) with 95% CI of the model were 0.838 [0.827–0.851], 0.849 [0.841-0.856], 0.767 [0.762–0.772], respectively. The AUCs of the multimodal model outperformed unimodal models in all test sets, and tabular data contributed to a higher discrimination. The medical history and physical exam were more useful than others in early assessment.

Conclusions:

The multimodal deep learning model for combining admission notes and clinical tabular data showed promising efficacy as a potential novel method in evaluating the risk of death in HF patients, providing more accurate and timely decision support.