Improving the Prognostic Evaluation Precision of Hospital Outcomes for Heart Failure Using Admission Notes and...

Area of Science:

• * Artificial Intelligence in Medicine
• * Clinical Informatics
• * Biomedical Data Science

Background:

• * Clinical notes offer rich, contextual patient information beyond structured electronic health records.
• * Accurate prediction of hospital outcomes for heart failure (HF) is crucial for timely intervention.
• * Integrating diverse data sources can improve predictive model performance.

Purpose of the Study:

• * To develop and validate a multimodal deep learning model for predicting in-hospital mortality in heart failure patients.
• * To assess the precision of this model using both clinical notes and electronic health record tabular data.
• * To compare the performance of the multimodal model against unimodal approaches.

Main Methods:

• * Retrospective derivation of data from MIMIC-III, MIMIC-IV, and eICU Collaborative Research Databases.
• * Development of a deep learning model integrating unstructured clinical notes (e.g., history of present illness, physical examination) and structured tabular data.
• * Internal, prospective, and external validation of the model; risk factor analysis using Integrated Gradients and SHAP.

Main Results:

• * The multimodal deep learning model demonstrated superior performance across internal, prospective, and external validation sets compared to unimodal models.
• * Area under the receiver operating characteristic curve (AUC) values ranged from 0.767 to 0.849, indicating strong predictive capability.
• * Tabular data significantly contributed to improved discrimination, with medical history and physical examination being key early assessment factors.

Conclusions:

• * The multimodal deep learning model effectively combines admission clinical notes and tabular data for predicting heart failure mortality.
• * This approach shows promise as a novel method for enhancing the accuracy and timeliness of decision support in HF patient care.
• * The findings suggest a valuable new tool for risk stratification and management of heart failure patients.