An Interpretable Deep-Learning Framework for Predicting Hospital Readmissions From Electronic Health Records

TL;DR

This work tackles predicting unplanned hospital readmissions from Electronic Health Records while meeting interpretability requirements. It introduces ConvLSTM1d, a dual-input neural framework that embeds ICD-9-CM codes and encodes time gaps between admissions to capture long-term dependencies, complemented by a model-dependent interpretability method that attributes predictions to individual codes. On MIMIC III data, ConvLSTM1d outperforms logistic regression, random forest, and Deepr in accuracy and F1 for both 30-day and 180-day readmission tasks, while providing interpretable code-level explanations. The approach offers a practical, transparent tool for clinicians and healthcare systems to anticipate readmissions and identify actionable factors, potentially reducing costs and improving patient care.

Abstract

With the increasing availability of patient data, modern medicine is shifting towards prospective healthcare. Electronic health records offer a variety of information useful for clinical patient characterization and the development of predictive models, given that similar medical histories often lead to analogous health progressions. One application is the prediction of unplanned hospital readmissions, an essential task for reducing healthcare costs and improving patient outcomes. While predictive models demonstrate strong performances especially with deep learning approaches, they are often criticized for their lack of interpretability, a critical requirement in the medical domain where incorrect predictions may have severe consequences for patient safety. In this paper, we propose a novel and interpretable deep learning framework for predicting unplanned hospital readmissions, supported by NLP findings on word embeddings and by ConvLSTM neural networks for better handling temporal data. We validate the framework on two predictive tasks for hospital readmission within 30 and 180 days, using real-world data. Additionally, we introduce and evaluate a model-dependent technique designed to enhance result interpretability for medical professionals. Our solution outperforms traditional machine learning models in prediction accuracy while simultaneously providing more interpretable results.

Figures (10)

• Figure 1: Representation in ConvLSTM1d framework of the clinical history regarding the first and second admission of a patient.
• Figure 2: Layers implemented by ConvLSTM1d.
• Figure 3: 2d graphical representation of medical concepts by using t-SNE with "cosine similarity" as evaluation metric. Diagnoses are represented by blue points, procedures by orange points. Clusters are surrounded by a purple dashed line. On the right side the most prominent clusters are listed.
• Figure 4: Implementation of the model-dependent interpretability approach -- first step.
• Figure 5: Implementation of the model-dependent interpretability approach -- second step.