Predicting 30-Days Hospital Readmission for Patients with Heart Failure Using Electronic Health Record Embeddings: Comparative Evaluation

TL;DR

This study compares embedding methods to predict hospital readmissions for heart failure patients, finding that word2vec outperforms other approaches.

Contribution

The study introduces and evaluates embedding-based feature generation for predicting heart failure readmissions using EHR data.

Findings

Word2vec embeddings outperformed pre-trained BERT and one-hot encoding in predicting readmissions.

XGBoost models achieved the best performance across all embedding approaches.

Embedding methods improved AUROC by 0.11 compared to baseline methods.

Abstract

Heart failure (HF) is a public health concern with a wider impact on quality of life and cost of care. One of the major challenges in HF is the higher rate of unplanned readmissions and suboptimal performance of models to predict the readmissions. Hence, in this study, we implemented embeddings-based approaches to generate features for improving model performance.

The objective of this study was to evaluate and compare the effectiveness of different feature embedding approaches for improving the prediction of unplanned readmissions in patients with heart failure.

We compared three embedding approaches including word2vec on terminology codes and concept unique identifier (CUIs) and BERT on descriptive text of concept with baseline (one hot-encoding). We compared area under the receiver operating characteristic (AUROC) and F1-scores for the logistic regression, eXtream gradient-boosting (XGBoost) and artificial neural network (ANN) models using these embedding approaches. The model was tested on the heart failure cohort (N=21,031) identified using least restrictive phenotyping methods from MIMIC-IV dataset.

We found that the embedding approaches significantly improved the performance of the prediction models. The XGBoost performed better for all approaches. The word2vec embeddings (0.65) trained on the dataset outperformed embeddings from pre-trained BERT model (0.59) using descriptive text.

Embedding methods, particularly word2vec trained on electronic health record data, can better discriminate HF readmission cases compared to both one-hot encoding and pre-trained BERT embeddings on concept descriptions making it a viable approach of automation feature selection. The observed AUROC improvement (0.65 vs 0.54) may support more effective risk stratification and targeted clinical interventions.

Figures