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

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Clinical prediction modeling

Background:

• Heart failure (HF) poses a significant public health challenge, marked by high unplanned readmission rates.
• Current models for predicting HF readmissions have suboptimal performance, necessitating improved feature engineering.
• Electronic health record (EHR) data offers a rich source for developing predictive models.

Purpose of the Study:

• To evaluate and compare the effectiveness of various feature embedding approaches for enhancing the prediction of unplanned readmissions in heart failure patients.
• To determine if embedding techniques can improve the accuracy of machine learning models in identifying patients at high risk for readmission.

Main Methods:

• Compared three embedding approaches: word2vec on terminology codes/concept unique identifiers (CUIs) and BERT on concept descriptions, against a baseline of one-hot encoding.
• Utilized logistic regression, eXtreme Gradient-Boosting (XGBoost), and Artificial Neural Network (ANN) models.
• Evaluated model performance using Area Under the Receiver Operating Characteristic (AUROC) and F1-scores on a heart failure cohort (N=21,031) from the MIMIC-IV dataset.

Main Results:

• Embedding approaches significantly improved prediction model performance across all tested algorithms.
• XGBoost demonstrated superior performance regardless of the embedding method used.
• Word2vec embeddings trained on the specific dataset achieved a higher AUROC (0.65) compared to pre-trained BERT embeddings (0.59) on concept descriptions.

Conclusions:

• Word2vec embeddings, trained on EHR data, effectively discriminate heart failure readmission cases, outperforming one-hot encoding and pre-trained BERT embeddings.
• These embedding methods offer a viable approach for automated feature selection in predicting readmissions.
• The observed AUROC improvement supports enhanced risk stratification and targeted clinical interventions for heart failure patients.