Predicting 30-day hospital readmissions using artificial neural networks with medical code embedding

Area of Science:

• Health Services Research
• Artificial Intelligence in Healthcare
• Clinical Informatics

Background:

• Reducing unplanned hospital readmissions is a key quality metric for hospitals.
• Current risk-standardization models using healthcare claims data often lack accuracy.
• Accurate prediction models are crucial to avoid unfair penalization of hospitals.

Purpose of the Study:

• To compare the predictive accuracy of four models for 30-day unplanned readmissions.
• To evaluate hierarchical logistic regression against gradient boosting and artificial neural network models.
• To assess the impact of novel embedding techniques on readmission prediction.

Main Methods:

• Utilized the Nationwide Readmissions Database (2014) for patients with acute myocardial infarction (AMI), congestive heart failure (HF), and pneumonia (PNA).
• Compared hierarchical logistic regression with gradient boosting and two artificial neural network models.
• Incorporated unsupervised Global Vector for Word Representations (GloVe) embeddings of administrative claims data into artificial neural network models.

Main Results:

• Artificial neural network models combined with GloVe embeddings demonstrated improved 30-day readmission prediction.
• The best models increased AUC from 0.68 to 0.72 for AMI, 0.60 to 0.64 for HF, and 0.63 to 0.68 for PNA compared to hierarchical logistic regression.
• Risk-standardized readmission rates using the novel model reclassified approximately 10% of hospitals.

Conclusions:

• Artificial neural network models with GloVe embeddings offer superior prediction of unplanned patient readmissions.
• Novel prediction models can significantly alter hospital performance classifications compared to traditional methods.
• Further investigation into advanced prediction models for hospital quality assessment is warranted.