Developing and validating a sequence-aware deep learning model for infection risk prediction in home care

Infections are a leading cause of hospitalizations and emergency department (ED) visits in home care. Existing prediction tools often underutilize longitudinal patterns of home care electronic health record (EHR) and narrative clinical notes.

To evaluate a sequence-aware deep learning approach that integrates longitudinal structured EHR data with Natural Language Processing (NLP)-derived features from clinical notes to predict infection-related hospitalizations or ED visits, derive a three-tier risk stratification tool from model outputs, and assess performance, interpretability, and fairness.

This retrospective cohort study analyzed 23,321 home care episodes (2015-2017), of which 1,528 (6.5%) involved an infection-related hospitalization or ED visit within 30 days of admission. We compared sequence-aware models against a non-sequential baseline. Models used 28 admission-level and 30 visit-level features (6 structured measures and 24 NLP-derived indicators) to represent longitudinal home visit data. Models estimated the risk of first infection-related hospitalization or ED visit over 2-, 3-, and 4-day windows. Evaluation metrics included the Area Under the Receiver Operating Characteristic (AUROC) curve, Area Under the Precision-Recall Curve (AUPRC), and F1-score.

Sequence-aware models consistently outperformed the non-sequential baseline, with larger gains observed when NLP-derived features were included. The best configuration (bidirectional long short-term memory at 4-day lead time) achieved AUROC 0.991, AUPRC 0.885, and F1-score 0.774. The three-tier stratification concentrated 77.7% of infection events within the highest-risk 5% of episodes. Interpretability analyses highlighted the importance of recent visit information, and fairness analyses showed equitable performance across most demographic and socioeconomic subgroups.

A sequence-aware deep learning model integrating longitudinal structured EHR data with NLP-derived features can accurately and equitably predict near-term risks of infection-related hospitalizations or ED visits among home care patients. The three-tier tool may support targeted triage and early intervention for patients at risk of infection-related acute care utilization.