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

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Informatics
• Predictive Analytics

Background:

• Infections are a major cause of hospitalizations and emergency department (ED) visits for home care patients.
• Current prediction tools often fail to fully leverage longitudinal data from electronic health records (EHR) and clinical notes.

Purpose of the Study:

• To evaluate a sequence-aware deep learning approach for predicting infection-related hospitalizations or ED visits.
• To integrate structured EHR data with Natural Language Processing (NLP)-derived features from clinical notes.
• To develop and assess a three-tier risk stratification tool for home care patients.

Main Methods:

• Retrospective cohort study of 23,321 home care episodes (2015-2017).
• Compared sequence-aware models with a non-sequential baseline using structured and NLP-derived features.
• Evaluated model performance using AUROC, AUPRC, and F1-score for predicting infection-related events within 2- to 4-day windows.

Main Results:

• Sequence-aware models, especially with NLP features, significantly outperformed the baseline.
• The best model achieved an AUROC of 0.991, AUPRC of 0.885, and F1-score of 0.774.
• A three-tier stratification tool identified 77.7% of infection events in the highest-risk 5% of episodes, with equitable performance across subgroups.

Conclusions:

• Sequence-aware deep learning effectively predicts near-term infection risks in home care.
• The integrated approach using EHR and NLP data offers accurate and equitable risk prediction.
• The developed tool can aid in targeted triage and early intervention to reduce acute care utilization.