Key result
Adding continuous vital signs and activity data improves machine learning prediction of 30-day readmissions by ~113%.
Why the study?
Accurate prediction and prevention of hospital readmission remains a clinical challenge, and the influence of continuous monitoring on machine learning performance for predicting 30-day readmissions needed examination.
Does the addition of continuous vital signs and activity data to episodic physiological monitoring improve machine learning prediction of all-cause 30-day hospital readmission in patients admitted to acute care?
Observational (n=354)
Does the addition of continuous vital signs and activity data to episodic physiological monitoring improve machine learning prediction of all-cause 30-day hospital readmission in patients admitted to acute care?
Adding continuous vital signs and activity monitoring to episodic physiological data significantly improves machine learning prediction of 30-day hospital readmission.
Answers are grounded in this paper and cite their sources.
May aid ML-based readmission risk stratification; hypothesis-generating and requires prospective validation before practice change.
Background: Accurate prediction and prevention of hospital readmission remains a clinical challenge. The influence of different data sources, including remotely monitored continuous vital signs and activity, on machine learning (ML) models' performances is examined for predicting all-cause unplanned 30-day readmission. Methods: Patients (n = 354) recruited in the emergency department and admitted to acute care at either hospital or home hospital settings are analyzed. Data sources included continuous vital signs and activity, electronic health record (EHR) data - episodic physiological monitoring of laboratory and vital signs, demographics, hospital utilization history, and quality of life survey measures. Five (5) machine learning classifiers were systematically trained by varying input data sources for readmission. Performances of ML models as well as the standard-of-care HOSPITAL score for readmissions were assessed with area under the receiver operating characteristic curve (AUROC) and area under precision-recall curve (AUPRC) statistics. Results: There were 29 patients readmitted out of the 354 total included patients (an 8.2 % readmission rate). The average five-fold cross-validation AUROC and AUPRC scores of the five readmission models ranged from 0.76 to 0.84 (P > .05) and 0.23-0.49 (P < .05), respectively. The model input with episodic physiological monitoring (vitals and labs) had an AUPRC of 0.23 ± 0.07, while the model input with continuous vitals and activity data and episodic vitals and laboratory measurements had an AUPRC of as 0.49 ± 0.10 (P < .005). The HOSPITAL score had an AUROC of 0.62 and AUPRC of 0.16 in this pilot study. Conclusions: The systematic ML modeling and analysis showcased diversity in predictive power and performances of patient data sources for predicting readmission. This pilot study suggests continuous vital signs and activity data, when added to episodic physiological monitoring, boosts performance. The HOSPITAL score shows low predictive power for readmission in this population. Predictive modeling of unplanned 30-day readmission improves with continuous vital signs and activity monitoring.
No takes yet. Share an insight, caveat, or question.
Pettinati et al. (2025) conducted an observational in Acute care admission (n=354). Continuous vital signs and activity monitoring vs. Episodic physiological monitoring was evaluated on All-cause unplanned 30-day readmission (p=<.005). Adding continuous vital signs and activity data to episodic physiological monitoring improved machine learning prediction of 30-day readmission (AUPRC 0.49 vs 0.23; P<0.005).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: