Machine learning on asynchronous clinical pages to predict clinical deterioration.

Objectives: Clinical deterioration in hospitalized patients is often preventable, but traditional early warning scores based on structured data are limited by delayed or inconsistent documentation. We developed and evaluated a machine learning pipeline that predicts clinical deterioration using real-time pager messages exchanged between clinicians. Materials and Methods: We conducted a retrospective study of adult non-ICU hospitalizations at Vanderbilt University Medical Center between January 2018 and June 2021. Using the content and frequency of messages, we trained long short-term memory models to predict rapid response activation, unplanned ICU transfer, or cardiac arrest within the next 6, 12, or 24 hours. Model performance was compared with the Epic's Deterioration Index and with a logistic regression ensemble combining predictions from both models. Results: There were 1 519 445 pages associated with 111 346 hospitalizations among 74 912 patients. Deterioration events were observed in 5114 (4.6%) hospitalizations. The model achieved moderate discriminative ability, with areas under the receiver operating characteristic curve (AUROC) of 0.684 (95% CI, 0.66-0.71), 0.724 (95% CI, 0.71-0.74), and 0.669 (95% CI, 0.65-0.68) for predicting deterioration within 6, 12, and 24 h, respectively. Predictions showed moderate correlation with EDI scores (Pearson's R, 0.297-0.323) suggesting complementary signals. The ensemble model consistently outperformed either approach alone, achieving an AUROC of 0.797 (95% CI, 0.78-0.82) for 12-h prediction. Discussion and Conclusion: Clinical pages represent an underutilized data source capturing clinicians' intuition and observations before they appear in formal documentation. Machine learning on pages can augment early warning systems by providing real-time information and intuition without increasing workload.