Enhancing prediction of inpatient deterioration by combining clinical and nurse concern features, with or without temporal clustering.

Background: Early warning systems (EWSs) help clinicians identify deteriorating patients using clinical data, such as vital signs. However, standard systems struggle to capture nuanced nursing concerns. The Healthcare Process Model-ExpertSignals (HPM-ExpertSignals) framework describes how nurses' concerns are reflected in their documentation patterns. While a recent trial showed positive outcomes, the predictive gain of combining both data types remains unquantified. Objectives: We evaluated improvements in F-measure by combining HPM-ExpertSignals with clinical data using the k-shape temporal clustering algorithm. Materials and Methods: Six models were compared based on their features and the inclusion of k-shape. Models were trained to predict patient deterioration (cardiac arrest and death) 12 h before the event using a large dataset. The primary outcome was the harmonic mean of precision and recall (F-measure). Results: The F-measure achieved by the model that uses both feature types was 0.25 (±0.01). The clinical features-only model was 0.16 (±0.01), and the HPM-ExpertSignals-only model was 0.19 (±0.02). F-measures for their corresponding k-Shape models were all at 0.06 (±0.0). Discussion: The combined model has the highest F-measure among the clinical-only and HPM-ExpertSignals-only models. The low performance of the k-Shape models suggests that k-Shape is not well suited to capturing the specific temporal patterns present in this problem set. Conclusion: Early warning systems leveraged both clinical data and HPM-ExpertSignals predictors, which may offer clinically significant improvements. Future research should explore alternative temporal pattern algorithms to further refine predictive accuracy.