Development and validation of a machine learning model integrated with the clinical workflow for early detection of sepsis.

Background: Sepsis is a life-threatening condition caused by a dysregulated response to infection, affecting millions of people worldwide. Early diagnosis and treatment are critical for managing sepsis and reducing morbidity and mortality rates. Materials and methods: A systematic design approach was employed to build a model that predicts sepsis, incorporating clinical feedback to identify relevant data elements. XGBoost was utilized for prediction, and interpretability was achieved through the application of Shapley values. The model was successfully deployed within a widely used Electronic Medical Record (EMR) system. Results: The developed model demonstrated robust performance pre-operations, with a sensitivity of 92%, specificity of 93%, and a false positive rate of 7%. Following deployment, the model maintained comparable performance, with a sensitivity of 91% and specificity of 94%. Notably, the post-deployment false positive rate of 6% represents a substantial reduction compared to the currently deployed commercial model in the same health system, which exhibits a false positive rate of 30%. Discussion: These findings underscore the effectiveness and potential value of the developed model in improving timely sepsis detection and reducing unnecessary alerts in clinical practice. Further investigations should focus on its long-term generalizability and impact on patient outcomes.

The Impact of a Hybrid Hospital at Home Program in Reducing Subacute Rehabilitation.

Purpose: The purpose of this study was to compare health outcomes for patients receiving acute care in their homes through a Hospital at Home (HaH) program to outcomes for inpatients in the traditional hospital setting. Patients and Methods: We compared outcomes for patients in a HaH program at Virtua Health in 2022 (N = 271) to traditional inpatients during the same year (N = 13,776) with the same diagnoses. We defined outcomes as recommendations for subacute rehabilitation (SAR) upon discharge as this recommendation indicates the need for additional therapy based on a physician's assessment of the patient. Specifically, we searched notes in the electronic medical records for terms related to recommendation for SAR using text mining algorithms and a natural language processing (NLP) model to confirm these recommendations. We then compared the proportion of patients within each group that had a SAR recommendation, and controlled for differences in sample size, age, and diagnosis using bootstrapping analyses. Results: We observed that the proportion of patients in the HaH program that were recommended for SAR (0.148) was significantly different from the proportion of patients who remained in the traditional hospital setting (0.363), with a reduced need for SAR for HaH patients. We obtained qualitatively similar results when we controlled for sample size and diagnosis. Controlling for age yielded an older control population, and the difference in the proportion of patients with SAR recommendations between the groups widened. Conclusion: The reduced need for SAR for HaH patients in this study suggests that HaH programs are a promising alternative care model. Future work may consider how health outcomes vary for patients with different diagnoses, clinical histories and demographics, which may inform how HaH programs operate moving forward.