Predicting Early Deterioration in Lower Acuity Telehealth Patients Using Gradient Boosting.

Timely recognition of physiological abnormalities is vital for early intervention, potentially preventing adverse outcomes and minimizing the need for transfer to a higher level of care. This is a primary focus of telehealth monitoring in which remote clinicians utilize population management to identify and prioritize patients of concern or instability. This work proposes an Early Warning Score model based on gradient boosting, emphasizing prompt deterioration detection, especially tailored to patients in lower acuity units (e.g. - medical/surgical) who are also receiving telehealth monitoring. Data included 36,963 patient encounters from the eICU Research Institute database. The model utilizes 35 features extracted from demographics, vital signs, and laboratory data. It showed enhanced performance in comparison to a version of the Modified Early Warning Score (MEWS*) that considers age and oxygen saturation instead of the level of consciousness. The model achieved an AUROC of 0.79 and AUPRC of 0.28, 24 hours before deterioration, surpassing MEWS* with values of 0.67 and 0.07, respectively. Within an hour before deterioration happens, the proposed model achieved an AUROC of 0.86 and AUPRC of 0.42 while MEWS* achieved 0.74 and 0.21, respectively. Future investigations will focus on exploring the impact of missing data, continuous performance for individual patients, and integration into clinical workflows.