Predicting Early Deterioration in Lower Acuity Telehealth Patients Using Gradient Boosting

Area of Science:

• Medical Informatics
• Clinical Decision Support
• Artificial Intelligence in Healthcare

Background:

• Early recognition of physiological abnormalities is crucial for timely intervention and preventing adverse patient outcomes.
• Telehealth monitoring uses population management for remote identification of unstable patients.
• Lower acuity units require specific tools for prompt detection of patient deterioration.

Purpose of the Study:

• To propose and evaluate a novel Early Warning Score (EWS) model using gradient boosting.
• To enhance the detection of patient deterioration, particularly for those in medical/surgical units under telehealth monitoring.
• To compare the proposed model's performance against a modified Early Warning Score (MEWS*).

Main Methods:

• Utilized a dataset of 36,963 patient encounters from the eICU Research Institute database.
• Developed a gradient boosting model incorporating 35 features from demographics, vital signs, and laboratory data.
• Compared the proposed model against MEWS* (considering age and oxygen saturation) for predicting patient deterioration.

Main Results:

• The proposed model achieved an AUROC of 0.79 and AUPRC of 0.28 at 24 hours pre-deterioration, outperforming MEWS* (AUROC 0.67, AUPRC 0.07).
• Within one hour pre-deterioration, the model reached an AUROC of 0.86 and AUPRC of 0.42, compared to MEWS* (AUROC 0.74, AUPRC 0.21).
• The gradient boosting model demonstrated superior performance in predicting patient deterioration.

Conclusions:

• The proposed gradient boosting EWS model shows significant promise for improving early detection of patient deterioration in telehealth settings.
• This model is particularly effective for patients in lower acuity units, offering enhanced predictive capabilities over existing scores like MEWS*.
• Future research should address missing data, continuous monitoring, and clinical workflow integration.