Cloud-Based Machine Learning Platform to Predict Clinical Outcomes at Home for Patients With Cardiovascular...

Area of Science:

• Digital Health
• Machine Learning in Healthcare
• Wearable Technology

Background:

• Hospital readmissions significantly contribute to US healthcare costs, exceeding $700 million annually.
• The Hospital Readmissions Reduction Program (HRRP) financially penalizes hospitals for excess readmissions.
• Predicting patient recovery and reducing readmissions requires intelligent monitoring of physical activity (PA).

Purpose of the Study:

• To develop a remote, low-cost, cloud-based machine learning (ML) platform for precision health monitoring of physical activity (PA).
• To validate the platform's ability to predict clinical outcomes in discharged patients through a clinical trial.

Main Methods:

• A wearable device with an accelerometer and Bluetooth sensor, connected to an iPhone and cloud-based ML interface, was used for remote PA analysis.
• Over 17,000 person-day data points were collected at a skilled nursing facility to train an extreme gradient boosting (XGBoost) ML model.
• A clinical trial, Activity Assessment of Patients Discharged from Hospital-I, tested the hypothesis that PA profiles predict clinical outcomes.

Main Results:

• The study enrolled 52 patients discharged from Stanford Hospital.
• A robust predictive system was developed, forecasting patient health outcomes based on PA data.
• The platform achieved 87% sensitivity, 79% specificity, and 85% accuracy in predicting clinical outcomes.

Conclusions:

• This study presents the first reliable clinical data using wearable devices for monitoring discharged patients to predict recovery.
• The developed platform offers a rigorous method for assessing outcome data for reliable remote home care.
• The technology has the potential to improve patient recovery monitoring and reduce hospital readmissions.