Facilitated Roundtable Discussion
CS2-05 - LINKHF-2 Study: Using Continuous Biosensor Data and Machine Learning Analytics to Predict Acute Care Events
Sunday, April 29
3:45 PM - 4:10 PM
Location: Education Zone, Booth 2416, Zone 3
Acute-on-chronic conditions like heart failure and COPD represent a disproportionate share of the costs associated with hospitalization and post-acute care. A majority of patients who deal with these diseases suffer from a complex set of co-morbidities that exacerbate the complexity of care and increase risk of unplanned acute care events. Recent advancements in wearable technology and advanced, machine learning-based analytics represent a potentially powerful new tool to manage ambulatory at-risk patient populations. The VA LINKHF-2 Study sought to validate how sophisticated, personalized, machine-learning, FDA-cleared analytics, when applied to continuous multivariate physiological data from wearable sensors, can detect vital sign anomalies that may be a precursor to an acute-care episode.
Within the LINKHF-2 study, 100 patients from across four different U.S. hospitals (Palo Alto, CA, Gainesville, FL, Houston, TX, and Salt Lake City, UT) were enrolled upon discharge related to heart failure decompensation. Patients were provided with a 90-day supply of wearable, disposable, bluetooth-enabled biosensors and a mobile phone with pre-paid data plan for data transmission. Data were collected via cloud-based IT platform across the 90 days and analyzed with multivariate, machine-learning analytics to detect changes from a personalized model of vital sign relationships.
The FDA-cleared analytics being evaluated build a personalized model of vital sign dynamics based on biosensor data captured the first 36 hours after hospital discharge. During that model training period, the system "learns" an individual patient's unique vital sign relationships (HR, RR, and activity) across the full spectrum of activities one would expect in an ambulatory environment. With this, the system analytics develop a personalized baseline by which to measure changes that may be a precursor to an acute care event. After the 36-hour model training period, the system automatically transitions from "learning" mode into "monitoring" mode where it indicates subtle changes that may indicate compensatory behavior within the cardiopulmonary control loop. Personalized anomalies are indicated via time-series plotted index of change from baseline. The higher the index, across a longer period of time, the greater the change from baseline.
This was an observational study. Current standard of care was delivered while data were being collected in the background. A retrospective analysis compared results of the personalized analytics relative to the medical record. The intent of the study was to evaluate 1) how well the personalized, machine learning-based physiology analytics detected acute care events in an ambulatory environment and 2) how early the analytics indicated health deterioration relative to the acute care event. The ultimate goal of the solution is to provide clinicians with a scalable tool that will allow them to manage a post-acute population using passive data collection and sophisticated analytics. The intended benefit is that large, at-risk patient populations can be managed by exception, as indicated by each individual's changes in cardiopulmonary physiology. A simplified ROC analysis yielded an AUC ~0.81, enabling very high specificity to alleviate false alert burden, while still delivering a potential cost-saving sensitivity.
- ...understand extended (90-day) continuous ambulatory multivariate physiology data capture using clinical-grade, disposable, wearable biosensors. Will review solution deployment across a 100-patient post-discharge heart failure population, inclusive of biosensor wear compliance.
- ...understand how machine learning-based, FDA 510k-cleared, Personalized Physiology Analytics (PPA) leverage multivariate continuous data (HR, RR, Activity) to detect physiological anomalies that can be a precursor to disease exacerbation.
- ...statistical analysis of performance of PPA analytics within this observational study in the context of A) sensitivity/specificity of detection of acute care events and B) quantification of early warning