BACKGROUND: Remote patient monitoring strategies in patients with heart failure (HF) are often based on manual readings and interpretation of various parameters by health care professionals. Automated multiparameter predictive models (MPMs) have the potential to improve early recognition of decompensated HF and to reduce the workload for both health care professionals and patients. To reduce costs and facilitate large-scale implementation, these models should preferably be based on noninvasive measurements, with user-friendly devices.
OBJECTIVE: This exploratory study aimed to evaluate whether an MPM, using various parameters from a wrist-worn device supplied with a photoplethysmography sensor and a triaxial accelerometer, contributes to the detection of decompensated HF and death in patients with unstable HF.
METHODS: Patients who were admitted to the hospital with acute decompensated HF, regardless of etiology or left ventricular ejection fraction, were instructed to wear a research-grade wrist-worn device from the moment of discharge. The device measured heart rate (HR), interbeat intervals (IBIs), respiration rate (RR), activity counts (AC), energy expenditure (EE), and sleep. Participants were instructed to wear the device 24 hours a day for 3 consecutive months. We evaluated 7 classifiers under four strategies for handling extreme class imbalance; the best model was then tested via leave-one-subject-out cross-validation on untouched data. The combined end point of interest was hospital readmission due to decompensated HF, decompensated HF treated at the outpatient clinic by increasing the loop diuretic dose, or death due to HF.
RESULTS: A total of 17 patients participated in the study (median age 77, IQR 70-84 y; n=9, 53% male). During follow-up, the device-wearing compliance was 78% (55%-81%). The activity-related parameters (EE and AC) performed best with respect to data quality: 72% and 79% of the data were of high quality, respectively. Concerning HR, 46% of the data were of high quality, whereas only 29% of the IBI and 14% of the RR data were of high quality. Sleep data were lacking 99% of the time during follow-up, resulting in exclusion from training the classifier. The most optimal model for the detection of the combined end point of HF deterioration showed a specificity of 97.2% and a sensitivity of 5.3% in the 2 weeks prior to an event (area under the curve=0.59) after leave-one-subject-out cross-validation analysis.
CONCLUSIONS: An MPM using a noninvasive wrist-worn device, measuring HR, IBI, RR, AC, and EE, showed high specificity but low sensitivity for the prediction of decompensated HF and HF-related mortality. Low sensitivity likely reflects the extreme class imbalance and sequences with low data quality (especially HR, RR, and sleep), resulting in exclusion from training the MPM in our older, real-world HF cohort. Future studies should improve data fidelity and enroll larger cohorts to address class imbalance and enhance predictive performance.
