Development of a Potentially Individualized Algorithm to Detect Heart Failure Events through Home Telemonitoring

Abstract

Aims: Home telemonitoring represents a promising approach to reduce heart failure (HF) patients’ hospital readmissions. The aim of the study was, in a first step, to identify the monitored parameters’ characteristics that are predictive of HF events. In a second step, it was to build a prediction score by combining both the identified characteristics and the patients’ clinical prognosis. Methods: Patients completed 6-month blind daily body weight, blood pressure and pulse measurements. A cardiac composite endpoint (CCE) of death, hospitalization or urgent visit was considered. A series of signal-derived statistics (SDS) were computed on 3, 5 and 7 days’ time windows. A signal score for CCE prediction was built by including SDS in a first logistic model using a subset of signal set (training) and its accuracy was assessed in another subset (testing). A clinical score was computed using the Meta-Analysis Global Group in Chronic Heart Failure formula. Both scores were combined using a second logistic model. We compared the three scores using ROC curves. Results: Monitoring was completed by 146 patients and 96 CCE occurred in 61 patients. The first logistic model resulted in a signal score which combined 7 SDS including body weight’s variability on 3 consecutive days, body weight’s increase on 3 and 7 consecutive days, pulse’s variability on 3 and 7 consecutive days, diastolic  blood pressure’s mean on 3 consecutive days, differential pressure’s variability on 3 consecutive days. The signal score had ability in predicting CCE occurrence (training set: AUC= 0.796, P < .001; testing set: AUC=0.830, P < .001). The second logistic model resulted in a combined score that improved CCE prediction (training set: AUC= 0.830, P < .001; testing set: AUC= 0.891, P < .001) with 92% sensitivity and 77% specificity. Conclusions: Signal data and clinical data provide additive information to risk prediction.