Mathematical software tool that helps clinicians to classify clinicand and spurious alerts in patients with heart failure and cardiac implantable devices

Abstract

Abstract Introduction Heart failure (HF) is the leading cause of cardiovascular hospitalisation in developed countries. More and more HF patients are receiving cardiac implantable electronic devices (CIED) that collect physiological data that may be used to predict HF decompensations and avoid hospitalisation through early medical interventions. Some CIEDs have a Multiparameter heart failure index (HF-I), which is an algorithm based on the information provided by five sensors (heart sounds, respiration, chest impedance, heart rate and activity level) capable of predicting 34 days before HF decompensation episodes when the reported index is ≥16 (HF-I alert) with a sensitivity of 70%, although with a significant rate of false positive HF diagnosis. Purpose Develop a computational tool to analyse HF-I curve dynamics to find the mathematical variables with the highest predictive accuracy of true clinical events in a tertiary hospital. Methods We developed a computational tool to analyse curves of 71 HF-I alerts in 25 patients with HF who were implanted with an implantable cardiac defibrillator (ICD) or cardiac resynchronisation-defibrillator (CRT-D) between 2017 and 2021, using mathematical software (fig 1). It includes peak, time to peak, area under the curve, increase and decrease slopes, variability and monotony. Alerts (HF-I ≥16) were clinically classified as real HF decompensation (clinical alerts) and false positive HF decompensation (spurious/subclinical alerts). Results Median age of the study group was 68 years, and 60% had non-ischemic cardiomyopathy. Median left ventricle ejection fraction (LVEF) was 29%; 60% had non-ischaemic cardiomyopathy. Patients with clinical alerts had higher raise slopes 7 days after alert onset (p=0.0055), higher HF-I values 3 days after crossing the alert, higher raise variability and less monotonic curves. Additionally, significant differences were observed in alert duration (p=0.004), HF-I maximum value (p< 0.0001), area under the overall alert curve (p< 0.0001), area under the alert curve at 3 / 7 days (p=0.04/0.002 respectively) and time to maximum peak of the alert (p=0,0032). HF-I value at the 7th day from the alert onset was an early and robust predictor of clinical events (P< 0.0001). A value at day 7 from alert onset ≥ 28 had a 95% sensitivity and 100% specificity for detecting clinical events (fig 2). Conclusion a computational mathematical tool can help to classify true positive clinical alerts versus spurious alerts in patients with a multiparameter heart failure index and CIED.Figure 1.Figure 2.