Abstract 18451: Point-of-Care Aid-to-Diagnosis for Heart Failure Using Artificial Intelligence Based on Seismocardiography Acquired With a Smartphone in the Emergency Department

Abstract

BACKGROUND The number of new heart failure (HF) diagnoses in hospitals constantly increases. The associated clinical signs lead to a delay in treatment. Reliable and rapid diagnostic assistance is needed at earlier stages. Kinocardiography (KCG), based on Seismocardiography (SCG), allows the computation of non-invasive cardiac kinetic energy that is used to characterize the cardiac function and could detect HF among patients at risk. PURPOSE In a multi-centric study, test the performance of smartphone-based KCG recording in diagnosing HF in the emergency department (ED). METHODS Adult patients presenting with dyspnea in the ED were recruited. Relevant clinical characteristics were recorded. A blood sample was collected for NT-proBNP measurement, and a 180-sec KCG was acquired with a smartphone. The patient is supine on the back; the smartphone is placed on his chest to acquire the precordial vibrations due to the heartbeats. Cardiac kinetic energy (CKE) metrics are derived from these vibrations, including diastolic gradient kinetic energy ( ΔiK diastolic ). The final HF diagnosis was assessed from the medical record at discharge from the ED. The primary outcome is the performance of an AI-based on CKE metrics at classifying HF, measured using metrics including the area under the receiver operating characteristic curve (AUROC), negative predictive value (NPV), and specificity, with two-sided 95% CIs. RESULTS 375 patients were recruited; among these, 137 patients were diagnosed with HF (78 [70.2;85.7] years, 26.8 [23.4;29.6] kg/m 2 , 45.7% male, 4843.0 [3199;7845] pg/ml, 88 HFpEF, 17 HFmrEF, 37 HFrEF) and 243 without HF (65 [54;74] years, 27.9 [25.3;31.3] kg/m 2 , 50.2% male, 291.0 [71;803] pg/ml). ΔiK diastolic was significantly lower in patients with HF (0.37 [0.11;0.72] vs. 0.06 [-0.05;0.29] %, p<0.005). The AI approach resulted in an AUROC of 0.85 (0.80-0.89), sensitivity of 85.0% (74.2-90.3), and NPV of 79.0% (77·7-82·9). CONCLUSION This study shows that an AI system applied to SCG acquired during a 3-mins examination with a smartphone can detect HF among patients presenting dyspnea. These findings highlight the potential for a non-invasive, workflow-adapted, point-of-care, aid-to-diagnosis for earlier diagnosis.