Detecting heart failure using novel bio-signals and a knowledge enhanced neural network. (March 2023)

• Record Type: Journal Article
• Authors: Nogueira, Marta Afonso; Calcagno, Simone; Campbell, Niall; Zaman, Azfar; Koulaouzidis, Georgios; Jalil, Anwar; Alam, Firdous; Stankovic, Tatjana; Szabo, Erzsebet; Szabo, Aniko B.; Kecskes, Istvan
• Is Part Of: Computers in biology and medicine. Volume 154(2023)
• Publication Date: 2023-03
• Subjects: Heart failure -- HFpEF -- HFmrEF -- HFrEF -- Bio-signal -- Primary care -- Early detection; AF Atrial Fibrillation -- ALVSD Asymptomatic LVSD -- AR Aortic Regurgitation -- AS Aortic Stenosis -- BNP B-type Natriuretic peptides -- CHART Cardio-HART™ from Cardio-Phoenix -- DCM Dilated Cardiomyopathy -- DD Diastolic Dysfunction -- DDIM Impaired Relaxation type DD -- ECG Electrocardiography -- ECHO Echocardiography (TTE) -- GLS Global longitudinal strain -- HART ECHO-like findings estimated by Cardio-HART AI system -- HF Heart Failure -- KENN Knowledge-enhanced neural networks -- LAE Left Atrial Enlargement -- LVSD Left Ventricular Systolic Dysfunction -- MCG Mechanical force bio-signal -- MI Myocardial Infarction -- MR Mitral Regurgitation -- PCG Phonocardiography -- RAE Right Atrial Enlargement -- RVE Right Ventricular Enlargement -- PH Pulmonary Hypertension -- WMA Wall Motion Abnormality; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2023.106547
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Background: Clinical decisions about Heart Failure (HF) are frequently based on measurements of left ventricular ejection fraction (LVEF), relying mainly on echocardiography measurements for evaluating structural and functional abnormalities of heart disease. As echocardiography is not available in primary care, this means that HF cannot be detected on initial patient presentation. Instead, physicians in primary care must rely on a clinical diagnosis that can take weeks, even months of costly testing and clinical visits. As a result, the opportunity for early detection of HF is lost. Methods and results: The standard 12-Lead ECG provides only limited diagnostic evidence for many common heart problems. ECG findings typically show low sensitivity for structural heart abnormalities and low specificity for function abnormalities, e.g., systolic dysfunction. As a result, structural and functional heart abnormalities are typically diagnosed by echocardiography in secondary care, effectively creating a diagnostic gap between primary and secondary care. This diagnostic gap was successfully reduced by an AI solution, the Cardio-HART™ (CHART), which uses Knowledge-enhanced Neural Networks to process novel bio-signals. Cardio-HART reached higher performance in prediction of HF when compared to the best ECG-based criteria: sensitivity increased from 53.5% to 82.8%, specificity from 85.1% to 86.9%, positive predictive value from 57.1% to 70.0%, the F-score from 56.4% to 72.2%, and area under curve from 0.79 to 0.91. The sensitivity of the HF-indicated findings is doubled by the AI compared to the best rule-based ECG-findings with a similar specificity level: from 38.6% to 71%. Conclusion: Using an AI solution to process ECG and novel bio-signals, the CHART algorithms are able to predict structural, functional, and valve abnormalities, effectively reducing this diagnostic gap, thereby allowing for the early detection of most common heart diseases and HF in primary care. Graphical abstract: Image 1 Highlights: ECG shows low sensitivity for heart structural abnormalities. ECG shows low specificity for functional abnormalities. The sensitivity of the HF-indicated ECHO-findings is doubled by the AI solution. CHART reaches 30% higher sensitivity than the best ECG-finding for HF detection. Bio-signal-based HART-findings are disease equivalent to ECHO-findings.