Machine learning-based models demonstrated a pooled sensitivity of 0.86 and specificity of 0.85 for detecting left ventricular dysfunction in 27,704 patients.
Does machine learning-based heart sound and ECG analysis accurately detect left ventricular dysfunction?
Machine learning-based analysis of ECG and heart sounds demonstrates high diagnostic accuracy for detecting left ventricular dysfunction, with heart sound analysis showing potentially superior performance.
Absolute Event Rate: 0% vs 0%
Abstract Background Patients undergoing routine electrocardiograms (ECG) may have undetected left ventricular dysfunction (LVD), leading to a delayed diagnosis. LVD progression can lead to ventricular remodeling, increased wall stiffness, reduced compliance, and further disease advancement. Non-invasive machine learning (ML)-based models applied to heart sound analysis or routine ECG may aid in the early diagnosis of LVD. Purpose We aimed to evaluate the diagnostic performance of ML-based ECG or heart sound analysis in detecting LVD. Methods PubMed, Embase, and Cochrane databases were systematically searched for studies evaluating the performance of ML algorithms in ECG or heart sound analysis for LVD diagnosis. A bivariate random-effects diagnostic meta-analysis was conducted to estimate the pooled sensitivity, specificity, and area under the summary receiver operating characteristic curve (AUC-SROC) with a 95% confidence interval (CI). A likelihood ratio test for bivariate meta-regression was performed to compare subgroups. Data synthesis was conducted using R software. Results Fourteen studies were included (Figure 1A), comprising 27,704 patients, of whom 5,408 (19.52%) had LVD. The algorithms used included convolutional neural networks (CNN) (n = 10), deep learning models (DLM) (n = 3), and support vector machines (SVM) (n = 1). The pooled sensitivity and specificity of ML-based models for predicting LVD were 0.86 (0.79–0.90) and 0.85 (0.80–0.89), respectively. The AUC-SROC was 0.92 (0.87–0.94). The diagnostic odds ratio (DOR) was 33.26 (18.35–60.28), the positive likelihood ratio (PLR) was 5.56 (4.08–7.56), and the negative likelihood ratio (NLR) was 0.18 (0.14–0.25). Assuming a 20% prevalence, the negative predictive value (NPV) and positive predictive value (PPV) were 95.9% and 58.2%, respectively (Figure 1B). In the subgroup analysis, ML-based ECG demonstrated a sensitivity of 0.841 (0.772–0.892), a specificity of 0.825 (0.772–0.869), and an AUC-SROC of 0.898 (0.851–0.925). ML-based heart sound analysis showed higher values, with a sensitivity of 0.916 (0.768–0.973), a specificity of 0.927 (0.879–0.957), and an AUC-SROC of 0.942 (0.884–0.980). The likelihood ratio test indicated a significant difference between subgroups (p = 0.0457) (Figure 2A). No statistically significant differences were found when comparing subgroups by ejection fraction (35% vs. 40%) (p = 0.06678) (Figure 2B) or by algorithm type (DLM vs. CNN) (p = 0.1023) (Figure 2C). Conclusion These findings support the high diagnostic performance of ML as a promising non-invasive tool for improving LVD care, aiding early diagnosis, and potentially impacting clinical practice. Innovative ML-assisted feature extraction from heart sounds showed promising performance compared to ECG-based algorithms. Further studies are needed to assess the impact of ML-based technologies on LVD prediction across different settings, cost-effectiveness, treatment patterns, and clinical outcomes.
Miranda et al. (Sat,) reported a other. Machine learning-based models demonstrated a pooled sensitivity of 0.86 and specificity of 0.85 for detecting left ventricular dysfunction in 27,704 patients.