A bayesian diagnostic-accuracy meta-analysis on hypertrophic cardiomyopathy diagnosis with the use of machine learning

Background Hypertrophic cardiomyopathy (HCM) is one of the most common inherited cardiac disorders, characterized by abnormal thickening of the heart muscle, which can lead to severe outcomes such as sudden cardiac death. Early detection is essential for effective management and risk stratification. Recent advancements in machine learning (ML) have shown promise in improving diagnostic accuracy for HCM, especially in clinical and imaging data. Purpose This study aimed to perform a Bayesian diagnostic-accuracy meta-analysis to comprehensively assess the diagnostic performance of ML models in identifying hypertrophic cardiomyopathy. The goal was to synthesize the results from various studies, evaluate the consistency of model performances, and provide actionable insights into the effectiveness of ML-based diagnostic tools for HCM. Methods A Bayesian diagnostic-accuracy meta-analysis was employed to pool the results of 18 distinct ML models used for the diagnosis of HCM. The analysis was conducted using non-informative priors, with results reported as effect sizes with 95% credible intervals (CrI) for diagnostic performance metrics, including sensitivity, specificity, and diagnostic odds ratio (DOR). Diagnostic accuracy data were extracted from studies that used various imaging modalities and clinical data inputs to train ML models. Statistical heterogeneity was assessed using the mean correlation between the studies and the funnel plot to detect any potential publication bias. Results The pooled sensitivity of the 18 ML models was 0.89 [0.85, 0.92]. The pooled specificity was 0.94 [0.90, 0.97]. The log DOR was 4.95 [3.83, 6.12], indicating a strong relationship between the diagnostic test results and the presence of HCM. The high log DOR reflects a notable ability of the models to discriminate between affected and unaffected individuals. No significant asymmetry was detected in the funnel plot, suggesting that the results were not biased by small-study effects. Furthermore, the mean correlation between the studies was 0.849, demonstrating high consistency across the included studies in terms of diagnostic performance. Conclusion Machine learning models exhibit high diagnostic accuracy for hypertrophic cardiomyopathy, with pooled sensitivity and specificity values of 0.89 and 0.94, respectively. The results support the potential of ML algorithms to serve as effective diagnostic tools in clinical settings for the early detection and management of HCM. These findings underscore the importance of further validating ML-based diagnostic tools across diverse populations and imaging modalities to optimize their application in clinical practice. As the field of ML continues to evolve, the integration of these models into routine diagnostic workflows could improve patient outcomes by facilitating earlier detection and more accurate risk stratification for individuals with hypertrophic cardiomyopathy.