Comprehensive Summary
This meta-analysis compiles the data of eight studies that used AI to screen for left ventricular hypertrophy (LVH) and compares their results to traditional diagnostic methods. Articles were selected based on article type, study design, population, and outcome of interest criterion, with the eight chosen articles all including reports of diagnostic accuracy, sensitivity, and specificity with 50% CIs. Among these studies, either ECG or MRI were utilized as diagnostic tools, and the AI models used included convolutional neural networks (NNs) (3 studies), ensemble NNs (2 studies), one back-propagation NN, and non-NN methods (2 studies). Statistical analysis was performed with R programming to calculate pooled sensitivity (69%) and specificity (87%) and generate a receiver-operating characteristic (SROC) curve (0.87). In comparison to Sokolow-Lyon’s and Cornell’s criteria for LVH, the pooled SROC was higher than both, indicating the AI had a higher diagnostic accuracy than the known criteria. However, the pooled data had a higher sensitivity and lower specificity than both traditional criteria, indicating AI may flag increased false positives.
Outcomes and Implications
AI has the potential to aid clinicians as a fast and relatively accurate tool to diagnose LVH. Additionally, AI can be used to assess multiple diverse forms of data, and by using non-black box models, clinicians can better understand which data points most influenced the AI’s prediction. However, due to this meta-analysis only including observational studies, the presence of residual confounding variables, and heterogeneity of the included studies, future research on the diagnostic capabilities of AI should be conducted to assess specifically for accuracy, as well as generalizability and cost-effectiveness.