Analyzing the Mechanism Behind Age-agnostic Prediction of Diastolic Dysfunction Using Echocardiography Variables in Deep Neural Networks

ObjectiveThis investigation delved into the inner workings of previously published Deep Neural Networks (DNNs) designed to detect changes in diastolic function related to age using echocardiographic parameters. The primary goal was to decipher the predictive mechanism and determine whether biological age and gender played concealed roles within the DNN model. MethodsWe conducted a secondary analysis of a previously published DNN model that was trained using data from 1,009 patients (average age: 62{+/-}17 years, 57% females) to forecast risk phenogroups based on nine echocardiographic parameters. This model was assessed on both an internal cohort (n=243, mean age = 62 {+/-} 17 years, [~]57% females) and an external validation cohort (n=5596, mean age = 76 {+/-} 5 years, [~]57% females). To forecast biological age and gender, we developed linear regression and classification models employing hidden layer activations from the DNN. Model performance was assessed using Pearsons correlation for regression, accuracy, and area under the curve (AUC) metrics for classification. ResultsUpon scrutinizing the hidden layer activations, we observed that the model accurately captured biological age in both younger and older populations, particularly in low-risk phenogroups, with robust correlations for the entire population (0.94, p<0.001), males (0.90, p<0.001), and females (0.94, p<0.001). In high-risk phenogroups, the correlations were lower, standing at 0.31 (p=0.274) for the entire population, 0.76 (p=0.003) for males, and 0.11 (p=0.723) for females. Predicting gender as an underlying factor resulted in an accuracy rate of 58.02% and 52.27%, accompanied by an AUC of 0.65 for both validation cohorts. ConclusionThis study underscores that the latent space within DNNs maintains a link with age in relation to diastolic functional parameters, offering a solution that is independent of age for predicting diastolic dysfunction. The dissection of the network can further enhance our comprehension of the information learned by DNNs, thereby providing novel pathophysiological insights for medical professionals.