Classificação da doença arterial coronariana usando a variabilidade da frequência cardíaca: uma abordagem de rede neural profunda com explicação agnóstica de Modelo Interpretável Localmente (LIME)

Abstract

Aim: We aimed to create a DNN capable of detecting subjects with coronary artery disease from healthy subjects, using heart rate variability indexes as a parameter, with as few variables as possible, increasing trust in model output and making the data collection for model deployment in a clinical setting. Methods: We used data collected from 24h holter data from Telemetric and Holter Electrocardiogram Warehouse (THEW) database of 354 patients. Heart rate variability data was extracted from the holters and worked as input of a deep neural network, and the most explanatory models’ variables were found using the LIME. Results: Time and frequency domain showed higher accuracy and lower loss. LIME was used to identify the five most explanatory variables, that were later reintroduced to the model separately. The accuracy and loss were maintained, indicating that the variables highlighted by LIME were the most important and most explanatory. Conclusion: Heart rate variability data can be used to evaluate coronary artery disease from healthy subjects using deep neural networks, and the LIME can simplify the model, increasing its trustworthiness.