Fully automated approach of machine learning combined with deep learning: How to predict the onset of major cardiovascular events in NAFLD patients. (March 2023)

• Record Type: Journal Article
• Authors: Cirella, A.; Sinatti, G.; Bracci, A.; Evangelista, L.; Bruno, P.; Santini, S.J.; Greco, G.; Guzzo, A.; Calimeri, F.; Di Cesare, E.; Balsano, C.
• Is Part Of: Digestive and liver disease. Volume 55(2023)Supplement 1
• Page Start: S32
• Publication Date: 2023-03
• Subjects: Digestive organs -- Diseases -- Periodicals; Liver -- Diseases -- Periodicals; 616.33005
• DOI: 10.1016/j.dld.2023.01.061
• Languages: English
• ISSNs: 1590-8658

Abstract

Abstract : Introduction: Growing evidence indicates that the presence of NAFLD increases cardiovascular (CV) morbidity and mortality. Coronary arteries disease (CAD) is detected by Coronary CT, moreover CT is used also for the determination liver steatosis. Aim: The aim of our study was to perform a prognostic's stratification risk of presence of CAD with a combined ML/DL approach in NAFLD patients Materials and Methods: Our retrospective study analyzed clinical data and CT images of 401 patients (217 males and 184 females), who underwent coronary CT between 2017 and 2021. Hounsfield Unit (HU), Agatston score, Fib-4 score were used to measure radiodensity, calcium score (CS) and the degree of fibrosis, respectively. Fully automated algorithms were trained with clinical data, among them the best were Support vector machine, Random Forest and XGBoost for ML, while Fully Convolutional Network and Long short-term memory for DL. Results: We performed a binary classification to compare the most used ML (XGBoost, RF, SVM) and DL (FCN, LSTM) algorithms for clinical data. Our algorithms predicted absent and severe CAD with a mean accuracy of 96% and a mean specificity of 97%. After using a multiclassification approach our algorithms were able to distinguish patients in 5 classes from healthy patients to patients affected by NAFLD and CVD, with a mean accuracy of 87% and a mean specificity of 86% for ML/DL. To improve the performance of this prediction models, we are integrating them with DL algorithms for liver CT images (e.g. 3D-UNet) Conclusion: Our integrated ML/DL approach could be used in practice to flag NAFLD patients at high risk of CVD.