Background/Purpose: Cardiac involvements cause of morbidity and mortality globally in rheumatoid arthritis (RA). Myocardial dysfunction may arise from a number of distinct processes, including myocardial inflammation and myocardial fibrosis, any of which may be active in RA.  Global longitudinal strain (GLS), using non-contrast feature tracking cardiac magnetic resonance (FT-CMR), has been reported to be significantly associated with the extent of myocardial fibrosis. Late gadolinium enhancement (LGE) corresponds to histopathologic zones of myocardial fibrosis and necrosis. In the last few years, it has been suggested that artificial neural networks (ANNs) approaches are an established method for analyzing large datasets. ANNs could be a useful tool for prediction in medical scenarios. We aimed to predict myocardial fibrosis in RA assessed by FT-CMR and LGE using ANNs models.

Methods: RA patients and controls with no known heart disease or risk factors were enrolled. The normal threshold was defined based on the statistical±2SD limits on healthy controls in FT-CMR. In this study, a three-layered feedforward neural network model was structured to detect a myocardial abnormality from GLS and LGE, respectively. Inputs for the network were totally 22 variables including attributes (e.g., Age, Sex, BMI, Duration, RF, SDAI, MMP-3, CRP, ESR, TC, TG, HDL, LDL, sBP/dBP, HbA1c, NTproBNP, MTX, PSL, biologics used) and observed values (e.g., DAS28, ACPA). The network’s output was existence or non-existence (1 or 0) of abnormity in each target index. The back-propagation learning algorithm was used to train the ANNs structure. As the hyper parameters of ANNs, the number of neurons (10, 50, 100, 200) in the hidden layer, optimization method (Stochastic Gradient Descent, Adaptive Moment Estimation), initial learning rate (0.001, 0.005, 0.05) in the optimization method and the number of iterations (10, 50, 100, 200, 400) were determined to get the best performance using leave-one-out cross-validation. We calculated accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) to evaluate the performance of the ANNs model.

Results: We evaluated 88 patients with RA (95% women; mean age, 59.5±9. 0 years) and 30 healthy controls (100% women; mean age, 55.7±4.5 years). All 118 subjects underwent FT-CMR and 51 patients underwent LGE. Abnormal GLS values and LGE were seen in 67/88 subjects (76.1%) and 19/51 subjects (37.3%), respectively. We created a mathematical model with an AUC value of 0.73 and 0.71, respectively, able to predict abnormal GLS and LGE. The accuracy, sensitivity, specificity, PPV and NPV for prediction of abnormal GLS values and LGE were 80.7%, 88.1%, 57.1%, 86.8%, 60.0% and 74.5%, 73.7%, 70.6%, 58.3%, 82.8%, respectively.

Conclusion: We applied ANNs to identify a prediction model for myocardial fibrosis in RA assessed by CMR. We could construct a mathematical model with laboratory and clinical items, and treatment, to potentially identify asymptomatic RA patients with myocardial fibrosis. This prediction tool could be used potentially in a clinical practice setting to stratify RA patients according to myocardial fibrosis.

« Back to 2019 ACR/ARP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/prediction-of-myocardial-fibrosis-in-rheumatoid-arthritis-assessed-by-cardiac-magnetic-resonance-imaging-using-artificial-neural-networks-models/