Löfgren's syndrome sarcoidosis and Non-LS sarcoidosis prediction using 1d-Convolutional neural networks

Abstract

Abstract Lofgren syndrome (LS) sarcoidosis and non-LS sarcoidosis (non-LS) can be very difficult to distinguish and identify due to their clinical and genetic similarities. However, these subclasses of sarcoidosis have different prognoses and thus require distinct treatments. For this reason, an accurate classification method is critical in diagnosing the correct form of sarcoidosis. We propose a method based on convolutional neural networks (CNN): we apply a one-dimensional CNN (1d-CNN) model to mass cytometry cell measurements and adapted feature map concatenation techniques to train on single-feature vectors. The model used the same 1d-CNN structure to predict on multiple cell input dimensions, including FCS files. The model achieved an area under the receiving operating characteristic curve score of 1.00 for the classification of LS versus non-LS patients on simulated multi-cell input data. We compared our model with nine state-of-the-art methods and showed that our model outperformed all others, with a mean F1 score of 0.98. In addition, the model's biomarker weights and molecular signature aligned with previous sarcoidosis biomarker research and yielded insight into the pathogenesis of the disease. The proposed model can be applied to provide the precise diagnosis and medical treatment for patients with both types of sarcoidosis.