Document resource
Introduction Scleroderma (SSc) is a rare fibrosing multisystem disease with high rates of morbidity and mortality. Delays in disease recognition are common, largely due to late referral, as internists and primary care physicians are typically less familiar with its clinical features.Although characteristic facial and hand changes are frequently observed, the diagnosis relies on additional clinical sign, immunological testing & organ involvement.Our goal is to examine the ability of an AI facial and hand recognition system to identify SSc related facial and hand features, and to evaluate whether the accuracy of such a system’s predictions can be improved by a larger carefully curated data set combined with the state-of-the-art AI models.Material and Methods participants from 4 participating centers, in the U.S. (2), Egypt and Czech republic. The participants were categorized into 3 groups: SSc patients (Group A), rheumatic non-SSc pt controls (Group B) and non-rheumatic healthy controls (Group C). We evaluated 112, 114 and 205 frontal face images belonging to the three groups respectively. 21, 21, and 41 of images from the respective groups are held out as the test set. The remaining images are used as the train set.Accuracy was the primary performance metric.We investigated 6 different AI models with this data, using 112, 114, and 205 frontal face images belonging to the three groups respectively. 21, 21, and 41 images from the respective groups were designated as the test set and the remaining images were the training setResults We recruited 112 SSc pts,with age ranging from 19-79 yrs, 84% females ( table 1: demographic distribution of SSc pts )ConvNeXt v2 and Eva02, were mainly used in our cohort, also represent two dominant computer vision architectures widely used today: Convolutional Neural Network (CNN) and Vision Transformer (ViT). When a good combination of effective batch size and learning rate are used, these two models achieved an accuracy of 84-89% and 86-90% respectively, typically with less than 20 epochs of training.Conclusions We demonstrated that modern AI models can identify SSc facial features from diverse set of images (84-90% accuracy). Further validation of these data with more ssc patients, in addition to hand images may improve accuracy. Thus, enabling Internists and GPs to timely diagnose and refer SSc patients to specialized SSc centres.Abstract P.225 Table 1Demographic characteristics of scleroderma (SSc) patients