Document resource
Introduction Early diagnosis of systemic sclerosis (SSc), especially in the edematous phase, is challenging. This study aimed to develop an AI model capable of distinguishing the edematous phase of SSc from other clinically similar conditions and to evaluate its accuracy.Material and Methods Short video clips were collected from two groups: (a) SSc patients with puffy skin in the hands and/or feet during the edematous phase and (b) non-SSc patients with skin edema requiring differentiation from SSc. The videos were captured and subsequently processed at a rate of 5 frames per second, resulting in 30 to 45 frames per video. Each frame was focused on a single area of interest, which was labeled and annotated using a bounding box. The images was resized to a standard input for AI training and testing, as illustrated in figure 1. Each frame was analyzed by the model and classified into one of three categories: SSc, non-SSc, or undetectable. AI development involved analyzing skin responses to finger pressure, recording the pressing and rebound phases. The collected videos were divided into three sets: 70% for AI training, 10% for validation, and 20% for accuracy testing. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated.Results A total of 2,080 videos from 22 SSc and 38 non-SSc patients were analyzed. Among the 38 non-SSc cases, the most common underlying conditions were renal disease and nephrotic syndrome (15 cases, 39.5%), followed by deep venous thrombosis (6 cases, 15.8%) and left- or right-sided heart failure (5 cases, 13.2%). Five AI models were evaluated at accuracy thresholds of 70%, 75%, 80%, and 85%. At thresholds of 70–80%, all models achieved sensitivity above 99% and specificity exceeding 98%. The extra-large model demonstrated the highest sensitivity, specificity, positive predictive value, and negative predictive value at the 85% threshold, outperforming the nano, small, medium, and large models.Conclusions The developed AI model demonstrated high accuracy in distinguishing the edematous phase of SSc from other similar skin edema across multiple confidence thresholds, supporting its potential as an effective diagnostic tool.Abstract P.396 Figure 1