Abstract
SAPHO syndrome is an inflammatory disorder with bone and cutaneous manifestations, for which whole-body bone scintigraphy (WBBS) is frequently used in diagnosis. The WBBS findings of SAPHO syndromes and secondary bone tumors (SBT) have overlapping features, posing diagnostic challenges. In this multicenter study, we aim to identify different bone and joint involvement patterns between the two disease entities through multiple methods to build machine-learning models and explore interpretable variables. The study included 1,193 patients, of which 593 were diagnosed with SAPHO syndrome and 600 with SBT. LASSO regression, logistic regression, and random forest techniques were applied in the training set to identify significant risk factors. Manual management and other methods were evaluated in the validation set to identify the top-performing model and the most interpretable terms. The study developed a model using 15 manually selected terms and multiple machine learning techniques, which demonstrated high diagnostic accuracy in the G1 dataset for (training AUC 0.934, testing AUC 0.929, accuracy = 88.3%, precision = 88.7%, Recall = 88.3%, F1 score = 0.882). The model was compared with logistic regression and random forest models and showed consistent results in the G2 dataset for external validation (AUC 0.957, Youden index = 0.806, sensitivity = 0.820, specificity = 0.986). The pelvis, femur, and ribs (excluding anterior ribs 1st-5th) and thoracic vertebrae 1st-8th were significant predictors of SBT, whereas the sacroiliac joints, sternum, foot, anterior ribs 1st-5th, and clavicle were indicative of SAPHO. This study assesses the effectiveness of WBBS terms in identifying SBT from SAPHO syndrome and utilizes machine learning to help screen features for patients. The final model demonstrates its dependability, providing a valuable tool for accurate and timely diagnosis.