02453nas a2200349   4500000000100000008004100001260005300042653002100095653001500116653002100131653001700152653001000169100001000179700001000189700001100199700001100210700000900221700001200230700001100242700001100253700000900264700000900273700001000282700001100292700001000303700000900313245013900322856006900461300001100530520154800541022001402089       2026                            d  c08/2026bSpringer Science and Business Media LLC10aLeprosy reaction10aPrediction10aMachine learning10aRisk factors10aChina1 aGuo Y1 aYin L1 aYang H1 aYang X1 aYu X1 aZhang C1 aZhou L1 aZhao F1 aLu S1 aHe Q1 aHan L1 aWang W1 aLiu Y1 aLi Y00aPredicting leprosy reactions: a machine learning framework incorporating clinical, demographic, and healthcare system factors in China  uhttps://www.nature.com/articles/s41598-026-65229-6_reference.pdf  a1 - 303 a<p>Our study evaluated nine machine learning algorithms using a Leprosy Management Information System dataset of 3,316 newly diagnosed leprosy patients admitted between 1985 and 2023 in Yunan Province, China. The model compared performance in a train-test split of 70 − 30 using the area under receiver operating characteristic curve (AUC-ROC), sensitivity, specificity, and F1 score. The SHapley Additive exPlanation technique ranked feature importance. Leprosy reaction was 12.85% (95% CI: 11.71–13.99%). Naive Bayesian Networks (NBN) achieved an AUC of 0.753 in training and 0.761 in testing, offering the highest sensitivity of 0.418 in training and 0.441 in testing among all models to minimize missed leprosy reaction cases. The model maintained specificity of 0.92 and F1 score of 0.4, indicating reliable performance across metrics. Extreme gradient boosting (XGB) showed superior overall discrimination (AUC: 0.824 training, 0.772 testing); however, lower sensitivity (0.251) limited clinical utility. The remaining seven models had competitive AUC with NBN. The potential key predictors revealed by interpretability analysis were sequentially ranked: family history of leprosy, education level, detection mode, source of infections, treatment regimen, bacterial index, gender, age, marital status, treatment duration, and time to initiate treatment. Bayesian model provided intuitive risk assessments for resource-limited settings, providing that model selection must weigh statistical precision against clinical needs.</p>
  a2045-2322