(P08-004-25) Predicting PV Severity: A Cross-Sectional Study Using Machine Learning Methods and Multifactorial Data Analysis From a High Prevalence Region

Objectives: According to LGBM Feature Importance, the top five features contributing to the PV severity prediction model for the mild PV, in descending order, were the consumption of molybdenum, fructose, vitamin C, biscuit, and burger, respectively.

Meanwhile, when assessing feature importance through SHAP, the top five features for the mild PV, in descending order, were associated with the consumption of fructose, vitamin C, biscuit, molybdenum and burger.

Methods: A total of 138 pemphigus vulgaris cases were studied, of which 108 had pemphigus disease area index (PDAI) < 15, and 30 had PDAI >15. Dietary intakes were assessed using a valid 168-item FFQ and the quartiles of dietary vegetables intakes were calculated. In this study, the analyzed dataset comprised 138 entries and 269 features. The Light Gradient Boosting Machine (LGBM) framework was employed, and optimized hyperparameter sets were obtained through Bayesian Optimization. Consequently, five sets of hyperparameters were utilized to create five distinct LGBM models, implemented in combination with Cross-Validation – Repeated Fold with five splits and 10 repeats, forming the so-called Megamodel. The log loss served as the evaluation metric, and KNN and XGB models were also explored. While each model has its own strengths and weaknesses, the LGBM framework is considered more promising for handling features, feature importance, and computation. For the test dataset, the log loss of the LGBM framework was 0.4187, with an AUC of 0.7913. Two distinct methods were employed to calculate feature importance within the LGBM framework: LGBM Feature Importance and SHAP. Feature Importance quantifies the contribution of a specific feature to the model's predictions.

Results: According to LGBM Feature Importance, the top five features contributing to the PV severity prediction model for the mild PV, in descending order, were the consumption of molybdenum, fructose, vitamin C, biscuit, and burger, respectively.

Assessing feature importance through SHAP showed that the top five features for the mild PV, in descending order, were associated with the consumption of fructose, vitamin C, biscuit, molybdenum and burger.

Conclusions: Our analyses have shown in principle the potential use of machine learning methods to identify items for potential evaluation for the prevention of severe PV.

Funding Sources: No funding