Machine Learning–Guided Feature Selection and Predictive Model Construction for Attention-Deficit/Hyperactivity Disorder (ADHD)

Haojie Meng¹Songtao Li²Xiwen Xing¹Ruyi Fu²Yang Li^2*Qianqi Liu^1*Xu Wang^2*

• ¹Department of Child Health, Children’s Hospital of Nanjing Medical University, Nanjing, China
• ²Children's Hospital of Nanjing Medical University, Nanjing, China

Background: Attention Deficit/Hyperactivity Disorder (ADHD) is a highly prevalent neurodevelopmental disorder, but its diagnosis remains constrained. This study aimed to identify potential candidate indicators and construct an interpretable machine learning model for the identification of ADHD. Methods: A total of 8,598 children were enrolled and classified into three groups: ADHD (n=3,678), subthreshold ADHD (s-ADHD) (n=1,495), and healthy controls (HC) (n=3,425). Data collection covered 40 variables, including demographics, routine blood counts, serum biochemical parameters, body composition and systemic inflammation markers. Analysis of Variance (ANOVA) compared differences among the three groups, and key predictors were selected via Least Absolute Shrinkage and Selection Operator (LASSO) regression. Five machine learning models (Decision Tree, Random Forest, Multilayer Perceptron, Extreme Gradient Boosting, and Light Gradient Boosting Machine [LightGBM]) were developed for three clinically relevant binary classification tasks. SHapley Additive exPlanations (SHAP) values were applied to interpret the optimal model. Results: ANOVA indicated significant differences (P < 0.05) in most parameters among the three groups. However, post-hoc Least Significant Difference (LSD) tests showed that compared with HC, the ADHD group showed elevated inflammatory markers (NLR, PLR, SII), glucose, body mass index(BMI), and body fat percentage, but reduced albumin, total cholesterol, and lymphocyte counts. Similar alterations were observed in the s-ADHD group, showing a pattern consistent with that of the ADHD group. LASSO regression (λ.1se=0.038) selected 11 core predictors, with age, RDW-SD, sex, calcium, glucose, and albumin among the most contributing variables. Among the models, LightGBM demonstrated the best performance when distinguishing ADHD from HC (AUC=0.924 with 36 features vs. AUC=0.885 with 11 features). However, the model failed to effectively distinguish between ADHD and s-ADHD. Conclusions: This study reveals potential candidate indicators of ADHD and establishes an interpretable, low-cost machine learning model based on routine clinical data, offering a promising tool for early screening and clinical decision support.