Feature Selection and Prediction of Pediatric Tuina in Attention Deficit/Hyperactivity Disorder Management: A Machine Learning Approach Based on Parent-Reported Children’s Constitution

Shu‐Cheng Chen, Guo-Tao Wu, Han Li, Xuan Zhang, Zihan Li, P. Y. Wong, Lefei Han, Jing Qin, Kwai-Ching Lo, Wing‐Fai Yeung, Ge RenBioengineering 2025 · DOI

TL;DR

This study leverages machine learning to personalize Traditional Chinese Medicine (TCM) treatments for children with ADHD, using parent-reported constitutional data. By identifying key features and employing models like MLP, they achieved high accuracy in predicting TCM patterns, paving the way for more effective, individualized pediatric tuina interventions.

What Problem Does This Paper Address?

Attention Deficit/Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder in children, often managed with conventional therapies that may have limitations or side effects. Pediatric tuina, a form of TCM massage, shows promise as a complementary therapy for ADHD, but its effectiveness hinges on personalized treatment based on individual TCM constitutions. Differentiating these constitutions and tailoring tuina prescriptions can be complex and relies heavily on experienced practitioners. This paper addresses the challenge of standardizing and enhancing this personalization process by applying machine learning to parent-reported data, aiming to make effective tuina treatments more accessible and consistent.

Methods

The researchers employed a two-phase machine learning workflow. First, a feature selection phase was conducted using logistic regression with five-fold cross-validation. This process aimed to identify the most informative questions from a parent-reported questionnaire that best predicted TCM patterns. Features were ranked based on their contribution to the Area Under the Curve (AUC) of predictive models.

In the second phase, TCM pattern identification was performed using four machine learning models: Support Vector Machines (SVM), Logistic Regression (LR), Multilayer Perceptron (MLP), and Random Forest (RF). These models were trained on the selected features to classify children into different TCM patterns associated with ADHD. The models were evaluated using AUC and Accuracy (ACC) metrics. Hyperparameter tuning was performed using a validation subset to prevent overfitting.

Datasets

The study utilized a dataset of 1005 parent-reported questionnaires from parents of children diagnosed with ADHD. These questionnaires contained 71 questions and corresponding labels, with the child's age also included as a demographic feature. The data was collected from multiple provincial administrative regions in China. The questionnaires were designed to assess various aspects of children's physical and psychological health according to TCM principles.

Key Insights and Ideas

  • Parental reports are valuable for TCM diagnosis: The study demonstrates that detailed parent-reported constitutional information can be effectively used to predict TCM patterns, bridging the gap between subjective parental observations and objective diagnostic tools.
  • Seven key features are crucial for ADHD TCM patterns: The research identified a core set of seven features from the questionnaire that are highly predictive of TCM patterns in children with ADHD. These features often relate to allergic predispositions, emotional regulation, and physical symptoms like mouth ulcers.
  • MLP excels in predicting TCM patterns: Among the tested models, the Multilayer Perceptron (MLP) showed the strongest performance, achieving the highest AUC and ACC, suggesting its suitability for complex pattern recognition in this domain.
  • Personalization is key for pediatric tuina: The findings underscore the importance of tailoring pediatric tuina treatments to individual TCM constitutions, moving beyond standardized approaches for better outcomes.

Key Results

  • The MLP model achieved the highest testing AUC of 0.90 and testing ACC of 0.74.
  • All four ML models (MLP, SVM, LR, RF) demonstrated strong predictive capabilities, with AUCs generally above 0.88 and ACCs above 0.70.
  • Seven features were consistently selected across five-fold cross-validation, indicating their high importance for TCM pattern identification.
  • The study successfully classified children into four TCM pattern categories, with higher accuracy observed for more prevalent patterns.

Limitations and Future Work

  • Generalizability: The study's reliance on parent-reported data from a Chinese population using a culturally specific questionnaire may limit the generalizability of findings to other populations or cultural settings. Future work should involve cultural adaptation and validation in diverse contexts.
  • Subjectivity: Parent reports, while valuable, can introduce inherent subjectivity and variability, potentially impacting data quality and model accuracy.
  • Data Imbalance: The study acknowledges challenges due to the imbalanced distribution of TCM patterns, particularly for rarer patterns, which can influence model performance. Future research should focus on expanding data collection for underrepresented patterns and exploring advanced data augmentation techniques.
  • Clinical Validation: While the ML models show promising predictive capabilities, their accuracy may not yet meet the highest standards for direct clinical application. Future work should include randomized controlled trials (RCTs) to compare the clinical efficacy of ML-guided prescriptions against those from traditional practitioners.
  • External Validation: Validating the model using independent datasets from different hospitals and regions is crucial for establishing broader generalizability.

Summary updated

Abstract

<b><i>Background</i></b>: Attention Deficit/Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder in children. Pediatric <i>tuina</i>, a traditional Chinese medicine (TCM) intervention, has shown potential in managing ADHD symptoms. Integrating machine learning (ML) into pediatric <i>tuina</i> could refine treatment personalization, allowing for a more feasible and better parent-administered use. <b><i>Methods</i></b>: We employed an ML-based model to analyze parent-reported constitutional features from 1005 children diagnosed with ADHD to predict individualized pediatric <i>tuina</i> treatments. This study focused on feature selection and the application of several ML models, including Support Vector Machines (SVM), Logistic Regression (LR), Multilayer Perceptron (MLP), and Random Forest (RF). The key task involved identifying the most relevant features for effective TCM pattern identification and diagnosis, which would guide personalized treatment strategies. <b><i>Results</i></b>: The ML models displayed strong predictive performance, with the MLP model achieving the highest Area Under the Curve (AUC) of 0.90 and an accuracy (ACC) of 0.74. Seven features were selected five times in cross-validation. This facilitated a more targeted and effective pediatric <i>tuina</i> application tailored to individual constitution. <b><i>Conclusion</i></b>: This study developed an ML-based approach to enhance ADHD management in children using pediatric <i>tuina</i>, informed by a parent-reported questionnaire. It identified seven key features for TCM pattern identification and personalized treatment strategies. MLP achieved the highest AUC and ACC.
BibTeX
@inproceedings{chen2025feature,
  author = {Shu‐Cheng Chen and Guo-Tao Wu and Han Li and Xuan Zhang and Zihan Li and P. Y. Wong and Lefei Han and Jing Qin and Kwai-Ching Lo and Wing‐Fai Yeung and Ge Ren},
  title = {{Feature Selection and Prediction of Pediatric Tuina in Attention Deficit/Hyperactivity Disorder Management: A Machine Learning Approach Based on Parent-Reported Children’s Constitution}},
  booktitle = {Bioengineering},
  year = {2025},
  doi = {10.3390/bioengineering12101012},
}
Figure