Poster #33 - Use of machine learning to predict ADHD in children: A multi-wave longitudinal study

• In Event: Poster Session 01

Abstract

Childhood attention-deficit/hyperactivity disorder (ADHD) can lead to serious consequences in later life, including school dropout, substance abuse, and criminal justice involvement. While there are many risk factors that have been implicated (e.g., psychological, psychosocial, and genetic), some risk factors may be more impactful or predictive than others. Having a better understanding of which of these risk factors are more predictive than others could make early detection of childhood ADHD more effective and efficient.

This study used machine learning to evaluate the predictive performance of thirty-two different psychological, psychosocial, and genetic risk factors for ADHD, informed by a recent meta-analysis of risk factors for ADHD (Shepherd et al., 2021). Data were from a small two-wave longitudinal sample of preschool-aged children (Time 1 mean age = 6, Time 2 mean age = 8), with and without ADHD (N = 110). We trained and validated three prediction models with repeated k-fold cross validation and evaluated their performance. We first ran an ordinary least squares (OLS) regression to predict ADHD symptom scores based on these risk factors. Then, we fit an elastic net model with the same risk factors. Elastic-net is known for its regularizations, preventing model overfitting via variable selection and beta shrinkage. Hyperparameters responsible for the two regularizations were tuned over an expanded grid to find the lowest RMSE. Finally, we ran an optimized OLS model containing the elastic-net-informed predictors.

We evaluated each model’s predictive performance based on root-mean-squared-error and r-squared values. The OLS model demonstrated moderate prediction accuracy (r-sq = .38, RMSE = .77). Predictive performance was increased by 67% in the elastic net model (r-sq = .62, RMSE = .59). The optimized OLS model, consisting of thirteen elastic-net-informed predictors, performed better than the full OLS model (r-sq = .52, RMSE = .64). In the OLS model, Time 1 ADHD symptom scores, sleep problems, planning/organization, and cognitive ability were significantly associated with ADHD symptoms at Time 2. However, only Time 1 ADHD and sleep problems were significantly associated with Time 2 ADHD in the optimized OLS model.

Our findings support the predictive value of several (but not all) known ADHD risk factors, as informed by a prior meta-analytic investigation (Shepherd et al., 2021). We also show that including more risk factors in the OLS regression did not lead to better prediction. Instead, elastic net improved individual-based ADHD prediction using fewer risk factors, pointing to the advantage of more advanced algorithms in ADHD prediction. Future studies will replicate these findings in a larger longitudinal study.

Authors

• Zheyue Peng, University of Wisconsin-Madison
  Presenting Author

• Quanfa He, University of Wisconsin - Madison

• James J Li, University of Wisconsin - Madison