Generalizable prediction of childhood ADHD symptoms from neurocognitive testing and youth characteristics

Abstract

Importance: Childhood Attention-Deficit/Hyperactivity Disorder (ADHD) symptoms are linked to many negative outcomes and widely believed to result from disrupted neurocognitive development. However, evidence for the clinical value of neurocognitive assessments in this context has been mixed and there have been no large-scale efforts to quantify the potential of neurocognitive abilities, along with data from other domains (e.g., child and family characteristics, demographics), for use in generalizable machine learning models that predict individual youths’ ADHD symptoms in independent data.Objective: To develop and test, in a large and diverse sample of youth, cross-validated machine learning models for predicting youths’ ADHD symptoms using demographics, child and family characteristics, and neurocognitive assessments. Design, Setting, and Participants: Data were from participants of the Adolescent Brain Cognitive Development Study (ABCD), a collaborative consortium that recruited a diverse sample of youth (ages 9-10 at baseline) across 21 sites in the United States. Participants with complete parent and teacher reports of ADHD symptoms at the 1-year follow-up (n = 5,900) were included.Exposures: Predictor variables were from the ABCD baseline timepoint and included basic demographic and biometric measures, geocoded community and neighborhood data, youths’ self-report about child and family characteristics, and neurocognitive tests.Main Outcome: Models predicted a latent cross-rater ADHD symptoms factor derived from a bifactor model of parent and teacher ratings at the 1-year follow-up.Results: Predictive models explained about 20% of the variance in ADHD symptoms for ABCD Study sites that were left out of the model-fitting process, with high generalizability across sites and only trivial loss of predictive power when applied to left-out data. Features from multiple domains contributed meaningfully to prediction, including neurocognition, sex, self-reported impulsivity, screen time, parental monitoring, school involvement, and socioeconomic disadvantage. A sparse model including only seven neurocognitive measures, male sex, screen time, and two impulsivity scales displayed comparable performance to larger models.Conclusions and Relevance: This work quantifies the informational value of neurocognitive abilities and other child characteristics for predicting individual children’s ADHD symptoms in unseen data and provides a foundational method for the prediction of ADHD symptoms in ABCD and across other research and clinical contexts.

A systematic review on the application of machine learning models in psychometric questionnaires for the diagnosis of attention deficit hyperactivity disorder

“…Table 1 summarizes the characteristics of the articles that refer to the use of ML on psychometric questionnaires for the diagnosis of ADHD. Of the 17 articles reviewed eight used random forest (RF) (Cordova et al, 2020; Davakumar & Siromoney, 2020; Goh et al, 2023; Grazioli et al, 2023; Haque et al, 2023; Kim et al, 2021, 2023; Tachmazidis et al, 2020), seven decision tree (DT) (Ardulov et al, 2021; Bledsoe et al, 2020; Chen et al, 2023; Christiansen et al, 2020; Grazioli et al, 2023; Haque et al, 2023; Tachmazidis et al, 2020), six support vector machine (SVM) (Bledsoe et al, 2020; Chen et al, 2023; Davakumar & Siromoney, 2020; Duda et al, 2016; Grazioli et al, 2023; Tachmazidis et al, 2020), four linear discriminant analysis (LDA) (Chen et al, 2023; Duda et al, 2016, 2017; Kim et al, 2021), three k‐nearest neighbours (KNN) (Chen et al, 2023; Kim et al, 2021; Tachmazidis et al, 2020), three Gaussian Naïve Bayes (Chen et al, 2023; Haque et al, 2023; Tachmazidis et al, 2020), three logistic regression (LR) (Chen et al, 2023; Duda et al, 2016; Tachmazidis et al, 2020), three artificial neural network (ANN) (Chen et al, 2023; Davakumar & Siromoney, 2020; Lin et al, 2023), two Lasso regression (Duda et al, 2016; Weigard et al, 2023), two gradient boosting (GB) (Chen et al, 2023; Kim et al, 2023), two elastic net (ENet) (Duda et al, 2017; Liu et al, 2023), one Q‐learning (Ardulov et al, 2021) and one principal components regression (PCR) (Weigard et al, 2023).…”

“…On the other hand, the questionnaires used in the different studies on which ML techniques were applied varied widely. Among them, 13 studies used scales that diagnosed ADHD (Bledsoe et al, 2020; Chen et al, 2023; Christiansen et al, 2020; Cordova et al, 2020; Davakumar & Siromoney, 2020; Goh et al, 2023; Grazioli et al, 2023; Haque et al, 2023; Kim et al, 2023; Lin et al, 2023; Liu et al, 2023; Tachmazidis et al, 2020; Weigard et al, 2023), four articles used questionnaires assessing social skills (Duda et al, 2016, 2017; Goh et al, 2023; Kim et al, 2021), an article administered a test diagnosing ASD (Ardulov et al, 2021), a study used a questionnaire that assessed personality (Kim et al, 2021), an article used a scale that measured intelligence (Grazioli et al, 2023), and an article used a test that assessed academic performance (Goh et al, 2023).…”

“…Cross‐validation is a technique for estimating the estimated performance of an ML model given a given set of parameters while reducing the risk of overfitting (Poldrack et al, 2020). A subset of 11 articles out of 17 employed k‐fold cross‐validation (Ardulov et al, 2021; Chen et al, 2023; Christiansen et al, 2020; Cordova et al, 2020; Davakumar & Siromoney, 2020; Duda et al, 2016, 2017; Goh et al, 2023; Haque et al, 2023; Lin et al, 2023; Liu et al, 2023), and four articles out of 17 used leave‐one‐out cross‐validation (LOOCV; Bledsoe et al, 2020; Grazioli et al, 2023; Tachmazidis et al, 2020; Weigard et al, 2023).…”

“…Focusing on the sample used, most of the studies work with a population of children and adolescents for their analyses (Ardulov et al, 2021; Bledsoe et al, 2020; Cordova et al, 2020; Davakumar & Siromoney, 2020; Duda et al, 2016, 2017; Faraone et al, 2021; Goh et al, 2023; Grazioli et al, 2023; Haque et al, 2023; Kim et al, 2023; Lin et al, 2023; Weigard et al, 2023). Only a few evaluate an adult sample (Chen et al, 2023; Christiansen et al, 2020; Grazioli et al, 2023; Kim et al, 2021; Liu et al, 2023; Tachmazidis et al, 2020).…”