Scientists at the University of Michigan have developed a machine-learning model that predicts symptoms of childhood attention-deficit hyperactivity disorder (ADHD) from neurocognitive tests and child characteristics.
This research translational psychiatry journal.
study: Generalizing and predicting childhood ADHD symptoms from neurocognitive tests and adolescent characteristics. Image credit: Ahlapot/Shutterstock.com
Background
Attention-deficit/hyperactivity disorder (ADHD) is a complication of childhood psychiatric disorders characterized by poor concentration, confusion, impulsive behavior, and excessive movement in inappropriate situations.
This disorder is associated with many negative consequences for young people, including poor academic performance, substance use and externalizing behavior, and persistent economic crises.
Impaired neurocognitive development is a major causative factor for childhood ADHD symptoms. However, evidence suggests that neurocognitive measures cannot distinguish ADHD patients or accurately characterize or predict ADHD phenotypes.
In the current study, scientists used neurocognitive tests and child traits to develop and validate a generalizable machine learning model for predicting trait ADHD symptoms from independent data.
research design
Scientists collected Adolescent Brain Cognitive Development Study (ABCD) data to develop a predictive model.
The ABCD study is a large longitudinal consortium study that enrolled over 10,000 children aged 9-10 years from 21 different centers in the United States. The vast amount of data from diverse communities in the United States makes the ABCD study a valuable resource for developing generalizable predictive models of her ADHD symptoms in young people.
Baseline demographics and biometrics, geo-coded neighborhood data, adolescent reports of child and family characteristics, and neurocognitive measures were incorporated into the model, with parent and teacher predicted the ADHD symptoms they reported. .
Two modeling strategies were used in the study. A comprehensive predictive modeling approach was used to estimate the predictive value of individual traits regardless of possible redundancy. In addition, we developed a more economical model using sparse predictive modeling techniques.
All research models were trained using one-site-out cross-validation to determine generalizability.
For each excluded study site, a model was trained using data from all other sites, which was then used to generate predictions of ADHD symptom scores for both training and excluded data. it was done.
key findings
The study model predicted approximately consistent ADHD symptoms across all study centers at 1 year. On average, both the comprehensive and sparse modeling methods produced identical predictions of his ADHD symptoms at this time point.
At sites excluded from the model fitting process, predictive models explained 15–20% and 12–13% of individual variability in ADHD symptoms at 1 and 2 years, respectively. These observations demonstrate that the model is highly generalizable to independent (unseen) data.
Analysis of the degree of predictive information for each trait tested revealed significant negative predictive correlations between neurocognitive measures (reasoning, memory, verbal ability, processing speed, cognitive efficiency) and ADHD symptoms. rice field.
Among the child-reported characteristics, impulsive behavior, screen time, high levels of family conflict, and low levels of parental supervision and school involvement were most associated with ADHD symptoms. showed a predictive effect.
Among demographic and geo-coded features, male gender and neighborhood poverty had the highest predictive effects.
Application of sparse predictive modeling methods identified 13 traits from multiple domains that significantly contributed to symptom prediction.
These characteristics included gender, neurocognitive measures, screen time, parental supervision, and child-reported impulsive behaviors.
Results obtained from predictive models that individually included or excluded neurocognitive data, child self-reported data, and demographic data revealed that neurocognitive testing can significantly improve the predictive power of these models. rice field.
The model was shown to lose significant predictive power when transferred from the training data to the exclusion data. The decline in efficacy was more pronounced at his 2 years.
However, cognitive indices, self-reported impulsive behavior, gender, and screen time remained the most important predictive features in these models.
Significance of research
In this study, a machine learning model generalized to unseen data from different samples using neurocognitive data, demographic data (gender), self-reported impulsive behavior, and screen time as key features. found to be predictive of childhood ADHD symptoms.
