Predicting Parent-Reported Sleep Problems Using Longitudinal Data and Machine Learning Methods in the Autism Speaks-Autism Treatment Network Registry
Objectives: To develop a model to predict sleep problems at first follow-up from baseline characteristics among children with no sleep problems at baseline.
Methods: A sample of children in the ATN registry without parent-reported sleep problems at baseline and with complete sleep data at first follow-up was randomly split into training and test samples, stratified by ATN site and age group. Children taking medications indicated for sleep at baseline or follow-up were excluded. Training sample baseline characteristics, including demographics, IQ, psychiatric diagnoses, health problems, and autism severity, behavior, and sensory scores, were tested for associations with subsequent sleep problems. Predictors were chosen based on statistical significance and clinical importance, correlation and multicollinearity considerations, and comparison of c-statistics from alternative logistic regression models. Given probabilities of sleep problems from the final model, a threshold for classifying children as at risk was selected that yielded at least 85% sensitivity and maintained maximum associated specificity. Each child in the test sample was then scored and assigned a predicted sleep problem status based on the model threshold. Comparison of the predicted and true sleep problem status of the test sample yielded sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy.
Results: In the training sample (n=527), children with less educated parents (28.0% vs. 15.9%, p=0.007), ear, nose, and throat (ENT) problems (27.4% vs. 14.6%, p=0.028), asthma (37.5% vs. 16.5%, p=0.023), higher Child Behavior Checklist (CBCL) Anxious/Depressed and Aggressive Behavior z-scores (0.0 vs. -0.3 for each, p=0.031 and p=0.029, respectively) have more sleep problems. In the final model using these predictors, aggressive behavior remains independently associated with higher odds of having sleep problems at first follow-up (OR 1.66, 95% CI 1.07-2.58, p=0.024), and having asthma is associated with higher odds but with borderline significance (OR 2.71, 95% CI 1.00-7.33, p=0.050). This model performed in the test sample (n=518) with sensitivity 80.3%, specificity 33.4%, PPV 18.5%, NPV 90.0%, and accuracy 40.9%.
Conclusions: Among children with ASD, those with more aggressive behavior, asthma, ENT problems, anxious/depressed behavior, and less educated parents at baseline may present with more sleep problems at first follow-up. In a multivariable model aggressive behavior independently predicts sleep problems. The multivariable model, with high sensitivity for identifying children at risk for sleep problems as well as accurate prediction of low risk of sleep problems, can help with treatment and prevention of sleep problems. Further data collection may provide better prediction through methods requiring larger samples.