Machine Learning Prediction of Autism Spectrum Disorder Through Linking Mothers' and Children's Electronic Health...

Area of Science:

• Neuroscience
• Computational Biology
• Pediatrics

Background:

• Autism spectrum disorder (ASD) is a neurodevelopmental condition often diagnosed in childhood, with girls frequently receiving later diagnoses.
• Early identification of ASD is crucial for improving long-term developmental outcomes in children.
• Leveraging electronic health records (EHRs) offers a promising avenue for developing predictive diagnostic tools.

Approach:

• Developed and evaluated machine learning models, including XGBoost and logistic regression, for ASD prediction using linked child-mother EHR data.
• Employed data balancing techniques such as Random Oversampling (ROS) and Random Undersampling (RUS) to address class imbalance inherent in ASD diagnosis.
• Incorporated maternal and child risk factors, including maternal health indicators, as features within the predictive models.

Key Points:

• Logistic regression combined with RUS and a three-year observation window demonstrated superior performance in predicting ASD across the general population (AUROC = 0.798), boys (AUROC = 0.786), and girls (AUROC = 0.791).
• SHAP values were utilized to interpret model predictions, quantifying the influence of significant clinical and sociodemographic risk factors on ASD diagnosis.
• The study highlights the potential of EHR data and machine learning for enhancing early ASD detection, particularly in underdiagnosed populations like girls.

Conclusions:

• Machine learning models utilizing EHR data can effectively predict autism spectrum disorder in children.
• Early ASD detection is feasible and can be improved through the analysis of maternal and child risk factors.
• The developed models show promise for clinical application in identifying children at risk for ASD, facilitating timely intervention and support.