Development and validation of a machine learning-based tool to predict autism among children

Area of Science:

• Public Health
• Machine Learning
• Developmental Pediatrics

Background:

• Early diagnosis of autism spectrum disorder (ASD) is crucial for timely intervention and improved social functioning.
• Current diagnostic methods can be time-consuming and may not be accessible to all populations.
• There is a need for novel, efficient approaches to identify children at risk for autism spectrum disorder (ASD) at the earliest possible stage.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting autism disorder (ICD10 84.0) using maternal and infant health administrative data.
• To identify key risk factors associated with autism spectrum disorder (ASD) diagnoses in a large population cohort.
• To explore the potential of routinely collected health data for early autism spectrum disorder (ASD) detection.

Main Methods:

• Utilized a retrospective cohort study design in New South Wales, Australia, including 262,650 mother-offspring pairs (2003-2005).
• Integrated data from three health administrative sources: Perinatal Data Collection (PDC), Admitted Patient Data Collection (APDC), and Mental Health Ambulatory Data Collection (MHADC).
• Applied machine learning algorithms to construct a predictive model for autism disorder (ICD10 84.0), assessing performance using the area under the receiver operating curve (AUC).

Main Results:

• The most successful machine learning model achieved an AUC of 0.73 for predicting autism disorder (ICD10 84.0).
• Identified significant risk factors for autism spectrum disorder (ASD) diagnoses including offspring gender, maternal age at birth, delivery analgesia, maternal prenatal tobacco use disorders, and low 5-minute APGAR scores.
• Demonstrated the feasibility of using linked administrative health data for autism spectrum disorder (ASD) risk prediction.

Conclusions:

• Machine learning combined with routinely collected maternal and infant health administrative data shows promise for the early detection of autism spectrum disorder (ASD).
• Further refinement and validation of these models could enhance accuracy and support early identification efforts.
• This approach offers a potential pathway to improve early intervention access for children with autism spectrum disorder (ASD).