Enhancing early autism prediction based on electronic records using clinical narratives

Area of Science:

• Pediatric medicine
• Computational psychiatry
• Machine learning in healthcare

Background:

• Predictive models using structured electronic health record (EHR) data can identify autism likelihood in infants.
• Integrating clinical narratives with structured EHR data can improve prediction accuracy in various medical applications.
• The added value of clinical narratives for early autism prediction remains underexplored.

Purpose of the Study:

• To enhance early autism prediction by integrating structured EHR data with clinical narratives.
• To evaluate the predictive performance of models using structured data alone, clinical narratives alone, and a combination of both.

Main Methods:

• Developed separate models for structured EHR data and clinical narratives.
• Created an ensemble model integrating both data sources for autism prediction.
• Assessed model performance in predicting autism diagnosis by age 4 years using data from ages 30 to 360 days in a 14-year Duke University Health System cohort (11,750 children).

Main Results:

• The ensemble model integrating both data sources demonstrated superior performance compared to models using only structured data.
• By 30 days of age, the ensemble model achieved 46.8% sensitivity and 28.0% positive predictive value (PPV) at 90% specificity (AUC4 = 0.769).
• By 360 days of age, the ensemble model achieved 44.5% sensitivity and 13.7% PPV at 90% specificity (AUC4 = 0.797).

Conclusions:

• Incorporating clinical narratives significantly improves early autism prediction accuracy, outperforming models based solely on structured EHR data.
• Promising predictive accuracy was achieved as early as 30 days of age.
• Features extracted from clinical narratives may offer novel insights into early autism development.