Abstract
Autism Spectrum Disorder (ASD) is a complex neurodevelopmental syndrome with diverse clinical presentations. Recent data integration has identified a specific subtype known as dyslipidemia-associated ASD (daASD), characterized by abnormal lipid profiles. This paper explores the clinical trajectories of daASD by analyzing Electronic Health Records (EHR) using a combination of Population Based Incremental Learning (PBIL) and Random Forest classifiers. The study specifically investigates the impact of maternal metabolic health—such as obesity and diabetes—on the risk of a child developing daASD, aiming to establish a robust framework for early prediction.
Introduction
The identification of ASD subtypes is critical for personalized intervention and understanding the underlying biological mechanisms. The daASD subtype represents a significant portion of the spectrum where neurodevelopmental irregularities coincide with metabolic dysfunction. This research aims to define the pattern sequence of medical events that lead to a daASD diagnosis, utilizing large-scale EHR data to identify individuals with both ASD and dyslipidemia. By comparing daASD patients against control groups matched for age, gender, and socioeconomic status, the study seeks to isolate the specific risk factors associated with this subtype.
Methodology
The researchers employed a multi-faceted computational approach to mine EHR data and predict disease outcomes:
Population Based Incremental Learning (PBIL)
PBIL is a semi-supervised learning technique used to model the probability of specific health markers within a population. By treating the population as a univariate distribution that evolves over generations, the algorithm refines the likelihood of identifying disease markers based on the most successful data patterns found in the medical records.
Random Forest Classification
A supervised Random Forest (RF) approach was implemented to classify patients into ASD and Non-ASD categories. This method is particularly effective for handling high-dimensional medical data and identifying the most significant clinical features for prediction, such as comorbid epilepsy or recurrent seizures.
Data Sources and Analysis
The study utilized EHR datasets from India and Clalit Health Services in Israel. The Israeli dataset was particularly valuable as it allowed for the linking of maternal health records to child outcomes through familial ties. A focused analysis was also conducted on a sample of 150 pregnant women to evaluate the correlation between Body Mass Index (BMI) and metabolic markers like blood pressure and glucose.
The analysis focused on comparing four groups:
- daASD: Individuals with both ASD and dyslipidemia.
- ASD+/dyslipidemia-: ASD without dyslipidemia.
- ASD-/dyslipidemia+: Dyslipidemia without ASD.
- ASD-/dyslipidemia-: Control group with no history of neurodevelopmental disorders or abnormal lipid levels.
Key Results
The study revealed a significant "metabolic synergy," where the combination of maternal obesity and diabetes posed a higher risk for ASD than either condition independently. Statistical analysis using Least Square Estimation Regression showed positive correlations between maternal obesity and several risk factors:
- Systolic Blood Pressure: R = 0.5
- Glucose Levels: R = 0.5
- LDL Cholesterol: R = 0.4
In terms of classifier performance, the Random Forest model achieved a Recall of 1.00, meaning it successfully identified all actual ASD cases in the test set. However, the Precision was low at 0.17, resulting in an F1-Score of 0.29. This indicates that while the model is highly sensitive and excellent for initial screening, it currently produces a high number of false positives.
Conclusion and Future Directions
This research underscores the potential of using maternal health markers in EHRs to predict ASD risk early in life—or even prenatally. While the current model demonstrates exceptional sensitivity, the low precision highlights the need for further refinement to reduce false positives in a clinical setting.
Future work will involve a larger-scale study in New Delhi, India, in collaboration with the National Centre for Autism. This joint India-Israel project aims to validate these findings across diverse populations and improve the specificity of the predictive framework to better serve clinical diagnostics.
Summary updated