Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session D

Poster A1 - Machine Learning Frameworks for Characterizing Phenotypic Variations in Autism Spectrum Disorder with Lifetime Medical Records

Friday, October 25, 2024

3:30 PM - 4:30 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

YC

Yunzhe Chen (he/him/his)

Undergraduate Researcher
Wake Forest University School of Medicine
Brooklyn, New York, United States

Co-Author(s)

AM

Arezoo Movaghar, PhD

Assistant Professor
Wake Forest University School of Medicine, United States

Introduction: Autism Spectrum Disorder (ASD) is a complex developmental condition marked by substantial clinical and biological heterogeneity across patients. Affecting approximately 1 in 36 children in the United States and over 78 million individuals worldwide, ASD has a significant impact on individuals, families, and healthcare systems. The wide range of variations in ASD symptoms and comorbidities causes crucial challenges in delivering timely and effective care. Identifying patient subgroups based on similar disease manifestations paves the way for providing more personalized and patient-centered interventions, treatments, and services.

This study proposes characterizing phenotypic variations within ASD populations utilizing machine learning techniques and patient medical history. We leverage Electronic Health Records (EHRs) from 6 million patients served by Atrium Health Wake Forest Baptist, covering rural and urban populations in North Carolina, South Carolina, Georgia, and Alabama. While most research primarily focuses on neuroimaging or specific behavioral traits, our study further includes the full spectrum of lifetime medical conditions in identifying subtypes within the ASD population. This comprehensive approach aims to provide a more holistic understanding of the diverse phenotypic presentations. By classifying patterns and clusters of comorbidities, intervention can be efficiently administered to the patient. Analysis of these patterns and clusters also offers new insights about ASD and its implications for patients and their families.

Materials and

Methods: The dataset contains primary demographics and historical diagnoses in International Classification of Diseases 10th Revision (ICD-10) codes spanning the past 23 years. Data preprocessing involved handling missing values and encoding categorical features into binary. We filtered outliers by including only those diagnosis codes that appeared at least in two separate visits (i.e., Rule of 2) and were observed in at least 5% of unique patients. Additionally, we mapped ICD-10 codes to Phenome-wide Association Studies (PheWAS) categories, such as congenital, neurological, genetic, mental, etc., to reduce noise. These PheWAS categories were selected based on their common prevalence in autistic patients compared to those who are not.

With the optimized dataset, we applied clustering methodologies, including K-Means, K-Medoids (with Hamming distance), Gaussian Mixture Models (GMM), hierarchical clustering, and Semi-Spectral Clustering, with and without dimensionality reduction, Principal Component Analysis (PCA). Clustering performance was evaluated using metrics: Silhouette Score, Calinski-Harabasz Score, Davies-Bouldin Index, and Dunn Index, facilitating the selection of cluster configurations with the Elbow Method. Visualization is done with t-distributed Stochastic Neighbor Embedding (t-SNE). To improve the clusterings, we iteratively adjusted parameters and categorical features. We use numeric table and pair plot analysis to further illustrate relationships between critical features within clusters. Random Forest analysis was also used to pinpoint significant contributors to cluster formation. Further clusterings were done in gender-specific cohorts.

Results, Conclusions, and Discussions: For this analysis, we identified 4,704,679 from 36,458 ASD patients, with 74.8% male and 25.2% female. The racial composition of the patients was 63.8% self-identified as White, 23.7% Black/African American, 4.1% Asian, and 8.4% from other or unspecified groups.

K-Means, especially with PCA, excelled in the evaluation. We identified seven (eleven with gender-specific) distinct clusters, with the smallest comprising 10.09% of patients and the largest containing 26.21%. Visualizations using t-SNE highlighted the phenotypic diversity within these clusters. Some clusters exhibited higher rates of gastrointestinal and neurological issues, while others did not have significant comorbidities.

Pair plot analysis revealed strong correlations between symptoms and gastrointestinal issues, particularly in clusters with high symptom prevalence. The Random Forest analysis underscored the importance of mental health and gastrointestinal issues as primary contributors to cluster formation. It also highlighted significant gender differences and the relevance of neurological, endocrine, and general symptoms in defining patient subgroups.

This study offers an approach to identifying significant phenotypic patterns within the ASD population, highlighting the high heterogeneity across patients and providing critical information for patients, healthcare providers, and policymakers. Differences in the size and characteristics of clusters demonstrate the variability in disease manifestation among subgroups. Mental health problems were identified as one of the primary factors differentiating subgroups. These findings signify the importance of diagnosing and treating mental disorders in patients. This study also demonstrated gender differences in autistic patients, with women having a more complex health profile and a higher rate of variability across subgroups compared to autistic men. These differences might be due to biological factors and the complexity of women's health. It also could reflect health inequalities often experienced by autistic women, resulting in a higher likelihood of those with more severe symptoms being diagnosed.

Further verification of these findings is needed. The validity of the clusters can be confirmed through longitudinal studies, replication in independent cohorts, and the integration of additional clinical and genetic trials. Understanding distinct phenotypic profiles allows healthcare providers to customize interventions to match the specific needs of each subgroup, replicating successful treatments and ensuring patients receive the most effective therapies.

Acknowledgements (Optional): This project was supported in part by the NIH/NLM (R25 LM014214) in the Department of Biomedical Engineering and Center for Artificial Intelligence Research at Wake Forest University School of Medicine.