Machine-Learning-based classification of Autism Spectrum Disorder using heterogeneous factors
Poster No:
394
Submission Type:
Abstract Submission
Authors:
Yong Jeon Cheong1, Jihyun Bae2, Seonkyoung Lee1, Ji Hyeong Ro3, Hirotaka Kosaka4, Minyoung Jung1
Institutions:
1Korea Brain Research Institute, Daegu, Republic of Korea, 2Korea Brain Research Institude, Daegu, Republic of Korea, 3Korea Brain Research Institute, Daegu, Korea, Republic of, 4University of Fukui, Fukui,Japan
First Author:
Co-Author(s):
Introduction:
Autism Spectrum Disorder (ASD) is characterized by a broad range of behavioral symptoms including atypical sensory responses, which reflects the etiologic heterogeneity of the disorder. There is increasing evidence on the effects of epigenetic modifications (e.g., DNA methylation; DNAm) on structure and function of the autistic human brain. Additionally, ASD is characterized by altered sensory response and altered thalamic-sensory hyperconnectivity.
Integrating various features derived from epigenetic (i.e., DNAm values of oxytocin receptor (OXTR) and arginine vasopressin receptor (AVPR) genes), brain-related (i.e., volumes of cortical and subcortical regions, and values of resting-state functional connectivity (rs-FC)), and sensory behavioral factors and applying a supervised machine learning (i.e., XGBoost), this study aims to identify core features of the disorder by building three different models : 1) Full model includes the three factors, 2) Brain model has the brain-related and behavioral factors, and 3) Epigenetic model contains the epigenetic and behavioral factors.
Integrating various features derived from epigenetic (i.e., DNAm values of oxytocin receptor (OXTR) and arginine vasopressin receptor (AVPR) genes), brain-related (i.e., volumes of cortical and subcortical regions, and values of resting-state functional connectivity (rs-FC)), and sensory behavioral factors and applying a supervised machine learning (i.e., XGBoost), this study aims to identify core features of the disorder by building three different models : 1) Full model includes the three factors, 2) Brain model has the brain-related and behavioral factors, and 3) Epigenetic model contains the epigenetic and behavioral factors.
Methods:
This study includes 34 individuals with ASD (F = 12, mean [SD] age = 26.0 [4.24] years old) and 72 IQ-matched neurotypical individuals (F = 39, mean [SD] age = 32.0 [12.73] years old). We extracted DNAm values of OXTR and AVPR genes from the participants' salivary samples. Using a 3-T MR scanner, structural and functional MRI data were collected. The participants completed Adolescent/Adult Sensory profile that allows us to assess the level of abnormality of sensory behavior.
Considering a small sample size with high dimensionality, we initially selected 30 baseline predictive features (i.e., 9 epigenetic, 11 brain-related, 7 behavioral and 3 demographic features) showing group difference. We split data for training (80%) and testing (20%). For each iteration, 80% of training data was used for validation. Next, we built training model while applying an ensemble feature selection procedure (i.e., feature occurrence frequency method) and an automatic hyperparameter tuning in an iterative manner. Last, the top 19 features showing the most discriminative power and best tuned hyperparameters were fed into XGBoost for testing. We estimated the area under the curve for receiver operator characteristic curves (ROC-AUCs) 1000 times for testing model, and differences in estimated accuracies were tested: 1) between Full and Brain model, 2) between Full and Epigenetic model, and 3) between Brain and Epigenetic model. To prevent the risk of overfitting and to minimize the problems derived from data imbalance, we performed our analyses using stratified 10-fold cross validation.
Considering a small sample size with high dimensionality, we initially selected 30 baseline predictive features (i.e., 9 epigenetic, 11 brain-related, 7 behavioral and 3 demographic features) showing group difference. We split data for training (80%) and testing (20%). For each iteration, 80% of training data was used for validation. Next, we built training model while applying an ensemble feature selection procedure (i.e., feature occurrence frequency method) and an automatic hyperparameter tuning in an iterative manner. Last, the top 19 features showing the most discriminative power and best tuned hyperparameters were fed into XGBoost for testing. We estimated the area under the curve for receiver operator characteristic curves (ROC-AUCs) 1000 times for testing model, and differences in estimated accuracies were tested: 1) between Full and Brain model, 2) between Full and Epigenetic model, and 3) between Brain and Epigenetic model. To prevent the risk of overfitting and to minimize the problems derived from data imbalance, we performed our analyses using stratified 10-fold cross validation.
·Figure 1. Process of Development of ASD classification model using epigenetic, brain-related, and sensory behavioral features
Results:
We achieved average F1-score 0.832 (median = 0.83, interquartile range (IQR) = 0.88 – 0.79) and average ROC-AUC 0.8395 (median = 0.8214; IQR = 0.8929–0.7857) for Full model. Brain model had average F1-score 0.809 (median = 0.82, IQR = 0.88 – 0.77) and average ROC-AUC 0.8119 (median = 0.8214, IQR = 0.8571–0.7500). Epigenetic model showed average F1-score 0.771 (median = 0.77, IQR = 0.83 – 0.72) and average ROC-AUC 0.7689 (median = 0.7857, IQR = 0.8214–0.7143).
The F1 score of Full model was significantly different than that of Brain model (U = 577147, p = 1.951x10-9) and that of Epigenetic model (U = 707230, p = 2.2 x10-16). The F1 score of Brain model was higher different from that of Epigenetic model (U = 633078, p = 2.2x10-16). The ROC-AUC of Full model was significantly different than that of Brain model (U = 596070, p = 4.989x10-14) and that of Epigenetic model (U = 739199, p = 2.2x10-16). The ROC-AUC of Brain model was higher different from that of Epigenetic model (U = 643998, p = 2.2x10-16).
The F1 score of Full model was significantly different than that of Brain model (U = 577147, p = 1.951x10-9) and that of Epigenetic model (U = 707230, p = 2.2 x10-16). The F1 score of Brain model was higher different from that of Epigenetic model (U = 633078, p = 2.2x10-16). The ROC-AUC of Full model was significantly different than that of Brain model (U = 596070, p = 4.989x10-14) and that of Epigenetic model (U = 739199, p = 2.2x10-16). The ROC-AUC of Brain model was higher different from that of Epigenetic model (U = 643998, p = 2.2x10-16).
·Figure 2. F1 and ROC-AUC scores of three models
Conclusions:
Full model showed higher performance in predicting ASD than Brain and Epigenetic model. Generalizable prediction of ASD can be achieved when considering interaction of epigenetic modification, brain function and structure, and sensory behaviors.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 1
Genetics:
Genetics Other 2
Modeling and Analysis Methods:
Task-Independent and Resting-State Analysis
Keywords:
Autism
Machine Learning
MRI
Saliva
Thalamus
1|2Indicates the priority used for review
Provide references using author date format
Chen, T. (2016), 'XGBoost: A scalable tree boosting system', Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Association for Computing Machinery, pp. 785–794
Fu, Z. (2019), 'Transient increased thalamic-sensory connectivity and decreased whole-brain dynamism in autism', NeuroImage, vol. 190. pp. 191–204
Gregory, S. G. (2009), 'Genomic and epigenetic evidence for oxytocin receptor deficiency in autism', BMC Medicine, vol. 7, no. 82, doi:10.1186/1741-7015-7-62
Habata, K. (2021), 'Relationship between sensory characteristics and cortical thickness/volume in autism spectrum disorders', Translational Psychiatry, vol. 11, no. 616, doi: 10.1038/s41398-021-01743-7.
Hopkins, W. D. (2023), 'Vasopressin, and not oxytocin, receptor gene methylation is associated with individual differences in receptive joint attention in chimpanzees (Pan troglodytes)', Autism Research, vol. 16, no. 4, pp. 713–722
Insel, T. R. (2010) 'The Challenge of Translation in Social Neuroscience: A Review of Oxytocin, Vasopressin, and Affiliative Behavior', Neuron, vol. 65, no. 6. pp. 768–779
Ladd-Acosta, C. (2014), 'Common DNA methylation alterations in multiple brain regions in autism', Molecular Psychiatry, vol. 18, no. 9, pp. 862-871
Nair, A. (2013). 'Impaired thalamocortical connectivity in autism spectrum disorder: A study of functional and anatomical connectivity,' Brain, vol. 136, no. 6, pp. 1942-1955
Thye, M. D. (2018). 'The impact of atypical sensory processing on social impairments in autism spectrum disorder,' Developmental Cognitive Neuroscience, vol. 29. pp. 151-167