<p>Autism Spectrum Disorder (ASD) is a neurodevelopmental disease marked by limited or repetitive behaviors as well as ongoing difficulties with social communication and interaction. The early detection of ASD remains a significant clinical challenge due to the subtle nature of symptoms linked to ASD. Eye is closely connected to human brain, and the electrical response of the retina in response to light stimulation can be measured by a diagnostic technique called electroretinogram (ERG). But the diagnosis from ERG signal is subjective to the expertise of clinicians. Therefore, it is crucial to develop a AI-enabled model that supports the clinical decision support system for ASD detection. We used ERG signal (N=106 participants) by leveraging Uniform Manifold Approximation and Projection (UMAP) that transforms ERG signals into discriminative patterns. To address the class imbalance between ASD and control samples, the workflow adopted SMOTE based oversampling, generating additional synthetic samples for the minority class during training. The transformed ERG signals were then fed into our application-driven integration framework UMAP-ERG, which combines UMAP, SMOTE, and conventional ML classifiers to differentiate ASD cases from the control group. We fine-tuned different models at different flash strengths, covering a range of flash strengths from low (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(-0.367\)</EquationSource></InlineEquation> cd<InlineEquation ID="IEq2"><EquationSource Format="TEX">\(\cdot\)</EquationSource></InlineEquation>s<InlineEquation ID="IEq3"><EquationSource Format="TEX">\(\cdot\)</EquationSource></InlineEquation>m<InlineEquation ID="IEq4"><EquationSource Format="TEX">\(^{-2}\)</EquationSource></InlineEquation>) to high (1.204 cd<InlineEquation ID="IEq5"><EquationSource Format="TEX">\(\cdot\)</EquationSource></InlineEquation>s<InlineEquation ID="IEq6"><EquationSource Format="TEX">\(\cdot\)</EquationSource></InlineEquation>m<InlineEquation ID="IEq7"><EquationSource Format="TEX">\(^{-2}\)</EquationSource></InlineEquation>). We conducted extensive set of experiments at flash strength as well as subject level. UMAP-ERG with SVM, achieved the best classification result with an area under the receiver operating characteristics curve (AUC-ROC) of 0.98, accuracy of 0.96, sensitivity of 0.93, and specificity of 0.98. Results vary across different flash strength of ERG signals and lower intensity flash strength (between 0.114 and – 0.119) provide the best classification results compared to mid and high flash strength. UMAP-ERG outperformed the existing state-of-the-art models, based on gMLP and VFCDM, by 7-15% accuracy highlighting the superiority of the proposed model. We believe the proposed framework , UMAP-ERG, will advance the current status quo for ASD detection and enable the ophthalmologists, neuro-ophthalmologists, electrophysiologists, etc., to utilize AI-enabled decision-making process.</p>

Machine learning model for the detection of autism spectrum disorder using electroretinogram signals

期刊论文
出版模式：

开放获取
发表日期：

2026年7月11日

查看全文

Muhammad Aqib Anwar,
Mohammad Tariqul Islam,
Tanvir Alam

摘要

Autism Spectrum Disorder (ASD) is a neurodevelopmental disease marked by limited or repetitive behaviors as well as ongoing difficulties with social communication and interaction. The early detection of ASD remains a significant clinical challenge due to the subtle nature of symptoms linked to ASD. Eye is closely connected to human brain, and the electrical response of the retina in response to light stimulation can be measured by a diagnostic technique called electroretinogram (ERG). But the diagnosis from ERG signal is subjective to the expertise of clinicians. Therefore, it is crucial to develop a AI-enabled model that supports the clinical decision support system for ASD detection. We used ERG signal (N=106 participants) by leveraging Uniform Manifold Approximation and Projection (UMAP) that transforms ERG signals into discriminative patterns. To address the class imbalance between ASD and control samples, the workflow adopted SMOTE based oversampling, generating additional synthetic samples for the minority class during training. The transformed ERG signals were then fed into our application-driven integration framework UMAP-ERG, which combines UMAP, SMOTE, and conventional ML classifiers to differentiate ASD cases from the control group. We fine-tuned different models at different flash strengths, covering a range of flash strengths from low (\(-0.367\) cd\(\cdot\)s\(\cdot\)m\(^{-2}\)) to high (1.204 cd\(\cdot\)s\(\cdot\)m\(^{-2}\)). We conducted extensive set of experiments at flash strength as well as subject level. UMAP-ERG with SVM, achieved the best classification result with an area under the receiver operating characteristics curve (AUC-ROC) of 0.98, accuracy of 0.96, sensitivity of 0.93, and specificity of 0.98. Results vary across different flash strength of ERG signals and lower intensity flash strength (between 0.114 and – 0.119) provide the best classification results compared to mid and high flash strength. UMAP-ERG outperformed the existing state-of-the-art models, based on gMLP and VFCDM, by 7-15% accuracy highlighting the superiority of the proposed model. We believe the proposed framework , UMAP-ERG, will advance the current status quo for ASD detection and enable the ophthalmologists, neuro-ophthalmologists, electrophysiologists, etc., to utilize AI-enabled decision-making process.