Deep Multimodal Edge-Deployable Neural Architecture for IoT-BasedAutism Detection Using EEG and Facial Biomarkers

Introduction: Autism Spectrum Disorder (ASD) requires early diagnosis to facilitate early interventions that increase developmental outcomes. The conventional forms of behavioral tests are usually time-consuming and qualitative, and inspire automated and objective forms of screening Methods: We suggest a multimodal model of neurophysiological biomarkers and visual biomarkers. P300 event-related potentials of the EEG BCIAUT dataset were correlated with 2,654 face images of children (1,325 autistic and 1,325 non-autistic). EEG band power characteristics were analyzed with the help of a one-dimensional convolutional neural network (1DCNN), and facial features were learned with the help of the EfficientNetB0 and InceptionV3. On the feature level, the modalities were fused together to provide the final classification. Results: The multimodal model was superior to unimodal baselines with the top bottleneck of EfficientNetB0 (128x128), EfficientNetB0 (180x180), and InceptionV3 with accuracies of 79, 77, and 74, respectively. Statistically significant improvement (p < 0.05) over the current multimodal methods, such as EEG-fMRI and eye-tracking techniques, was established by crossvalidation and paired t-tests. Discussion: EfficientNetB0 (128×128) was associated with efficient performance and computational efficiency, and low levels of overfitting and stable F1-scores. The non-invasive and lightweight model is IoT-compatible and has low latency (around 72 ms) and low memory footprint (around 98 MB), which can be scaled and used in real-time to detect ASD in intelligent healthcare systems. Conclusion: An MDL system utilizing EEG and facial video information in a deep-learning system demonstrates a higher ASD detection rate and EfficientNetB0 (128x128) with 79 percent accuracy and high applicability to real-life IoT-based healthcare implementation.