A hybrid deep learning framework for early autism screening
Abstract: Objectives Early diagnosis of autism spectrum disorder (ASD) remains a significant challenge due to the time-consuming and subjective nature of traditional diagnostic methods. This study proposes a reliability-oriented hybrid deep learning framework that provides a low-cost, scalable, non-invasive, and AI-assisted pre-screening tool for early ASD risk indication and referral support. Methods The proposed system integrates two independent deep learning architectures: (1) a ResNet18 model optimized with 10-fold cr…

"Mental Patients Against Stigma" by Mindy Fisher is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.
Researchers developed a hybrid deep learning system for early autism spectrum disorder risk indication that fuses a ResNet18 model trained on static periocular images with a multi-CNN facial ensemble of ResNet50, EfficientNet-B0 and DenseNet121, combining outputs through an OR-based reliability rule and using Grad-CAM to highlight decision regions.
The approach matters because traditional ASD diagnosis is described as time-consuming and subjective, and a low-cost, scalable, non-invasive AI pre-screening tool could accelerate referral; uncertainty remains because the highest system-level sensitivity is analytically estimated under a conditional-independence assumption and the tool is explicitly not a replacement for clinical evaluation.
- Periocular pathway uses fine-tuned ResNet18 with 10-fold cross-validation and softmax decision boundary.
- Facial pathway uses weighted probabilistic averaging across ResNet50, EfficientNet-B0, and DenseNet121.
- Explainable AI via Grad-CAM was employed to visualize decision-relevant regions.
A reliability-oriented hybrid framework combining a periocular ResNet18 and a facial ensemble of ResNet50, EfficientNet-B0 and DenseNet121 increased early ASD risk indication to 90% sensitivity in the periocular pathway, 87.1% sensitivity with 0.948 AUC in the facial pathway, and an analytically estimated 98.71% system
The rundown
The methods integrate two independent architectures: a ResNet18 optimized with 10-fold cross-validation on static periocular data, and a multi-CNN facial ensemble combining ResNet50, EfficientNet-B0 and DenseNet121 with weighted probabilistic averaging, followed by OR-based reliability fusion of the two subsystem outputs.
Results reported by 2026-09-15 include 90% sensitivity for the periocular model, 87.1% sensitivity and 0.948 AUC for the facial ensemble, and a joint false-negative probability of approximately 1.29% derived from the fusion rule, with Grad-CAM used for transparency.
Sources
- Peer-reviewedPrimary Health Care Research & Development2026-09-15
How should this claim be treated?
ace
The debate