Hybrid deep learning model for autism spectrum disorder diagnosis.

Researchers developed computer programs that can help diagnose autism by analyzing photos of children's faces. They tested five different programs and found one that was 95.5% accurate at identifying autism. This technology could potentially help diagnose autism earlier and more objectively than current methods, which rely heavily on doctor observations and can be time-consuming.

This study developed and tested five hybrid deep learning models to diagnose autism spectrum disorder using facial image analysis. The researchers compared MobileNetV2+BiLSTM, ResNet50+LSTM, EfficientNetB4, InceptionV3, and MobileNetV2+GRU models using a facial image dataset from Kaggle. The MobileNetV2+GRU hybrid model demonstrated superior performance with 95.5% test accuracy, 95.94% precision, 95.45% F1-score, and 98% ROC value. The study aimed to address limitations of traditional diagnostic approaches, which are clinician-dependent, subjective, and time-consuming, potentially hindering early detection of ASD in children.

While promising, this technology requires extensive validation before clinical implementation. The approach could potentially support earlier autism detection and reduce diagnostic delays, but must be validated against gold-standard clinical assessments and tested across diverse populations before clinical adoption.

Study does not report sample size, validation methodology, or demographic characteristics. No comparison with clinical diagnosis standards. Limited information about dataset quality and representativeness. Lacks details about model generalizability across different populations.

Autism spectrum disorder (ASD) is a neurodevelopmental condition pertaining to the communication, social connectivity and conduct of individuals. ASD individuals develop symptoms such as recurrent actions, atypical facial expressions and challenges in social engagement. ASD prediction depends on various measures such as functional Magnetic Resonance Imaging (fMRI) data, game-based assessments, kinematic traits, questionnaires, head activity analysis, motor activities and eye-tracking. Traditional diagnostic approaches are subjective.

These approaches are clinician-dependent and time-consuming. This has resulted in various challenges for the early detection of the condition. This work evaluated the performance of five hybrid approaches such as MobileNetV2+BiLSTM, ResNet50+LSTM, EfficientNetB4, InceptionV3 and MobileNetV2+GRU. Each model was meticulously refined to achieve optimal performance on the facial image dataset obtained from the Kaggle repository.

The hybrid MobileNetV2+GRU model showed high performance with 95.5% test accuracy, 95.94% precision, and 95.45% F1-score. When the suggested hybrid model was compared with the remaining models, the latter outperformed with a ROC value of 98%. The findings highlight the optimal performance and generalizability of the proposed MobileNetV2+GRU model in ASD diagnosis in children.