GeneticNAS: a novel self-evolving neural architecture for advanced ASD screening

Area of Science:

• Neurodevelopmental research
• Artificial Intelligence in Healthcare
• Machine Learning for Diagnostics

Background:

• Early identification of Autism Spectrum Disorder (ASD) is hindered by subjective assessments and limited resources.
• Current diagnostic processes can be lengthy and resource-intensive.
• There is a critical need for efficient and accurate ASD diagnostic tools.

Purpose of the Study:

• To develop a memory-efficient Neural Architecture Search (NAS) framework for autonomous ASD classification.
• To reduce the computational resources required for NAS while maintaining high performance.
• To enable practical clinical deployment of advanced diagnostic tools.

Main Methods:

• Developed a novel search space integrating simple, residual, and bottleneck operations.
• Implemented a memory-efficient genetic algorithm reducing GPU memory consumption by 76%.
• Utilized an adaptive fitness function balancing model performance and computational complexity.

Main Results:

• Achieved a classification accuracy of 96.5% (95% CI: 94.89-98.11) and an ROC AUC of 0.986.
• Significantly outperformed traditional CNN, ResNet-based, and LSTM models.
• Demonstrated practical viability with 2.8M parameters and 15ms processing time per sample.

Conclusions:

• The memory-efficient NAS framework offers a superior, resource-conscious approach to ASD classification.
• This methodology holds significant potential for improving early ASD identification in clinical practice.
• The framework addresses current limitations in diagnostic speed and resource accessibility.