Autism Spectrum Disorder Detection with 2 Stage AI Model
This project addresses the critical challenge of early Autism Spectrum Disorder (ASD) detection using a two-stage approach combining deep learning for preliminary diagnosis and ensemble machine learning for confirmatory diagnosis. The researchers, Aashish Acharya and Dennish Karki, have developed a system that leverages both facial image analysis and behavioral data to provide a more accurate and accessible diagnostic tool. This project demonstrates a promising approach to ASD detection that combines the strengths of deep learning and traditional machine learning techniques. The two-stage system provides a comprehensive method that could potentially improve early detection and diagnosis, particularly in resource-limited settings. The implementation details reveal a thoughtful approach to data processing, model architecture, and evaluation. While the current accuracy levels are promising (76.54% for the preliminary stage and 92% for the confirmatory stage), there remains room for improvement through further refinement of the models and expansion of the datasets.
Autism Spectrum Disorder Detection with 2 Stage AI Model
- 1. Prediction of Aashish Acharya | Dennish Karki
- 2. A complex neurodevelopmental condition characterized by difficulties in social interaction, communication, and atypical behaviors ASD symptoms typically manifest in early childhood and continue throughout life, with varying degrees of severity. About one in every 100 children is diagnosed with ASD. The transition from childhood to adulthood is particularly challenging for autistic individuals The condition significantly impacts the quality of life of individuals and their families, often leading to social isolation and emotional strain.
- 3. The exact cause of this disease is unknown, symptoms widely vary among individuals Many individuals, especially in low-income regions, struggle to access timely ASD diagnosis and care. Traditional diagnostic methods for ASD are costly and slow The diverse range of ASD symptoms complicates the development of universally effective solution. The burden on families is immense, as they often struggle with emotional and financial pressures.
- 4. The Verdict United Nations , 2013
- 5. • Autism Questionnaire (Simon Baron-Cohen and team.) • 50-item measure assessing social skills, communication, imagination, attention to detail, and attention switching. • AQ10 - a compacted version of the above to make diagnosis quicker. • Capture detailed images of brain structures and monitor brain activity. • Used to study atypical neural connectivity patterns • Allows researchers to delve into the underlying brain mechanisms. • Hypothesis: certain facial features and expressions may be indicative of ASD • Features may not be easily noticeable to the human eye • Study has shown to aid early detection.
- 6. 92.26% Random Forest (Thabtah & Peebles, 2020) 97.70% Random Forest & XGBoost (Hemu et al, 2022) 90% (Naik et al, 2023) 83.42% Decision Tree (Sun et al, 2023) 85.30% Deep Belief Network (Li et al, 2019) 87% Mobile Net (Khoshla et al, 2021) 88% EfficientNet (Narala et al, 2023) 97.48% CNN (Jaffar & Abdulbaqi,2022)
- 7. Many ML based models, most with significant performance measures. Random Forest has been a consistent perform in multiple studies Naik et al, 2023 | Sethi et al 2024 RF & XGBoost outperformed other ML techniques with 97% accuracy Hemu et al, 2022 Ensemble approach has evolved to become the most reliable approach Sun et al 2023 While very simple, Logistic Regression has proven effective in several studies (Hemu et al 2022, Naik et al 2023, Sethi et al 2024)
- 8. CNNs have shown significant potential in ASD detection. Custom CNN have been developed and tested, promising results on fMRI data Jaffar & Abdulbaqi, 2022) MobileNet outperformed ResNet50 with 87% accuracy Khosla et al 2021. ResNet50's capabilities in extacting complex features from images make it a promising candidate for analyzing facial images (Narala et al, 2023)
- 9. The BIG Challenge Sethi et al, 2023
- 10. ResNet50 Preliminary Diagnosis using Facial Images Confirmatory Diagnosis using Behavorial Data Image Data 1st Diagnosis AQ-10 Data Ensemble Model 2nd Diagnosis
- 12. 2652 images
- 13. 2652 images
- 14. 2652 images
- 15. 2652 images
- 17. 4 csv 2497 records Single Dataset
- 18. Single Dataset
- 19. Single Dataset
- 26. Bias Risk: The system may produce unfair or inaccurate predictions due to demographic imbalances and varying data quality, especially in ASD diagnosis. Privacy Concerns: The use of facial recognition and neuroimaging data raises issues around privacy, consent, and data security. Need for Inclusivity: Transparency and diverse stakeholder input are crucial to prevent harm and ensure fair distribution of the technology's benefits.
- 27. Model Improvement Architectural Enhancements Dataset Expansion Explainability and Trust
- 28. Thank you! Aashish Acharya | Dennish Karki