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.

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