top of page

A-System

Autism Spectrum Disorder Diagnosis and Treatment System

Group Publication 

2021 Summer

 

Personal Designed System

2022 Spring

Role:

Group Leader, Data Visualization, Product Design, UI/UX 

​

Tools:

Python, Fusion 360, SolidWork, Adobe Xd

Advisor:

Manolis Kellis

Recognition:

Published in the 4th International Conference on Computing and Data Science CONF- CDS (2022)
 

 Machine Learning Early Stage

Clinical Diagnosis

AllIsometricOct2-10 copy.png
new2-12.png

Independent Emotion

Regulation Technology

UIUX System Interface

Autism spectrum disorder (ASD) is a neurodevelopment disorder difficult to diagnose and recover. In addition to the ASD patient being affected, there are several other stakeholders will impact the disease including school, family and hospital. A-System is designed to create a coherent network dedicate for the ASD children's diagnose accuracy to 75% and a better independent emotion regulation treatment through assistive technology. 

Background 

In each main scenario for the ASD children, there are different painpoints with various stakeholders such as: specialists, doctors, teachers, parents and children themselves. Different stakeholders will have the timely relationship in order to ensure a fair treatment overall. 
Application portfolio  (28).png
Application portfolio  (30).png
Application portfolio  (31).png
Painpoints
Scenario 
Stakeholder map
AllIsometricOct2-10 copy.jpg

Machine Learning Early Stage
Clinical Diagnosis

The lack of diagnosis in early stages cause the ASD treatment to be harder. The current psychiatric diagnostic process is based simply on observing the behavioral symptoms, nor suitable for quantitative diagnosis and may lead to false diagnosis. Advances in neuroimaging technologies have made it easy to measure those pathological changes related to the brain with autism spectrum disorder. Nevertheless, the differences between the brain with ASD and the healthy controls difficult to apply.

pone.0049172_edited.jpg
Introduce the abnormality in ASD brain images

​

​

(A) Structural covariance patterns are poorly connected in autism

​

(B) Corresponding MRI map in normal controls, connectivity marked in cool colors 

​

(C) MRI maps from both groups shown in comparison as a single anatomic volume. 

ABIDE Dataset with Sufficient fMRI Data for Early Age Group

To face the complexity and heterogeneity of ASD, large-scale samples are essential. The ABIDE initiative has aggregated brain imaging data collected from laboratories around the world. In the MRI data, different brain volumes are represented by voxels and the activity is tracked over time:

​

  • 539 individuals 

  • ages 7-64 years

  • median 14.7 years 

  • 573 typical controls 

abide_logo.jpg
Application portfolio  (27).png
1. Feature Extraction:
Increases the Accuracy of Learned Models by Diagonal Symmetry from the Input

The first flowchart illustrates the process of the feature extraction. Calculate all pairwise Spearman correlation to approximate the functional connectivity in fMRI data to generate a matrix. The generated matrix calculated is diagonal symmetric, the matrix’s upper triangle transform into a one-dimensional vector as the features of Autoencoder.

Copy of Flowchart - Flowchart.png
2. Data Augmentation:
Mixup Reduced Insufficient Training

The dataset applied in the study is preprocessed by a pipeline called C-PAC (Configurable Pipeline for the Analysis of Connectomes), which parcellated the brain into 200 functionally homogeneous regions using a spatially constrained spectral clustering algorithm (CC-200). To reduce the impact of the insufficient training set on the generalization ability of the model, this work uses a linear data augmentation method called ‘Mixup’ to expand our dataset

Copy of Flowchart - Flowchart (3).png
Copy of Flowchart - Flowchart (7).png
Application portfolio  (26).png
3.1. Sparse Autoencoder with Deep Neural Network: Prevent Overfitting Data and Improve Generalization

The benefit of sparse autoencoder is that the model can acquire finer representations and activations are sparser, preventing the data from overfitting. A rectified linear unit (ReLU) activate the units expanding the network with more representative and improved generalization.

Copy of Flowchart - Flowchart (8).png
Application portfolio  (29).png
3.2. Variational Autoencoder with Deep Neural Network: Reduce Parameters and Network Errors

Variational autoencoders with objective Kullback-Leibler Divergence (KL divergence) mapped input to distribution and the bottleneck vector is replaced by two various vectors including the standard deviation and the mean of the distribution. This reduce the generative modal parameters and reconstruction error within the network.

Copy of Flowchart - Flowchart (12).png
Application portfolio  (30).png
4. Cross Validation: Model Performance Evaluation

To test model performance, adapted the model validation method K-folding cross validation. The original data is divided into K folds for each individual site and for all the sites together.

Copy of Flowchart - Flowchart (3).png
Calculation of Result:
Individual Sites Result Different from the Sites as a Whole

​

The ABIDE dataset, including 1035 subjects, has patients from various sites. Therefore, the result is summarized from both individual sites and the generalization across different data acquisition sites.

Application portfolio  (32).png
Application portfolio  (28).png
Compare with Other Methods:
Mixup+SAE+DNN with Highest Accuracy and Sensitivity

​

In the proposed experiment, there are two different approaches named method 3.1 Mixup +SAE+DNN and 3.2. Mixup+VAE+DNN. Different performance compared to the past models through the three significant values Accuracy, Sensitivity and Specificity.

​

  • Accuracy: correctly classified subjects.

  • Sensitivity: accurately considered carrying the ASD.

  • Specificity: successfully classified healthy subjects.

GSD_MDes Application portfolio (9).png
new2-12.png

Independent Emotion Regulation Technology

In addition to the accurate clinical diagnosis from the brain imaging, the ASD children often suffer from the symptoms of hypersensitivity. The goal is to train the children in order to be more independent and regulate the emotion by themselves which could benefit them when they get older.  

Persona in Journey Map

little-boy-800x533_edited.jpg
  • ASD student

  • Age: 8

  • Hyper-sensitivity: light and noise

​

Jason is a student in the ASD Special school and had been diagnosis ASD with sensory processing disorder.

 

How can child cope with the daily activities that happens in the school?

Application portfolio  (29).png
GSD_MDes Application portfolio (10).png
Application portfolio  (30).png
System User Interface

After the development of the early stage clinical diagnosis and the product design of the assistive technology. Now it is the time to combine all the stakeholder in one system interface to contribute to the better treatment and independent emotion regulation of the ASD children. 

Application portfolio  (33).png

Schedule clinical diagnosis on the User Interface

Teacher document the regulation process 

Parents take their kids to hospital for brain imaging

Activate the motion regulation through device and ambient sounds

Child sketch creatures in the Island of Safety 

The AR features brings the creature to live

Independent emotion regulation benefits all 

The assistive earphone is delivered to home

Go to school alone with the earphone 

Experience hypersensitivity when there is loud sou

earphone together 1.png

Intro Page

GSD_MDes Application portfolio (11).png
2.png

Parent Service Pages: personal assistant and doctor appointment 

4.png
6.png

Doctor Evaluation Pages: Patient profile and diagnosis 

Nata iPad profile setting.png
5.png

Children Service Pages: Assistant, Keyboard and AR Visuals

5.png
9.png

Teacher Service Page: ASD training & Children Record 

7.png
bottom of page