What is Autism Test with Artificial Intelligence?
For several years, I’ve been working on artificial intelligence and machine learning methods for detecting autism spectrum disorders. Artificial intelligence and machine learning have achieved success in many areas today. One of these developments; To determine the risk ratio of autism spectrum disorder from face photographs of children aged 2-8 years. Currently, the risk detection success rate with this method is 85%. Diagnosing ASD is a complex and expensive process that not every family can experience. Considering that detecting autism spectrum disorder with other methods is difficult and expensive, the ratio to be obtained from the test results will be an important guide. I want to emphasize that my goal is not to diagnose Autism Spectrum Disorder. Instead, it is to use the power of artificial intelligence and machine learning to help parents decide whether to take their children to the Autism Spectrum Disorder specialist. This method is not for diagnostic purposes. It is a method used only to determine the probability of risk in children. And in the presence of this risk, it is aimed to take early measures and to warn families about whether their children are in the risk group. In order to support the risk results obtained by artificial intelligence and machine learning methods, question-answer tests and Computer Vision Scanning Classification applications are also performed.
Autism Spectrum Disorder (ASD) is a neurological and developmental disorder that begins in early childhood and lasts throughout a person’s life. Specialists believe that people with ASD have distinct facial features that can be used to help diagnose their ASD and even correlate with ASD severity. This correlation can be used to train a CV model to detect ASD using those unique facial features. The goal is to provide a preliminary diagnostic tool that can aid parents in their decision to pursue further ASD testing. It’s important for models that have significant real-world impacts to represent their results responsibly and use Bayesian statistics to explain their meaning.
Why this is important?
Approximately 25% of children with autism are undiagnosed. Diagnosing ASD is a complicated and expensive process that not every family can go through. Getting the right kids to the right specialists is crucial to reducing the number of undiagnosed children and the burden on families. I want to emphasize that my goal is not to diagnose ASD. Rather, I use the power of CV with the rationale of Bayesian statistics to help parents decide if they should take their children to an ASD specialist.
Version 9 of the Kaggle data set has 2536 images that are evenly split between two classes: autistic and not autistic. The images are of varying sizes but they are already cropped to show only the child’s face.
The data set has an age distribution of approximately 2 years to 14 years, but a majority of the pictures are from younger kids ages 2 years to 8 years.
The gender ratios are close to their respective populations. Males are diagnosed with autism 3 times more than females. Thus, the distribution of male to female pictures in the autistic class is close to 3:1. In the not autistic class, the ratio is much closer to 1:1.
The ratio of white children to children of color is 10:1. This is close to the real distribution in America at 7:1.
The model will be based on VGG Face, Oxford’s deep facial recognition model.
To extract the facial features needed to detect autism, the new model will use the same architecture and weights as VGG Face. This is an example of feature-representation transfer learning. This type of transfer learning is used when the source and the target domains look for similar features but make different inferences based on those features.
The Oxford model is trained to identify underlying facial features to recognize unique faces. The new model also wants to identify underlying facial features but to detect autism.
Sensitivity and Specificity
Sensitivity and specificity are used to quantify how often the model predicts false negatives and false positives.
Sensitivity is the true positive rate. It measures how often a positive prediction is correct.
Specificity is the true negative rate. It measures how often a negative prediction is correct.
The model achieves 76% sensitivity and 85% specificity.
Impact: Bayes Rule
Bayes rule is critical to understanding results produced by any Machine Learning (ML) model.
The probability of a random male having autism is very low, around 2.7%. This is the probability of a male having autism prior to knowing any other information. Thus, it is called the prior probability.
A model prediction provides new information that should be considered in conjunction with the prior knowledge when making a decision.
Bayes theorem can be used to quantify how much this new information should affect the prior probability. This updated probability is called the posterior.
Using the model’s sensitivity and specificity, the posterior probability for a positive prediction can be calculated. For males, a positive result increases the probability of having autism from 2.70% to 12.34%. For females, the probability increases from 0.66% to 3.27%.
Diagnosing autism is a difficult and expensive process. Fortunately, the correlation between facial features and autism means that a CV model can be trained to detect it. Using Bayesian statistics, the results from this model can be reported responsibly and parents can make more informed decisions about their child’s health.