AI-enhanced autism therapy: How multi-stage ensemble learning is changing the game By Pooja

By Pooja Toshniwal PahariaReviewed by Lily Ramsey, LLMNov 17 2023

In a recent study published in Scientific Reports, researchers examined the capacity of ensemble learning to anticipate and identify characteristics that impact or contribute to autism spectrum disorder therapy (ASDT) for intervention purposes.

Study: On effectively predicting autism spectrum disorder therapy using an ensemble of classifiers. Image Credit: Chinnapong/Shutterstock.com

Background

ASD is a developmental condition that interferes with social interaction, communication, and learning. Early identification and treatment can prevent diseases from deteriorating and save money. Ensemble learning, which mixes many single classifiers, has been found to enhance predicted accuracy by reducing variation.

The concern is improving early ASD diagnosis and testing decisions, which may result in considerable time, expense, and even death savings.

The type of ensemble learning system, also known as multiple-classifier learning systems (MCLS), should be determined to offer the most benefits concerning size or variety. Robots could facilitate short-term treatments since ensemble models are more stable and better predictors than single classifiers.

About the study

In the present study, researchers used five single classifiers versus several MCLS algorithms to predict ASD in autistic children.

The researchers evaluated the efficacy of machine learning algorithms in predicting ASDT for children with autism receiving robot-assisted care against a control group receiving just human interaction. They also investigated ways in which ensemble learning could increase ASDT forecasting accuracy.

They proposed using MCLS to enhance ASD therapy and assess whether it could overcome the predictive limitations of single-classifier learning systems (SCLS) due to their incapacity to handle complicated tracking circumstances with high accuracy.

All feasible classifier combinations for each ensemble were assessed to compare single- and multiple-classifier performances. The physical parameters most important in ASDT therapy were identified via feature selection using decision tree (DT)-based techniques.

The research utilized data including behavioral information and robot-enhanced treatment (intervention) vs. regular human treatment (control) based on 3,000 sessions and 300 hours of therapy recorded from 61 autistic children over the age of three.

Both groups used the applied behavior analysis (ABA) procedure, which uses behavioral principles and scientific observations to enhance and modify socially relevant behaviors. Both group participants were subjected to an initial evaluation, eight interventions for ASD, and a final evaluation.

Treatment effects were evaluated using the Autism Diagnostic Observation Schedule (ADOS) based on the differences between the initial and final assessments.

Five base classifiers were designed for the simulations, with default hyper-parameters for each classifier, utilizing various kinds of parametric estimation or learning. The training dataset (60%), validation dataset (30%), and test dataset (10%) were analyzed to evaluate base classifier performance.

The study investigated wait time, social contact, communication, behavioral and emotional consequences, and the effectiveness of social robot-enhanced treatment in autistic children.

The dataset includes characteristics for head position, body motion, body motion, eye gaze, age, gender, goal ability, therapy condition, therapy date, ASD diagnosis, and a three-dimensional skeleton.

Results

The experimental findings revealed considerable variations in performance among single classifiers for ASDT prediction, with decision trees being the most accurate. DT outperformed other base classifiers with a 36% smoothed error rate.

Other base classifiers showing superior performance were artificial neural networks (ANN), k-nearest neighbor (k-NN), and logistic discrimination (LgD), with smoothed error rates of 36%, 39%, and 42%, respectively.

For the single classifiers, eye contact (cross-validation error, 7.5%) and social communication (cross-validation error, 13%) were the most essential contributing elements to the ASDT issue among children.

For ASDT prediction, MCLS performed much better than single classifiers. In particular, ensembles with three classifiers showed the best performance among MCLS systems, followed by two-classifier ones, with 21% and 31% smoothed error rates, respectively.

The lowest error rates were reported for bagging ensemble classifiers (23%) and boosting (26%), followed by feature selection (31%), and randomization (35%). MCLS classifiers using multi-stage designs showed the most significant effects (74% accuracy rate), followed by the static-parallel and dynamic architecture designs (72% and 68% accuracy rates, respectively).

Bi-directional interactions were found between resampling methods, multi-classifier systems, and resampling techniques.

Conclusion

Overall, the study findings showed that static parallel MCLS with three classifiers built through bagging and incorporating decision trees, k-nearest neighbor, and logistic discrimination were the most effective for predicting ASD.

Eye contact and social interaction seemed to impact ASD-enhanced treatment more than stereotypes, non-verbal speech, and social touch.

Future studies could compare autistic infants to autistic adults and explore particular cognitive systems that could be targeted or altered by robot vs. human interactions.