From Data to Insights: AI Predicts Aggressive Behavior in Autism

Make it stand out

Host: Hermione He

Research Analyst: Raquel Chen, Ekki Lu

Guest: Ahmet Demirkaya

In the realm of autism research, the quest to understand and address the complexities of behavioral patterns has been both profound and challenging. However, at the intersection of technology and compassionate inquiry, groundbreaking strides are being made. Today, we have the privilege of delving into the pioneering work of a researcher who stands at the forefront of this endeavor.

Meet Ahmet Demirkaya, a PhD candidate at Northeastern University in the field of machine learning and artificial intelligence. His project focuses on the development and application of biosensors capable of predicting aggressive behavior in profoundly autistic children. With a blend of technological innovation and deep empathy for those affected by autism, Ahmet along with his team is unraveling the mysteries surrounding aggression in autism, offering hope and tangible solutions to families and caregivers.

In this interview, we will explore the motivations, methodologies, and transformative implications of Ahmet's work. Join us as we journey into the realm where cutting-edge technology meets compassionate care, reshaping our understanding of autism and opening new avenues for early intervention and support.

Can you give us an introduction of your project?

One of my projects is a wearable biosensor designed to predict potentially aggressive behaviors, particularly in children with autism. Although applicable across various domains, our primary focus remains on autistic children due to the critical need for early detection and intervention. By anticipating behavioral changes a few minutes in advance, we can proactively engage with them to prevent undesirable outcomes.

For this project, we leverage smartwatches equipped with diverse sensors to capture crucial data points. These include analyzing accelerometer data for rapid movements, monitoring skin conductance to gauge sweat levels, and tracking heart rate per minute. Our machine learning model processes this information to forecast aggressive behavior, often detecting it a full minute before it occurs within a three-minute window.

How do you collect data? Do you collaborate with third-party organizations?

We collaborate with five universities and multiple clinics where physicians collect and label data before sending it to us. This labeled data is used to train our machine learning models. Our study comprises data from over 50 individuals, some observed for up to four years, while others more recently. These observations occur exclusively during hospitalization, as our focus is on severe cases of autism where individuals exhibit higher rates of aggressive behavior or self-injury. The goal of our research is to mitigate harmful behavior, enabling these individuals to integrate more effectively into society.

You mentioned this research can be applied to other realms, can you give us an example? 

Generally it can be applied to people who show high risk of aggressive behavior or self injury. For example, we can predict some criminals that are under police or imprisonment.

How do machine learning and AI work for your project and help you develop the biosensor? 

As mentioned, our analysis focuses on three distinct inputs, which we gather and process within a three-minute window. This involves analyzing time-series data, which encapsulates intricate temporal signals. Machine learning algorithms excel at discerning nuanced patterns within such data, patterns that may not be immediately discernible to human observation. Through training, these algorithms become adept at recognizing patterns indicative of potential aggressive behavior.

By leveraging labeled data, the model is trained to associate specific patterns with the likelihood of aggression. Importantly, machine learning models are capable of discerning multiple patterns within the data, each contributing to the predictive outcome. Additionally, the model may identify patterns suggesting a lower likelihood of aggression, thus offering a spectrum of predictions. Essentially, the model learns to detect specific signal changes that correlate with behavioral patterns, enhancing its predictive capabilities.

How accurate and reliable are the predictions made by your machine learning models?

So far, we can detect aggressive behaviors with 80% accuracy, so it’s like four out of five times.

What challenges have you encountered while developing and implementing this technology? 

The challenge lies in the variability of patterns observed among different individuals, making it difficult to develop a universal model applicable to all. However, we also face limitations due to insufficient data from each individual to train a personalized model. Therefore, our approach aims to create a generalized model while accommodating unique cases for individuals with distinct data profiles. Balancing the need for a broad applicability with the consideration of individual nuances poses a significant challenge in model training.

Where do you see the biggest research opportunities or gaps in the application of machine learning to healthcare, and how do you plan to address them in your future work?

I recognize the significance of personalized models, which have the potential to yield more precise results tailored to each individual's needs. Furthermore, the role of machine learning models in healthcare is poised to expand in the future. These models are increasingly adept at addressing real-world challenges, as evidenced by their evolving capabilities. Consequently, the importance of machine learning in healthcare is expected to grow as it becomes increasingly effective in tackling diverse problems compared to previous approaches.

As a PhD student in machine learning and AI, do you have any suggestions for people who want to step into this area? 

Certainly, I suggest not confining your focus solely to particular aspects of machine learning but rather embracing a broader approach that encompasses various machine learning-related domains. The challenges we encounter often include diverse facets of machine learning, transcending specific fields.

Sometimes, what appears to be solely a healthcare issue may involve complexities related to time-series data or computer vision, both of which fall within the realm of machine learning. Therefore, it is beneficial to consider the wider scope of machine learning applications rather than restricting oneself to narrow domains.

Quick take away

As we wrapped up the interview, Ahmet mentioned he is also doing another project called Modeling Biological Dynamical Systems. This project is meant to figure out how organs communicate with each other, such as understanding the timing of bladder voiding, which holds significance for individuals dealing with bladder issues. In this context, comprehending bladder voiding entails not only assessing bladder liquid volume but also considering signals originating from other organs. These signals play a crucial role in controlling bladder voiding. Therefore,  various bodily measurements are employed to estimate the timing of bladder voiding accurately. This study is still at an early stage and predictably will take 15 years. 

Through this interview, we see the remarkable potential of machine learning and AI in revolutionizing our approach to understanding and addressing complex behavioral challenges, particularly within the realm of autism research.

The innovative use of biosensors, coupled with sophisticated machine learning algorithms, offers a glimmer of hope for profoundly autistic children and their families. By predicting aggressive behavior and facilitating timely interventions, these technologies have the power to enhance the quality of life for individuals with autism, fostering greater inclusion and support within society.

Read more: https://news.northeastern.edu/2024/01/18/autism-aggression-prediction-biosensors/