Autism AI-IMB

From Knoesis wiki
Jump to: navigation, search


Autism AI-IMB is an interdisciplinary project between the Artificial Intelligence Institute (AIISC@UofSC) and the Institute of Mind and Brain (IMB) at the University of South Carolina .


Faculty - AI Institute: Prof. Amit P. Sheth Prof. Christian O'Reilly

Faculty - IMB: Prof. Jessica Bradshaw

Graduate Students: Deepa Tilwani


Pilot Grant Funding: $40,000


Autism spectrum disorder (ASD) is a congenital disorder, present from birth, but a diagnosis of ASD is now only possible through behavioral observation and assessment after the age of 18-24 months. There is no known biomarker for autism (e.g., blood test) and current behavioral measures lack the sensitivity and specificity for reliable diagnosis in the first year of life.


Specific Aims:

Aim 1: Determine the feasibility of utilizing wearable sensors to automate the acquisition of infant behavior during infant interactions, including: (i) synchronization of streams of heterogeneous signals (e.g., electrocardiogram (ECG) and gaze behavior); and (ii) storage and visualization of signals at various temporal resolutions utilizing a HIPAA-compliant cloud-based web application.

Hypothesis 1a: Infants and parents will find the wearable sensors acceptable and the quality and content of infant interactions will not be affected by the use of sensors.

Hypothesis 1b: Sensors and human-guided automation will classify infant physiological phases of attention with respect to a particular target that is > 85% in agreement with hand-coded behavior.

Aim 2: Develop computational models to predict the emergence of ASD, based on gaze behavior and heart activity (ECG), utilizing machine learning and time-series analysis techniques. This will include a data science pipeline to process and analyze the collected pilot data that will reveal preliminary differences in gaze behavior during infant interactions.

Hypothesis 2: Analysis of pilot data will show context-specific and group differences in gaze behavior and heart activity (ECG) among infants at high and low likelihood for ASD during social and nonsocial Interactions.

Results: This study pipeline helps in automating the feature extraction,analysis and predicting the likelihood of the autism in infants. Analysis of ECG and heart rate variability have become an important avenue of biomarker research because it is non-invasive and simple to implement. In addition, HRV is a useful measure for studying sympathetic and parasympathetic functions, particularly in infancy. This study demonstrated how HRV, sample entropy, DFA, CSI, and CVI extracted from ECG can predict the familial likelihood of ASD in 3-6-month-old infants. In the future, we plan to build upon this preliminary study to explore further HRV and sample entropy as potential biomarkers of ASD in infancy and replicate these results in a larger sample that includes infants and children with confirmed ASD diagnoses and with non-ASD developmental disorders (e.g., language disorder, attention deficit hyperactivity disorder). We plan to develop a system that can be used by parents to monitor the infant anytime using ECG and have a easy GUI and visualization to track performance.

More Information: For more information on data contact The Early Social Development & Intervention Lab and for analysis pipeline contact


Tilwani, Deepa, Jessica Bradshaw, Amit Sheth, and Christian O’Reilly. "ECG Recordings as Predictors of Very Early Autism Likelihood: A Machine Learning Approach." Bioengineering 10, no. 7 (2023): 827 [1]

O’Reilly, Christian, Sai Durga Rithvik Oruganti, Deepa Tilwani, and Jessica Bradshaw. "Model-Driven Analysis of ECG Using Reinforcement Learning." Bioengineering 10, no. 6 (2023): 696. [2]