A multi-task learning model with reinforcement optimization for ASD comorbidity discrimination

How to discriminate the comorbidities in autism spectrum disorder (ASD) population has long been an intriguing and challenging issue in neuroscience and neurology practices. Taking attention deficit hyperactivity disorder (ADHD) for example, electroencephalogram (EEG) analysis has alleviated the problem caused by the task of evaluation of similar behaviors of subjects with ASD, ADHD and ASD+ADHD, which requires a very high expertise to reach any concrete conclusions. However, the performance of ASD comorbidity discrimination is still limited by two major difficulties 1) crucial EEG features regarding ASD and ASD+ADHD largely overlap, and 2) reliable data for model training are routinely insufficient. This study proposes a multi-task learning method with "reinforcement optimization" (namely RO-MLT) working in a two-fold manner: 1)Modeling for Discrimination: a multi-task CNN model maintains the target discrimination task (ASD vs. ASD+ADHD) with the aid of the auxiliary task (ASD vs. Typically Developed (TD)), which is designed to mitigate the aforementioned difficulties on model training; and 2) Reinforcement Optimization: a reinforcement learning algorithm enhances the model's feature extraction and fusion capabilities by optimizing its shared structure. Experimental results based on resting-state EEG that collected from 150 ASD, ASD+ADHD or TD children with the RO-MLT method against the state-of-the-art counterparts indicate that RO-MLT is far superior in terms of all performance indicators (e.g., accuracy). Ablation experiments also show that introduction of multi-task learning and reinforcement optimization can achieve a performance boost-up by 11.07%, a gain even higher than the sums of introduction of two individual techniques to the model design.