- Researchers at Massachusetts General Hospital used artificial intelligence (AI) to train convolutional neural networks (CNN) to diagnose attention-deficit hyperactivity disorder (ADHD)
- In this study of 20 adults with ADHD and 20 healthy controls, a CNN trained with EEG event-related spectrograms was 88% accurate in classifying the participants
- If validated in larger datasets, CNNs could be used to support the diagnosis of ADHD on a single-patient basis
Clinical diagnosis of attention-deficit hyperactivity disorder (ADHD) is inherently uncertain because of its multiple different cognitive profiles. In addition, many other conditions present with disordered attention, impulsivity and executive dysfunction, so differential diagnosis is complicated. A biomarker for diagnosis of ADHD would be of great value.
Subscribe to the latest updates from Psychiatry Advances in Motion
In previous studies, artificial intelligence has been used to classify ADHD by analyzing EEGs, but the disease-characterizing features of the EEGs had to be programmed manually. Now, Joan A. Camprodon, MD, PhD, chief of the Division of Neuropsychiatry in the Department of Psychiatry at Massachusetts General Hospital, and research fellow Laura Dubreuil-Vall, MSc, and colleagues have used deep learning to train a convolutional neural network (CNN) to do the classifying.
In Frontiers of Neuroscience, they report the accuracy of the network and the potential for CNNs to be applied someday to individual patients.
Participants and Methods
Deep learning refers to the process of computer systems training themselves to perform tasks using deep neural networks, which have multiple layers of artificial neurons. A CNN is a particular type of neural network that analyzes visual data.
The EEG data in this study came from 20 healthy adults and 20 adults with ADHD. CNNs were trained with:
- Spontaneous EEG data recorded while participants rested with eyes closed
Event-related spectral perturbations (ERSPs) recorded while participants performed a task that assesses sustained attention, conflict monitoring and response inhibition
ERSP reflect changes in the brain oscillations that are time-locked to an external stimulus or internal mental process, such as making decisions or holding an incorrect response
The CNN trained with ERSPs classified 88% of the participants correctly. The other CNN correctly classified 66%.
By using feature visualization techniques popularly known as DeepDream, the researchers determined the main features the CNNs used to classify participants as having ADHD were:
- Decreased activity in the alpha band, which suggests deficits in response inhibition
- Increased power in the delta–theta band around 100 ms, which suggests ADHD patients had to shift more attention to the task to accomplish as many correct responses as the healthy group did
These findings provide further evidence of the pathophysiology underlying ADHD.
Informing Clinical Decisions
If validated with larger datasets, CNNs could be used to support the diagnosis of ADHD in individual patients.
These CNNs were trained with low-resolution EEG datasets (seven channels) of short duration (three minutes), which would make it easy to implement them in EEG clinics and possibly also by outpatient clinicians.
If deep learning systems do become used routinely, clinicians should view their output merely as statistical predictions. They should judge whether the prediction applies to each specific patient and decide whether additional data or expertise are needed to inform the diagnosis.
Visit the Department of Psychiatry
Learn more about research in the Department of Psychiatry