A group of U.K. scientists has developed a machine-learning-based method to detect ADHD by analyzing the actions of individuals in video clips. These videos included recordings of study participants working on specific tasks, captured using multiple cameras from different angles. The authors report that this method outperformed alternative diagnostic systems in differentiating between individuals with and without ADHD. The research was published in Neuroscience Applied.
Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder characterized by persistent patterns of inattention, hyperactivity, and impulsivity that interfere with functioning or development. Individuals with ADHD struggle to focus on tasks, follow instructions, or organize activities and are easily distracted by external stimuli. Hyperactivity symptoms can include excessive fidgeting, restlessness, or an inability to remain seated or quiet when appropriate.
The disorder typically begins in childhood and can continue into adulthood. It adversely affects academic performance, work responsibilities, and social relationships. ADHD is most often diagnosed when a child starts school, as their behaviors are generally seen as disruptive and frequently result in poor academic performance. To mitigate these and other adverse consequences, timely diagnosis is of utmost importance.
Study author Yichun Li and his colleagues aimed to create an automated ADHD detection system. Their plan involved designing a trial to assess the actions and reactions of individuals with ADHD. Findings from this trial would then be used to develop a detection system based on recognizing human actions from video recordings. The system would classify individuals in the videos as either having ADHD or not.
The researchers first recorded videos of 10 adults diagnosed with ADHD and 12 without the disorder performing designated tasks. Among participants with ADHD, five were male and five were female. Of the participants without ADHD, eight were male and four were female. Participants’ ages ranged between 18 and 45 years. Individuals with ADHD were recruited by CNTW-NHS Foundation Trust, while healthy participants volunteered from Newcastle University in the U.K.
The videos were recorded from three fields of view—front, left, and right—using GoPro cameras. Additionally, the researchers recorded audio and used a keypad’s touch signal to capture tactile data. A screen displaying posters was placed within the participants’ line of sight, and various small objects, such as pens and spinners, were placed on the desk to serve as distractions, which individuals with ADHD are generally more susceptible to.
During the recordings, participants conducted a series of activities, including a 10-20 minute interview, the Cambridge Neurological Test Automated Battery, the beep reaction task (where participants respond to randomly generated beeps), and watching videos labeled as exciting. The entire process lasted about 1 to 1.5 hours.
The researchers created a machine-learning system that recognized elements and movements of the human body from the videos and identified the actions individuals were performing. The extracted information was used to generate various indexes indicating how much the behavior of the person in the video aligned with that expected of individuals with ADHD. Ultimately, the system classified individuals in the videos as having ADHD or not. The authors tested the system using different processing options and selected the best-performing one.
In the final tests, the system achieved a classification accuracy of 95.5%, outperforming similar classification systems based on magnetic resonance imaging (MRI), electroencephalography (EEG), or trajectory analysis. Additionally, the testing procedure was reported to be significantly less expensive.
“Experimental results demonstrate that our system outperforms state-of-the-art methods in terms of F1 score [a measure of prediction precision], accuracy, and AUC [area under the curve, another measure of how good a diagnostic system is]. Compared to conventional EEG [electroencephalography] and fMRI-based techniques [functional magnetic resonance imaging], our system is cost-effective, highlighting its potential for clinical practice. The collected data and results can be shared with doctors to support their diagnosis and follow-up procedures,” the study authors concluded.
The study presents a novel system for recognizing ADHD based on machine learning. However, the authors note that the system was less accurate in identifying females with ADHD. They attribute this to behavioral differences between males and females, with females exhibiting “prolonged small actions” that are more easily overlooked. Furthermore, the system’s performance on shorter video recordings was not as robust as on longer ones.
The paper, “ADHD Detection Based on Human Action Recognition,” was authored by Yichun Li, Rajesh Nair, and Syed Mohsen Naqvi.
