Within-cohort discrimination and cross-cohort transfer of EEG-based ADHD classifiers: An exploratory two-cohort evaluation
This exploratory study reveals that while EEG-based classifiers achieve strong within-cohort discrimination for adults with ADHD, they fail to generalize to children and often produce reverse rankings across cohorts, highlighting significant limitations in cross-cohort transferability due to unseparated differences in age, task, and acquisition.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
The human brain is a complex electrical organ, constantly firing signals that can be measured from the scalp using sensors. These signals, known as electroencephalography or EEG, create a unique electrical fingerprint for every person. For decades, scientists have hoped to use these fingerprints to identify specific brain conditions, such as attention-deficit/hyperactivity disorder, or ADHD. The idea is that if a computer can learn to recognize the electrical patterns of a healthy brain versus one with ADHD, it could eventually help doctors make faster, more accurate diagnoses. However, a major hurdle exists: the brain changes as we grow, and the way we record these signals changes depending on what a person is doing at the moment. A pattern found in a child sitting still might look very different from a pattern found in an adult, or even from the same child doing a different task. The big question is whether a computer program trained to spot ADHD in one group of people can actually work on a completely different group of people, or if it only works for the specific group it was taught on.
A recent study set out to test this exact problem using two very different groups of people. The researchers gathered data from 121 children who were performing a visual task and 79 adults who were simply resting with their eyes open. They used a standard computer learning method to build a model that could distinguish between people with ADHD and those without, but they did this in a very strict way. They trained the model on one group and then immediately tried to use it on the other group without making any changes or adjustments. This is like teaching a student to recognize a specific type of bird in a forest and then asking them to identify that same bird in a completely different forest, with different lighting and different trees, without giving them a new guide.
The results were striking and revealed a deep divide between what happens inside a single group versus what happens when you try to cross over to another. When the researchers tested the computer models on the adults they had trained them on, the system worked exceptionally well. It correctly identified the adults with ADHD about 95% of the time, a level of accuracy that suggests the electrical patterns in the adult group were very distinct and easy for the computer to learn. However, when that same computer model, trained on the adults, was applied to the children, it failed completely. In fact, it did worse than random guessing. The model consistently ranked the children with ADHD as if they were healthy, and the healthy children as if they had the disorder. The reverse was also true: a model trained on the children performed poorly when tested on the adults, often ranking the groups in the opposite order of what was expected.
This finding is not just a simple failure of the technology; it highlights a fundamental difference in how the brain signals behave across ages and tasks. The researchers found that the specific electrical features that made the adults easy to classify were completely different from the features that might have helped with the children. In the adult group, the computer relied heavily on the frequency of the brain waves, a measure of how fast the signals were oscillating. In the children, these same frequency measures provided almost no useful information. The study showed that even when the researchers used the exact same sensors on the same two spots on the head and recorded for the exact same amount of time, the underlying data was so different that a model trained on one group became useless for the other.
The researchers also tested how the amount of data used for training affected the results. They found that giving the computer more data from the children did not fix the problem. Even when they fed the model every single available recording from the children, the model trained on the children remained confused, though its performance improved slightly to approach random guessing rather than staying at the bottom. The model trained on the adults stayed highly accurate for adults but remained useless for children. This suggests that the issue is not simply a lack of data or a need for more training time. Instead, the nature of the signal itself changes so drastically between a child doing a task and an adult resting that a single set of rules cannot apply to both.
The study concludes that the success of a brain-computer model is deeply tied to the specific group of people it was built for. A model that works perfectly for one population does not automatically work for another, even if the goal is the same medical diagnosis. The researchers emphasize that this does not mean the technology is broken or that ADHD cannot be detected with brain waves. Rather, it means that the current methods are not yet robust enough to cross the gap between different ages and different activities. The findings serve as a necessary reality check for the field, showing that high accuracy in one setting does not guarantee success in another. Until scientists can find a way to build models that understand these deep differences between children and adults, or between different tasks, these tools will remain limited to the specific groups they were designed for, rather than becoming a universal diagnostic tool.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.