Depression Between the Electrodes: Why AI Misreads the Brain's Broken Network

“Fine.”

It’s the most common lie told in a psychiatrist’s office and the reason half of all depression cases walk out undiagnosed (Mitchell et al., 2009). A psychiatrist sits across a patient and asks how they’ve been sleeping. The patient likely responds with “not well”, “fine”, or shrugs. The conversation that follows is watered down through a structured questionnaire and whatever the clinician’s training and intuition can extract from it. From this, a diagnosis either emerges or doesn’t. Catching it takes time: six to twelve months on average for formal diagnosis (Ghio et al., 2014). First-line antidepressants, typically SSRIs such as sertraline or fluoxetine, work for barely a third of the people who take them (Rush et al., 2006); the other two out of three are, pharmacologically speaking, guessing.

A better signal has been available the whole time. For decades, electroencephalography (EEG) has been placing electrodes across the scalp and recording the brain’s electrical activity in real time, a standard tool for diagnosing epilepsy and studying sleep for nearly a century. Unlike brain imaging machines that cost millions of dollars and fill entire rooms, an EEG system costs only a few thousand dollars and is portable and fast enough to deploy in a primary care office. Buried in those recordings are patterns, from frontal asymmetries to breakdowns in how brain networks communicate, that decades of EEG studies have linked to depression, sitting undetected in that conversation. Artificial intelligence (AI) designed to read those signals should have changed everything, but it hasn’t because it’s been looking at the brain the wrong way.

Start with the earliest finding: a simple imbalance. In the late 1990s, researchers using EEG noticed that in depressed patients, the left side of the brain’s frontal region was quieter than the right. This is the area of the brain that drives motivation and engagement with the world, and it looked as though someone had reached in and turned it down at the source. The finding survived decades of replication, and raw EEG could reveal it without sophisticated processing (Thibodeau et al., 2006). For a field that in over a century of clinical psychiatry never had an objective diagnostic marker, this felt like the breakthrough. Then a second pattern complicated the picture. Excessive slow-wave activity showed up in the brain’s frontal regions, meaning that the mind would be suspended in a feedback loop, replaying the same regret, rehearsing the same dread like a broken record (Pizzagalli et al., 2003). Now there were two biomarkers instead of one, and both showed up at individual electrode sites, which made detection look like a matter of better engineering, or so the field thought.

Then, researchers changed the question. Instead of asking which region was malfunctioning, they began asking how those regions communicate. They stopped looking at individual regions and started listening to the conversations between them, and what they heard back changed what depression was. The brain runs a default mode network, the system responsible for self-reflection and inward thought, the one that switches on when you stare out a window and replay a conversation from six years ago. Normally a second system, the salience network, acts as its emergency brake, pulling attention outward and overriding the loop before rumination takes hold. In depressed patients, that balance collapses. The default mode network runs hyperconnected, rumination looping with no off switch, while the salience brake stays disconnected from the wheels and never engages. Kaiser et al. (2015) mapped these disruptions across both the default mode and salience networks, showing that the problem wasn’t any single region but the collapse of communication between them (Kaiser et al., 2015; Fingelkurts, 2007). Fingelkurts (2007) traced the same failure in the brain’s electrical rhythms: the alpha and theta waves, the slow oscillations that normally keep the front and back of the brain in sync, had fallen out of step. Depression had never been a region producing the wrong signal, but a conversation between regions breaking down. Reading that conversation was like diagnosing a failing phone network by checking whether each tower’s light was on without ever testing whether the towers could talk to each other.

Convolutional neural networks (CNNs) are the dominant AI used for reading EEG, and they were built for exactly the framework that neuroscience had just overturned. A CNN works like a technician inspecting a phone network one tower at a time, confirming that each tower has power and a blinking light but never asking whether the towers can reach each other. Each electrode channel is the equivalent of one tower, scanned on its own, its patterns learned in isolation. How those signals relate to one another never enters the picture (Schirrmeister et al., 2017). The model confirms that every tower is lit but never tests the lines between them. Most of the time, checking the towers alone is enough to get the answer right.

Recent studies report that these models correctly classify patients 85 to 89 percent of the time (Ay et al., 2019; Saeedi et al., 2020), but those numbers carry a caveat. Most EEG depression datasets contain fewer than 100 subjects, and some fewer than 40. High accuracy on small samples is notoriously inflated by the model memorizing specific patients rather than learning the disease itself, where the model essentially learns to recognize a specific patient’s brain rather than depression’s brain. Inflated or not, that success may have been the worst thing to happen to the field. 89 percent are seductive. High enough to publish, all the while being too high to question. So researchers kept optimizing the approach instead of asking whether it could see the disease at all.

So what happens when the model built to find the connections can’t actually see them? Graph-based models were supposed to be the answer. They treat electrodes like stations on a transit map, connected by tracks representing their relationships. As a framework, this was a genuine conceptual advance. It acknowledged what neuroscience had been saying, that the brain is a network and the disease lives in that network’s wiring. Before the model ever sees a patient’s data, researchers have to decide which electrodes are connected to which. They measure how correlated the activity between each pair of electrodes is and draw a line between the pairs that pass a threshold. Pairs that don’t pass get no line and become invisible to the model (Li et al., 2022). If a connection between two regions falls just below the threshold, the model will never know it existed. Even if that connection is exactly where the disease lives. That same pesky blind spot kept moving.

Consider what changes when the model turns into a cartographer and draws its own map. In 2020, Song’s team did what the depression field had not, they built a model that learns the brain’s connections during training (Song et al., 2018). The model figures out which electrodes matter to each other by listening to the data itself. Say a study builds its map from 60 healthy controls and 60 depressed patients. That map is an average. Patient 61 walks in with a connectivity disruption that the average model never saw, in a connection the group data said didn’t matter. When the model learns its own connections from each individual recording, the map changes with the patient. Every connection in that map carries a weight, a number representing how much it mattered to the classification. A clinician can open it up and ask: which connections drove this diagnosis? Do they line up with what we know about the disease? That’s the difference between a tool that hands an answer and a tool that shows how it got there.

Catching on has been slow, held back by scarce training data and by methods that have worked well enough for long enough to discourage change. No published study as of 2026 has applied fully adaptive graph learning to depression EEG classification on a clinical-scale dataset. Only in preprints have the first attempts begun to surface (Wang et al., 2024). Half a decade of models published, optimized, and celebrated, without ever being asked to see the disease the way neuroscience describes it.

Electrodes were never the bottleneck, and neither were the biomarkers. Neuroscience had identified the network disruptions underlying depression years ago; what lagged behind was the model reading the signal. Even when the model gets it right, catching the disease is only half the answer. Depression is not one uniform disease. A patient trapped in the same spiral of self-critical thought may respond to cognitive behavioral therapy. Someone whose attention-control system has weakened may need an entirely different intervention. Drysdale et al. (2017) took connectivity data from over 1,000 patients, all diagnosed with depression, and found that those patients sorted into four distinct subtypes. Patients in two of the four subtypes responded to transcranial magnetic stimulation at roughly three times the rate of patients in the other two (Drysdale et al., 2017). Same diagnosis on paper. Wildly different brains underneath. Anyone building the next model needs to answer a harder question: where in the brain to look, and once you’ve found something, what it actually tells you about which treatment might work.

The psychiatrist is still asking. The patient is still answering. The brain is generating a signal the entire time. The question is whether the tools we build will be designed to see it, or whether we’ll keep optimizing the ones that can’t.

References

1. Mitchell, A. J., Vaze, A., & Rao, S. (2009). Clinical diagnosis of depression in primary care: A meta-analysis. The Lancet, 374(9690), 609-619.
2. Ghio, L., Gotelli, S., Marcenaro, M., Amore, M., & Natta, W. (2014). Duration of untreated illness and outcomes in unipolar depression: A systematic review and meta-analysis. Journal of Affective Disorders, 152-154, 45-51.
3. Rush, A. J., Trivedi, M. H., Wisniewski, S. R., et al. (2006). Acute and longer-term outcomes in depressed outpatients requiring one or several treatment steps: A STAR*D report. American Journal of Psychiatry, 163(11), 1905-1917.
4. Thibodeau, R., Jorgensen, R. S., & Kim, S. (2006). Depression, anxiety, and resting frontal EEG asymmetry: A meta-analytic review. Journal of Abnormal Psychology, 115(4), 715-729.
5. Pizzagalli, D. A., Oakes, T. R., & Davidson, R. J. (2003). Coupling of theta activity and glucose metabolism in the human rostral anterior cingulate cortex. Psychophysiology, 40(6), 939-949.
6. Fingelkurts, A. A., & Fingelkurts, A. A. (2007). Impaired functional connectivity at EEG alpha and theta frequency bands in major depression. Human Brain Mapping, 28(3), 247-261.
7. Schirrmeister, R. T., et al. (2017). Deep learning with convolutional neural networks for EEG decoding and visualization. Human Brain Mapping, 38(11), 5391-5420.
8. Ay, B., et al. (2019). Automated depression detection using deep representation and sequence learning with EEG signals. Journal of Medical Systems, 43(7), 205.
9. Saeedi, A., Saeedi, M., Maghsoudi, A., & Shalbaf, A. (2020). Major depressive disorder diagnosis based on effective connectivity in EEG signals. Cognitive Neurodynamics, 15(2), 239-252.
10. Li, Y., et al. (2022). EEG-based depression recognition using improved graph convolutional neural network. Computers in Biology and Medicine, 148, 105815.
11. Song, T., Zheng, W., Song, P., & Cui, Z. (2018). EEG emotion recognition using dynamical graph convolutional neural networks. IEEE Transactions on Affective Computing, 11(3), 532-541.
12. Wang, Y., Wu, Y., Luo, G., & Chen, W. (2024). A hybrid graph neural network for enhanced EEG-based depression detection. arXiv preprint, arXiv:2410.18103.
13. Drysdale, A. T., et al. (2017). Resting-state connectivity biomarkers define neurophysiological subtypes of depression. Nature Medicine, 23(1), 28-38.