The review, conducted by Atefeh Abedzadeh Attar, Mohammad Hossein Moattar and Esmaeil Kheirkhah of Islamic Azad University in Mashhad, Iran, systematically analyzed 69 studies published between 2020 and 2026 that apply machine learning (ML) and deep learning (DL) techniques to electroencephalography (EEG) and event-related potentials (ERPs) for depression diagnosis. The authors searched Scopus and Web of Science, screened 187 initial records down to the final set, and organized the field around a complete analytical pipeline: signal preprocessing, feature engineering, feature selection, and classification. Their central message is that artificial intelligence can extract objective, reproducible biomarkers of depression from brain signals that clinicians currently cannot see with the naked eye.

Why EEG? Unlike MRI and fMRI, which are expensive, immobile and offer poor temporal resolution, EEG is cheap, portable and captures neural activity at millisecond precision. It records five canonical frequency bands—delta (0.1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz) and gamma (30–40 Hz)—each tied to different cognitive states. ERPs, the brain’s stereotyped voltage responses to specific stimuli, add another layer: depressed patients show characteristic abnormalities in components such as P1 and P2 and heightened sensitivity to negative stimuli. Together, resting-state EEG and stimulus-evoked ERPs offer a complementary window into the neural signatures of the disorder, something the review emphasizes is often missed by studies that examine only resting-state data.

Before any algorithm can learn, the raw signals must be cleaned. EEG recordings are contaminated by eye movements, muscle activity and power-line interference, so the reviewed studies relied heavily on finite impulse response filters, discrete wavelet transforms, independent component analysis, notch filters and even Kalman filters to strip out noise. A second, less obvious problem is dimensional explosion: a single recording can yield thousands of features from only dozens of patients, inviting overfitting. The authors document how researchers tamed this with principal component analysis, locally linear embedding, correlation and variance thresholds, and in one extreme case compressed a set of 10,800 features before modeling even began.

Feature engineering emerges as the heart of the enterprise, and the review organizes it into three families. Handcrafted approaches extract interpretable quantities directly: power spectral density across frequency bands, synchronization measures such as the self-synchronization index and the phase lag index, functional connectivity matrices, and nonlinear descriptors like Lempel–Ziv complexity and various entropies. End-to-end deep learning models—convolutional neural networks, CNN-LSTM hybrids, attention mechanisms, graph convolutional networks and even spiking neural networks—learn their own representations from raw or transformed signals. Between the two sit hybrid methods, which feed manually engineered features such as STFT spectrograms or directed connectivity measures into deep networks, combining human domain knowledge with automated pattern discovery.

The performance numbers are striking. On the machine learning side, accuracies ranged from 82.68 to 100 percent, with support vector machines and K-nearest neighbors dominating the literature; boosting methods such as XGBoost reached 98.92 percent, and one lightweight LightGBM framework hit 97.42 percent using only three prefrontal electrodes—evidence that wearable-scale screening is technically feasible. Deep learning models ranged from 77.78 to 100 percent, with CNN-LSTM architectures reaching 99.9 percent in a real-time wearable system called DepCap, and attention-based and graph-based models consistently exceeding 95 percent. Reported biomarkers converge on frontal and temporal regions, elevated theta and alpha power, interhemispheric asymmetry in delta, alpha and beta bands, and disrupted functional connectivity in parietal-occipital networks.

But the review delivers a sharp caution about those headline numbers. Several studies reporting perfect 100 percent accuracy relied on validation schemes that may leak subject-specific information: when EEG recordings are chopped into segments and randomly split into training and test sets, segments from the same person can appear on both sides, letting the model memorize individuals rather than learn the disorder. One study’s accuracy fell from 100 percent under ordinary 10-fold cross-validation to 83.96 percent under leave-one-subject-out cross-validation, a dramatic illustration of the problem. The authors argue that subject-independent validation, transparent data partitioning and strict leakage controls should be mandatory before any accuracy claim is taken at face value.

The two AI paradigms also carry distinct trade-offs. Machine learning models are interpretable and computationally cheap, letting clinicians trace which biomarkers drove a decision, and they perform well on small datasets—but they depend on laborious, error-prone manual feature pipelines. Deep learning models are robust to noise and capable of real-time monitoring, early detection and severity tracking, yet they demand large datasets, heavy computation, and suffer from black-box opacity that undermines clinical trust. The review also flags ethical concerns: most studies enrolled fewer than 100 participants, raising doubts about generalizability across populations, and over-reliance on automated systems risks eroding the human empathy central to mental health care. The authors call for explainable AI techniques, adherence to reporting standards such as TRIPOD and PROBAST, and patient-centered deployment.

Looking forward, the review sketches a roadmap for the field: larger and more diverse public datasets with standardized acquisition protocols, multimodal models that fuse EEG with speech, eye-tracking and clinical records, federated learning to share knowledge across hospitals without moving sensitive patient data, and foundation models pre-trained on large biomedical corpora to overcome the small-data bottleneck. It also highlights a practical dual trend—increasingly sophisticated architectures on one hand, and radically simplified systems using as few as two frontal electrodes on the other—that could bring objective, AI-assisted depression screening out of the laboratory and into clinics, wearables and underserved communities worldwide. If those validation and equity challenges are met, the authors conclude, brain-signal-based AI could transform depression diagnosis from a subjective art into an objective, scalable science.

Subject of Research: Machine learning and deep learning approaches for diagnosing depressive disorders from EEG and ERP signals

Article Title: A review on machine learning and deep learning approaches for depressive disorder diagnosis based on EEG and ERP signals

Article References: Abedzadeh Attar, A., Moattar, M. H., & Kheirkhah, E. (2026). A review on machine learning and deep learning approaches for depressive disorder diagnosis based on EEG and ERP signals. Discover Artificial Intelligence, 6(1), Article 1308. https://doi.org/10.1007/s44163-026-02334-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02334-5

Keywords: depression, EEG, ERP, machine learning, deep learning, brain signals, diagnosis, biomarkers, neural networks, mental health, feature engineering, clinical AI

The research team recruited 23 younger adults and 22 older adults and asked them to perform a demanding dual-task paradigm. The first task was an audiovisual discrimination task, designed to measure how effectively participants combined what they heard with what they saw. The second was a multiple object tracking task, a classic laboratory probe of sustained visual attention in which observers must keep track of several moving targets among identical distractors. By varying the number of objects that participants had to track, the researchers could systematically manipulate sustained attentional load, from a light demand that left plenty of cognitive resources free, through a medium load, to a high load that consumed most of the observer’s attentional capacity. Crucially, the audiovisual stimuli had to be processed simultaneously with the tracking task, so any change in multisensory performance across load levels could be attributed to how attention was being allocated.

The logic of the experiment rests on a long-standing theoretical debate in cognitive neuroscience. According to classic resource theories of attention, dating back to Daniel Kahneman’s influential 1973 work, attention behaves like a finite pool of energy that can be divided among competing tasks. When one task consumes more of the pool, less remains for everything else. Earlier studies had shown that audiovisual speech integration falters when listeners are asked to carry out concurrent demanding tasks, and that attentional load in one modality can alter how signals from other modalities are combined. But most of this evidence came from young participants, and the question of how aging reshapes the load-integration relationship remained open. Older adults experience declines in sensory processing speed, visual acuity, auditory sensitivity, and attentional capacity, all of which could plausibly change the efficiency with which their brains merge cross-modal signals.

The behavioral results revealed a striking divergence between the age groups. Younger adults displayed an inverted U-shaped pattern of audiovisual integration across attentional loads. Their multisensory advantage, the performance benefit gained from presenting sound and vision together rather than separately, peaked when the tracking task imposed a medium load. Under both low-load and high-load conditions, integration was weaker. This pattern suggests that a moderate attentional demand may actually be optimal for binding the senses in young brains, perhaps because light loads leave attention under-engaged while heavy loads drain the resources needed for cross-modal processing. Older adults, by contrast, showed greater audiovisual integration under low and medium loads than under high load, without the same sharp peak at intermediate demand. Their integration profile was flatter and shifted toward the easier conditions, consistent with the idea that their overall pool of attentional resources is smaller and is exhausted earlier as load increases.

Perhaps the most intriguing finding emerged when the researchers examined the time course of integration using electroencephalography. By recording event-related potentials, the tiny voltage fluctuations elicited by stimuli at the scalp, the team could track when in the processing stream multisensory interactions occurred. They identified distinct integration components at different latencies: an early window reflecting relatively automatic sensory-level convergence, an intermediate window, and a late window spanning roughly 420 to 460 milliseconds after stimulus onset. Remarkably, older adults showed enhanced integration during this late window compared with younger adults. Late components of the evoked potential are typically associated with higher-order cognitive operations, including stimulus evaluation, decision formation, and the allocation of working memory. Enhanced late integration in older adults therefore hints at a compensatory strategy in which the aging brain recruits additional, later-stage processing to achieve the multisensory binding that younger brains accomplish earlier and more automatically.

The electrophysiological data also exposed a second age-related difference: timing. The onsets of both the early and intermediate integration components were delayed in older adults relative to their younger counterparts. In other words, even the relatively automatic, sensory-level stages of combining sound and sight began later in the older brain. This delay fits with a broad literature documenting generalized slowing of sensory and neural conduction with age, and it echoes earlier work by the same research group showing that the temporal window within which older adults bind audiovisual stimuli differs from that of young adults. Together, the delayed onsets and the enhanced late activity paint a coherent picture: the aging multisensory system starts later but works longer, extending its integration into time windows that younger brains have already moved past.

These findings carry significant theoretical weight for understanding how the mind changes across the lifespan. One influential framework in cognitive aging neuroscience, the posterior-to-anterior shift in aging, proposes that older adults compensate for reduced efficiency in posterior sensory cortices by recruiting frontal regions typically associated with executive control. The enhanced late audiovisual integration observed here is consistent with such a compensatory account, suggesting that multisensory processing in aging becomes increasingly entangled with higher-order cognitive resources. It also aligns with prior reports that older adults sometimes show stronger multisensory enhancement than young adults in simple detection tasks, a phenomenon that has been interpreted as the aging brain exploiting every available source of information to offset noisier individual senses. When both vision and hearing degrade, combining them yields a proportionally larger benefit, and the present study suggests that this benefit is actively sustained by late-stage neural processing even under attentional pressure.

The study also refines our understanding of the relationship between attention and multisensory integration more generally. Researchers have debated whether binding across the senses requires attention at all, or whether early integration proceeds automatically and only later stages are modulated by attentional focus. The inverted U-shaped load function in young adults supports a nuanced middle position: integration is neither wholly automatic nor wholly dependent on attention, but instead flourishes within an optimal band of attentional engagement. Too little demand may mean the multisensory system is not prioritized; too much demand starves it of resources. The fact that this band shifts with age indicates that the optimal engagement level is not fixed but calibrated to the organism’s total cognitive capacity, a conclusion that resonates with recent work showing that attentional demands in the visual field modulate audiovisual interactions in the temporal domain.

There are, of course, important caveats. The experiments were not preregistered, and the sample sizes, while adequate for detecting the reported effects, are modest. The multiple object tracking task manipulates sustained visual attention specifically, so the findings speak most directly to situations in which vision is the attentionally taxed modality. The authors note that stimulus materials and de-identified data are available from the corresponding author upon reasonable request, and the study received support from the National Natural Science Foundation of China. Future work will need to determine whether the enhanced late integration observed in older adults reflects genuinely compensatory neural recruitment, perhaps in frontal cortex, or a slower and less efficient form of the same computations that younger brains perform earlier.

For everyday life, the implications are tangible. Older adults routinely navigate environments, such as busy streets, crowded restaurants, and conversations in noisy rooms, in which they must divide attention between visual monitoring and auditory-visual communication. If high attentional load degrades multisensory integration more severely in aging, then situations that demand intense visual vigilance may disproportionately impair the ability to combine what older individuals see and hear, with potential consequences for driving safety, social interaction, and fall risk. At the same time, the demonstration that the aging brain can enhance late-stage integration offers an encouraging message: the multisensory system retains plasticity and can adapt its temporal dynamics to meet the demands imposed by a changing cognitive landscape. Understanding and perhaps training these compensatory mechanisms could one day inform interventions designed to preserve effective perception in older age, turning a laboratory curiosity about event-related potentials into a practical tool for healthy aging.

Subject of Research: Age-related changes in how sustained visual attentional load modulates audiovisual integration in younger and older adults

Article Title: Age-related changes in sustained visual attentional modulation of audiovisual integration

Article References: Ren, Y., Xue, H., Yang, M., Wu, Y., Li, Y., & Yang, W. (2026). Age-related changes in sustained visual attentional modulation of audiovisual integration. Attention, Perception, & Psychophysics, 88(7), Article 197. https://doi.org/10.3758/s13414-026-03346-0

Image Credits: AI Generated

DOI: 10.3758/s13414-026-03346-0

Keywords: audiovisual integration, sustained attention, attentional load, aging, event-related potentials, multisensory processing, cognitive aging, multiple object tracking, neural compensation, psychophysics, ERP, older adults

The study was inspired by an earlier hypothesis in the field sometimes called sound-induced visual segmentation. When two two-dimensional objects approach one another and reach their point of closest approach, a coincident sound appears to lead observers to overestimate the distance between them, as though the auditory event slices the visual scene into distinct objects. The research team, led by Wenxuan Song, Xiaoying Li, Yuan Guo, and Song Zhao of Soochow University in China, reasoned that if a sound can segment approaching objects in this way, it might also amplify the non-contact illusion that arises when two moving squares disappear precisely at the instant their edges meet. Across two experiments, they tested whether a sound presented at contact would increase the probability that observers report the squares as having never touched.

The results were clear and consistent. In both Experiment 1 and Experiment 2, participants were significantly more likely to give a non-contact response when a sound accompanied the moment of visual contact than when the squares collided in silence. This sound-induced increase in the illusion occurred even though the sound carried no information about the visual stimulus and participants were explicitly instructed to ignore it. The effect is therefore a genuine cross-modal influence: an auditory signal, meaningless on its own, reshapes the conscious perception of a visual event in a systematic and reproducible way.

What makes the finding particularly striking is that the effect runs in the opposite direction from what a simple collision interpretation would predict. The classic stream-bounce literature, dating back to Sekuler and colleagues’ 1997 demonstration that a sound can make two crossing discs appear to bounce off one another, has often been interpreted through the lens of causal inference: the sound suggests a collision, so the brain infers a collision. But in the new study, the sound produced more non-contact judgments, not more contact judgments. If the sound had merely signaled that two objects struck one another, it should have biased observers toward reporting contact. Instead, it deepened the impression that the objects remained apart, ruling out this response-bias account.

The researchers also tested and rejected an attentional explanation. One possibility is that the abrupt sound simply distracts observers, impairing their ability to encode the visual event and thereby pushing them toward a default or guess response. If that were true, non-contact responses in the sound-present condition should have been slower than contact responses, reflecting a disruption of processing. The reaction-time data told a different story: the non-contact response was never slower than the contact response when the sound was present. In fact, supplementary analyses showed that reaction times for non-contact responses were significantly shorter in the sound-present condition than in the sound-absent condition, while contact responses did not differ between sound conditions. This pattern contradicts any account based on attentional distraction or slowed processing.

To probe the neural mechanisms underlying the effect, the team recorded high time-resolution event-related potentials, or ERPs, during Experiment 2. The critical comparison focused on an early cross-modal component known as PD170, a positivity peaking between roughly 125 and 175 milliseconds after sound onset that has previously been associated with low-level audiovisual interactions in early sensory cortex. The results showed that the PD170 was significantly larger on sound-present trials in which participants ultimately reported the non-contact percept than on sound-present trials in which they reported contact. This dissociation, emerging within the first fifth of a second after the sound, indicates that the sound-induced enhancement of the illusion is rooted in early, low-level cross-modal interactions rather than in later cognitive evaluation or decision processes.

The PD170 finding carries substantial theoretical weight. Early cross-modal components of this kind have been linked to interactions between auditory and visual cortex that occur automatically, before attention and higher-order cognition can shape the percept. By tying the strength of the behavioral illusion to the amplitude of this early component, the study substantiates the perceptual nature of the effect: the sound does not merely change what people say about the display but genuinely alters what they see. This aligns the new phenomenon with a family of well-documented sound-induced visual illusions, including the sound-induced flash illusion, in which a single flash accompanied by two beeps is perceived as two flashes, and the audiovisual bounce-inducing effect, in which a sound at the crossing point of two moving discs promotes a bouncing percept.

At the same time, the non-contact illusion and its sound-induced enhancement differ in an important way from the stream-bounce paradigm. The non-contact illusion does not require a sound to emerge; it arises purely from the visual statistics of the display, specifically the sudden disappearance of two squares at the moment their edges make contact. The sound merely strengthens an already-existing visual illusion rather than resolving an ambiguous motion event. This makes the paradigm a particularly clean tool for studying how auditory signals modulate visual spatial perception, because baseline and sound-modulated conditions can be compared within the same unambiguous geometric event. The authors note, following recent work by Zeljko and Grove, that there is no inherent bias toward any particular percept in the sound-absent condition unless it is intermixed with sound-present trials, further underscoring the importance of careful experimental design in this domain.

The methodological rigor of the study strengthens its conclusions. The researchers used mixed-effects logistic modeling with maximal random-effects structures, in line with contemporary best practices for categorical data analysis, and complemented frequentist tests with Bayesian analyses to evaluate null results. All trial-level data, subject-level data, analysis scripts, and experimental files for both experiments are openly available on the Open Science Framework, allowing independent verification and reuse. The study was approved by the Institutional Review Board of Soochow University, informed consent was obtained from all participants, and the work was supported by the National Natural Science Foundation of China and an undergraduate innovation training program at Soochow University.

Looking forward, the authors suggest that the sound-induced enhancement of the non-contact illusion constitutes a promising new paradigm for investigating multisensory processing. Because the effect is behaviorally robust, mechanistically traceable to an identifiable early ERP component, and free from the confounds that complicate stream-bounce designs, it offers researchers a versatile instrument for dissecting when, where, and how the brain integrates sound and sight. Some limitations remain: the design’s contact-duration constraints restricted reaction-time analyses to a single duration condition, and ERP requirements limited the number of participants meeting trial-count criteria in some conditions, pointing toward richer future experiments. Even so, the demonstration that a meaningless beep can tip the brain’s reconstruction of a visual collision toward the conviction that no contact ever occurred is a vivid reminder that perception is not a passive recording of the world. It is an active, multisensory construction, one that sound can quietly rewrite within a fraction of a second.

Subject of Research: Sound-induced enhancement of the non-contact illusion, a cross-modal audiovisual perceptual effect studied with behavioral and ERP measures.

Article Title: A novel cross-modal perceptual effect: Sound-induced increase of the non-contact illusion

Article References: Song, W., Li, X., Guo, Y., & Zhao, S. (2026). A novel cross-modal perceptual effect: Sound-induced increase of the non-contact illusion. Attention, Perception, & Psychophysics, 88(7), Article 183. https://doi.org/10.3758/s13414-026-03332-6

Image Credits: AI Generated

DOI: 10.3758/s13414-026-03332-6

Keywords: non-contact illusion, cross-modal interaction, audiovisual perception, multisensory integration, event-related potential, PD170, sound-induced illusion, visual motion perception, stream-bounce effect, perceptual psychology, ERP, attention perception psychophysics