- Homepage
- Programme
- Registration
- Practical Guide
- About Neuronus
24.04.2026, Friday, 16:00-17:30
Sano Centre for Computational Medicine, Czarnowiejska 36, 30-054 Kraków, Poland.
"Interpretable Machine Learning Framework Reveals Event-Related and Oscillatory EEG."
Early detection of dementia remains a clinical challenge and often relies on subjective cognitive assessments. Electroencephalography provides objective biomarkers of neuronal dysfunction, yet clinical adoption is limited by lack of standardisation and limited interpretability of machine learning approaches.
This study aimed to develop and evaluate a framework combining portable EEG devices with explainable artificial intelligence for the early detection of cognitive impairment.
A proof-of-concept study included 24 participants, 15 with Mild Cognitive Impairment and 9 healthy controls, stratified using MMSE and MoCA. EEG was recorded with a 4-channel portable device during a 5-minute passive emotional visual paradigm designed to elicit event-related potentials related to emotion, attention, and memory processing. Extracted features were analysed using tree-based classifiers within a Leave-One-Out Cross-Validation scheme. Robustness and interpretability were evaluated using SHAP analysis, bootstrap validation, and permutation testing.
The Random Forest model achieved an area under the ROC curve of 0.93. At the optimised threshold of 0.56, accuracy reached 87.5 percent, with 86.7 percent sensitivity and 88.9 percent specificity. Bootstrap validation confirmed stability, yielding a mean AUROC of 0.925 with a 95 percent confidence interval from 0.790 to 1.000. Permutation testing indicated statistical significance with p equal to 0.032. SHAP analysis revealed increased N2 latency, increased delta power, and reduced theta and beta activity, indicative of an early sign of cognitive decline, neurodegeneration, early hippocampal dysfunction and loss of synaptic connectivity, respectively.
The framework captures neurophysiological signatures of early cognitive decline using a non-invasive and scalable approach. It provides objective measures supporting early cognitive assessment and further investigation of dementia-related neural mechanisms.
Uniwersytet Warszawski, Wydział Fizyki
"Machine and Deep Learning Approaches for Automated EEG Neuroscreening."
High-dimensional EEG data exhibit high redundancy and low signal-to-noise ratios, complicating automated neuroscreening. This research develops ML-assisted methods to construct reduced, interpretable EEG representations that preserve clinically relevant information.
To integrate dimensionality reduction, connectivity modeling, and generative methods to build interpretable ML-supported systems that enable rapid, objective neuroscreening and serve as clinical decision support tools.
Phase 1 (Bachelor’s): Feature-Based Analysis Investigated 2,850-dimensional handcrafted time–frequency features from the ELM19 dataset. Applied PCA and ICA for dimensionality reduction and developed a Selective Activation method to back-project latent components, enabling transparent inspection of original signal drivers. Phase 2 (Master’s): Network-Based Analysis Extended work to directed connectivity via multichannel autoregressive models, including Directed Transfer Function (DTF) and frequency-resolved variants. Clustering and aggregation were used to stabilize high-dimensional connectivity vectors. Variational Autoencoders (VAEs) were employed to learn low-dimensional latent spaces, identifying stable connectivity motifs as potential early biomarkers of dysfunction.
In the feature-based phase, applying PCA and ICA to the ELM19 dataset achieved a 90% reduction in dimensionality with only a 2-percentage-point decrease in AUC (0.86 to 0.84) for a Gradient-Boosted Ensemble classifier. The Selective Activation method successfully enabled the identification of specific EEG fragments driving model decisions. Phase 2 is being developed.
With variance-based component selection, dimensionality reduction preserved GBE performance while improving interpretability and generalizability across heterogeneous clinical contexts. Building on this, connectivity-based features introduce network-level representations to further strengthen robust, physiologically meaningful EEG assessment.
Jagiellonian University, Centre for Cognitive Science, Ingardena 3, Krakow, Poland
"Predicting Response Inhibition: A Deep Learning Approach Using Pre-Response Single-Trial EEG Data."
Response inhibition, the ability to suppress automatic actions in favor of goal-directed behavior, is crucial for self-regulation. It is commonly studied using the stop-signal task (SST), in which frequent go responses must be inhibited when an infrequent stop signal appears. Electroencephalography (EEG) studies have identified neural markers of inhibitory control, yet brain-behavior correlations cannot determine whether observed neural activity reflects inhibition specifically or more general processes that occur in the same time window. Machine-learning approaches may address this limitation by predicting stopping behavior directly from neural signals, but most EEG-based models rely on late post-stimulus activity that may reflect post-response monitoring rather than inhibition itself.
This study used deep neural networks to predict stopping outcomes from EEG activity related to go and stop stimuli while excluding post-response signals.
EEG was recorded from 225 volunteers (113 female, 1 non-binary), aged 18–39, performing an SST. Signals were segmented into go-locked and stop-locked windows and classified using the EEGNet architecture. To compare information across time windows, we trained three models on the following signals: go, stop, and combined go-stop.
Only the go-stop model performed above chance (accuracy = 57.8%, AUC = 57.7%). Saliency analyses indicated that both go- and stop-related activity within the first 100 ms contributed to classification result. This suggests that successful inhibition depends on interactions between response initiation and stopping processes. Importantly, this effect cannot be explained by stop-signal delay (SSD) differences between trial types across participants (t(224)=.81, p=.41).
Together, these findings show that early perceptual EEG activity predicts stopping success beyond behavioral timing differences, highlighting the role of early interactions between go and stop processes in response inhibition.
Machine Learning Group, Luleå University of Technology, Sweden
"Bandwise Dynamic EEG Microstate Connectivity and Transition Graph Modeling for Alzheimer’s Disease and Frontotemporal Dementia"
Electroencephalography (EEG) microstates offer a non-invasive window into large-scale brain connectivity and its disruption in neurodegenerative disorders such as Alzheimer’s disease (AD) and frontotemporal dementia (FTD).
We aimed to develop a band-wise dynamic microstate and microstate transition framework that characterizes connectivity changes in AD and FTD relative to healthy controls (HC) and links these changes to the cognitive disease stage.
Dynamic microstates were extracted for AD, FTD, and HC across five EEG bands (delta–gamma) using a quantile-based filtering procedure. For each band, brain connectivity-based microstate transition graphs were built by integrating Granger causality and spatial power variations. Local and global graph metrics were quantified using clustering coefficient (CC) and information flow (IF). Between-group differences and associations with AD severity (mild, moderate, severe) were assessed with Mann–Whitney U tests (p ≤ 0.05).
Compared with HC, both AD and FTD showed significant band-specific alterations in microstate transition graphs. Delta and alpha bands exhibited the strongest CC reductions, whereas beta and gamma bands showed marked IF deficits. In FTD, delta, alpha, and gamma abnormalities were most pronounced. CC and IF changes increased with AD severity.
Band-wise dynamic microstates and their transition graphs capture disease and stage-specific connectivity disturbances, supporting their use as interpretable, non-invasive EEG biomarkers for monitoring neurodegeneration in clinical settings.
University of Silesia in Katowice, Faculty of Humanities, ul. Bankowa 12, 40-007 Katowice, Poland
"Designing Artifact-Aware Virtual Reality Environments for Mobile EEG Research in Human Cognition."
The integration of virtual reality (VR) with mobile electroencephalography (EEG) enables ecologically valid cognitive research but introduces technical constraints related to rendering latency, frame-dependent timing variability, motion artifacts, and cross-device synchronization. Inadequate control of these factors can compromise event-related potential (ERP) integrity.
The aim of this work was to design a technically robust VR environment architecture optimized for mobile EEG experiments, emphasizing deterministic stimulus scheduling, hardware-level synchronization, and artifact-aware scene construction.
The environment was developed in Unity using a modular control architecture separating stimulus presentation logic, locomotion handling, behavioral input detection, and trigger transmission. Stimuli were preloaded and activated or deactivated at runtime to eliminate instantiation-related latency. Temporal control relied on coroutine-based scheduling executed on the main thread to reduce timing variability. Event markers were transmitted through a USB TTL hardware interface directly to the EEG acquisition system. Marker signals were dispatched at stimulus onset and response detection to ensure temporal correspondence. To stabilize performance, environmental geometry was constrained and lighting was fully baked to reduce rendering load. Head-position tracking was used to monitor vertical displacement during movement, and stimulus flow could be paused to mitigate motion-related artifacts.
System validation demonstrated stable trigger transmission and consistent temporal alignment between VR events and EEG acquisition across stationary and movement conditions.
Careful VR–EEG integration requires explicit control of stimulus timing, rendering performance, and hardware synchronization. The presented architecture provides a technical framework for constructing artifact-aware immersive environments suitable for mobile EEG research in cognitive neuroscience.
