01 — Research
Research
Two projects I care about. Both are about building analysis pipelines that stay honest when the sample size is small.
Manuscript
first author · in preparation
Distilling Trial-Level EEG Variability into Participant Representations — Neurophysiologically Constrained Graph Learning in ADHD
open
Distilling Trial-Level EEG Variability into Participant Representations — Neurophysiologically Constrained Graph Learning in ADHD
Summary
ADHD shows up in the trial-to-trial variability of EEG more than in any single averaged peak. We built a graph neural network that learns from each trial as a set of phase-synchrony graphs and combines them into a single representation per participant. It does better than a standard baseline, and — more usefully — we can see where the model gets its information from, which makes the approach easier to trust.
Approach
- ERP-informed windows — five canonical stages: Early (0–200 ms), N2 (200–350 ms), P3 (350–500 ms), Late (500–1000 ms), Slow wave (1000–2000 ms).
- PLV graph priors — a phase-locking-value adjacency per window (720 edges each), so message passing follows functional synchronization, not electrode geometry.
- Trial → participant — hierarchical transfer learning plus a differentiable patient aggregation layer that fuses mean signatures with intra-individual variability.
- Controls — 5-seed × 5-fold patient-level cross-validation against EEGNet; attention patterns and embedding geometry both interrogated.
Findings
- ST-GAT reaches 75.3% ± 6.34% accuracy vs EEGNet's 70.0% ± 3.92% — a +5.3 pp advantage.
- Group structure did not appear in spatial attention allocation — it emerged in the geometry of participant-level embeddings, which distilled trial-level variability.
- The graph architecture compressed discriminative information into few, non-redundant dimensions that CNN features do not capture.
- Window anchors verified against the grand-average NoGo ERP: N2 at −6.2 µV / 324 ms, P3 at +4.0 µV / 496 ms.
Metrics
75.3%
mean accuracy · ST-GAT
+5.3 pp
vs EEGNet baseline
0.68–0.84
AUC across seeds
5
ERP-informed windows
Why it matters
Clinical models are usually judged on accuracy alone, which can hide shortcuts. Our argument is simple: you should also check whether the model is learning something real and robust. Work with Christian Beste and Negin Gholamipourbarogh at the Department of Child and Adolescent Psychiatry, TU Dresden. The manuscript is in preparation.
M.Sc. thesis
DZNE BIONICS · 2026
Unsupervised Computational Framework for Data-driven Phenotyping of Large-scale Neural Dynamics in Stress-related Brain Networks
open
Unsupervised Computational Framework for Data-driven Phenotyping of Large-scale Neural Dynamics in Stress-related Brain Networks
Summary
Prenatal stress changes the hippocampus in every exposed animal — but only some develop stress-related problems. My thesis separated those two questions, what stress exposure does to the brain and what makes some animals resilient, using an unsupervised pipeline on very large electrode recordings.
Data & context
- 4,096-electrode HD-MEA — acute hippocampal slices from adult offspring of stressed dams, under a 4-AP functional challenge.
- ~35 billion voltage samples per slice — Control (CTR), stress-Resilient (RES), and stress-Vulnerable (VUL) cohorts.
- N is small — N = 14 effective (8 animals) for control-vs-pathology; N = 10 (6 animals) for resilient-vs-vulnerable. The protocol must not overfit.
Approach
- Two-stage target — Stage 1 detects the exposure signature (control vs pathology); Stage 2 attempts resilience separation (RES vs VUL).
- Gaussian mixture models on spectral features with stability-based model selection.
- Strict validation — leave-one-slice-out cross-validation and permutation null models.
- Exploratory epoch-level microstate analysis for dynamic state-switching signals.
Findings
- Stage 1 recovered a robust exposure signature — elevated gamma-band power, suppressed spectral entropy — with perfect leave-one-out stability (1.000), present in both RES and VUL offspring.
- Stage 2 hit a hard ceiling: static slice-level features could not separate phenotypes above chance (≤70%, ARI ≈ 0) — exposure shapes the substrate, not the outcome.
- Exploratory microstate analysis found a descriptive network–energy ordering: Control < Resilient < Vulnerable, suggesting resilience is regulated, dynamic state-switching rather than a fixed trait.
- Contributions: a quantified exploratory-to-confirmatory performance gradient and a stability-first template for credible small-sample neural phenotyping.
Metrics
4,096
electrodes per array
~35B
voltage samples per slice
1.000
leave-one-out stability · stage 1
≤70%
static-feature ceiling · stage 2
Conducted at
BIONICS lab (Biohybrid Neuroelectronics), German Center for Neurodegenerative Diseases (DZNE), Dresden — supervised by Dr.-Ing. Hayder Amin, with Prof. Dr. Andreas Deutsch (TU Dresden). The full thesis summary is available in the project workspace.
In progress
updatesI'll link the ADHD paper here once it's out.
Other things I'm thinking about: how the brain combines multiple senses, and whether resilience shows up consistently across different types of recordings.