Peer-reviewed papers, clinical abstracts, and conference posters spanning clinically validated biosignal ML, cognitive neuroscience and language processing, and language model evaluation. Full list on Google Scholar.
Clinical biosignal entries reference three study programmes: SEPARATE (in vivo peripheral arterial clot trial, 17 patients), E-SEPARATE (ex vivo thrombus validation versus histology, 15 patients), and INSPECT (first in human lung tissue classification during bronchoscopic biopsy, 26 patients).
Clinically validated machine learning (ML) on biomedical signals: peripheral arterial clot trials, lung biopsy tissue classification, and multimodal psychiatric biomarkers from clinical trial cohorts.
SEPARATE and E-SEPARATE evaluated impedance based clot characterisation with the Clotild® smart guidewire (Sensome). Bioimpedance targets clot tissue rich in red blood cells.
Clinical abstract
In vivo clinical validation (n = 17); 100% primary endpoint success for lesion impedance data. Clot identification in peripheral arterial disease (tissue rich in red blood cells).
Also presented at Paris Vascular Insights (PVI) 2024, Paris
Clinical abstract
Ex vivo machine learning (ML) versus histology gold standard (n = 15); coefficient of determination R² = 0.79 in peripheral arterial disease.
Also presented at Paris Vascular Insights (PVI) 2024, Paris
INSPECT used bioimpedance on a bronchoscopy stylet for tool-in-lesion confirmation during biopsy of central and peripheral lung lesions (Sensome).
Clinical abstract
Machine learning (ML) tissue classification during bronchoscopic lung biopsy; 26 patients across Australia and France.
Multimodal model predicting treatment response across four psychiatric conditions in pooled clinical trial cohorts. Christos led biosignal and imaging feature extraction (EEG, ECG, galvanic skin response).
Conference poster
Major depressive disorder, ADHD, OCD, and PTSD; external validation on unseen clinical cohorts (TRIPOD Type 4); ranked first in the TDBRAIN international competition.
MEG and EEG during sentence reading, contrasted with a recurrent language model (*Cortex*), and grammatical agreement in humans versus language models (EMNLP 2023).
PhD work at NeuroSpin / Sorbonne University: how humans and language models process sentence structure. MEG/EEG during reading versus a two-layer long short-term memory (LSTM) language model.
Peer-reviewed journal
n=22; combined MEG and EEG during sentence reading. Only hierarchical structure was decodable from brain signals; transition and congruity stayed at chance. A two-layer LSTM on the same sentences showed all three effects decodable.
Human grammatical agreement versus language models. First-author EMNLP 2023 main-track paper with Meta AI and NeuroSpin co-authors.
Conference paper
Rapid serial visual presentation agreement task (n=34): humans and language models err more with proximal attractors; linear response-time effect of distance; GPT-Neo-1.3B and grammar-corrected T5 compared to humans.
Computational studies of large language model (LLM) behaviour and internal representations: personality-trait expression and semantic-violation detection in causal LMs.
Personality-trait probing and layer-wise semantic-violation decoding in causal language models.
Conference paper
Six LLMs, BFI-2, temperature 0–2: four traits differ across models; Neuroticism and Extraversion track temperature (R² = 0.35 / 0.25).
Conference paper
Phi-2, 1520 sentence pairs: per-layer AUC shows semantic violations decoded in layers 18–30 (cluster p < 0.001); early layers at chance; participation ratio expansion then collapse.
Earlier contributions before the clinical-AI and NeuroSpin research lines.
Cross-sectional psychometrics (Aristotle University of Thessaloniki): valence and arousal norms across the adult lifespan.
Conference abstract
Cross-sectional study (n = 84): older adults rated Hellenic words more positively and with higher arousal than younger groups; age-by-valence interactions across pleasant, neutral, and unpleasant word sets.