I develop compact learning methods for physiological signals. I want representations that preserve the information needed to study human function and remain useful across tasks, people, and recording conditions.
Long recordings contain many samples, but the information relevant to a particular process may lie in a few temporal, spectral, or spatial patterns. I design models around these properties, and study how their training objectives shape what they learn. This connects efficiency with generalization: I ask which aspects of the signal remain useful outside the training setting.
In WildECG, I trained a self-supervised state-space model on 275,000 ECG segments recorded in everyday settings. Its training task identifies transformations applied to a signal, encouraging sensitivity to distortions while the architecture efficiently models long sequences. [1]
I examine this question at the signal level with BioCodec, using quantization to turn continuous biosignals into discrete codes. This bottleneck learns compact signal structure: a tokenizer trained on single-channel EEG retained information useful for clinical, sleep, and neural-decoding tasks. [2]
I then used these codes to develop BrainCodes, a lightweight EEG foundation model over discrete neural codes, pretrained on recordings from over 1,000 participants. It matches or exceeds larger models on several evaluated tasks, with up to 97% fewer trainable parameters. [3]
To understand this transfer, I also examine the representations themselves. Predictive accuracy alone does not establish which aspects of physiology a model uses. My work examines generalization across recording settings [4] and bias toward aperiodic and low-frequency activity [5]. I also developed spatial and spectral probes to distinguish information encoded in a representation from information that influences its predictions [6].
I am interested in how reconstruction, contrastive, and predictive objectives preserve different aspects of physiology. I want to develop hierarchical representations that make both transient events and stable individual traits accessible, with evaluations that test signal morphology, oscillations, and spatial relationships.
@article{avramidis2023scaling,
title={Scaling Representation Learning From Ubiquitous ECG With State-Space Models},
author={Kleanthis Avramidis and Dominika Kunc and Bartosz Perz and Kranti Adsul and Tiantian Feng and Przemysław Kazienko and Stanisław Saganowski and Shrikanth Narayanan},
year={2024},
journal={IEEE Journal of Biomedical and Health Informatics},
publisher={IEEE},
url={https://ieeexplore.ieee.org/abstract/document/10574308/}
}
@article{neural-codecs-as-biosignal-tokenizers,
title={Neural Codecs as Biosignal Tokenizers},
author={Kleanthis Avramidis and Tiantian Feng and Woojae Jeong and Jihwan Lee and Wenhui Cui and Richard M Leahy and Shrikanth Narayanan},
year={2025},
journal={arXiv preprint arXiv:2510.09095},
url={https://arxiv.org/abs/2510.09095}
}
@misc{avramidis2026braincodes,
title={BrainCodes: A Lightweight EEG Foundation Model over Discrete Neural Codes},
author={Kleanthis Avramidis and Ivan Tashev and Dimitra Emmanouilidou},
year={2026},
eprint={2610.29320},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2610.29320}
}
@inproceedings{a-multi-dimensional-framework-for-evaluating-generalization-in-eeg-foundation-models,
title={A Multi-Dimensional Framework for Evaluating Generalization in EEG Foundation Models},
author={Aditya Kommineni and Emily Zhou and Kleanthis Avramidis and Tiantian Feng and Shrikanth Narayanan},
year={2026},
booktitle={Conference on Health, Inference, and Learning (CHIL)},
url={https://arxiv.org/abs/2605.28563}
}
@article{aperiodic-and-low-frequency-spectral-bias-in-reconstruction-based-eeg-foundation-models,
title={Aperiodic and Low-Frequency Spectral Bias in Reconstruction-Based EEG Foundation Models},
author={Aditya Kommineni and Emily Zhou and Kleanthis Avramidis and Simon Bock Segaard and Jeppe Roden Münster and Andreas Peter Juhl Hansen and Takfarinas Medani and Tiantian Feng and Richard Leahy and Shrikanth Narayanan},
year={2026},
journal={arXiv preprint arXiv:2605.26434},
url={https://arxiv.org/abs/2605.26434}
}
@misc{avramidis2026grounding,
title={Probing Spatial and Spectral Grounding in EEG Foundation Models},
author={Kleanthis Avramidis and Ivan Tashev and Dimitra Emmanouilidou},
year={2026},
eprint={2610.29367},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2610.29367}
}