software

Two of these libraries are used outside my own group.

PhysioEx — physiological signal explainer

A PyTorch library that puts 12 clinical cohorts behind one declarative interface, so that training a state-of-the-art model on physiological signals is a command and the research effort goes into explaining it. It fixes subject-wise splits and seeds, ships staging architectures and pretrained encoders, and carries an explainability subsystem. Roughly 46,500 lines across 207 modules, with continuous integration on two Python versions and automated releases to PyPI. Described in Physiological Measurement (2025).

Models distributed through PhysioEx were used as external-validation baselines by Thapa et al., Nature Medicine (2026).

Docs · PyPI · GitHub · Paper

ProtoSleepNet — prototype-based interpretable sleep staging

Reading the prototypes is reading the decision: the model classifies a sleep epoch by matching it against learned micro-structural patterns, and those patterns are the explanation. Trained and evaluated across 11 polysomnography datasets and 12,317 subjects. MIT licensed and built on PhysioEx; the pretrained weights are released on Hugging Face together with the paper.

The demo below is the fastest way to see what the model is doing — pick an epoch, see its nearest prototype, and see the time–frequency evidence behind the match.

Live demo · Docs · GitHub · Preprint (v2) · Research Square (v1)

Spectral Gradients — disentangled time–frequency attributions (manuscript in preparation)

An attribution method that separates a network’s evidence in time from its evidence in frequency, so that an explanation over a physiological signal can be read as “this rhythm, at this moment” rather than as an undifferentiated heatmap. It combines progressive band ablation in the frequency domain with path-integrated gradients, and is benchmarked against nine STFT-explainer configurations across synthetic, audio, arrhythmia and sleep data. Manuscript in preparation with Prof. W. Samek (Fraunhofer HHI).

Code to be released with the paper.

Also in development

Not yet public — these are released as the corresponding papers appear.

EEGBenchmarks — an extension-oriented benchmark harness for EEG: datasets, wrappers for foundation and supervised models, downstream evaluation tasks, and preprocessing and reporting workflows.

agentic-aasm-staging — neuro-symbolic explainable sleep staging: retrieval over the AASM manual selects the rule, a deterministic engine decides the stage, and a local language model writes the justification. A demonstrator for auditable clinical reasoning, not a state-of-the-art classifier.

eeg-signal2text-survey — a survey and empirical probe of signal-to-text models applied to single EEG epochs, across four adaptation strategies. The finding is negative and worth stating: no open-weight model is credibly zero-shot on EEG.