Biosignals

Analyzing temporal spike patterns in C-fibers recorded via microneurography is challenging due to the use of a single recording electrode, waveform variability, and high similarity of spike shapes across neurons, limiting the interpretation of sensory coding, such as pain and itch. We present a computational pipeline combining peak detection and supervised classification for spike sorting to improve the analysis of discharges, identified through activity-dependent conduction velocity changes. In the knowledge-driven step, we extract spike templates from electrically evoked spikes obtained during low-frequency stimulation and focus on the “best” template as the fiber of interest. Spike detection is further restricted to intervals showing activity-dependent latency shifts, substantially reducing the search space compared to unsupervised clustering. In the data-driven steps, we systematically evaluate three feature sets and machine learning models: One-class SVM, SVM, and XGBoost. For the evaluation, we created a specialized stimulation protocol, providing reliable ground truth labels for all electrically evoked spikes, allowing precise spike time-locking. Compared to Spike2 software, our approach achieved higher F1-scores and reduced false positives, indicating improved spike sorting. Although XGBoost achieved the highest median F1-scores, optimal performance was dependent on individual combinations of feature sets and models for each recording. In some recordings with many nerve fibers and a low signal-to-noise ratio, reliable sorting was not feasible. This highlights the necessity to determine sortability and optimal configurations for individual recordings. To illustrate the potential of our approach to sensory spike train analysis, we present a proof-of-concept application of the pipeline to chemically induced C-fiber activity. These findings represent an important step toward reliable analysis of activity associated with pain and itch signaling.

Underreporting of seizures, particularly focal onset impaired awareness seizures (FIAS), compromises the effectiveness of patient care and condition management in patients with epilepsy. Traditional reliance on patient self-reporting can lead to inaccuracies, hindering effective treatment. Wearable-based seizure detection algorithms offer a promising solution, however, developing an efficient method for detecting FIAS remains a challenge. Additionally, as data quality can vary in wearable settings, the absence of continuous data quality assessment poses a concern for the reliability of such algorithms

The objective of our study is to develop and evaluate the performance and feasibility of FIAS detection algorithm with automatic data quality assessment (ADQA) using a wearable electrocardiography (ECG) device. We will also conduct an exploratory analysis of inter-individual variability in autonomic seizure signatures to identify potential future candidates, or “responders” to this system. Performance will be evaluated using sensitivity, false alarm rate per 24 h (FAR/24), positive predictive value, and F1-Scor

A multicenter study was conducted across three epilepsy centers and recruited patients of all ages who were admitted to video-EEG monitoring for a minimum of 24 h consecutively. Data were collected using a wearable ECG device. The algorithm involved R-peak detection to identify heartbeats, extraction of knowledge domain heart rate variability features, ADQA, heart rate (HR) filter to address class imbalance, and a deep learning model for the final detection step. The algorithm was validated in a leave-one-patient-out (LOPO) approach using expert-labeled ictal events from video-EEG monitoring as ground truth.

A total of 236 patients were recruited, of whom 49 patients experienced at least one FIAS, resulting in 3278 h of ECG data and 260 seizures. Two patients with 33 seizures were excluded due to a technical error in the recording files, leaving 47 patients for analysis. After data quality screening, 161 seizures from 38 patients met the quality criteria. In this group, the median sensitivity was 66.6% (95% CI:33.3%-100%) with a median FAR/24 of 5.2 (95% CI:3.5-8.2). An exploratory responder analysis identified 20 patients with a detection sensitivity of ≥66.6%, for whom the median sensitivity was 100% (95% CI: 92%–100%) and the median FAR/24 was 4.3 (95% CI: 3–7). Finally, removing ADQA from the test data reduced the algorithm’s reliability, while removing it from training and test data reduced sensitivity, robustness, and reliability.

The proposed algorithm demonstrated reasonable performance in patients whose wearable ECG data met the ADQA quality criteria (n = 38), with the highest detection performance observed in an exploratory responder subgroup (n = 20). These findings highlight the potential of ECG-based wearable systems for improving FIAS monitoring and underscore the importance of data quality in ensuring reliable algorithm performance.

The Medical Information Mart for Intensive Care (MIMIC) waveform databases provide high-resolution physiological signals from intensive care units but remain underutilised compared to the relational MIMIC core databases. One barrier is the lack of accessible tools for transforming waveform data stored in the Waveform Database (WFDB) format into machine learning-ready datasets. This work introduces WavePrep, an open-source software framework that addresses this gap by enabling flexible, transparent, and efficient preprocessing of MIMIC waveform data. WavePrep allows users to define fine-grained, channel-specific preprocessing pipelines via a user-friendly JSON configuration, supporting common signal processing steps such as downsampling, data cleaning, imputation, windowing, and dataset splitting. The framework automatically detects dependencies between preprocessing steps, prevents data leakage through patient-level splitting, and exploits parallelisation to accelerate dataset creation. We evaluate WavePrep by creating multiple datasets derived from the MIMIC-III matched waveform database, including numeric and non-numeric signals. Experimental results on a multi-core server demonstrate robust and scalable performance, with processing times between 33.10 and 50.53 milliseconds per hour of recording. Datasets created using WavePrep were successfully applied in machine learning tasks. By lowering technical barriers and providing transparent preprocessing metadata, WavePrep facilitates broader adoption of clinical waveform data and improves the AI-readiness of signal datasets. GitHub: https://github.com/BI-K/WavePrep

Keywords intensive care, MIMIC, AI-ready, open source software, signal processing, biosignal

Sorting spikes from noisy single-channel in-vivo extracellular recordings is challenging, particularly due to the lack of ground truth data. Microneurography, an electrophysiological technique for studying peripheral sensory systems, employs experimental protocols that time-lock a subset of spikes. Stable propagation speed of nerve signals enables reliable sorting of these spikes. Leveraging this property, we established ground truth labels for data collected in two European laboratories and designed a proof-of-concept open-source pipeline to process data across diverse hardware and software systems. Using the labels derived from the time-locked spikes, we employed a supervised approach instead of the unsupervised methods typically used in spike sorting. We evaluated multiple low-dimensional representations of spikes and found that raw signal features outperformed more complex approaches, which are effective in brain recordings. However, the choice of the optimal features remained dataset-specific, influenced by the similarity of average spike shapes and the number of fibers contributing to the signal. Based on our findings, we recommend tailoring lightweight algorithms to individual recordings and assessing the “sortability feasibility” based on achieved accuracy and the research question before proceeding with sorting of non-time-locked spikes in future projects.

The increasing amount of open-access medical data provides new opportunities to gain clinically relevant information without recruiting new patients. We developed an open-source computational pipeline, that utilizes the publicly available electroencephalographic (EEG) data of the Temple University Hospital to identify EEG profiles associated with the usage of neuroactive medications. It facilitates access to the data and ensures consistency in data processing and analysis, thus reducing the risk of errors and creating comparable and reproducible results. Using this pipeline, we analyze the influence of common neuroactive medications on brain activity.

The pipeline is constructed using easily controlled modules. The user defines the medications of interest and comparison groups. The data is downloaded and preprocessed, spectral features are extracted, and statistical group comparison with visualization through a topographic EEG map is performed. The pipeline is adjustable to answer a variety of research questions. Here, the effects of carbamazepine and risperidone were statistically compared with control data and with other medications from the same classes (anticonvulsants and antipsychotics).

The comparison between carbamazepine and the control group showed an increase in absolute and relative power for delta and theta, and a decrease in relative power for alpha, beta, and gamma. Compared to antiseizure medications, carbamazepine showed an increase in alpha and theta for absolute powers, and for relative powers an increase in alpha and theta, and a decrease in gamma and delta. Risperidone compared with the control group showed a decrease in absolute and relative power for alpha and beta and an increase in theta for relative power. Compared to antipsychotic medications, risperidone showed a decrease in delta for absolute powers. These results show good agreement with state-of-the-art research. The database allows to create large groups for many different medications. Additionally, it provides a collection of records labeled as “normal” after expert assessment, which is convenient for the creation of control groups.

The pipeline allows fast testing of different hypotheses regarding links between medications and EEG spectrum through ecological usage of readily available data. It can be utilized to make informed decisions about the design of new clinical studies.

This work aims to improve FAIR-ness of the microneurography research by integrating the local (meta)data to existing research data infrastructures. In the previous work, we developed an odML based solution for local metadata storage of microneurography data. However, this solution is limited to a narrow community. As a next step, we propose the integration into the Local Data Hubs, data-sharing services within NFDI4Health infrastructure. We outline a first concept, that streams chosen data from the established odMLtables GUI.