EASELAN: Biosignal Annotation Framework
- EASELAN is an open-source framework for multimodal biosignal annotation that extends ELAN with integrated preprocessing, Git-based collaboration, CI/CD validation, and dashboard annotation.
- It structures workflows into three components—pre-processing, human-in-the-loop annotation, and post-processing—to support synchronized audio, video, and sensor data.
- The framework underpins robust provenance tracking and semantic export, enabling interdisciplinary research in adaptive cognitive systems and machine learning applications.
EASELAN is an open-source annotation framework for multimodal biosignal annotation and data management that extends the ELAN environment with pre-processing, Git-based collaboration, CI/CD validation, dashboard-based human-in-the-loop annotation, and post-processing/export facilities for analysis, machine learning, and semantic reasoning (Rammohan et al., 17 Oct 2025). In the provided literature, the name also has a terminological ambiguity: the paper on automated longitudinal network annotation formally names its method Expert-Augmented LLM Annotation (EALA) rather than EASELAN (Liu et al., 3 Mar 2025), and the paper on bounded-support scalar annotation formally names its method EASL rather than EASELAN (Sakaguchi et al., 2018). Accordingly, the term “EASELAN Annotation Framework” most precisely denotes the ELAN-based multimodal biosignal framework introduced in 2025, while related works use similar strings only as variant or clarifying labels rather than as the paper’s formal framework name (Rammohan et al., 17 Oct 2025).
1. Terminology, definition, and scope
The formal EASELAN framework is introduced as “EASELAN: An Open-Source Framework for Multimodal Biosignal Annotation and Data Management” (Rammohan et al., 17 Oct 2025). It is motivated by the increasing need for richly annotated multimodal datasets in machine learning and adaptive cognitive systems, especially where fusion models combine audiovisual channels with biosignals. The framework is positioned as a biosignal-centered extension of ELAN rather than a replacement for ELAN itself (Rammohan et al., 17 Oct 2025).
Its stated purpose is end-to-end support for annotation workflows spanning preparation of annotation files, setup of additional channels, integrated version control with GitHub, and simplified post-processing. The framework explicitly targets complex human-behavior recordings involving audio, video, text, and biosignals from eyes, body, skin, muscles, and brain, and it is presented as suitable for interdisciplinary teams and distributed research infrastructures such as LabLinking (Rammohan et al., 17 Oct 2025).
The literature provided also records two adjacent naming clarifications. First, the climate-negotiation annotation paper states that its framework is named Expert-Augmented LLM Annotation (EALA) and that “EASELAN” does not appear in that paper; if the two are linked, EASELAN is only a synonym or variant of EALA in that context (Liu et al., 3 Mar 2025). Second, the scalar-label annotation paper formally names its method EASL, pronounced “easel,” while the supplied notes use “EASELAN” only as a referential variant emphasizing bounded support and online operation (Sakaguchi et al., 2018). This suggests that, for precise scholarly usage, EASELAN should be distinguished from EALA and EASL even when secondary summaries juxtapose them.
2. Core architecture and workflow
EASELAN’s architecture is organized into three tightly integrated components: a pre-processing component (“Box A”), a human-in-the-loop annotation component (“Box B”), and a post-processing component (“Box C”) (Rammohan et al., 17 Oct 2025). This division gives the framework an explicitly pipeline-oriented structure, with each stage tied to concrete tools and file transformations.
In the pre-processing stage, EASELAN ingests synchronized multimodal recordings including audio, video, and biosignals such as EEG, ECG, EMG, and IMU. Audio and video are converted with FFMPEG according to ELAN’s recommended specifications. Speech transcription is performed with Whisper using word-level timestamps (word_timestamps = TRUE), producing text, word segments, and detected language. High-dimensional biosignals may optionally be converted into CSV for downstream interoperability. EASELAN adopts ELAN’s EUDICO Annotation Format (.eaf) and integrates biosignals through ELAN’s secondary data functionality, with Python tooling via Pympi automating preparation (Rammohan et al., 17 Oct 2025).
The annotation stage centers on versioned ELAN documents, CI/CD-mediated quality control, and a dedicated dashboard. GitHub synchronization is implemented through pygit2, with repository URLs, author credentials, and branch names managed in configuration files. CI/CD jobs are defined in YAML and executed by runners to validate annotations and transcriptions. Reported checks include spell-checking transcripts, verifying annotations against predefined ELAN templates, checking for empty tiers, and enforcing controlled vocabularies. Automated feedback is published as HTML on GitHub Pages (Rammohan et al., 17 Oct 2025).
The EASELAN Dashboard provides a multitier “partitur” view that aligns multiple video perspectives, Whisper transcripts, and time-aligned biosignals within ELAN. The interface supports synchronized playback across tiers, and the paper describes example views containing multiple camera streams, think-aloud transcripts, and tiered biosignals such as raw EEG, ICA components for artifact removal, acceleration, speech, and hierarchical action/object/motion annotations (Rammohan et al., 17 Oct 2025).
In post-processing, EASELAN generates HTML pages via mkdocs from the repository to provide project overviews and validation summaries. These reports include spell-check results and template conformance checks and can be exported as CSV for further analysis, for example through pandas. The framework also supports semantic export to Narrative Enabled Episodic Memories (NEEMs) and related knowledge bases (Rammohan et al., 17 Oct 2025).
3. Data model, schemas, modalities, and synchronization
EASELAN inherits ELAN’s tier-based annotation model. It supports independent and linked tiers, tier types, and constraints stored within .eaf, while controlled vocabularies in .ecv are used to ensure label consistency and may be autogenerated from domain ontologies (Rammohan et al., 17 Oct 2025). Secondary data, including biosignals and dashboard configurations, are handled through _tsconf.xml and .pfsx files. Because these files encode GUI state, they are not version-controlled; instead, they are reconstituted during post-processing to keep repositories clean (Rammohan et al., 17 Oct 2025).
The framework supports multimodal inputs and outputs across several layers:
| Category | Formats / modalities | Notes |
|---|---|---|
| Primary annotation | .eaf, .ecv |
ELAN-based tier model |
| Secondary data | _tsconf.xml, .pfsx, CSV |
Biosignals and GUI configuration |
| Media and speech | Audio, video, Whisper transcripts | Multi-camera RGB and timestamped ASR |
| Semantic/analysis export | .rfd, HTML, CSV |
Knowledge bases, reports, analytics |
The explicitly supported modalities include audio and video, speech via Whisper transcripts, and biosignals such as EEG, EMG, ECG, EDA, IMU or acceleration, and eye-tracking. In LabLinking contexts, the framework can visualize fMRI alongside other tiers. Interoperability is organized around ELAN-compatible media, CSV-based biosignal conversion, HTML and CSV reports, and .rfd semantic export targeting ontology-backed knowledge bases such as OpenEASE (Rammohan et al., 17 Oct 2025).
Time alignment is operational rather than formula-driven. In the EASE Table Setting Database, all biosignals are synchronized and recorded using the Lab Streaming Layer (LSL), which provides precise multimodal time coordination across distributed devices. EASELAN then uses ELAN’s common time axis to display annotations, transcripts, and secondary data together. The paper explicitly states that it does not provide mathematical formulas for time mapping, resampling, or drift correction, and it does not introduce quantitative alignment metrics. Instead, synchronization depends on LSL-based capture, ELAN’s time-aligned display model, and quality-control checks such as empty-tier detection and controlled-vocabulary conformance (Rammohan et al., 17 Oct 2025).
This design choice places acquisition-level synchronization and signal sanitation outside the core annotation framework. Artifact removal and de-noising are assumed to be addressed in the acquisition stack or in prior preprocessing, although EASELAN provides hooks for integrating custom scripts (Rammohan et al., 17 Oct 2025). A plausible implication is that the framework prioritizes annotation provenance and structural consistency over algorithmic signal-processing innovation.
4. Collaboration, provenance, validation, and semantic export
A defining property of EASELAN is its use of Git-based provenance and CI/CD automation as first-class annotation primitives. GitHub integration through pygit2 allows branch-based work, commits, pushes, and traceable history on .eaf annotations, transcripts, and metadata. The paper emphasizes diffs and provenance as audit mechanisms showing who changed which files and when (Rammohan et al., 17 Oct 2025).
Validation is formalized through CI/CD pipelines configured in YAML. These jobs concurrently verify annotations, enforce templates, check controlled vocabularies, and run spell-checkers, with logs visible in the pipeline UI. Automated feedback is compiled into HTML reports and published via GitHub Pages. The framework therefore connects local annotation activity to centralized, continuously updated quality-control artifacts (Rammohan et al., 17 Oct 2025).
EASELAN also supports semantic interoperability through ontological grounding. For export to NEEMs and semantic querying, it connects to the SOMA ontology for everyday activities. SOMA concepts are used to generate .ecv controlled vocabularies for ELAN, and repositories can be exported to .rfd files. These RDF-like outputs can be loaded into semantic databases such as Apache Jena or OpenEASE or accessed through libraries such as rdflib (Rammohan et al., 17 Oct 2025).
The paper frames this export path as enabling queries over activities, objects, locations, and agent movements, and as supporting cross-experiment alignment. In the reported downstream usage, EASE-TSD annotations exported through this mechanism contributed to pipelines including 3D object tracking, gait parameter estimation, and EEG-based attention detection in human–robot interaction (Rammohan et al., 17 Oct 2025). This suggests that EASELAN is not only a front-end annotation environment but also a bridge from human-produced multimodal labels to symbolic and hybrid reasoning infrastructures.
5. Table Setting Database case study
The main case study is the EASE Table Setting Dataset (EASE-TSD), collected to understand human everyday activities in a table-setting scenario for cognitive robots (Rammohan et al., 17 Oct 2025). Data were acquired in controlled laboratory conditions and integrated into EASELAN for annotation, validation, and dissemination within the DFG Collaborative Research Center 1320 “Everyday Activities Science and Engineering (EASE)” (Rammohan et al., 17 Oct 2025).
The dataset has substantial scale. It includes recordings from 100 participants, with data from 78 sessions available; each session contains six table-setting trials. The setup comprises eight synchronized biosignal streams captured by 22 sensors, including marker-based motion capture, 7 RGB cameras, eye-tracking, 16-channel EMG on both hands, EDA, acceleration via IMU, 2 microphone channels, and 16-channel EEG. All biosignals were synchronized using LSL, and the collection contains more than 430 hours of sensor data in total (Rammohan et al., 17 Oct 2025).
The annotation scheme is explicitly multi-level. Participants produced concurrent and retrospective think-aloud protocols, transcribed with Whisper and cross-checked manually. Within ELAN, the annotation tiers comprise three levels: phase, action, and motion, subdivided into hands, head, and body, together with object tiers per hand and additional think-aloud codes. EASELAN supported this workflow through Git-based synchronization, a GitLab Issue system for task prioritization, automated verification against ELAN tier templates, controlled-vocabulary checks, and spell-checking (Rammohan et al., 17 Oct 2025). The paper reports that more than 260 hours of EASE-TSD data are annotated (Rammohan et al., 17 Oct 2025).
The verification dashboard displays per-file tier presence or absence using checkmarks and red Xs, and post-processing creates per-run “partitur” files that merge annotation tiers and biosignals for structured overview. Dissemination proceeds through HTML and CSV summaries, and the finalized dataset is made available for research purposes through OSF (Rammohan et al., 17 Oct 2025).
Notably, the paper states that it does not report formal inter-annotator agreement metrics for EASE-TSD, such as Cohen’s or F1, and it does not provide LaTeX formulas for such metrics (Rammohan et al., 17 Oct 2025). This absence is consequential for methodological interpretation: the framework strongly emphasizes process reproducibility, validation automation, and semantic export, but the case study does not quantify annotation reliability in the conventional coder-agreement sense.
6. Positioning, limitations, and relationship to similarly named frameworks
EASELAN is positioned against VIA, CVAT, WebAnno, BORIS, ANVIL, and ELAN itself (Rammohan et al., 17 Oct 2025). The paper’s central comparative claim is not that these tools are generically inadequate, but that they do not jointly provide direct biosignal support, integrated version-controlled collaboration, CI/CD-backed validation, and the biosignal-centric dashboard architecture delivered by EASELAN. ELAN is treated as the backbone being extended rather than a competitor (Rammohan et al., 17 Oct 2025).
Several limitations and future directions are stated explicitly. Although EASELAN supports semi-automated, human-in-the-loop workflows, the integration of more accurate AI-assisted event detection is left for future improvement. Continuous training and validation loops for machine-learning pipelines are currently limited, and planned extensions include direct export to frameworks such as TensorFlow and PyTorch and better compatibility with external databases (Rammohan et al., 17 Oct 2025). The framework was also designed to address data heterogeneity, distributed annotation and provenance, and privacy concerns in multimodal fusion, but acquisition-level synchronization and preprocessing remain upstream responsibilities (Rammohan et al., 17 Oct 2025).
The broader literature provided makes the terminological distinction especially important. The climate-negotiation system in “Automated Annotation of Evolving Corpora for Augmenting Longitudinal Network Data: A Framework Integrating LLMs and Expert Knowledge” is formally EALA, a framework for annotating evolving text corpora into longitudinal network quadruplets
using codebooks, historically annotated data, and LLMs (Liu et al., 3 Mar 2025). By contrast, the bounded-support scalar annotation method in “Efficient Online Scalar Annotation with Bounded Support” is formally EASL, which models items with Beta posteriors and updates them through scalar observations,
while using uncertainty and match quality for online selection (Sakaguchi et al., 2018). These are substantively different frameworks with different objects of annotation, statistical assumptions, and output structures.
A plausible implication is that “EASELAN” functions in the supplied material as a naming collision across at least three annotation traditions: multimodal biosignal annotation grounded in ELAN (Rammohan et al., 17 Oct 2025), expert-augmented LLM annotation for longitudinal relational data (Liu et al., 3 Mar 2025), and efficient online scalar annotation with bounded support (Sakaguchi et al., 2018). For precise scholarly reference, the 2025 multimodal framework is the one whose formal title actually includes EASELAN (Rammohan et al., 17 Oct 2025).
