Abstract
Multimodal learning analytics (MMLA) has shown that integrating heterogeneous data yields a fuller picture of learning, yet most studies remain confined to laboratory settings with fewer than fifty participants, rely on research-grade sensors costing thousands of dollars, and produce one-off prototypes that never reach real classrooms. Meanwhile, generative AI has introduced a new kind of learning data — the dialogue trace, which records not just what a student clicked but how a student reasons — that existing MMLA frameworks have yet to incorporate. We present Educational Omics, a six-dimensional framework that borrows the integrative paradigm of the life-science "omics" and organizes the learning experience into six analytically independent dimensions. The framework is operationalized through Uedu, a production platform deployed across eighteen universities over four semesters, demonstrating that sustainable, classroom-scale multimodal analytics is achievable with consumer-grade devices and a modular, multi-tenant architecture.
Problem & Motivation
MMLA has proven that combining cognitive, linguistic, physiological, social, and environmental signals produces a more complete understanding of learning than any single stream. In practice, however, the evidence base is dominated by small-sample laboratory studies using expensive research-grade instrumentation, and the resulting systems rarely survive past the study that produced them. At the same time, generative-AI tutoring has made the dialogue trace a routine by-product of learning — a record of reasoning, not just behavior — yet no established MMLA framework treats it as a first-class modality. The open question is how to integrate cognitive, linguistic, physiological, social, environmental, and ethical data in real, sustainable classrooms rather than in the lab.
Method
We propose Educational Omics, which organizes the learning experience into six analytically independent dimensions: Cognomics (cognitive processes captured through LLM dialogue and automated Bloom's-taxonomy classification), Linguomics (language expression via speech-to-text and semantic embeddings), PhysioNeuromics (physiological signals from consumer-grade Garmin wearables), Sociomics (social interaction through forums and collaborative tools), Environomics (the physical learning environment via IoT sensors), and Ethicomics (ethical governance, treated as a first-class dimension through consent management and privacy-by-design). The framework is not merely conceptual: it is operationalized through Uedu, a production-grade platform built with Python Flask and MySQL from sixty-seven modular blueprints, whose subdomain multi-tenancy lets a new institution join through configuration alone, with no code changes. Uedu has been deployed across eighteen universities over four semesters.
Findings
- Higher-order cognition in the majority: automated Bloom classification over 136,000+ student messages found higher-order activity (apply, analyze, evaluate, create) accounting for 50.2% of messages.
- Disciplinary variation: the higher-order proportion ranged from 51.3% in business to 77.6% in medicine and life sciences — to our knowledge the first disciplinary-stratified Bloom baselines in AI-assisted learning, letting instructors reflect on their teaching design against disciplinary norms.
- Trustworthy automation: at the binary higher-/lower-order level, LLM–human agreement (κ ≈ .40) was comparable to agreement among human annotators, indicating that automated classification is as reliable as trained human raters at the consequential granularity.
- Physiologically-aware dialogue: PALM integrates wearable data from 41 users over 3,000+ person-days, adapting its dialogue strategy to each learner's physiological baseline rather than surfacing yet another analytics dashboard.
Implications
The value of Educational Omics is structural: it lets a study state explicitly which dimensions it does and does not cover — a transparency that ad-hoc modality selection cannot provide — and makes cross-dimensional hypotheses, such as whether stress pushes students toward lower Bloom levels, available to be posed and discussed. Critically, placing Ethicomics on equal footing with the other dimensions ensures that multimodal sensing is not surveillance: each dimension can be opted into or out of independently, consent is per-dimension and revocable at any time, and physiological data is never exposed directly to instructors. The framework makes each dimension usable, not mandatory. Physiological signals are reported as correlates of learning states, not as causal or affective claims.
Citation
BibTeX
@inproceedings{chang2026uedu_eo_platform,
author = {Chia-Kai Chang and Kuei-Hao Li},
title = {Uedu: A Six-Dimensional Educational Omics Platform Bridging AI, Physiological Sensing, and Classroom Practice},
booktitle = {Proceedings of the 21st European Conference on Technology Enhanced Learning (EC-TEL 2026), Industry and Practitioner Track},
address = {Valencia, Spain},
publisher = {Springer},
series = {Lecture Notes in Computer Science},
year = {2026},
note = {in press},
}