Diverse by design
Combine conventional contact PPG with remote PPG reconstructed from video and data from multiple devices and populations.
ARIS postdoctoral project · 2026–2028
Transparent intelligence for the pulse wave.
LANTERN-PPG is developing a general-purpose, explainable AI foundation model for photoplethysmography—the optical pulse signal captured by wearables, pulse oximeters, and cameras.
The research challenge
One signal.
Many contexts.
PPG makes continuous cardiovascular and psychophysiological monitoring accessible, but models trained for a single dataset or task often struggle across people, sensors, and real-world conditions. LANTERN-PPG aims to replace isolated models with a reusable, transparent backbone.
Learn without abundant labels
Use self-supervised learning to capture waveform morphology and physiological structure from large, mostly unlabelled collections.
Explain what matters
Connect model decisions to meaningful parts of the pulse wave through attention, gradient, and reconstruction-based explanations.
System architecture
From diverse signals to reusable insight.
The pipeline standardizes and quality-indexes diverse PPG sources, pre-trains a hybrid CNN–Transformer with complementary self-supervised objectives, and adapts the learned encoder to downstream tasks with explanation methods integrated throughout.
Research programme
A connected
five-part plan.
Each stage produces the foundation for the next, while explainability and dissemination connect the technical work to scientific and practical use.
24months
2026–2028
2026–2028
01
Months 1–5
Build the data foundation
Integrate diverse contact PPG and video-derived remote PPG, standardize signals, and annotate quality and artefacts.02
Months 6–18
Learn general representations
Train an artefact-aware hybrid CNN–Transformer using morphology-, subject-, and physiology-aware self-supervision.03
Months 12–18
Make the model explainable
Combine attention, gradient-guided attribution, and masked reconstruction to reveal influential waveform regions.04
Months 18–24
Evaluate across tasks
Benchmark transfer to physiological and mental-state tasks, including limited-label and out-of-distribution settings.05
Throughout
Share reusable outputs
Publish methods, benchmarking protocols, and research outputs following open-science and responsible-data principles.Expected outcomes
Built to transfer.
Designed to be trusted.
The project will assess whether a shared PPG backbone can generalize across sensors, populations, and tasks while requiring fewer labels and offering explanations aligned with meaningful waveform characteristics.
Potential downstream tasks
Heart rate & variabilityCuffless blood pressureStress detectionCognitive load
Reusable research
Dataset & benchmark
A curated, quality-aware data resource and a standard suite of tasks for comparable evaluation, where licensing and ethics permit.
Transparent AI
Model & explanations
A hybrid foundation encoder paired with interpretable maps that show which waveform regions contribute to its representations and predictions.
Project details
Research at the intersection of AI and physiological sensing.
LANTERN-PPG is an ARIS-funded postdoctoral research project hosted by the Department of Intelligent Systems at the Jožef Stefan Institute in Ljubljana, Slovenia.
- Project lead
- Dr. Gašper Slapničar
- Grant
- Z2-70066
- Period
- 1 Mar 2026 – 29 Feb 2028
- Field
- Computer science & informatics
Project documents
Resources
Project governance and communication documents are available below as downloadable PDFs.