Foundation model

Wearable Foundation Model (WFM / WEAR; Ramamoorthy)

project-page sensor-language FM

PPG IMU EDA Temp HR HRV Text Project Page Only Weights not released

PreprintCatalog paper date: 2026Added to atlas: Nov 10, 2026

Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.

Early research project by Santhakumar Ramamoorthy (doctoral student at RMIT University on his public profile — the project page itself does not state an affiliation). There is no paper, preprint, DOI, arXiv id, public code or weights as of 2026-11-10. Numbers below are taken from the project page and its architecture explorer, not from a peer-reviewed manuscript.

Parameters: 9.5M
Signals: PPG, IMU, EDA, Temp, HR, HRV, Text
Year: 2026

Architecture: WEAR (Wearable Evidence-Aware Reasoning). Built WFM encoder/decoders: 10-minute window of 18 derived channels × 60 steps (0.1 Hz); 9 channel-pairs × 10 one-minute patches = 90 tokens × 256-d; Sensor Transformer is 4 layers, 4 heads, width 256, with learned pooling for a match embedding; text decoder 3 causal layers width 256; multimodal decoder 3 layers with cross-attention to the 90 sensor tokens + tied LM head; forecast head is a 512-unit layer over the 256-d pooled sensor vector plus a 36-d wrist summary (18 channel means + 18 observed flags), producing low/mid/high pinball forecasts for heart rate, HRV, movement and glucose at two horizons. Training losses already built: symmetric InfoNCE match, next-token caption CE, pinball forecast. Architecture page marks the WEAR evidence store, exact tools, memory, question encoder, reasoning core (planned 4-layer width-256 Transformer), claim verifier, claim bank and composed outputs as 'Not built yet'; event heads 'in progress'.
Pretraining data: Author-claimed 3.0 million hours prepared from 12 publicly available research datasets, 25,000+ people, split by person. The Datasets section lists 16 corpora plus the PhysioNet umbrella citation; which 12 make up the 3.0 million hours is not stated.
Form factor: wrist-derived 10 s features (architecture demo); listed sources also include ICU bedside waveforms (MIMIC)
License: None public. Project page: 'Code not public yet.' 'No raw data or model weights are shared.'

At a glance

Type: Foundation model
Modalities: PPG IMU EDA Temp HR HRV Text
Form factor: Wrist derived 10 s features (architecture demo); listed sources also include ICU bedside waveforms (MIMIC)
Parameters: 9.5M
Architecture: WEAR (Wearable Evidence-Aware Reasoning). Built WFM encoder/decoders: 10-minute window of 18 derived channels × 60 steps (0.1 Hz); 9 channel-pairs × 10 one-minute patches = 90 tokens × 256-d; Sensor Transformer is 4 layers, 4 heads, width 256, with learned pooling for a match embedding; text decoder 3 causal layers width 256; multimodal decoder 3 layers with cross-attention to the 90 sensor tokens + tied LM head; forecast head is a 512-unit layer over the 256-d pooled sensor vector plus a 36-d wrist summary (18 channel means + 18 observed flags), producing low/mid/high pinball forecasts for heart rate, HRV, movement and glucose at two horizons. Training losses already built: symmetric InfoNCE match, next-token caption CE, pinball forecast. Architecture page marks the WEAR evidence store, exact tools, memory, question encoder, reasoning core (planned 4-layer width-256 Transformer), claim verifier, claim bank and composed outputs as 'Not built yet'; event heads 'in progress'.
Venue: Project page only; no preprint or venue found as of 2026-11-10
Year: 2026
arXiv: Not reported
DOI: Not reported
Code: Not released
Weights: Not released
Licence: None public. Project page: 'Code not public yet.' 'No raw data or model weights are shared.'
Caveats: Not reported
Datasets: PPG-DaLiA
CAPTURE-24
WESAD
+11 more datasets
Devices: ActiGraph GT3X+
Axivity AX3
Empatica E4
+6 more devices
Added: Nov 10, 2026
Updated: Nov 10, 2026

Wearable Foundation Model (WFM / WEAR; Ramamoorthy): ChatPPG research summary card

How it works

Datasets linked in the hub; pretraining vs evaluation roles follow the paper.

Pretraining data: Author-claimed 3.0 million hours prepared from 12 publicly available research datasets, 25,000+ people, split by person. The Datasets section lists 16 corpora plus the PhysioNet umbrella citation; which 12 make up the 3.0 million hours is not stated.
Architecture: WEAR (Wearable Evidence-Aware Reasoning). Built WFM encoder/decoders: 10-minute window of 18 derived channels × 60 steps (0.1 Hz); 9 channel-pairs × 10 one-minute patches = 90 tokens × 256-d; Sensor Transformer is 4 layers, 4 heads, width 256, with learned pooling for a match embedding; text decoder 3 causal layers width 256; multimodal decoder 3 layers with cross-attention to the 90 sensor tokens + tied LM head; forecast head is a 512-unit layer over the 256-d pooled sensor vector plus a 36-d wrist summary (18 channel means + 18 observed flags), producing low/mid/high pinball forecasts for heart rate, HRV, movement and glucose at two horizons. Training losses already built: symmetric InfoNCE match, next-token caption CE, pinball forecast. Architecture page marks the WEAR evidence store, exact tools, memory, question encoder, reasoning core (planned 4-layer width-256 Transformer), claim verifier, claim bank and composed outputs as 'Not built yet'; event heads 'in progress'.
Inputs: PPG, IMU, EDA, Temp, HR, HRV, Text
Approach: sensor language · project page · no code · no weights · quality gating · RMIT

Devices· 9

Photo: Open Movement, Newcastle University / project docs, CC BY 3.0

Empatica E4

Empatica · Wrist

PPGIMUEDATemp

Product image: Empatica

Fitbit Sense

Google / Fitbit · Wrist

PPGIMUEDATemp

Fitbit Versa 2

Fitbit (Google) · Wrist

PPGIMU

Fitbit wrist devices (PPG irregular-rhythm family)

Fitbit (Google) · Wrist

Photo: OMG Plc / Wikimedia Commons, CC BY 3.0

RespiBAN chest band

BiosignalsPlux / RespiBAN-class · Chest

ECGIMURespTemp

Shimmer3 GSR+

Shimmer Sensing · Wrist

PPGIMUEDA

Product image: Shimmer Sensing

Full photo credits and licences are on each device page. An icon means we have no openly licensed or manufacturer-cleared photo of that exact device yet.

Datasets used· 14

Show 2 more datasets

Cite this model

APA 7BibTeX

Cite the publication recorded for this model. Check it against the publisher before submitting.

Wearable Foundation Model (WFM / WEAR; Ramamoorthy). (2026). Project page only.

@article{wfmramamoorthy2026wearable,
  title         = {{Wearable Foundation Model (WFM / WEAR; Ramamoorthy)}},
  year          = {2026},
  journal       = {Project page only}
}

History

Nov 10, 2026
Added to the ChatPPG research hub

Newsletter

Get new PPG datasets, papers and models

Get alerts when Wearable Foundation Model (WFM / WEAR; Ramamoorthy) is benchmarked on new datasets, plus the weekly digest.

New datasets, papers and models
Practical guides from the hub
Unsubscribe anytime

The waveform is an illustration of a PPG pulse, not a recording.

Foundation model

Wearable Foundation Model (WFM / WEAR; Ramamoorthy)

Properties

Type: Foundation model
Modalities: PPG IMU EDA Temp HR HRV Text
Form factor: Wrist derived 10 s features (architecture demo); listed sources also include ICU bedside waveforms (MIMIC)
Parameters: 9.5M
Architecture: WEAR (Wearable Evidence-Aware Reasoning). Built WFM encoder/decoders: 10-minute window of 18 derived channels × 60 steps (0.1 Hz); 9 channel-pairs × 10 one-minute patches = 90 tokens × 256-d; Sensor Transformer is 4 layers, 4 heads, width 256, with learned pooling for a match embedding; text decoder 3 causal layers width 256; multimodal decoder 3 layers with cross-attention to the 90 sensor tokens + tied LM head; forecast head is a 512-unit layer over the 256-d pooled sensor vector plus a 36-d wrist summary (18 channel means + 18 observed flags), producing low/mid/high pinball forecasts for heart rate, HRV, movement and glucose at two horizons. Training losses already built: symmetric InfoNCE match, next-token caption CE, pinball forecast. Architecture page marks the WEAR evidence store, exact tools, memory, question encoder, reasoning core (planned 4-layer width-256 Transformer), claim verifier, claim bank and composed outputs as 'Not built yet'; event heads 'in progress'.
Venue: Project page only; no preprint or venue found as of 2026-11-10
Year: 2026
arXiv: Not reported
DOI: Not reported
Code: Not released
Weights: Not released
Licence: None public. Project page: 'Code not public yet.' 'No raw data or model weights are shared.'
Caveats: Not reported
Datasets: PPG-DaLiA
CAPTURE-24
WESAD
+11 more datasets
Devices: ActiGraph GT3X+
Axivity AX3
Empatica E4
+6 more devices
Added: Nov 10, 2026
Updated: Nov 10, 2026

Papers

Models

RAE-PPGAlso uses PPG-DaLiAFoundation modelPPG2026