Research scientist @GoogleAI and visiting academic @Cambridge_Uni. Multimodal foundation models × AI for health. Past @Nokia @BellLabs, @Microsoft, @Telefonica.
🦜PaPaGei: an open foundation models for the most widely used biosignal🦜
It is pre-trained on over 57,000 hours of PPG signals (like those from smartwatches & pulse oximeters) from publicly available sources.
Paper: arxiv.org/abs/2410.20542
Models/code: github.com/nokia-bell-labs/p…
we announce SensorFM, a foundation model pretrained on the largest personal health dataset (1T+ mins, 5 mil people, 100 countries)
- embeddings transfer across 35 tasks via systematic performance scaling
- agents search for predictive heads
- grounding LLMs for clinical answers
Introducing SensorFM, a large-scale Sensor Foundation Model that learns from 1 trillion-minutes of unlabeled wearable data drawn from five million consented participants.
SensorFM learns a single, reusable representation of sensed human physiology that transfers across cardiovascular, metabolic, sleep, and mental health, as well as lifestyle and demographic factors.
More →goo.gle/4ycJvot
Proud to be part of this work! We trained a foundation model on 40M hrs of wearable data where gaps from charging, motion artifacts etc. are inevitable. Rather than imputing or discarding, it uses a new attention masking mechanism.
Don't fight the messiness, embrace it!
Introducing LSM-2, our newest foundation model for wearable sensor data. LSM-2 uses Adaptive & Inherited Masking, a novel self-supervised framework, to learn from incomplete data & achieve strong performance without requiring explicit imputation. More → goo.gle/4kP6Ncc
If you are working on AI for health×timeseries please consider submitting to our NeurIPS 2025 workshop below. The team has prepared a great lineup of speakers and topics — stay tuned for updates!
🩺📈 The "Learning from Time-series for Health (TS4H)" workshop is BACK at NeurIPS 2025 🥳!
This workshop unites researchers across health time-series domains (from wearables to clinical systems) to tackle shared challenges. Details: timeseries4health.github.io/
🧵 (1/6)
I'll be at #ICLR2025 this week in Singapore and would love to connect with researchers interested in foundation models, representation learning, multimodal pretraining, & health sensing!
Feel free to catch me at any of these events or ping me here to schedule a time to chat.
Great news from Vancouver, happy to share that 🦜PaPaGei received the 🏆 Best Paper Award at the Time Series in the Age of Large Models (TSALM) workshop in #NeurIPS2024! Huge congratulations to @ArvindPillai10 and the team. Stay tuned for the v2 of the paper with more results.
Some news: after 2.5 years at @BellLabs, I’m moving on. It was a truly rewarding experience, letting me experiment with the latest in AI during a seismic shift. Huge thanks to @raswak and the team.
I've now joined @GoogleAI where I'll be working on AI for Health and Fitness.
We'd love to hear from researchers and developers who try it out!
Your feedback will be invaluable in improving and expanding its capabilities.
Tremendous work by @ArvindPillai10 and team @malekz4deh@raswak!
PaPaGei includes extensive testing for data-efficiency and robustness to biases, including analysis across diverse skin tones, setting a standard for future models.
In tests across 20 health tasks, PaPaGei outperformed other timeseries foundation models and contrastive methods in analyzing PPG signals for cardiovascular health, sleep disorders, and (a lot!) more.
PaPaGei's unique architecture learns detailed representations from PPG morphology and is designed to specifically pay attention to these morphological features. We use contrastive learning and multi-tasking to point the model to these important parts of the signal waveform.
Exercising regularly can be hard and there's often conflicting information on the best ways to measure progress and fitness levels. But one stat that's increasingly seen as important to track is 'VO2 max' and Bloomberg Prognosis explains why.... trib.al/Tswcryv
Check out this recent article on the tokenization problem - a quite accessible read for anyone interested in how LLMs represent inputs. Happy to see our work featured alongside cutting-edge research 😊
Additionally, Google's recent Personal Health LLM resorted to encoding signals with a separate adapter and providing them as a prefix to the LLM, further confirming the challenges we identified.
Today on the blog, read about the latest from our two new research papers on how AI, particularly fine-tuned Gemini models, can create personalized health experiences that cater to individuals’ unique health journeys. →goo.gle/3RnwHbl#AI#healthcare#personalizedhealth
Since our preprint in Sep 2023, multiple studies have empirically tested some of the issues we raised. For instance, UW's study below highlighted limitations in reasoning with timeseries data, possibly hinting at input representation issues among others.