I am a research scientist at Google and a visiting
academic at the University of Cambridge, focusing on learning robust
representations from noisy temporal data and building
foundation models. My work bridges the
gap between fundamental research and real-world application, developing models that are currently used by
millions of people worldwide.
Multimodal foundation models & agents: Building large-scale models and agents for
health data, including SensorFM
(pre-trained on 1T+ minutes of wearable data), LSM-2 (handling
incomplete multimodal data), PaPaGei (first open foundation
model for biosignals), WEQA (query-adaptive agentic reasoning),
and Time2Lang (bridging timeseries and LLMs).
Real-world product impact: I've led model training, evaluation, or post-training
improvements for some of Google Health's core physiological models. This includes the heart sensing and
activity recognition models powering the Pixel Watch and Fitbit Air
devices, as well as driving 'blood pressure' and 'insulin resistance trends' features via SensorFM
and other models.
My research has been published in top venues in artificial intelligence, AI for health, and human-centered
signal processing, while recent projects have been featured in international media such as the New York Times,
BBC, CNN, Guardian, Washington Post, Forbes, and Financial Times (see more in Press).
Analyzing the pitfalls of representing and tokenizing temporal data for Large
Language
Models, particularly from sources like mobile sensors or medical records.
We developed some of the first pre-trained models using self-supervised objectives
and applied them to various health-related
downstream tasks.
News
aug 2026
The recently announced Health Guardian tool includes 'blood pressure' and 'insulin resistance trends' and
is based
on our SensorFM
and other models. You can read more here.
jul 2026
We released SensorFM, a foundation model trained on the largest personal health dataset. You can read more
here.
We released the paper for Large Sensor Model 2 (LSM-2), a
foundation model trained on 40M hours of wearable data.
may 2025
We are organizing a new workshop EvalComp @
Ubicomp'25 focusing on the future of evals. Consider submitting your relevant work!
apr 2025
Time2Lang, a new method to use timeseries foundation models
with LLMs, was accepted at CHIL 2025. I also gave an invited talk at
Singapore Management University on foundation models for personal health.
feb 2025
Our work on model averaging for label noise mitigation was published in Scientific Reports.
Our 🦜 PaPaGei work received the best paper award at the
NeurIPS’24 workshop on Time
Series in the Age of Large Models. I also gave an invited lecture at the
Aristotle University of Thessaloniki on the topic of foundation models for personal health. Our SoundCollage paper was accepted at ICASSP'25.
nov 2024
I joined Google in London, working within the Consumer Health
Research team.
oct 2024
We released 🦜 PaPaGei, the first open foundation model for
biosignals (PPG). You can read more here.
I was also interviewed by Bloomberg for a
newsletter
about VO2max.
StatioCL, a new non-stationary self-supervised model for
timeseries, was accepted to CIKM 2024.
jul 2024
Our work on how Large Language Models struggle with temporal data was published in JAMIA, and was covered by
Techcrunch
and LG AI Research. You can read more in this post.
jun 2024
Our work on how Self-Supervised Learning improves fairness was accepted at KDD 2024. We released the paper,code, and a project website. I was also interviewed by Runner's World
magazine for a feature article about VO2max - you can read more here.
may 2024
My MedAI talk from earlier this year is now available on Youtube.
apr 2024
I was interviewed by the New York Times for an article on cardio
fitness and wearables. Also launched a new Short Papers section at IEEE Pervasive
journal - consider submitting your works! In addition, my first patent from a few years ago became
public; you can read more here.
Co-chaired the Human-Centric Representation Learning
workshop at AAAI 2024 in Vancouver, with a great set of papers and keynotes - you can read some
highlights of the day at AIhub.org.
I also gave an invited keynote at the
Health Intelligence workshop of the same conference (here are the slides from the talk).
jan 2024
Gave an invited talk at Cambridge Biomedical
Campus as part of the MedAI seminar series.
I have published over 60 papers in top-tier venues including NeurIPS, ICLR, KDD, Nature Digital Medicine,
UbiComp, and ICASSP. You can also see the full list on Google Scholar.
Martin Maritsch, Timo Stoffregen, Thomas Kaar, Behsad Riemer, Maxwell A. Xu, Max
Rosenblattl, Juncheng Liu, Nicolas Zumarraga, Yu Yvonne Wu, Denys Herasymuk, Sparsh Rastogi, Hyungjun Yoon,
Bosong Huang, Arvind Pillai, Dmytro Lopushanskyy, Tony Chen, Robin Deuber, Yichen Liu, Shvat Messica, Dan Li,
Jian Lou, Yuwei Zhang, Jaeho Kim, Renée Rosillo Garcia, Fan Wu, Robert Müller, Elgar Fleisch, Flora D. Salim,
Dimitris Spathis, Yuzhe Yang, Aaqib Saeed, Daniel McDuff, Ming Jin, Markus Kreft, Kevin O'Sullivan,
Robert Jakob, Azul Garza, Paul Schmiedmayer, Patrick Langer
Girish Narayanswamy, Maxwell A. Xu, A. Ali Heydari, Samy Abdel-Ghaffar, Marius Guerard,
Kara Vaillancourt, Zhihan Zhang, Jake Garrison, Levi Albuquerque, Dimitris Spathis, Hong Yu, Hamid
Palangi, Xuhai "Orson" Xu, David G. T. Barrett, Joseph Breda, Jed McGiffin, Yubin Kim, Yuwei Zhang, Naghmeh
Rezaei, Samuel Solomon, Karan Ahuja, Tim Althoff, Jake Sunshine, Ming-Zher Poh, Benjamin Yetton, Ari Winbush,
Nicholas B. Allen, James M. Rehg, Isaac Galatzer-Levy, Yun Liu, John Hernandez, Anupam Pathak, Conor Heneghan,
Yuzhe Yang, Ahmed A. Metwally, Pushmeet Kohli, Mark Malhotra, Shwetak Patel, Xin Liu, Daniel McDuff
Xiao Gu, Yuxuan Shu, Jinpei Han, Yuxuan Liu, Zhangdaihong Liu, James Anibal, Veer
Sangha, Edward Phillips, Bradley Segal, Hang Yuan, Fenglin Liu, Kim Branson, Patrick Schwab, Danielle
Belgrave, Lei Clifton, Dimitris Spathis, Vasileios Lampos, A Aldo Faisal, David A Clifton
Time2Lang: Bridging Time-Series Foundation Models and Large Language Models for Health
Sensing Beyond Prompting
Arvind Pillai, Dimitris Spathis, Subigya Nepal, Amanda C Collins, Daniel M
Mackin, Michael V Heinz, Tess Z Griffin, Nicholas C Jacobson, Andrew Campbell
Conference on Health, Inference, and Learning (CHIL 2025)
Learning under label noise through few-shot human-in-the-loop refinement
Aaqib Saeed, Dimitris Spathis, Jungwoo Oh, Edward Choi, Ali Etemad
Alican Akman, Harry Coppock, Christian Bergler, Maurice Gerczuk, Chloë Brown, Jagmohan
Chauhan, Andreas Grammenos, Apinan Hasthanasombat, Dimitris Spathis, Tong Xia, Pietro Cicuta, Jing Han,
Shahin Amiriparian, Alice Baird, Lukas Stappen, Sandra Ottl, Panagiotis Tzirakis, Anton Batliner, Cecilia
Mascolo, Björn Wolfgang Schuller
Reviewer: NeurIPS,
ICLR,
ICML,
AAAI,
IJCAI,
KDD,
CHI,
Ubicomp/IMWUT,
CHIL, Nature Digital Medicine, WACV, Nature
Scientific Reports,
ICASSP,
Expert Systems with Applications, Neurocomputing,
WWW/The Web Conference, Engineering Applications of Artificial Intelligence,
ICWSM,
and more.
I have also been a teaching assistant for the following undergraduate courses:
Over the years, I've had the privilege of supervising and closely collaborating with a talented group of
researchers, both through industry internships and university PhD co-supervision. We focus on building
intelligent and personalized agents for health. Here are the amazing researchers who have worked with me:
Chi Ian Tang (University of Cambridge): Self-supervised and continual learning
Benjamin Searle (University of Cambridge): Capturing compulsive behaviours w/ wearables
Kevalee Shah (University of Cambridge): Benchmarking contrastive learning algorithms
Chuen Low (University of Cambridge): Attention models for timeseries
Yu Yvonne Wu (University of Cambridge): Weakly-supervised and self-supervised learning
Multimodal, data-efficient, and robust AI for real-world biosignals & the role of generative
models Cambridge MedAI Seminar
Series, Biomedical Campus, Cambridge, UK
nov 2023
Multimodal AI for real-world signals – does the key to specialized models lie in
language?
Microsoft AI & Pizza talk -
Cambridge ELLIS Unit, Cambridge, UK
Self-Supervised Learning for Health Signals Rising Stars in AI, KAUST, Saudi Arabia
nov 2022
Representation learning for cardio-fitness prediction in free-living environments
King's College London, Precision Health Informatics Data Lab, London, UK
jun 2022
AI-powered Wearables Transforming Mobile Health
AI Summit, London Tech week, London, UK
In statistics, a "prior" is your baseline belief before observing new data. These are the current assumptions
driving my work: strongly held, but always ready to be updated when reality proves otherwise.
LLMs commoditized human language. The next massive unlock is deciphering the
language of the body.
The Bitter Lesson proved that scale wins, but current AI is sensory-deprived.
The next scaling frontier is not just more parameters/data, it's more senses.
Stop treating wearables like step counters. They are the non-invasive
precursor to Brain-Computer Interfaces (BCI) and the primary data ingestion layer for
real-time/closed-loop human-AI symbiosis.
Most of the industry is obsessing over fine-tuning text and image wrappers.
The actual frontier is training multimodal foundation models on raw, unannotated, and unpaired signals.
Software people are spoiled by perfectly clean data in the cloud. The real
alpha is processing noisy and complex signals directly on constrained edge devices.
AI is eating software; biology will eat AI.
Personal
Non-academic things about me: I love music, both playing and listening. I am mostly into art rock and indie
folk, with the occasional exception of some well-crafted pop. Although I am an accordionist by training, over
the last few years I've been playing mostly piano and ukulele. In a previous life, I performed with the
critically acclaimed band The Children of the Oldness (aka Kore Ydro) and recorded the album "Consortium in
Amato" (listen here).
I also enjoy street photography and in particular playing with light—photography comes from Greek φως (light)
and γραφή (writing), or drawing with light. A sample of my shots is on Flickr and one of my landscapes was featured in the Huffington Post.
Lastly, and perhaps most importantly, I'm always on the lookout for ways to move items from the "non-academic
list" to the "academic list"—let me know if you'd like to help!
Playground
“The next big thing in technology often starts off looking like a toy”
Quantifying name-dropping
Communitypoprefs.com is a data visualization website, where
we present every pop-culture reference over the course of 5 seasons of the TV series Community.
Map out your music taste on Spotify
Visualizing my favourite songs on Spotify with dimensionality reduction and anomaly detection. Data essay published in Cuepoint Magazine, Medium's
premier music publication.
Children's books and childish language?
Text mining Game of Thrones, Harry Potter, Hunger Games and Lord of the Rings books. Data essay featured in Medium's Editor Picks.
Anonymize kids' faces before posting online
Mobile app with face recognition, age estimation,
& emotion recognition to blur
kids or replace their faces with emotion-based emojis. Developed during HackZurich 2018.
Discover top local news globally
Glocalne.ws was a mashup of Google News and Google Maps.
Unfortunately it is now defunct due to API discontinuance.
Composing music and text with Recurrent Neural Networks
Training neural networks on massive amounts of musical notation and literature and letting them create their
own art. The essay
is in
Greek, but you can still see/listen to the results.
Bonus points to anyone who can decipher all the landmarks and chapters of my life sketched in the footer
below.
