Healthcare
Care settings generate vital signs and lab values continuously. Catching deterioration hours earlier changes preparation and staffing alike. We have researched biosignal forecasting, such as ICU pain prediction, jointly with clinical partners.
- What we forecast
- Onset, deterioration, vital signs.
- Decisions that change
- When to intervene.
2026-08-18
PENGUIN: General Vital Sign Reconstruction from PPG
A generative model that reconstructs ECG, respiratory and blood-pressure waveforms from PPG, the optical pulse sensor found in smartwatches. Accepted at ICASSP 2026, it achieved the best accuracy on 7 of 8 metrics across six datasets.
2026-08-18
ICU Pain Risk Estimation from Bedside Signals
A model that estimates elevated nurse-assessed pain scores (CPOT) from automatically recorded ICU bedside signals alone. In joint research with Tohoku University, we curated a clinical dataset of 891 patients and over 23,000 samples, achieving the highest AUROC among compared models.
2026-08-18
DecompSSM: A Decomposition-Based State Space Model for Multivariate Time-Series Forecasting
Our forecasting model learns to decompose series into trend, seasonal and residual components end-to-end. Accepted at ICASSP 2026, it achieved the best accuracy in 28 of 32 standard benchmark settings.
2026-08-17
Treatment Side-Effect Forecasting AI
With a healthcare business, we developed a multivariate time-series model that forecasts treatment side effects and a small language model that serves expert knowledge.