Abstract
Everyday wearable devices, such as smartwatches, smart rings, and fitness trackers, track valuable health metrics, but analyzing this data becomes tricky when sensors record at different speeds, miss measurements, or get bumped during exercise. Traditional processing methods attempt to fix this by dropping or padding data, which visibly destroys important physiological details. This publication describes a generative artificial intelligence (AI) system that uses machine learning to seamlessly combine health signals recorded at different speeds, automatically reconstruct corrupted data gaps (data imputation), and dynamically generate a biometric profile or digital twin of the user’s physiology to run interactive “what-if” simulations and predict future health anomalies. Instead of discarding messy data, the system described relies on motion sensors to identify corrupted data windows. Furthermore, the system intelligently cross-references other clean sensor readings to accurately reconstruct the missing waveforms in real time. The system continuously learns and maintains this physiological profile to power externally observable features, such as interactive mobile simulations that dynamically adjust 48-hour recovery curves based on user inputs (e.g., adjusting a slider for a hypothetical additional 2 hours of sleep), glanceable, real-time notifications for stress and recovery score metrics, and proactive system alerts generated 24 hours before symptoms appear (e.g., predicting cardiovascular strain spikes).
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Shin, Dongeek, "MULTIMODAL GENERATIVE BIOSIGNAL SYNTHESIS USING CONTINUOUS-TIME DIFFUSION", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12083