Wearable foundation models: a brief history (www.empirical.health)

The article provides an overview of the evolution of wearable foundation models (WFMs), which are advanced neural networks designed to analyze physiological signals from wearable devices. These models leverage vast amounts of unlabeled sensor data to learn general health-related patterns, enabling them to be fine-tuned for specific medical tasks with minimal labeled data. The significance of WFMs lies in their potential to enhance health monitoring through devices like the Apple Watch and Fitbit, offering capabilities such as detecting hypertension, sleep apnea, and estimating biometrics more accurately than traditional methods. Key developments include Apple and Google's innovative approaches to WFM architecture and training. Apple's models utilized raw photoplethysmography (PPG) and electrocardiogram (ECG) signals, demonstrating that each can significantly outperform demographic-based models in health condition detection. Google’s LSM series further explored scaling laws and developed a model capable of learning from fragmented real-world data. Additionally, Empirical Health's JETS model employed a novel joint embedding approach, achieving impressive results in identifying various health conditions. This lack of reliance on large teams underscores the potential for smaller entities to contribute to impactful advancements in healthcare AI. Overall, these developments exemplify the ongoing intersection of AI, healthcare, and wearable technology, offering promising implications for personalized medicine.