Also of course connections with weather / food / medication logs and other sensors (continous glucose monitoring, CGM) would be great. 6/6

In my opinion, a more pro-active data collection loop with the watch detecting trends / outliers live and prompting for context would be a game-changer ("Your heart rate is spiking. / I noticed you cannot sleep. / Your skin is much warmer than usual. What's going on?"). 5/6

The only interesting labels are the "semantic captions" based on logged activity & mood and detected sleep. The "statistical & structural captions" are rather boring, their sole role is to describe the sensor data in words. This highlights once again the issue of missing annotations/context. 4/6

The actual PPG and ACC raw data sensor streams (probably around 50 Hz) would allow much more interesting analyses of course. This might be a pipeline / hardware limitation as I imagine raw data are not transferred to the cloud / stored on-device long-term. 3/6

As the Discussion and Conclusion in the paper are only 5 lines each (arxiv.org/abs/2506.09108), let me add some thoughts: There's heavy pre-processing and feature-extraction of the raw sensor data (probably on-device), leading to a fairly low-dimensional input (26 features × 1440 minutes/day). 2/6

SensorLM: Learning the language of wearable sensors

Finally a large scale sensor-language model for wearables by Google incorporating heart rate (photoplethysmography, PPG), activity (accelerometer, ACC), altimeter (ALT), skin temperature (TEMP), and electrodermal activity (EDA). It's a pity the model is not openly released! goo.gle/4lSLwQi 1/6