Transferring learned features to ultra-wideband radar
We then ran a study that collected UWB radar data, along with electrocardiogram (ECG) and photoplethysmogram (PPG) data as our ground truth for heart rate, using a setup that placed the UWB radar sensor in positions where users typically hold their phone, i.e., on a table in front of them or on their lap. Compared to the FMCW dataset, which was 980 hours of data, the UWB radar dataset was much smaller, with 37.3 hours. As the UWB radar configuration was close to what is feasible on a mobile phone, with a much lower bandwidth, its range resolution was far lower than the FMCW dataset.
To ensure that our model was optimized to transfer to the UWB dataset, we retrained it after performing additional pre-processing steps to modify the mm-wave FMCW radar data to better resemble the target IR-UWB data, effectively lowering its range resolution. We then fine-tuned this model on the IR-UWB dataset, achieving an MAE of 4.1 bpm and mean absolute percentage error (MAPE) of 6.3%, a 25% reduction over the baseline error rate. Our baseline for performance on UWB radar was 5.4 bpm MAE and 8.4% MAPE, achieved by selecting the best model trained from scratch on our UWB dataset. With transfer learning, we enabled the UWB radar to meet the Consumer Technology Association standards for heart rate measurement for consumer devices: an accuracy of up to 5 bpm MAE and 10% MAPE.
