Development and Validation of Deep Learning Models for Motion Artifact Mitigation in Wearable PPG Devices

Wearable devices are assuming a larger role in remote healthcare, fitness tracking, and athletics. Predominantly based on photoplethysmography, these devices use light to non-invasively detect changes in blood flow and oxygenation within peripheral circulation – leading to estimates of heart rate, pulse oximetry, and the identification of arrhythmias. However, the presence of motion greatly reduces the quality and interpretability of data from wearable devices and limits the development of ML/AI models. Currently, accelerometers are used to detect excessive motion and dispose of ‘contaminated’ data. However, accelerometers capture global motion, not the relative motion at the sensor-skin interface – the predominant source of motion artifacts. Here, we propose the reconstruction of physiological waveforms in the presence of motion through the development of a deep learning model, PPG-MART, trained through a unique multimodal sensor platform. The platform consists of two synchronized multiwavelength photoplethysmography sensors, a pressure sensor that captures relative motion at the sensor interface, and a linear actuator that can apply arbitrary motion-mimicking waveforms. Systematic application of basis stimuli generate a perturbed multiwavelength data channel which is compared to a non-perturbed control dataset obtained on the opposite hand. We demonstrate the training of multiple deep learning models that reconstruct photoplethysmography waveforms based on subsets of data, down to a single PPG channel. Results show improving performance with an increasing number of wavelength channels and inclusion of a pressure channel, but that significant improvements are accessible with only a single PPG channel. Comprehensive testing on five external PPG datasets verifies the generalizability of the model and applicability to all optical devices which perform PPG. This work enables real-time motion artifact cancellation in wearable optical devices and will lead to more robust remote health care devices, athletic performance trackers, and algorithms. The training concept can also be applied to any wearable measurement modality that supports multiple, independent measurements sites of a systemic biomarker or vital sign.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.