Pulse2AI: An Adaptive Framework to Standardize and Process Pulsatile Wearable Sensor Data for Clinical Applications

Abstract

Goal: To establish Pulse2AI as a reproducible data preprocessing framework for pulsatile signals that generate high-quality machine-learning-ready datasets from raw wearable recordings. Methods: We proposed an end-to-end data preprocessing framework that adapts multiple pulsatile signal modalities and generates machine-learning-ready datasets agnostic to downstream medical tasks. Results: a dataset preprocessed by Pulse2AI improved systolic blood pressure estimation by 29.58%, from 11.41 to 8.03 mmHg in root-mean-square-error (RMSE) and its diastolic counterpart by 26.01%, from 7.93 to 5.87 mmHg in RMSE. For respiration rate (RR) estimation, Pulse2AI boosted performance by 19.69%, from 1.47 to 1.18 breaths per minute (BrPM) in mean-absolute-error (MAE). Conclusion: Pulse2AI turns pulsatile signals into machine learning (ML) ready datasets for arbitrary remote health monitoring tasks. We tested Pulse2AI on multiple pulsatile modalities and demonstrated its efficacy in two medical applications. This work bridges valuable assets in remote sensing and internet of medical things to ML-ready datasets for medical modeling.

Keywords: Bridge2AI; IoMT; ML for Healthcare; signal processing; wearable pulsatile signals.

Fig. 1.

Overview of the Pulse2AI framework: the framework optimizes preprocessing specific to the dataset, confidence, and medical task. Each filter assesses data qualities individually, a cardiac cycle will be rejected if failed to pass the threshold.

Fig. 2.

Example of pulsatile signals with annotated fiducial points; morphologies are derived from the amplitudes and longitudes of fiducial points.

Fig. 3.

Maximum slope extraction on BIDMC PPG signals. Maximum slope is one of the more complex fiducial point used for inter-beat-interval (IBI) calculation. A data point with its first gradient higher than 95th percentile and its third gradient lower than 5th percentile of the sliding window is considered a maximum slope candidate. A candidate is dropped if it's less than 0.2 seconds to its previous candidate.

Fig. 4.

Real-world example of limitations on IBI matching. A 240-second PPG sequence found the correct segment on the arterial blood pressure (ABP) continuous stream. Although the two IBIs were closely aligned in the sample domain (upper right plot), the PPG IBI was still behind for 1 second in the time domain as shown in the lower left plot. At last, we performed the second shift via the cross-correlation phase to correct the misalignments .

Fig. 5.

IBI and blood pressure recordings with Bio-Z protocols highlighted. While ice water (cold pressor) keeps BP elevated, the absolute SBP declines over time while the IBI increases.

Fig. 6.

An example of noise in the Bio-Z dataset. Bio-Z1, the first Bio-Z sensor, experienced sensor misplacement. Additionally, motion artifacts, another form of noise, affected the recordings on other sensors, particularly between 135 to 140 seconds.