[Paper Review] Wrist-worn blood pressure tracking in healthy free-living individuals using neural networks

Mustafa Radha, Koen de Groot+10 more|arXiv (Cornell University)|May 23, 2018

Non-Invasive Vital Sign Monitoring45 references3 citations

TL;DR

This study proposes a wrist-worn photoplethysmography (PPG) sensor combined with a long short-term memory (LSTM) neural network to unobtrusively estimate the nocturnal systolic blood pressure (SBP) dip in free-living healthy individuals. Using 5,111 matched PPG and ambulatory BP monitor readings over 226 days, the LSTM model achieved an RMSE of 3.12±2.20 mmHg for SBP dip estimation, outperforming traditional machine learning and non-machine learning methods.

ABSTRACT

Objective: Evaluate a method for the estimation of the nocturnal systolic blood pressure (SBP) dip from 24-hour blood pressure trends using a wrist-worn photoplethysmography (PPG) sensor and a deep neural network in free-living individuals, comparing the deep neural network to traditional machine learning and non-machine learning baselines. Approach: A wrist-worn PPG sensor was worn by 106 healthy individuals for 226 days during which 5111 reference values for blood pressure (BP) were obtained with a 24-hour ambulatory BP monitor and matched with the PPG sensor data. Features based on heart rate variability and pulse morphology were extracted from the PPG waveforms. Long- and short term memory (LSTM) networks, dense networks, random forests and linear regression models were trained and evaluated in their capability of tracking trends in BP, as well as the estimation of the SBP dip. Main results: Best performance for estimating the SBP dip were obtained with a deep LSTM neural network with a root mean squared error (RMSE) of 3.12$\pm$2.20 $\Delta$mmHg and a correlation of 0.69 $(p=3*10^{-5})$. This dip was derived from trend estimates of BP which had an RMSE of 8.22$\pm$1.49 mmHg for systolic and 6.55$\pm$1.39 mmHg for diastolic BP (DBP). While other models had similar performance for the tracking of relative BP, they did not perform as well as the LSTM for the SBP dip. Significance: The work provides first evidence for the unobtrusive estimation of the nocturnal SBP dip, a highly prognostic clinical parameter. It is also the first to evaluate unobtrusive BP measurement in a large data set of unconstrained 24-hour measurements in free-living individuals and provides evidence for the utility of LSTM models in this domain.

Motivation & Objective

• To evaluate the feasibility of estimating the nocturnal systolic blood pressure (SBP) dip using a wrist-worn PPG sensor in free-living conditions.
• To compare deep learning, traditional machine learning, and non-machine learning models for tracking 24-hour blood pressure trends.
• To assess the performance of long short-term memory (LSTM) networks in capturing long-term BP variability and estimating clinically relevant parameters like the SBP dip.
• To provide evidence for the clinical utility of continuous, unobtrusive BP monitoring using wearable PPG sensors in real-world settings.

Proposed method

• A wrist-worn PPG sensor collected continuous photoplethysmographic signals from 106 healthy individuals over 226 days.
• Feature extraction from PPG waveforms included heart rate variability and pulse morphology characteristics.
• Long short-term memory (LSTM) networks were trained to model temporal dependencies in PPG-derived features for predicting 24-hour blood pressure trends.
• Dense neural networks, random forests, and linear regression models were also trained and compared for BP trend estimation and SBP dip prediction.
• Reference blood pressure values were obtained via 24-hour ambulatory BP monitoring, with data matched to PPG sensor readings.
• Model performance was evaluated using root mean squared error (RMSE) and correlation coefficients against reference SBP dip values.

Experimental results

Research questions

• RQ1 Can a wrist-worn PPG sensor combined with a deep neural network accurately estimate the nocturnal systolic blood pressure (SBP) dip in free-living individuals?
• RQ2 How does the performance of an LSTM-based model compare to traditional machine learning and non-machine learning models in estimating the SBP dip from PPG data?
• RQ3 To what extent can PPG-derived features capture long-term blood pressure trends necessary for SBP dip estimation?
• RQ4 Is the SBP dip, a key prognostic indicator, reliably estimable using unobtrusive, continuous PPG monitoring in unconstrained daily environments?

Key findings

• The LSTM neural network achieved the lowest root mean squared error (RMSE) of 3.12±2.20 mmHg for estimating the nocturnal SBP dip, significantly outperforming other models.
• The SBP dip estimation correlated with reference values at r=0.69 (p=3×10⁻⁵), indicating strong statistical significance.
• BP trend estimation using the LSTM model had an RMSE of 8.22±1.49 mmHg for systolic and 6.55±1.39 mmHg for diastolic blood pressure.
• Other models showed comparable performance in tracking relative BP trends but were less accurate in estimating the SBP dip.
• This study provides the first evidence of unobtrusive, continuous estimation of the clinically prognostic nocturnal SBP dip using wearable PPG and deep learning.
• The results demonstrate the utility of LSTM models in capturing complex temporal patterns in PPG data for cardiovascular risk assessment.