Imbalanced Power Spectral Generation for Respiratory Rate and Uncertainty Estimations Based on Photoplethysmography...

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Signal Processing for Biosignals

Background:

• Accurate respiratory rate (RR) estimation is crucial for cardiopulmonary function monitoring in the elderly.
• Machine learning (ML) algorithms in health monitoring systems are susceptible to errors due to biosignal data imbalance.
• Significant sample size differences between normal breathing, dyspnea, and hypopnea data negatively impact ML model training and accuracy.

Purpose of the Study:

• To propose a novel methodology combining bootstrap-based imbalanced continuous power spectral generation (IPSG) with ML to accurately estimate RRs and associated uncertainty.
• To address the challenge of biosignal data imbalance in home health monitoring systems.
• To improve the reliability of ML-driven medical decision-making for respiratory conditions.

Main Methods:

• Developed an Imbalanced Continuous Power Spectral Generation (IPSG) technique utilizing a nonparametric bootstrap approach.
• Generated artificial feature curves to augment sparse data segments (dyspnea, hypopnea) in respiratory signals.
• Integrated original and artificial feature curves, training ML models like Gaussian Process Regression (GPR) on a combined dataset.

Main Results:

• The IPSG method effectively addressed data imbalance, enabling efficient training for predicting the relationship between photoplethysmography signal features and RR.
• The proposed GPR-IPSG model demonstrated improved prediction accuracy and uncertainty quantification compared to standard methods.
• Achieved a mean absolute error of 0.79 and 1.47 breaths per minute (brpm) for RR estimation on the Beth Israel Deaconess Medical Center dataset.

Conclusions:

• The GPR-IPSG model offers a robust and stable framework for accurate RR estimation, significantly enhancing home-based health monitoring systems.
• The methodology successfully mitigates the impact of data imbalance in biosignal analysis for respiratory monitoring.
• This approach provides a reliable foundation for designing advanced clinical home-based monitoring solutions.