Research on Multiple Spectral Ranges with Deep Learning for SpO₂ Measurement.

Area of Science:

• Biomedical Engineering
• Optical Sensing
• Machine Learning

Background:

• Pulse oximetry (SpO₂) is crucial for symptom diagnosis.
• Traditional SpO₂ measurements have inherent errors due to limited wavelengths and algorithms.
• Advancements in machine learning and spectral analysis offer potential for improved accuracy.

Purpose of the Study:

• To develop a more accurate SpO₂ measurement technique using multi-wavelength spectral absorption and deep learning.
• To investigate the impact of different spectral regions on SpO₂ measurement accuracy.
• To optimize a 1D-CNN model for enhanced SpO₂ prediction.

Main Methods:

• Utilized 12-wavelength spectral absorption data across various spectral regions.
• Constructed three datasets for training and verification.
• Employed data augmentation to improve model generalization.
• Applied 1D-CNN architecture for SpO₂ measurement model development.
• Optimized hyperparameters using GridSearchCV and Bayesian optimization.

Main Results:

• The 1D-CNN model achieved optimal accuracies of 89.3% (GridSearchCV) and 99.4% (Bayesian optimization).
• The best model, optimized by Bayesian optimization, used six specific wavelengths and achieved a total relative error of 0.46%.
• Analysis indicated that using a shorter set of six wavelengths was sufficient, and longer spectral ranges were redundant for this model.

Conclusions:

• A 1D-CNN model based on multi-wavelength spectral measurement offers a novel and feasible approach for accurate SpO₂ determination.
• Selecting appropriate spectral ranges and wavelengths is critical for effective deep learning model construction in spectral measurements.
• The study highlights the importance of hyperparameter optimization for achieving high accuracy in SpO₂ measurement.