Transformer-based nonlinear modeling in dynamic spectrum for noninvasive human blood component analysis.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Spectroscopy and Photonic Sensing

Background:

• Photoplethysmography (PPG) derived dynamic spectrum (DS) offers non-invasive blood component analysis potential.
• Nonlinearity from blood component scattering currently limits prediction accuracy in DS-based methods.
• "M + N" theory suggests incorporating nonlinear factors can enhance calibration models.

Purpose of the Study:

• To develop a nonlinear modeling approach for DS data to improve blood component concentration prediction.
• To leverage transformer-based neural networks for fitting complex nonlinear relationships in spectral data.
• To evaluate the performance of transformer models against baseline deep learning models for non-invasive blood analysis.

Main Methods:

• Extraction of arterial blood absorbance at various wavelengths to generate dynamic spectrum (DS) data.
• Application of transformer-based neural networks to model the nonlinear relationship between DS and blood component concentrations.
• Comparative analysis against Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models using data from 1400 participants.

Main Results:

• Transformer modeling significantly improved prediction accuracy (R2, MAE, MSE) for red blood cells (RBC), urea, and glucose compared to CNN and LSTM.
• The transformer model effectively captured nonlinear mappings within the spectral data, demonstrating its capability for complex data analysis.
• Consistent performance improvements across multiple blood components validate the model's effectiveness in non-invasive spectral analysis.

Conclusions:

• Transformer-based neural networks show significant potential for enhancing the accuracy of non-invasive dynamic spectrum (DS)-based blood component analysis.
• The model's ability to account for both nonlinear and general factors in DS data improves the mapping of component concentration to spectral features.
• Further clinical validation is necessary to establish the widespread applicability of this advanced modeling technique.