Optimized spectroscopic regression for white blood cell quantification using metaheuristic feature selection

Area of Science:

• Biomedical Engineering
• Spectroscopy
• Hematology

Background:

• Accurate White Blood Cell (WBC) count is crucial for clinical decisions.
• Existing spectroscopic methods face challenges with signal fidelity and confounding factors.
• In-blood spectral acquisition offers enhanced signal integrity by reducing surface artifacts.

Purpose of the Study:

• To develop an optimized spectroscopic framework for enhanced WBC count estimation.
• To integrate metaheuristic feature selection with biologically informed sub-models.
• To minimize spectral interference from hemoglobin and platelet variations for improved WBC prediction.

Main Methods:

• Utilized high-dimensional optical spectrum data from 468 patients via fiber-optic probes.
• Employed Genetic Algorithm (GA), Firefly Algorithm (FA), and Grey Wolf Optimization (GWO) for feature selection.
• Developed biologically informed sub-models based on M+N theory, partitioning data by hemoglobin and platelet levels.
• Trained Random Forest regression models within each sub-model.

Main Results:

• The Firefly Algorithm with Random Forest achieved the highest R² (0.864) and lowest MAE (0.694) in the Low-HG/High-PLT sub-model.
• GWO and GA also showed strong performance in specific sub-models (R²=0.809 and R²=0.889, respectively).
• Sub-model-specific modeling significantly outperformed global regression (R²=0.73, RMSE=1.46), improving prediction precision.

Conclusions:

• Biologically guided partitioning and metaheuristic optimization effectively enhance spectroscopic WBC diagnostics.
• The proposed framework offers a practical and accurate alternative for point-of-care hematological analysis.
• This method demonstrates potential for rapid and reliable clinical decision support.