Lightweight Machine Learning Pipeline for WGM Micro-Laser Sensing: A Feasibility Study on IFN-γ Detection

This work analyze the feasibility of lightweight machine learning for automated analyte detection in whispering gallery mode (WGM) micro-laser biosensors. While recent approaches increasingly rely on deep learning architectures to process spectral data, these architectures are often too computationally complex for embedded or edge-based sensing systems. In this regard, the proposed framework uses a resource-efficient pipeline based on handcrafted spectral features and traditional classification algorithms. This enables real-time inference with a reduced computational footprint. The approach is validated through a case study on the detection of human interferon-gamma (IFN-γ) using functionalized polymeric microlaser beads with a diameter of ~ 30μm. Spectral features capturing wavelength shifts and intensity variations are extracted from temporal WGM resonance evolution. These features are used to formulate a multi-class classification problem aimed at discriminating IFN-γ concentration levels ranging from 31.3 to 500 pg mL-1. Model performance is evaluated using a bead-wise leave-one-out cross-validation strategy to assess generalization across different sensing units. The dataset exhibits class imbalance and limited sample size, resulting in modest classification performance that nonetheless demonstrates the potential of lightweight ML approaches to discriminate analyte concentrations in WGM systems. These preliminary results suggest a viable pathway toward practical deployment in distributed sensing infrastructures for environmental and medical monitoring applications.