Machine Learning-Driven Classification and Quantification of E. coli Using Photonic Crystal Fiber Biosensor Data: A COMSOL-Integrated Approach

Abstract

1. coli contamination of drinking water remains a serious public health issue in areas where there is no proper water treatment system. The photonic crystal fiber (PCF) biosensors generate a number of measurable results when simulated using COMSOL these include the effective refractive index (n_eff), confinement loss, sensitivity, spectral shift, and capacitance but it is not easy to convert these results into a decision regarding concentration, particularly when there are eight concentration levels that are very close together. In this paper, the COMSOL Multiphysics 6.1 simulation data for PCF is directly linked to a machine learning (ML) pipeline, with seven algorithms being tested for binary detection of contamination and for classifying concentration into eight levels (from 0 to 100,000 CFU/mL). A dataset consisting of 3,000 samples was created from the simulation outputs with realistic measurement noise included. Random Forest achieved a binary accuracy of 95.17% (AUC=0.985); the accuracy for the eight-class concentration classification was 80.33% (F1-macro=80.37%); and the regression analysis yielded an R² of 0.918. McNemar's test (p=0.860) and the Wilcoxon signed-rank test (p=0.063) showed that the best models are statistically equivalent. Both sensitivity (32.3%) and capacitance (31.7%) contributed the most to the classification a result which suggests that dual-readout PCF designs should be considered in future research. This method eliminates the requirement for experimentally calibrated laboratory datasets while at the same time producing concentration decisions fast enough for real-time field applications.