Deep learning optimized dual-analyte detection-based biosensor for monitoring pregnancy stage using a urine sample

Area of Science:

• Photonics
• Biosensing
• Artificial Intelligence

Background:

• Surface Plasmon Resonance (SPR) biosensors offer high sensitivity for analyte detection.
• Photonic Crystal Fibers (PCFs) provide a versatile platform for SPR sensor development.
• Optimizing multi-analyte SPR biosensor design and performance prediction remains challenging.

Purpose of the Study:

• To develop and optimize a hybrid deep learning (DL) approach for designing PCF-based SPR multi-analyte biosensors.
• To accurately predict sensor performance metrics like confinement loss (CL).
• To enhance sensitivity, wavelength sensitivity (WS), and sensor resolution (SR) for multi-analyte detection.

Main Methods:

• Utilized Finite Element Method (FEM) simulations to generate a comprehensive dataset of sensor parameters and refractive index (RI) values.
• Developed a hybrid Recurrent Neural Network Long-Short-Term Memory (RNN-LSTM) model for CL prediction.
• Performed ablation studies and SHAP-based explainability analysis for model validation.

Main Results:

• The RNN-LSTM model achieved superior prediction performance with Mean Squared Error (MSE) of 0.0014, Mean Absolute Error (MAE) of 0.0188, and R-squared (R²) of 0.9510.
• Achieved high sensor performance: maximum Amplitude Sensitivity (AS) of 3102.41 RIU⁻¹, WS of 10,000 nm/RIU, and SR of 1×10⁻⁵.
• Demonstrated the model's effectiveness through rigorous validation techniques.

Conclusions:

• The proposed hybrid DL approach significantly improves the design and performance prediction of PCF-based SPR multi-analyte biosensors.
• Deep learning offers a powerful tool for advancing biosensor technology.
• This work paves the way for more efficient and accurate multi-analyte biosensing applications.