Sensitivity, Prediction Uncertainty, and Detection Limit for Artificial Neural Network Calibrations.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Machine Learning

Background:

• The increasing use of artificial neural networks (ANNs) in multivariate calibration necessitates standardized metrics for performance evaluation.
• Current methods for evaluating ANN calibrations, such as average prediction error, lack the comprehensive detail of linear methods.
• There is a need for quantifiable figures of merit for nonlinear multivariate calibration to ensure comparability.

Purpose of the Study:

• To derive and present novel expressions for calculating key figures of merit for ANN-based multivariate calibration.
• To enable the computation of sensitivity, prediction uncertainty, and detection limits for nonlinear calibration models.
• To facilitate a more rigorous and comparable evaluation of ANN calibration models.

Main Methods:

• Adapted established methods from linear multivariate calibration, specifically partial least-squares regression.
• Developed new expressions for calculating sample-dependent sensitivity.
• Formulated methods for estimating prediction uncertainty and detection limits within ANN frameworks.

Main Results:

• Successfully derived expressions for sensitivity, prediction uncertainty, and detection limits applicable to ANN calibrations.
• Demonstrated that sensitivity is sample-dependent, aligning with theoretical expectations.
• Validated the proposed expressions using both simulated and real near-infrared spectral data.

Conclusions:

• The developed expressions provide essential tools for characterizing ANN multivariate calibration performance.
• This work bridges the gap between linear and nonlinear calibration evaluation, enhancing model comparability.
• The findings are crucial for advancing the application and standardization of ANN-based analytical methods.