A Comparative Evaluation of Sample Selection Algorithms for Multivariate Calibration in Near-Infrared Spectroscopic Analysis of Pharmaceutical Formulations

Technical Summary: Comparative Evaluation of Sample Selection Algorithms for Multivariate Calibration in NIR Spectroscopy

Problem Statement
The development of robust multivariate calibration models for Near-Infrared (NIR) spectroscopic analysis in pharmaceutical quality control relies critically on the rational partitioning of available samples into calibration (training) and validation sets. Inadequate sample selection leads to model overfitting, where excellent calibration statistics fail to translate into predictive accuracy for novel samples. While various algorithms exist to optimize this partitioning—such as Kennard–Stone, Honigs, Duplex, and Naes—previous literature often lacks rigorous statistical validation of performance differences, systematic investigation of distance metric effects (Euclidean vs. Mahalanobis), and comprehensive analysis of training set proportion impacts. This study addresses these gaps by evaluating four established algorithms applied to a pharmaceutical NIR dataset.

Methodology
The investigation utilized a dataset of 58 commercial paracetamol tablets (500 mg) acquired from pharmacies in Bamako, Mali. Reference paracetamol concentrations were determined via validated pharmacopoeial UV-Vis spectrophotometry. NIR spectra were acquired in reflectance mode (800–1700 nm) and preprocessed using Standard Normal Variate (SNV), derivatives, and Savitzky–Golay smoothing, with smoothed spectra selected for optimal performance.

The core experimental design involved:

1. Algorithm Implementation: Four sample selection algorithms were implemented:
   • Kennard–Stone: Selects samples to maximize spectral space coverage.
   • Honigs: Iteratively selects spectrally distinct samples based on residual projection.
   • Duplex: Simultaneously constructs balanced calibration and validation sets.
   • Naes: Uses k-means clustering to ensure subtype representation.
2. Modeling Framework: Gaussian Process Regression (GPR) with a radial basis function kernel served as the modeling framework.
3. Performance Metrics: Models were evaluated using the coefficient of determination ($R^2$) and Root Mean Square Error of Prediction (RMSEP).
4. Statistical Analysis: Non-parametric statistical methods were employed to validate performance differences. The Kruskal–Wallis test assessed global differences, followed by post-hoc Mann–Whitney U tests with Bonferroni correction for pairwise comparisons.
5. Variables: The study systematically varied training set proportions (60% to 90%) and compared Euclidean versus Mahalanobis distance metrics where applicable.

Key Results

• Algorithm Performance: The Kennard–Stone algorithm employing the Mahalanobis distance metric demonstrated superior performance, yielding a validation $R^2$ of 0.99999 and an RMSEP of $1.74 \times 10^{-6}$. The Honigs algorithm showed comparable performance ($R^2 = 0.999996$, RMSEP = $1.76 \times 10^{-6}$). Both significantly outperformed the Duplex, Naes, and random splitting methods.
• Statistical Significance: Kruskal–Wallis testing confirmed highly significant differences among algorithms ($p < 0.001$). Post-hoc analysis revealed that while Kennard–Stone and Honigs were statistically equivalent to each other ($p = 1.000$), both were significantly superior to Duplex, Naes, and random partitioning.
• Overfitting Indicators: Naes and Duplex algorithms exhibited excellent calibration statistics but degraded validation performance, indicating a tendency toward overfitting and inadequate representation of validation set variability.
• Training Set Size: A monotonic relationship was observed between training set size and predictive accuracy. Performance improved as the training proportion increased from 60% to 90%, with asymptotic behavior suggesting diminishing returns beyond the 80–90% range.
• Distance Metrics: For the Kennard–Stone algorithm, the Mahalanobis distance metric yielded superior validation results compared to Euclidean distance, attributed to its ability to account for inter-wavelength correlations in spectral data. For the Honigs algorithm, both metrics performed comparably.

Significance and Claims
The paper asserts that its findings provide evidence-based guidance for optimizing sample selection protocols in pharmaceutical NIR applications. The primary contributions include:

1. Validation of Rational Selection: Demonstrating that systematic sample selection algorithms significantly outperform random partitioning, necessitating their inclusion in method development workflows.
2. Statistical Equivalence: Establishing that Kennard–Stone and Honigs methods are statistically equivalent for this dataset, offering practitioners flexibility in algorithm choice based on computational or implementation preferences.
3. Methodological Rigor: Highlighting the critical importance of external validation over calibration metrics alone, as evidenced by the overfitting observed in Naes and Duplex methods.
4. Metric Selection: Underscoring the necessity of selecting appropriate distance metrics, specifically the advantage of Mahalanobis distance for Kennard–Stone in correlated spectral spaces.

The authors conclude that the Kennard–Stone algorithm with Mahalanobis distance represents a well-validated approach for sample set partitioning in this domain, while acknowledging that future work should extend these findings to diverse pharmaceutical product classes.