Benoît Igne

Evaluation of Spectral Pretreatments, Partial Least Squares, Least Squares Support Vector Machines and Locally Weighted Regression for Quantitative Spectroscopic Analysis of Soils

Soil testing requires the analysis of large numbers of samples in the laboratory that is often time consuming and expensive. Mid-infrared spectroscopy (mid-IR) and near infrared (NIR) spectroscopy are fast, non-destructive and inexpensive analytical methods that have been used for soil analysis, in the laboratory and in the field, to reduce the need for measurements using complex chemical/physical analyses. A comparison of the use of spectral pretreatment as well as the implementation of linear and non-linear regression methods was performed. This study presents an overview of the use of infrared spectroscopy for the prediction of five physical (sand, silt and clay) and chemical (total carbon and total nitrogen) soil parameters with near and mid-infrared units in bench top and field set-ups. Even though no significant differences existed among pretreatment methods, models using second derivatives performed better. The implementation of partial least squares (PLS), least squares support vector machines (LS-SVM) and locally weighted regression (LWR) for the development of the calibration models showed that the LS-SVM did not out-perform linear methods for most components while LWR that creates simpler models performed well. The present results tend to show that soil models are quite sensitive to the complexity of the model. The ability of LWR to select only the appropriate samples did help in the development of robust models. Results also proved that field units performed as well as bench-top instruments. This was true for both near infrared and mid-infrared technology. Finally, analysis of field moist samples was not as satisfactory as using dried-ground samples regardless of the chemometrics methods applied.

Online Monitoring of Pharmaceutical Materials Using Multiple NIR Sensors—Part I: Blend Homogeneity

IntroductionThe present article discusses the implementation of a semi-automated blend homogeneity control system by two near-infrared spectrometers.MethodsA statistic was introduced to combine blend trends output by individual instruments based on the root mean squared error from the nominal value calculation. The necessity to monitor homogeneity at more than one location of a V-blender is highlighted and the impact of sensor and model differences on blend trends was evaluated. Using two different formulations, classical least-squares based models were developed to monitor blending. Calibration transfer between the two sensors was demonstrated as a useful approach when more than one sensor is used. Several classical transfer methods were implemented (optical, post-regression correction, and orthogonalization based) to balance the two sensors.Results and ConclusionResults showed that the use of only one calibration model, transferred to all units monitoring the process was highly beneficial to achieving consistent results. Specifically, standardization methods targeting instrument differences were demonstrated to be the most successful. However, results showed that the optimization of a given transfer method was formulation-dependent.

Effect of Experimental Design on the Prediction Performance of Calibration Models Based on Near-Infrared Spectroscopy for Pharmaceutical Applications

Near-infrared spectroscopy (NIRS) is a valuable tool in the pharmaceutical industry, presenting opportunities for online analyses to achieve real-time assessment of intermediates and finished dosage forms. The purpose of this work was to investigate the effect of experimental designs on prediction performance of quantitative models based on NIRS using a five-component formulation as a model system. The following experimental designs were evaluated: five-level, full factorial (5-L FF); three-level, full factorial (3-L FF); central composite; I-optimal; and D-optimal. The factors for all designs were acetaminophen content and the ratio of microcrystalline cellulose to lactose monohydrate. Other constituents included croscarmellose sodium and magnesium stearate (content remained constant). Partial least squares-based models were generated using data from individual experimental designs that related acetaminophen content to spectral data. The effect of each experimental design was evaluated by determining the statistical significance of the difference in bias and standard error of the prediction for that models prediction performance. The calibration model derived from the I-optimal design had similar prediction performance as did the model derived from the 5-L FF design, despite containing 16 fewer design points. It also outperformed all other models estimated from designs with similar or fewer numbers of samples. This suggested that experimental-design selection for calibration-model development is critical, and optimum performance can be achieved with efficient experimental designs (i.e., optimal designs).

Modeling strategies for pharmaceutical blend monitoring and end- point determination by near-infrared spectroscopy

The implementation of a blend monitoring and control method based on a process analytical technology such as near infrared spectroscopy requires the selection and optimization of numerous criteria that will affect the monitoring outputs and expected blend end-point. Using a five component formulation, the present article contrasts the modeling strategies and end-point determination of a traditional quantitative method based on the prediction of the blend parameters employing partial least-squares regression with a qualitative strategy based on principal component analysis and Hotellings T(2) and residual distance to the model, called Prototype. The possibility to monitor and control blend homogeneity with multivariate curve resolution was also assessed. The implementation of the above methods in the presence of designed experiments (with variation of the amount of active ingredient and excipients) and with normal operating condition samples (nominal concentrations of the active ingredient and excipients) was tested. The impact of criteria used to stop the blends (related to precision and/or accuracy) was assessed. Results demonstrated that while all methods showed similarities in their outputs, some approaches were preferred for decision making. The selectivity of regression based methods was also contrasted with the capacity of qualitative methods to determine the homogeneity of the entire formulation.

Blending process modeling and control by multivariate curve resolution.

The application of the Multivariate Curve Resolution by Alternating Least Squares (MCR-ALS) method to model and control blend processes of pharmaceutical formulations is assessed. Within the MCR-ALS framework, different data analysis approaches have been tested depending on the objective of the study, i.e., knowing the effect of different factors in the evolution of the blending process (modeling) or detecting the blend end-point and monitoring the concentration of the different species during and at the end of the process (control). Data analysis has been carried out studying multiple blending runs simultaneously taking advantage of the multiset mode of the MCR-ALS method. During the ALS optimization, natural constraints, such as non-negativity (spectral and concentration directions) have been applied for blend modeling. When blending control is the main purpose, a variant of the MCR-ALS algorithm with correlation constraint in the concentration direction has been additionally used. This constraint incorporates an internal calibration procedure, which relates resolved concentration values (in arbitrary units) with the real reference concentration values in the calibration samples (known references) providing values in real concentration scale in the final MCR-ALS results. Two systems consisting of pharmaceutical mixtures of an active principle (acetaminophen) with two or four excipients have been investigated. In the first case, MCR results allowed the description of the evolution of the individual compounds and the assessment of some physical effects in the blending process. In the second case, MCR analysis allowed the detection of the end-point of the process and the assessment of the effects linked to variations in the concentration level of the compounds.

Effects and Detection of Raw Material Variability on the Performance of Near-Infrared Calibration Models for Pharmaceutical Products

The impact of raw material variability on the prediction ability of a near-infrared calibration model was studied. Calibrations, developed from a quaternary mixture design comprising theophylline anhydrous, lactose monohydrate, microcrystalline cellulose, and soluble starch, were challenged by intentional variation of raw material properties. A design with two theophylline physical forms, three lactose particle sizes, and two starch manufacturers was created to test model robustness. Further challenges to the models were accomplished through environmental conditions. Along with full-spectrum partial least squares (PLS) modeling, variable selection by dynamic backward PLS and genetic algorithms was utilized in an effort to mitigate the effects of raw material variability. In addition to evaluating models based on their prediction statistics, prediction residuals were analyzed by analyses of variance and model diagnostics (Hotellings T(2) and Q residuals). Full-spectrum models were significantly affected by lactose particle size. Models developed by selecting variables gave lower prediction errors and proved to be a good approach to limit the effect of changing raw material characteristics. Hotellings T(2) and Q residuals provided valuable information that was not detectable when studying only prediction trends. Diagnostic statistics were demonstrated to be critical in the appropriate interpretation of the prediction of quality parameters.

Online Monitoring of Pharmaceutical Materials Using Multiple NIR Sensors—Part II: Blend End-point Determination

The effect of various blend end-point algorithms on the observed blending time of pharmaceutical powders and content uniformity of subsequent tablets was investigated for a five-component system. The blending process was monitored online, in real time, using two near-infrared sensors. Algorithms based on the standard deviation, average, and distributions of concentration predictions were tested for all major blend constituents, respectively. The potential to combine sensor outputs in the end-point decision was contrasted by the consideration of the two sensor outputs individually or simultaneously. The algorithms employed demonstrated highly variable end-points when compared with the final tablet quality, although blends were deemed to have reached homogeneity faster when only the active ingredient was considered. Some algorithms proved to be either too sensitive to local mixing and demixing phenomena or not sensitive enough, yielding results not consistent with the observed tablet content uniformity. Results showed that the choice of an end-point algorithm must be directed by the product of interest (nature of the active, therapeutic window, etc.), the particular characteristics that the delivery forms should have (immediate release, sustained release), and most significantly, the purpose of blending. A single algorithm is not expected to be adequate across all formulations. However, as the complexity of the blending process increases (multiple sensors, trends of multiple ingredients to follow, etc.) the decision process becomes more complex with not only calibration maintenance issues to consider, but also calibration transfer, relevance of the criteria for the actives, and the desired final product properties.

Efficient Near-Infrared Spectroscopic Calibration Methods for Pharmaceutical Blend Monitoring

Near-infrared (NIR) spectroscopy is an important analytical tool for online process monitoring of pharmaceutical unit operations. Traditionally, the development and maintenance of robust, precise, and accurate quantitative NIR calibrations requires a substantial investment for the creation of sample sets. This study demonstrates the ability to develop efficient NIR calibrations using reduced sample sets. Prediction performance of several multivariate algorithms was compared on two different NIR spectrometers for pharmaceutical blend monitoring. Classical least-squares (CLS)-based algorithms took advantage of pure component scans to produce the most sensitive quantitative calibrations using reduced sample sets when compared to partial least squares (PLS) regression and two nonlinear methods. The PLS algorithm and the nonlinear methods produced models with low error but lacked the sensitivity needed to model subtle blending trends. The CLS-based methods produced models with adequate sensitivity for blend monitoring. The robustness of the CLS-based methods was further demonstrated in the ease of transfer between instruments using only a bias correction of the predictions.

Triticale Moisture and Protein Content Prediction by Near-Infrared Spectroscopy (NIRS)

ABSTRACT The use of near-infrared spectroscopy (NIRS) for the prediction of whole-grain triticale moisture and protein content was evaluated. Because triticale is genetically close to wheat, commercially available wheat prediction models for Foss Infratec analyzers were applied in a year-by-year basis to triticale samples harvested in Iowa between 2002 and 2006. Wheat models were not applicable to moisture prediction (SEPavg = 0.37% pt; expected SEP on wheat samples 0.15% pt), but usable for screening for protein (SEPavg = 0.38% pt; expected SEP on wheat samples 0.25% pt). Dedicated triticale calibrations were developed from 2002 to 2005 data. Prediction results for 2006 samples only were compared. Triticale calibrations performed better than wheat calibrations for 2006 samples (moisture SEPtriticale = 0.29% pt, SEPwheat = 0.50% pt; protein SEPtriticale = 0.30% pt, SEPwheat = 0.68% pt).

Standardisation of near infrared spectrometers: evaluation of some common techniques for intra- and inter-brand calibration transfer

Six standardisation methods (direct standardisation, piecewise direct standardisation, robust standardisation, slope and bias and bias only post regression correction and no standardisation or spectral pre-treatment only) were compared for the transfer of whole soybean protein, oil and linolenic acid models. Two Foss Infratec and two Bruins OmegAnalyzerGs instruments were used to evaluate the standardisation methods in intra- and inter-brand scenarios. Each instrument was calibrated on its own calibration set for comparison. Partial least squares was used to develop all models. For each brand, a master was selected and the calibrations developed on it were transferred to the secondary unit of its own network (intra-brand transfer) and to the two units of the other brand (inter-brand transfer). Foss Infratec models were transferable to Bruins OmegAnalyzerG units with a similar or better precision than when all instruments were calibrated on their own calibration sets and vice versa. Optical standardisation methods (techniques that modify spectra of secondary units to match those of the master unit) performed significantly poorer than other methods. Other techniques provided similar results. Pre-processing of the spectra with or without simple treatment of the predictions was sufficient to erase absorption differences and wavelength shifts.