Statistical Experimental Design for Pat Method Development

lytical technology (PAT) guidance has reinforced the commitment by the FDA to pursue a new strategy of regulatory oversight in the pharmaceutical industry. Key ideas in the guidance include riskbased design and implementation of processes supported by real-time information. The framework requires that PAT implementations integrate design, measurement and control of a process. Within this philosophy, it is important to recognise that the guidance requires a new approach to analytical methods validation. Process measurements integral to PAT, such as NIR spectroscopy, must be validated in a manner consistent with this new approach. In this column we explore the relationship between design of experiments (DOE) and a multivariate calibration. The “design” component of a PAT application often includes DOE studies of the manufacturing process. Data collected during these experiments are used as the basis for developing process models. In this column we examine how the structure of a DOE can affect the validation of a multivariate calibration. Pharmaceutical scientists, like all NIR spectroscopists, face similar challenges during calibration development. It is possible using NIR spectroscopy to calculate a complex multivariate calibration model which is apparently sensitive, robust and specific, but is a veritable house of cards. Typical calibration development errors include: spurious correlations among constituents, insufficient leverage or poorly distributed samples, and insufficient demonstration of linearity. Specifically, the effect of aliasing components (where the model inadequately describes the two sources of variation—two sources of variation are described as one) due to an active pharmaceutical ingredient (API) “spiking” experimental design is examined. The specific relevance of such a relationship with respect to validation is the inherent lack of specificity of the resulting model. Such flaws in a DOE/calibration may cause inaccurate results causing incorrect processing of materials and are likely to be spotted by an experienced scientist, including FDA inspectors and reviewers, causing regulatory difficulties. The two experimental designs considered for this column are illustrated in Figure 1. Design 1, represented by open circles in Figure 1, is a typical design employed by many NIR practitioners. The three levels of API are generated from this experiment by addition of API or an excipient blend to a nominal batch. A second design (design 2) is illustrated in Figure 1 in which both the excipient ratio and the API level is varied. Design 2 represents a greatly improved set of samples for calibration of a three-component system. The variation of constituents in the samples is as close to orthogonal as can be expected in a ternary system. Simulated spectra representing samples in the two designed experiments are used to compare the effects of DOE on the resulting calibration. The spectra were created using linear combinations of a set of artificial pure-component spectra according to the designs in Figure 1. The pure component spectrum for each constituent is composed of three Gaussian peaks of different sensitivity, location and bandwidth as illustratStatistical experimental design for PAT method development