Our previous study (Tian et al., 2018) showed that differences between α-CEZ-Na samples sourced from three vendors (Gosun, Kyongbo, and Fujisawa) could be directly observed by PXRD; however, the above technique did not allow the detection of subtle variations, which was mitigated by the use of ssNMR.

Distinct differences were observed for the shape of peak C19 (i.e., for the intensity of the shoulder peak of C19) and the chemical shifts of peaks C14 and C9 (Figures 1A,B). Based on the computed conformations of CEZ-Na, the shoulder peak of C19 observed in ¹³C ssNMR spectra was assigned to the formation of intramolecular hydrogen bonds between N22 and N18 in conformation 2 (Tian et al., 2018). Comparison of the relative intensity of this shoulder peak for α-CEZ-Na 1–5 revealed that the fractions of conformation 2 in Gosun samples were significantly lower than those in Kyongbo samples, but close to those in Fujisawa samples. The chemical shift differences of C14 between different polymorphs equaled ~0.1 ppm, which was probably ascribed to the corresponding variation of van der Waals forces, whereas a difference of 0.4 ppm observed for the chemical shift of C9 may be attributed to changes of electrostatic interactions between sodium ion and the COO⁻ residue of C9, as well as intermolecular hydrogen bonds between crystal water and the COO⁻ residue of C9. All of these differences were influenced by the proportions of different conformations in the product and could be defined as the CQAs of α-CEZ-Na to assess the variability of the crystallization process. The spectra of α-CEZ-Na 1–3 revealed that Gosun samples featured similar ssNMR signals; however, the above CQAs were still suitable for investigating process consistency.

Standard deviations (SDs) of pretreated spectra were calculated to reveal the dispersion of response values at a given wavelength (Figure 2A). As a result, notable differences were observed in the range of 5,500–4,000 cm⁻¹, especially around 5,280, 4,431, 4,381, and 4,339 cm⁻¹. The signals at 5,280, 4,431, and 4,339 cm⁻¹ could return to their original positions around 5,164, 4,405, and 4,335 cm⁻¹, respectively; however, this behavior was not observed for the signal at 4,381 cm⁻¹. According to a previous study (Li, 2006; Lu, 2007), absorbances at 5,164, 4,405, and 4,335 cm⁻¹ can be used to extract information on O–H stretching (str.) and O–H deformation (def.) vibrations of H₂O and C–H str./def. vibrations of CH₃ and CH₂ moieties, respectively.

CEZ-Na contained only one methyl group (C17), the vibration of which was probably described by the absorption band at 4,405 cm⁻¹. The above methyl group was connected to C14 and was therefore expected to feature a vibration performance dependent on the variation of C14 electron cloud density, which was likely to happen according to the results of ssNMR analysis. Figure 2B shows that the peak at 4,431 cm⁻¹ shifted to 4,427 cm⁻¹ only for α-CEZ-Na 5, i.e., for the sample characterized by a significantly high proportion of conformation 2. On the other hand, substituent folding at C7 was expected to induce a vibrational change of C20 CH₂, and another CH₂ group (C4) might be also affected by the concomitant space narrowing. However, the vibration mode of C10 CH₂ was expected to be unaffected and was concluded to correspond to the peak at 4,339 cm⁻¹ that retained its position in all cases.

The higher proportion of conformation 2 in Kyongbo samples with lower NIR intensities (Figure 2B) indicated that folding arrangements differing in the pattern of intramolecular hydrogen bonds weakened the interaction between specific groups and NIR photons. Herein, the intensities observed for pretreated spectra at wavenumbers of 5,280, 4,431, and 4,339 cm⁻¹ were regarded as three variables (x₁, x₂, and x₃) to establish an MRL-DA model. To highlight crystallization process differences for the same vendor (Gosun), α-CEZ-Na samples of different processing periods were chosen for consistency analysis. Fifty samples from Period 1 were defined as Set 1 (limiting value: y₁ = 0 ± 0.5), and 100 samples from Period 2 were defined as Set 2 (limiting value: y₂ = 1 ± 0.5). Figure 3 shows that the above method achieved a 94% classification accuracy for the tested samples. Comparison of predicted value SDs (SD₁ = 0.31, SD₂ = 0.16) indicated that the Gosun crystallization process was modified and that its consistency improved upon going from Period 1 to Period 2. The fact that lower NIR intensities were observed for Period 2 than for Period 1 implied that the proportion of conformation 2 increased in α-CEZ-Na upon going from Period 1 to Period 2. Therefore, it would be excellent if the folded structure was needed to improve the efficacy of α-CEZ-Na. Taking into account the physical property differences between various solid-state forms, the proportion of discrepant conformations should be strictly controlled in view of safety considerations, which could be achieved by adopting the above NIR strategy to evaluate the consistency of the crystallization process and monitor the corresponding changes on-line.

In the case of CETD, the Raman spectra of specified positions were scanned to reflect the differences of solid-state forms (Figure 4). Under a microscope, CETD samples appeared as a combination of plate-like and rod-like crystals that could be distinguished based on the relative intensities of peaks at 1,658 and 1,644 cm⁻¹. For plate-like crystals, the peak at 1,658 cm⁻¹ was always more intense than that at 1,644 cm⁻¹, whereas the opposite was true for rod-like crystals. The absorption peaks around 1,658 and 1,644 cm⁻¹ were ascribed to the vibrations of free and associated amide C = O moieties at C15, respectively, and the lower absorption frequency and larger width of the latter peak were ascribed to the formation of hydrogen bonds. The above findings indicated the presence of at least two crystal types in CETD. However, Raman microscopy simply provided local information on polymorphic CETD.

The ¹³C NMR data obtained in dimethyl sulfoxide-d₆ (DMSO-d₆) were typically similar to those extracted from ssNMR spectra. Hence, the data assigned based on ¹³C NMR results served as guide for the assignment of ¹³C ssNMR data. Specifically, the ¹³C ssNMR signals of CETD were accurately assigned using the techniques of distortionless enhancement of polarization transfer, ¹H-¹H homonuclear correlation spectroscopy, heteronuclear single quantum coherence, and heteronuclear multiple bond correlation (Table 1).

Figure 5 shows that two distinguishable ssNMR peaks were observed for a number of CETD carbon sites, which implied the coexistence of two molecules in the asymmetric unit, i.e., chemical shift differences were ascribable to the asymmetric crystal packing in the solid (Wang et al., 2012). Comparison of the spectra of CETD 1–5 revealed distinct differences in the shape of C7 and C23 peaks (i.e., in the intensities of the corresponding shoulder peaks) and in the chemical shift of peaks C2 and C18. Notably, the spectra of CETD 1, 2, and 3 were similar and exhibited the following common features (1) the chemical shifts of C2 and C18 equaled 107.3 and 161.1 ppm, respectively; (2) the shoulder peak of C7 was weak or even non-existent; and (3) the shoulder peak of C23 was clearly visible. Conversely, the shoulder peak could be obviously observed for peak C7 but not for peak C23 in the cases of CETD 4 and 5, which demonstrated the possible existence of hydrogen bonds separately connecting the NH2+ residue of N22 or the CONH residue of C15 in different conformations. Furthermore, the C2 and C18 chemical shift differences between CETD 1–3 and CETD 4–5 both equaled ~0.4 ppm (Δδ = 107.3–106.9 = 0.4 ppm and Δδ = 166.1–165.7 = 0.4 ppm, respectively), which suggested that these changes might reflect variations of electrostatic interactions or hydrogen bonding connecting the COO⁻ residue of C9 and the NH2+ residue of N22. The above findings indicated the presence of two or more different structural orientations (lattice structures and/or molecular conformations) in the studied CETD samples, the proportions of which varied between batches. Thus, these theoretical results identified subtle variations in different batches of CETD and demonstrated that these variations can be regarded as CQAs of the corresponding crystallization process.

All NIR spectra of CETD were transformed by SG(17, 3) 1st Der and vector normalization, and SDs at each wavelength were calculated for the intensities of these pretreated spectra (Figure 6). The largest variations were observed for signals at 5,211, 5,284, and 5,369 cm⁻¹, which indicated the need to scrutinize peaks located at 5,168, 5,238, and 5,273 cm⁻¹ in original spectra (O–H str. and O–H def. of H₂O, C = O str. second overtone of CONH, and C = O str. second overtone of COOH, respectively) (Li, 2006; Lu, 2007). According to the molecular structure of CETD, the transformed intensities at 5,284 and 5,369 cm⁻¹ could be assigned to the vibration of CONH at C15 and COOH at C9, and the above results were therefore in good agreement with those of theoretical CQAs analyses.

All samples were classified according to the intensities of peaks at 5,211, 5,284, and 5,369 cm⁻¹ in pretreated spectra using HCA. Based on the results in Figure 7, 96 samples were well divided into two classes, namely CETD 1–3 and CETD 4–5, in agreement with the results of ssNMR spectroscopy. Microscopy-coupled Raman imaging revealed that CETD 1–3 had a higher proportion of rod-like crystals than CETD 4–5, as reflected by the higher I₁₆₄₄/I₁₆₅₈ > 1.00 value of the former. Thus, NIR was proven to be a suitable technique for determining the CQAs of CETD.