Analytical Method Development

Requirement USP General Chapter ⟨921⟩, “Water Determination,” describes three families of Karl Fischer method: Method Ia/Ib (volumetric — titrant delivered by burette, endpoint detected electrometrically; suited to moderate-to-higher water content), Method Ic (coulometric — iodine generated in situ, water calculated from the electrical charge passed via Faraday's law; suited to very low water content and small samples), and the older Method II, azeotropic distillation, retained for matrices where Karl Fischer chemistry itself is unsuitable.

C1.1  The chemistry, briefly

Karl Fischer titration works by having iodine oxidize sulfur dioxide in a buffered alcohol solvent system, in a reaction that consumes water stoichiometrically — count the iodine consumed (by titrant volume, for volumetric methods, or by charge, for coulometric methods) and you have the water content. Modern reagents typically use imidazole-type buffering rather than the older pyridine-based chemistry, but the underlying iodine/sulfur-dioxide redox reaction is the same one Fischer described in 1935.

Figure C1.1

C1.2  Four stages, four ways to be quietly wrong

Practice None of the failure modes below trip an instrument alarm. Each one produces a number that looks like an ordinary result, which is exactly why they matter:

• Titrant or generator-reagent standardization drift. Karl Fischer titrant (or, for coulometric methods, the generator reagent) has to be standardized against a certified water standard on a justified schedule. A titer that has drifted since the last standardization silently shifts every result calculated from it.
• Carbonyl side-reactions. Aldehydes and ketones in the sample can react with the methanol solvent to manufacture water that was never in the original sample, inflating the result. Alternate-solvent KF reagent formulations exist specifically to suppress this side reaction for carbonyl-containing matrices.
• Sample matrix interference. Amines, mercaptans, strong oxidizers or reducers, and metal oxides or hydroxides can all interfere with the underlying redox chemistry directly, in either direction.
• Atmospheric moisture ingress. Poor technique in a humid environment, or a worn septum on the titration vessel, lets ambient water into the system during sample handling — a purely procedural failure with nothing wrong with the chemistry or the instrument.
• Endpoint-detection problems. Cloudy, colored, or poorly soluble samples can make the electrometric or visual endpoint genuinely hard to detect correctly, independent of everything else being done right.

---

Requirement Diffuse-reflectance or transmittance NIR spectra, roughly across the 700–2500 nm range, contain O–H overtone and combination bands sensitive to water content. A chemometric model — typically partial least squares (PLS) regression — is built to translate that spectral signal into a predicted percent water, in seconds, without consuming the sample. USP ⟨1119⟩ governs the instrument-qualification and chemometric model-validation principles behind this.

C2.1  The critical point: NIR doesn't measure water directly

Practice An NIR moisture result is a prediction from a model, not a direct physical measurement of water content the way Karl Fischer titration is. That model has to be calibrated and validated against a trusted reference method — almost always Karl Fischer, sometimes Loss on Drying — across the full range the method will be used for. This is the same “you must prove it against a trusted method first” idea that recurs, differently dressed, in Arc F's calibration traceability and Arc G's method suitability. An NIR result is only ever as trustworthy as the reference dataset its model was calibrated against.

C2.2  What that calibration actually looks like, worked as real numbers

Ten samples spanning roughly 1% to 6% water were measured by both the Karl Fischer reference method and the NIR model under development, and the two sets of results were plotted against each other:

Figure C2.1

A regression of NIR-predicted values against the Karl Fischer reference gives a slope of 0.996 and an intercept of 0.030 — close to the ideal 1.000 and 0.000 of perfect agreement — with R² = 0.9977. But R² alone is a weak accuracy metric here, the same way it would be for any calibration; the number a validation report actually leans on is the root mean square error of prediction (RMSEP), computed directly from the residuals between what NIR predicted and what Karl Fischer measured: RMSEP = 0.073 percentage points, with a mean bias of +0.018 percentage points across the calibration set.

Practice That RMSEP is the practical, defensible answer to “how good is this NIR method,” and it is entirely inherited from the Karl Fischer dataset it was computed against. If even one or two of those ten reference values were wrong — say, from one of Module C1's quiet failure modes — the NIR model would be calibrated to reproduce that error, and its RMSEP would look reassuring while the model itself learned the wrong answer. Building a model on a bad reference dataset does not average the error away; it teaches the model to repeat it.

---

C3.1  Intas Pharmaceuticals Limited — destroyed titration records

“An analyst destroyed CGMP records by pouring acetic acid in a trash bin containing analytical balance slips for testing the standardization of (b)(4). A QC employee stated he observed the same analyst destroy KF titration curves and balance printouts.”Intas Pharmaceuticals Limited — Warning Letter, July 28, 2023. fda.gov

The same letter documents a second, related gap: “The KF instrument used for water content testing and assay testing is capable of storing electronic data; however, this capacity was not utilized, and you did not save this data electronically.” Read against Module C1: the records destroyed here were precisely the titrant-standardization and titration-curve evidence needed to prove a KF result was valid in the first place — the instrument had a way to preserve exactly that evidence automatically, and it was never turned on.

C3.2  DuPont Nutrition USA Inc. — undocumented retesting until LOD passed

“Your personnel improperly retested pH, conductivity, loss on drying, and particle size samples.”DuPont Nutrition USA Inc. — Warning Letter, December 2, 2022. fda.gov

Practice Failing Loss on Drying (LOD) results were changed to passing results through undocumented, unjustified retesting — a different failure mode than Intas's, but the same underlying gap: a moisture result that cannot be trusted because the evidence trail behind it was not preserved or was manipulated after the fact. Whether the missing evidence is destroyed (Intas) or simply absent because an inconvenient result was quietly retested away (DuPont), the effect on trust in the reported number is the same.