PATProcess Analytical Technology

FDA's 2004 PAT guidance defines PAT as 'a system for designing, analysing and controlling manufacturing through timely measurements (i.e. during processing) of critical quality and performance attributes of raw and in-process materials and processes, with the goal of ensuring final product quality.' The phrase 'timely measurements' is the operative one — PAT is not necessarily in-line; it covers in-line (sensor in the stream), on-line (sample diverted, measured, returned) and at-line (sample pulled, measured at the line within seconds-to-minutes).

Real-Time Release Testing (RTRT) is the highest expression of PAT: the in-process measurements, combined with a validated process model and the control strategy, demonstrate that each batch meets all CQAs at the moment of manufacture, eliminating the need for end-product testing of those CQAs (the finished-product spec stays in the file; the test is replaced by the in-process measurement that drives it). RTRT requires the model and the measurement to be qualified to the same standard as the lab method they replace.

PAT is a system, not a sensor. The FDA framework identifies four tool categories that work together:

1. Process analysers — NIR, Raman, UV/Vis, FBRM, PVM, mass spec, in-line HPLC, in-line particle sizing, in-line pH/conductivity/dissolved O₂, capacitance-based biomass, off-gas analysis. Pick the analyser that responds to the CQA's underlying physics/chemistry.
2. Multivariate data acquisition + analysis — chemometric models (PLS, PCA, OPLS, MCR) that transform raw spectra/sensor signals into a CQA prediction. The model is the bridge between 'the sensor reads 0.234' and 'the assay is 99.6 %'.
3. Process / endpoint monitoring + control — feedback control loops that act on the analyser output (e.g. terminate a blending step when NIR-derived blend uniformity SD drops below threshold; trigger a granulation endpoint when the in-line particle-size D50 hits target).
4. Continuous improvement + knowledge management — every batch's PAT trace updates the knowledge base, feeds CPV trends, and informs the next iteration of the control strategy.

The chemometric model is the regulated artefact at the heart of every PAT application. Its lifecycle has four phases:

• Calibration — build the model from a designed calibration set spanning the expected variation in CMAs, process conditions and instrument behaviour. Use a reference method (HPLC, Karl Fischer, lab assay) of known performance. Document the pre-processing (SNV, MSC, first derivative, mean centring), the rank/latent variables and the calibration figures of merit (RMSEC, RMSECV, R²).
• External validation — challenge the model with an independent test set, ideally including new lots, instruments and operators. Report RMSEP (root mean square error of prediction), bias, slope, intercept and the working range. ICH Q2(R2) applies — the model is an analytical procedure.
• Implementation + control — install the model on the production analyser; commission with PQ runs. Define the in-control envelope (Hotelling's T², Q residuals, prediction interval) so the model self-reports when the spectrum is unlike the calibration set.
• Monitoring + maintenance — run scheduled calibration verifications (CalCheck batches), monitor T²/Q drift, recalibrate or augment the model when a CMA range expands (new supplier, new excipient grade) or when bias drifts beyond the change-control threshold. Recalibration is a change-control event with a defined re-validation requirement.

PAT is high-leverage where (a) the CQA is driven by a fast-changing in-process state, (b) end-product testing is slow, destructive or insensitive, or (c) the cost of a failed batch is high. Examples: blend uniformity by NIR (replaces stratified sampling with thousands of measurements per minute); granulation endpoint by torque + NIR moisture; coating thickness by Raman/NIR; lyophilisation endpoint by tunable diode laser O₂; bioreactor critical-feed control by capacitance + off-gas; continuous tableting RTRT for assay + content uniformity per ICH Q13.

PAT is overkill where the process is short, the CQA is robust to process variation, the lab method is fast and cheap, and validation cost outweighs the value of real-time data. Forcing PAT onto every unit operation is a common over-correction; the value is in targeting the steps that drive the highest-risk CQAs.

1. Chemometric model treated as engineering output, not a validated method — no Q2(R2)-style validation, no working range, no acceptance criteria for ongoing performance.
2. RTRT claim in the filing, but the in-process measurement is not actually controlling the batch — the operator still releases on the lab result.
3. No procedure for model maintenance — bias drifts, T²/Q residuals trend out of control, no one acts.
4. Calibration set built from a single campaign — model has no robustness to lot-to-lot variation in API or excipient.
5. Analyser PQ done once, never re-qualified after lamp/laser replacement or detector swap.
6. Real-time data not archived in an ALCOA+-compliant repository — spectra and predictions live on the instrument PC, not in a managed data lake.
7. PAT data not linked to the batch record — operator and reviewer can't see the in-process trace alongside the lab CoA.

• Streaming analyser drivers (OPC-UA, Modbus, REST) land time-series and prediction data on the batch record automatically — every PAT sample is e-signed by the system, time-stamped to the millisecond and tied to the work-order, equipment and operator on shift.
• Chemometric models are versioned as controlled documents — calibration set, RMSEC/RMSEP, working range, latent-variable count, pre-processing chain are part of the model record; deploying a new model version is a change-control event with two-person e-signature.
• Per-batch PAT trace renders alongside the BMR — the auditor sees the spectrum, the model prediction, the in-process spec, the OOS flag and the operator's response in one timeline view.
• Endpoint logic (e.g. 'stop blending when NIR-SD < 1.5 % over 20 consecutive samples') is configured in the master recipe and enforced at the kiosk; deviating requires a controlled deviation with QA e-signature.
• T² / Q-residual monitors flag when the live spectrum looks unlike the calibration set, so the operator (and QA) know in real time when the model is being asked to extrapolate.
• CPV dashboards pull PAT data alongside batch yield, deviations and APR metrics — no separate spreadsheet, no manual extracts.