PAT in the Regulatory Context: FDA Guidance and the Intent of Real-Time Data
FDA's Process Analytical Technology guidance (2004) describes PAT as a system for designing, analyzing, and controlling manufacturing through timely measurements of critical quality and performance attributes of raw and in-process materials and processes. The intent is real-time knowledge of process state — not just endpoint confirmation. In bioreactor manufacturing, this framing positions PAT not as an add-on monitoring layer but as a core element of a science-based process control strategy.
In practice, most bioprocess PAT deployments generate far more data than the deviation investigation workflow was designed to consume. A near-infrared (NIR) or Raman spectroscopy probe installed on a 200L bioreactor running a 14-day fed-batch process can generate spectral readings every 5–60 minutes — hundreds of predictions across the batch for glucose concentration, lactate, and in some implementations viable cell density and osmolality. An in-line capacitance probe logging VCD generates data at 1–5 minute intervals. These data streams, if logged and retained, are available for retrospective investigation. The challenge is that most DCS historians are not configured to ingest and store high-frequency PAT data efficiently, and most deviation investigation procedures do not specify how PAT data should be incorporated into the evidence package.
The Data Architecture Problem: DCS Historians and High-Frequency PAT Streams
OSIsoft PI (now AVEVA PI) and DeltaV Continuous Historian use exception-reporting storage algorithms — data is stored when the value changes beyond a defined compression threshold, not at every time point. For slowly varying parameters like temperature (which might be stored a few hundred times per batch at 1°C compression), this works well. For a Raman-based glucose prediction with natural spectral noise producing ±0.05 g/L reading-to-reading variability, the exception-reporting algorithm may store near-every-point or may over-compress depending on how the deadband is configured.
The consequence for deviation investigation is that the PAT data quality in the historian may not match the source data quality from the PAT analyzer. If the historian was configured with an aggressive compression deadband — to reduce storage load — the archived PAT trend may appear smoother than the actual measurement, potentially masking short-duration anomalies in the prediction. Conversely, if the compression deadband was set too tight, the historian may be storing near-continuous data for every PAT tag, creating performance issues in historian queries.
The correct approach is to configure PAT tag storage separately from process control tag storage, with a compression deadband appropriate for the natural noise floor of each PAT measurement. For Raman-based glucose predictions with ±0.1 g/L noise, a deadband of 0.15 g/L captures meaningful changes without storing noise. This configuration should be documented as part of the PAT method validation and the DCS system configuration specification.
Raman Spectroscopy in Investigation Context
Raman spectroscopy for at-line or in-line glucose and lactate measurement has moved from research-stage PAT to manufacturing-capable PAT in the last decade. In a deviation investigation context, in-line Raman glucose data provides a continuous record of nutrient availability between the 24-hour offline sampling points that most processes rely on for metabolic characterization.
Consider a scenario: a 200L pilot mAb bioreactor program in a CDMO's process development group captures a DO% excursion at hour 40. The offline glucose sample at hour 24 showed 3.2 g/L — within the target range of 2.5–5.0 g/L. The offline glucose sample at hour 48 showed 6.8 g/L — above the process target, indicating that a feed event had overloaded the culture. Without in-line glucose data, the investigation would attribute the DO% excursion to either culture metabolic acceleration (fast growth, increased OUR) or the hour-48 measurement point and work backward from there.
With Raman glucose data logged continuously, the investigation can precisely identify when glucose crossed 5.0 g/L (at approximately hour 36, based on the continuous prediction), confirm that the DO% decline began within 2 hours of that threshold crossing, and calculate that the DO% response lag of 90–120 minutes is consistent with the known OUR response dynamics for this cell line at this density. The investigation is not just more complete — it is more accurate. The root cause timeline can be established with 2-hour resolution instead of 24-hour resolution.
In-Line VCD Probes and Their Limitations as Investigation Evidence
Capacitance-based in-line VCD probes (measuring viable cell density via dielectric properties of the cell membrane) are increasingly common in mammalian cell culture bioreactors. In theory, they provide continuous VCD data that eliminates the information gap between daily sampling points. In practice, using in-line capacitance data as primary investigation evidence requires some caution.
Capacitance probes measure permittivity — a correlate of VCD that requires a calibration model built from the specific cell line, the specific medium formulation, and the specific process conditions. When any of these change — which can happen during a deviation event, since a metabolic shift or media perturbation changes the cells' dielectric properties — the probe's VCD prediction may drift from the true VCD. The in-line probe may show a stable VCD while the actual culture density is declining due to increased cell death, because dead cells temporarily retain their dielectric properties before lysing.
We are not saying in-line VCD probes are unreliable as a monitoring tool — they are valuable for real-time process control and trend monitoring. For deviation investigation purposes, in-line VCD data should be treated as a leading indicator that directs sampling rather than as a replacement for offline cell count. When a deviation investigation cites in-line VCD, the investigation record should include the most recent offline VCD confirmation and a note on the agreement between in-line and offline methods at the time of the deviation.
Incorporating PAT Data into the GMP Deviation Record
The practical question for a QA reviewer is: what PAT data should appear in the deviation record, and in what form? The FDA PAT guidance does not prescribe this specifically; the general requirement is that the deviation investigation be complete and that the evidence support the root cause conclusion under 21 CFR 211.192.
A workable framework treats PAT data as supporting evidence rather than primary record. The primary record elements are: deviation description, DCS historian tags (DO%, pH, RPM, temperature), batch record references, and offline analytical results (VCD, titer, lactate, glucose). PAT data appears as supplementary exhibits — labeled plots of the PAT time series for the relevant parameter, with the deviation window highlighted, and a note on the PAT method qualification status (is this an at-spec method with a defined acceptable prediction error, or is this a development-stage PAT measurement with exploratory data status?).
The qualification status matters because it affects how the data is weighted in the investigation. A fully qualified in-line glucose method with a RMSEP of ±0.3 g/L and documented statistical process control limits carries evidentiary weight in a deviation investigation. A research-stage Raman method used during pilot-scale process development carries different weight — valuable for hypothesis generation, not appropriate as the sole quantitative basis for a root cause conclusion.
Integration With the Deviation Investigation Workflow
The most practical integration point for PAT data in a deviation investigation workflow is at the hypothesis-testing stage: after the initial parameter family and equipment/process classification is done using DCS data, PAT data is pulled to sharpen the timeline and quantify the process state during the deviation window.
The Fermentile integrations layer supports PAT data ingestion from multiple sources — OSIsoft PI historian tags that carry PAT predictions, IDBS PIMS exports from spectroscopy software, and direct API connections where the PAT software supports it. The analytics platform aligns PAT time series with DCS process data on a unified batch timeline, allowing the investigator to see Raman glucose predictions, in-line VCD, and classical DCS parameters in a single coordinated view rather than in separate export windows. This is the technical prerequisite for the kind of evidence-based timeline reconstruction that converts a PAT-equipped investigation from a two-data-source analysis into a five-or-six-data-source analysis — without requiring the investigator to manually align timestamps across different export formats. Additional context on how PAT data integrates with the process verification framework is in the CPV batch comparison Field Report.