Chromatographic Integration: The Step Nobody Audits

Between the detector's electrical signal and the percentage printed on the certificate sits a software algorithm and a set of human choices. The algorithm decides where peaks begin and end. The choices decide where the baseline runs, what counts as a peak, and which numbers reach the report.

None of this appears on the certificate. All of it decides what the certificate says.

A small peak is a subtraction

A large peak's area is robust, because the signal is far above the noise and the boundaries barely matter. A small peak's area is the difference between two nearly equal quantities, and every choice in the calculation moves it.

Raise the peak threshold slightly and a 0.04 per cent impurity disappears from the report without any human decision, because the software classifies it as noise. Lower the baseline sensitivity and a valley between two peaks becomes a saddle, splitting one impurity into two or merging two into one. Shift the baseline start point and the area under a tailing peak changes by an amount that is a rounding error for the assay and a specification breach for the impurity.

None of these operations is exotic. All of them are available as parameters in every data system, and most of them are left at defaults that nobody has reviewed.

Manual integration is normal. Undocumented manual integration is not.

There is a persistent belief that manual integration is a red flag. It is not. Real chromatography produces shoulders, solvent fronts, baseline steps after gradient changes and peaks that a fixed algorithm handles badly. The pharmacopoeial general chapters, published by the EDQM and the USP, acknowledge that intervention is sometimes necessary.

What separates a controlled laboratory from an unreliable one is not whether the button exists. It is whether the use of the button is a defined event: permitted by the method, recorded with a reason, attributed to a named person, and visible in the audit trail beside the original result.

The failure pattern is characteristic and easy to describe. Automatic results everywhere, manual integration only on the runs that were failing, and no procedure that says when the intervention is allowed. The pattern is invisible on the certificate and obvious in the audit trail, which is why the audit trail is the first thing an inspector opens after a disputed result.

Defaults differ between systems

The same raw data file integrated on two different data systems, with each at its own defaults, can produce impurity totals that differ by more than the specification limit. Peak detection thresholds, valley treatment and minimum area settings are vendor choices, not industry standards.

That is why the parameters belong in the method document, named and fixed, rather than in the instrument's local settings. A method that specifies the column, the mobile phase and the wavelength but not the integration parameters has described three quarters of a method, and the missing quarter is the one nearest the result.

It is also why a purity by area normalisation is so sensitive to these choices: every peak's share of the total depends on how many peaks the algorithm found and where it put the boundaries.

What the audit trail has to show

Event	What the trail records
Original acquisition	File, date, instrument, operator
Automatic integration	Parameters applied, as a set
Manual change	What changed, who, when, why
Reprocessing	That a new result replaced an old one, with both retained
Report generation	Which version of the result reached the report

The European expectations for these records are set out in EudraLex volume 4, and the critical one is the fourth row: a reprocessed result must keep its predecessor visible. A report that shows only the final number, with the earlier versions deleted, has converted a measurement into an assertion.

The data integrity consequences run deeper than integration alone, but integration is where they become cheapest to check, because one raw file and its trail tell the whole story.

A test that costs one question

Ask for the integration parameters of the method that produced a certificate in front of you. A laboratory that can produce them in a paragraph has them written down, which means they are controlled. A laboratory that answers "the system default" has a method with an uncontrolled step sitting directly between the signal and the number.

The same question resolves most disputes about a result sitting at the limit, because the first thing to establish is whether the number is a property of the batch or of the baseline.

Where this leaves a supply-chain partner

We fix the integration parameters in the method, lock them for routine use, and record any intervention with a reason that a stranger could read. Where a customer's method specifies its own parameters, we run theirs, because the parameters are part of the method's identity.

That position produces fewer clean-looking reports and more defensible ones. The difference between the two is invisible until the day somebody asks, and on that day only one of them still exists.

Where the parameters belong

The method document is where integration parameters live, named as a set: threshold, minimum area, minimum height, peak width, baseline treatment, and the rules for shoulders. Stated as a set, they make the method reproducible across data systems and across years. Left as instrument defaults, they make the method a property of whichever workstation ran the batch.

Vendor defaults differ for good reasons: each system balances noise rejection against small-peak detection differently. That is precisely why the choice has to be made by the method rather than by the software, because the software does not know where the specification limit sits.

The compendial general chapters published by the EDQM and the USP acknowledge the variability and place the responsibility on the laboratory to define and document its settings. The European expectations for the records these settings generate are in the GMP annexes on computerised systems, and the short version is that a reprocessed result keeps its history.

Three scenarios worth knowing by shape

The shoulder. A main peak with a shoulder is either one impurity eluting too close to integrate or a column problem. Integrating the shoulder as a separate peak reports an impurity that may be an artifact; flattening it into the main peak reports a clean result that may be a miss. The method should say which treatment applies, because both are defensible and neither is self-evident.

The solvent front. Early-eluting material overlapping the injection peak is usually excluded by a start-time rule rather than by integration skill. Where the excluded region is specified in the method, the exclusion is a design choice; where it is decided run by run, it is a degree of freedom nobody wants near an impurity result.

The baseline step. A gradient run's baseline steps at the gradient change, and a small peak sitting on the step is measured against a baseline whose position is a choice. Drop lines and tangent skims give different answers, and the difference is largest exactly where the specification is tightest.

Each scenario has a right answer for a given method. None has a universal one, which is why the method document, not the analyst's judgement, is the place the answer lives. A result near the limit that was produced under an ambiguous scenario is a question regardless of what the number says.

The one-paragraph test

Ask a laboratory to describe its integration control in a paragraph. A controlled one says: parameters are fixed in the method, manual changes are permitted under a written rule, every change carries a reason, and the audit trail keeps the original. An uncontrolled one says: the system handles it, or the senior analyst handles it.

Both laboratories produce certificates that look identical. Only one of them can reproduce last year's number on last year's file, which is the only test that matters when the question arrives.

Reprocessing and the version question

Reprocessing a chromatogram is ordinary: a parameter set was applied incorrectly, or a peak requires a defined treatment. What distinguishes a controlled reprocessing from an uncontrolled one is the version trail. The original result stays visible, the new result is dated and attributed, and the reason is recorded in terms a reader who was not present can follow.

Two failure modes are common and both are documentary rather than technical. The first is overwriting, where the original result is replaced with no trace, which converts a correction into a deletion. The second is silent proliferation, where five versions exist in the system and the report cites whichever was most convenient, with nothing to say which is authoritative.

The remedy in both cases is a rule that names which version reaches the certificate: the latest approved version, approved by a named role, with its predecessor retained. That sentence belongs in the method or the data system's procedure, and it is the sentence auditors look for after a disputed result.

Integration and the customer's file

For a buyer, the practical relevance is narrow and real. An impurity result that sits just inside a limit is more sensitive to integration choices than any other line on the certificate, because the peak is small and the baseline decision dominates its area. When such a result is disputed, the first document to request is the integration history of that run, not the method.

The second request is the parameter set the method specifies, so that the two can be compared. A run that deviated from the method's parameters is a finding whether or not the deviation helped. A run that followed them is a measurement, and the argument moves to where it belongs: the specification and the batch.

Frequently asked questions

Is manual integration allowed?

Yes, where the method defines when it may be used, the change is recorded with a reason, and the audit trail shows both the original and the modified result. What is not acceptable is manual integration as an undocumented habit, applied selectively to the runs that need it.

Why does the baseline matter so much for small peaks?

Because a small peak's area is the difference between two nearly equal quantities: the signal and the baseline beneath it. Moving the baseline a fraction of a per cent can double or halve a 0.05 per cent impurity, which is why integration parameters deserve the same control as the method itself.

What are the parameters that decide the result?

Peak threshold, minimum area, peak width, baseline sensitivity, and the treatment of shoulders and valleys. Each has a default in the data system, and the defaults differ between systems, which is why the same raw file can produce different numbers on two instruments.

Who should be allowed to change integration?

A defined role, under a procedure that distinguishes a permitted adjustment from a redefinition. The distinction is written down, and every change carries the name of the person and the reason.

Can integration hide an impurity entirely?

A threshold set above a small peak's height will exclude it from the report without any manual intervention, which is why the parameters belong in the method and not in the operator's preferences. An audit trail that shows only manual changes misses this case entirely.

What should a buyer ask for?

The integration parameters as written in the method, and confirmation that they are locked for routine use. Asking for the raw files is possible but rarely necessary; the parameters answer most of the question.