Requirement A compendial or regulatory requirement, sourced at the end of each module. Practice Established practice or a working heuristic — not a requirement.

Module 1

What integration actually computes

Before any question of technique or compliance, a plain mechanical one: what did the data system do to turn a detector trace into the number on the certificate of analysis?

By the end of this module you should be able to

Describe, for any chromatogram, the three independent decisions that produce a peak area.
Explain how slope thresholding sets the start and end of a peak, and why a tailing peak's integration window is not symmetrical about its apex.
State the effect of sampling rate on area accuracy, and distinguish the practice guidance from anything compendial.
Choose between area and height for a given measurement and justify the choice.

1.1 From detector to number

A chromatographic detector does not produce a curve. It produces a sequence of numbers — response sampled at a fixed rate, typically a few hertz to a few tens of hertz. Everything drawn on screen is interpolation between those samples, and everything calculated from it is arithmetic on them.

Integration is that arithmetic. The reported area is the sum, across the integration window, of the sampled response minus a constructed baseline, multiplied by the sampling interval. Written out, there are exactly three independent decisions inside it:

Where the peak starts.
Where the peak ends.
What the baseline does between those two points.

Nothing else. Every dispute about integration — every warning letter observation, every disagreement in a data review — reduces to one of those three, and usually to the third. It is worth being precise about this because the phrase “the peak area” suggests a property of the sample. It is not. It is a property of the sample and three decisions.

1.2 How the data system finds the peak

Peak detection is slope-based. The system computes the first derivative of the signal over a short smoothing window, then compares it against a threshold. The peak is declared to start when the rising slope first exceeds that threshold, the apex is where the slope crosses zero, and the peak ends when the slope on the descending side returns to within the threshold.

Peak start and end are found by slope, not by eye. The data system marks the start where the rising slope first exceeds a threshold, the apex where slope crosses zero, and the end where the descending slope returns within threshold. The shaded region is the reported area: signal minus a baseline the system drew.

Two things in that figure are worth dwelling on. First, the detected window (2.85–3.27 min for an apex at 3.00) is not symmetrical about the apex. The peak tails, so its descending edge is shallower than its rising edge and takes longer to fall back within the slope threshold. An integration window that looks lopsided is usually correct, not evidence of tampering.

Second, the baseline in that figure is a straight line the system drew. It is not something the detector measured.

The parameters that control detection

Parameter	What it actually does	Symptom when wrong
Slope sensitivity / threshold	Sets how steep the signal must be rising before a peak is declared	Too sensitive: noise integrated as peaks. Too insensitive: small peaks missed, or peak start clipped inward
Peak width	Sets the width of the smoothing window used to compute slope	Too narrow: noise fragments one peak into several. Too wide: narrow peaks smoothed away
Area / height reject	Discards detected peaks below a cut-off	Set carelessly it can silently delete a reportable impurity
Bunching / data rate	How many raw points are averaged per stored point	Over-bunching distorts narrow peaks and biases area

None of these values is specified by any pharmacopoeia. They belong to the method, and their justification belongs in the method development record.

1.3 Sampling rate

A peak has to be described by enough points for its area to be reliable. The commonly cited working range is fifteen to thirty points across the peak, measured at half height.Practice Below roughly ten, area precision degrades and the apex — and therefore the retention time — starts to move between injections. Far above thirty, nothing improves except file size, and detector noise is captured more faithfully.

This is a rule of practice, not a requirement. USP <621> sets no data-rate specification. If an inspector asks why the method runs at 10 Hz, the answer must come from your development data, not from a chapter.

A trap worth knowing

Detector response time (or time constant, or filter setting) is a separate control from data rate, and on some instruments changing one does not change the other. A fast data rate combined with a slow response time gives you many points describing a peak the detector has already smoothed. The two must be set together.Practice

1.4 Area or height

Area is the default measure for quantitation because it is proportional to the amount of analyte regardless of how the peak is shaped. Height is proportional only if the peak shape is constant.

The same peak on a flat baseline (dashed) and a rising one (solid). Height is read vertically from the constructed baseline beneath the apex, so only the baseline at that one point affects it. Area accumulates whatever the baseline gets wrong across the entire integration window.

The trade-off is where each is vulnerable. Area accumulates baseline error across the whole integration window, so a baseline that is wrong by a small amount over a wide peak produces a large error. Height depends on the baseline at a single point, directly beneath the apex, so baseline error affects it far less.

Against that, height is sensitive to anything that changes peak shape: column ageing, temperature drift, injection solvent mismatch, overload. Area is largely indifferent to those.

Height is therefore sometimes the better choice for very small peaks on a drifting or noisy baseline, or for peaks that are partly merged.Practice What matters for defensibility is that the choice is stated in the method, justified during development, and applied consistently — not selected per injection after seeing the result.

1.5 What the number is not

The reported area is the peak minus an assumption. Two competent analysts given identical raw data and no further instruction can produce materially different numbers, both defensible in isolation. That is not a failure of either analyst. It is a property of the measurement, and it is precisely why the assumption has to be fixed by the method rather than chosen at the screen.

Check yourself

Why is a tailing peak's integration window longer after the apex than before it?

Because peak detection is slope-based. The descending edge of a tailing peak is shallower than its rising edge, so the slope takes longer to fall back within the threshold that closes the peak. In the figure above the apex sits at 3.00 minutes but the window runs 2.85–3.27 — about 1.8 times as long after the apex as before it. An asymmetric window is normally correct.

An analyst reports that switching the method from area to height ‘fixed’ a drifting result. What should a reviewer ask?

Height is less sensitive to baseline error than area, so switching can certainly make a drift-affected result look more stable — but it treats the symptom. The question is why the baseline is drifting. Switching the measurement basis after seeing results, rather than deciding it during development, also changes the method: it needs the same change control, and the validation data must support the basis actually used.

Where in the compendia is the required number of data points across a peak specified?

Nowhere. The fifteen-to-thirty guidance is well-established practice, not a compendial requirement. USP <621> is silent on data rate. The justification for the rate your method uses has to come from your own development data.

Sources for this module

USP General Chapter <621> Chromatography — for what it does and does not specify. It contains no data-rate, slope-sensitivity or peak-width requirement.

Module 2

Baseline construction, and what each one costs

The baseline is the third decision, and the one that moves the number furthest. It is also the only part of integration about which the pharmacopoeia expresses any preference at all.

By the end of this module you should be able to

Name and draw the five baseline constructions in routine use.
Predict the direction and rough magnitude of the error each one introduces on a given peak geometry.
State what USP <621> actually says about skimming, and identify the decision it leaves open.
Justify a baseline construction in writing, in terms a reviewer can check.

2.1 There is no baseline in the data

There is a detector signal. Beneath a peak, the contribution that would have been there in the absence of the analyte cannot be observed, because the analyte is there. Every baseline is therefore an interpolation across a region where no measurement exists.

This is not a philosophical point. It is the reason baseline construction is a method decision requiring justification, rather than an observation requiring only care.

2.2 The constructions

The four constructions you will meet. Top row, a fused pair of comparable size: a group baseline carried beneath both peaks with a vertical drop line at the valley, versus a valley-to-valley baseline that follows the signal through the valley. Bottom row, a small rider on the tail of a much larger peak, skimmed with a straight line and with an exponential fitted to the parent's decay. All four traces are exponentially modified Gaussians, so the underlying areas are known exactly.

A fifth, forced or manual baseline, is not a construction so much as the absence of one: the analyst places the anchors by hand. It has legitimate uses and is the subject of Module 7.

2.3 What each one costs

The bottom row of that figure is the case worth measuring. The rider's true area is 14.0 units by construction. The straight skim returns 11.1, an error of -20.4%. The exponential skim returns 13.8, an error of -1.3%.

The reason is geometric rather than procedural, and it generalises. An exponential tail is convex. A straight line drawn between two points on a convex curve lies above the curve everywhere in between. So a straight skim sets the baseline too high across the whole rider and loses area — systematically, in a predictable direction, every time.

The same geometry explains the opposite error. A baseline carried flat beneath a rider — a drop line, or a projected baseline — sits below the parent's tail, so the rider is credited with a slab of tail that belongs to the peak it is sitting on. That error is an over-report, and on a small rider it can exceed the rider itself.

The shape of it

Flat baseline under a rider → over-report. Straight chord across a curved tail → under-report. A baseline that matches the parent's curvature → approximately right. If you remember one thing from this module, remember the directions; you can then predict which way any given construction will move a result before you apply it.

2.4 What the pharmacopoeia says — and what it leaves open

USP <621> addresses integration in a single sentence, under Other Considerations:

“Integration of the peak area of any impurity that is not completely separated from the principal peak is preferably performed by tangential skim.” USP General Chapter <621> ChromatographyRequirement

That is the entire compendial position on integration technique. Note carefully what it does not resolve: it expresses a preference for skimming without specifying which skim. On the geometry above, that unspecified choice is worth 19 percentage points on the reported result.

A method that says only “riders are tangentially skimmed” has therefore not actually specified the integration. It has deferred it to whatever the data system happens to default to. That is a finding waiting to happen, and it is entirely avoidable by naming the skim.

2.5 Choosing, and then not choosing again

The working heuristics below are widely used and reasonable. None of them is compendial.Practice

Geometry	Usual construction	Why
Peak returns to baseline on both sides	Baseline-to-baseline	No interpolation needed under the peak; the easy and unambiguous case
Two peaks of comparable size sharing a valley	Group baseline with a drop line at the valley	Neither peak is a rider on the other; the error is shared rather than loaded onto the smaller
Small rider on the flank of a much larger peak	Tangential skim — and specify straight or exponential	The parent's flank is the local baseline for the rider
Baseline drifting through the peak	Construction that follows the drift	A flat baseline under a rising signal counts drift as analyte

A frequently quoted trigger for skimming rather than dropping is a parent-to-rider height ratio of roughly ten to one or greater.Practice Treat it as an orientation, not a rule: it appears in no compendium, and the honest justification is the geometry, not the ratio.

At the bench

The construction is a property of the method, not of the injection. It should be determined during development against representative worst-case chromatograms, written into the processing method, and then applied to every injection in the sequence without further decision. The moment the construction varies between injections of the same sequence, someone has to explain why — and that explanation is what Module 7 and Module 8 are about.

▸

Sandbox, scenario 1 — Rider on the tail. Cycle the three constructions on a real rider and watch the reported figure move from −12% to +78%. Scenario 2 shows the fused-pair case, where a straight chord does not merely bias the result but drives it negative.

Check yourself

A method states that partially resolved impurities are ‘integrated by tangential skim’. Is that sufficient?

No. It follows the USP <621> preference, but it does not say whether the skim is straight or exponential. On the geometry in this module those two differ by about 19 percentage points on the reported impurity. Unspecified, the answer is set by whatever the data system defaults to — which is not a method decision, and cannot be defended as one.

Without calculating anything, which way will a flat projected baseline bias a small rider sitting on a large peak's tail, and why?

It will over-report it. The parent peak is still descending beneath the rider, so a flat baseline sits below the parent's tail and the rider is credited with the area between the two. On a small rider that slab of borrowed tail can be larger than the rider itself.

Why does a straight skim under-report rather than over-report?

Because an exponential tail is convex. A straight line between two points on a convex curve passes above the curve everywhere in between, so the baseline is set too high across the rider and area is lost. The error is systematic and always in the same direction, which is what makes it predictable — and correctable by fitting the curvature instead.

Sources for this module

USP General Chapter <621> Chromatography, Other Considerations — the tangential skim preference.
The height-ratio heuristic for skimming appears in vendor and industry practice guidance, not in any pharmacopoeia.

Module 3

Difficult peaks, and telling defects from features

Most integration disputes happen on a small number of recognisable geometries. Knowing them by sight — and knowing which are chromatographic defects rather than sample properties — settles most arguments before they start.

By the end of this module you should be able to

Relate resolution to the overlap it implies, and explain why the same resolution is far worse for a minor component.
Recognise shoulders, riders, fused pairs, fronting, tailing and split peaks, and name the usual causes.
Distinguish a baseline disturbance from an analyte peak.
Decide when an integration difficulty should be escalated as a chromatography problem.

3.1 Resolution and what it implies

USP <621> defines resolution from peak widths at half height:Requirement

R_S = 1.18 (t_R2 − t_R1) / (W_h1 + W_h2)

The chapter gives the formula but sets no universal numeric limit — acceptance criteria belong to the individual monograph or method. The familiar figures below are classical results for equal, symmetrical Gaussian peaks, quoted here as orientation only.Practice

Resolution	Conventional description (equal, symmetrical peaks)
1.5	Essentially baseline resolved; overlap of the order of 0.1%
1.0	Visibly separated with a distinct valley; overlap of the order of 2–3%
< 1.0	Valley does not return toward baseline; integration construction begins to dominate the result

Equal-sized peaks

Minor peak at 5% of the major — vertical scale expanded to make it visible

The same resolution is not the same problem. Top row, two equal peaks at R_S = 0.80, 1.20 and 1.50. Bottom row, identical separations, but the later peak is 5% of the size of the earlier one. At R_S = 1.50 the equal pair is comfortably separated while the minor peak is still sitting on its neighbour's tail. Resolution alone does not tell you whether a measurement is safe.

This asymmetry is the single most useful thing in the module. A system suitability criterion of “R_S not less than 1.5” is a reasonable guarantee for two major components. For a 0.1% impurity eluting on the tail of the API, it guarantees very little, because the quantity being measured is small compared with the interference being interpolated away.

3.2 The geometries

What you see	Usual causes	Integration consequence
Rider — small peak on a flank, baseline never returns	Impurity or degradant closely related to the main component	Skim or drop decision dominates; see Module 2
Shoulder — inflection with no distinct apex	Very poor resolution; sometimes two unresolved components	Detection may not declare a peak at all; area highly construction-dependent
Tailing — T above about 1.5	Secondary silanol interaction, column overload, extra-column volume, contaminated inlet frit, strong sample solvent	Extends the integration window, buries later small peaks in the tail
Fronting — T below about 0.9	Column overload, sample solvent stronger than mobile phase, void or channelling	Peak start difficult to locate; area biased by baseline placement on the leading edge
Split peak	Column void, partially blocked frit, injection solvent mismatch	Detection may report two peaks where one analyte exists

The tailing and fronting thresholds above are working descriptions, not limits. The compendial limit is the symmetry factor range in Module 6.

3.3 Disturbances that are not peaks

Four things routinely appear in a chromatogram that are not analyte, and all four are integrable if the method lets them be:

The solvent front / unretained peak. Everything not retained elutes together near the void volume, often with a large disturbance.
Gradient drift. The mobile phase composition changes, so the baseline rises or falls continuously. A flat baseline under a peak eluting in a gradient counts drift as analyte — a large error, and a consistent one across the whole sequence.
Negative excursions. Injection artefacts, refractive index effects, or a sample diluent that absorbs less than the mobile phase. If the excursion has not recovered by the time the analyte elutes, no baseline drawn through it is a measurement.
Carryover and ghost peaks. Material from a previous injection, or from the mobile phase itself, eluting in the current run.

The blank settles most of these

A blank injection processed with the same processing method is the fastest way to separate a disturbance from an analyte. If the feature appears in the blank, it is not sample. If it appears in the blank at a different size, it is partly sample and partly not, and that needs resolving before any integration decision is meaningful. The method should state which peaks are excluded on the basis of the blank, and why.

3.4 When to stop integrating and escalate

The judgement this module is really training is the decision to stop. Escalate rather than integrate when:

the difference between two defensible constructions exceeds the method's precision;
the peak of interest cannot be distinguished from a baseline disturbance;
a system suitability descriptor is outside its limit (Module 6);
the geometry has changed from what the method was developed and validated against.

In each of those cases the integration is not the problem, and adjusting it will not produce a reportable result — only a plausible-looking one.

▸

Sandbox, scenarios 2 and 5. The fused pair at R_S ≈ 1.2 shows two defensible constructions disagreeing by more than the analytical variability. The negative dip shows a disturbance reaching the peak, where a forced flat baseline loses roughly two fifths of the analyte.

Check yourself

System suitability requires R_S ≥ 1.5 between the API and its nearest impurity, and the run passes. Is the 0.08% impurity on the API's tail safely measurable?

Not necessarily. Resolution criteria behave very differently for unequal peaks. At the same R_S, a minor component sitting on a much larger neighbour's tail is still heavily dependent on the baseline interpolated beneath it — as the lower row of figures shows. Passing R_S tells you the peaks are separated; it does not tell you the interference under the small one is negligible relative to the quantity being measured.

A peak appears at 2.1 minutes in every injection including the blank, but is larger in the samples. How should it be handled?

It is partly sample and partly not, so neither integrating it wholesale nor excluding it wholesale is correct. This needs resolving before any integration decision: identify the blank contribution, decide on a documented basis how it is handled, and write that basis into the method. A decision taken at the integration screen, per injection, is exactly the pattern that later reads as manipulation.

Give two reasons a peak might tail, one chromatographic and one avoidable at the injection step.

Chromatographic: secondary interaction with residual silanols on the stationary phase, or a contaminated or partially blocked inlet frit. At the injection step: a sample solvent stronger than the mobile phase, or simple mass overload. The distinction matters because the first needs a column or method change and the second can often be corrected the same day.

Sources for this module

USP General Chapter <621> Chromatography — resolution formula; note that no universal numeric acceptance criterion is given.
Overlap percentages for given resolutions are classical results for equal, symmetrical Gaussian peaks and are orientation only for real, tailing peaks.

Module 4

The processing method, and the honest use of timed events

Integration parameters are not settings. They are a controlled document that determines a reported result, and they should decide every injection in a sequence identically.

By the end of this module you should be able to

Describe the processing method as a controlled object with a lifecycle.
Distinguish a legitimate timed integration event from one that exists to reach a result.
Explain why an event applied to some injections and not others is the reviewable signal.
Specify what a processing method must contain to make manual intervention unnecessary.

4.1 The processing method is the real deliverable

Everything in Modules 1 to 3 — slope sensitivity, peak width, rejects, baseline construction, area or height — is stored together as a processing method. That object, not the individual chromatogram, is what determines the reported result for every injection it is applied to.

It follows that the processing method has all the properties of a controlled document. It has a version. Changes to it are changes to the analytical procedure. Who may change it, and on what authority, is a question with a regulatory answer, and it is the subject of Module 7. Here we are concerned only with building one that works.

4.2 The goal: no decisions left at the bench

A well-built processing method integrates the entire sequence — standards, samples, blanks, system suitability injections — correctly and identically, with no intervention. That is not an aspiration; for a well-behaved method it is routinely achievable, and where it is not achievable the reason is usually the chromatography rather than the software.

The practical test is simple. Apply the processing method to the full sequence, then look at every chromatogram. If any of them needed a hand, the method is not finished.

4.3 Timed events

A timed event applies a change to the integration during a defined retention window: integration off, baseline hold, forced skim, altered sensitivity, and so on. Used properly they are how a processing method copes with predictable, characterised features of the chromatogram.

Legitimate use	The same tool, misused
Integration off across the solvent front, because nothing there is reportable and the disturbance is characterised	Integration off across a window that happens to contain an impurity, applied after the impurity was seen
Forced tangential skim in a window where a known rider always elutes	Forced skim applied to one injection because the drop-line result was out of specification
Baseline hold through a reproducible gradient artefact demonstrated in the blank	Baseline hold used to flatten a disturbance nobody has characterised
Altered sensitivity in a late region where peaks are broad, set during development	Altered sensitivity tuned per injection until a peak appears or disappears

Notice that the left and right columns contain the same four tools. Nothing distinguishes them technically. What distinguishes them is when the decision was made and against what evidence: during development, against representative chromatograms, applied to everything — or after the result was seen, against this injection, applied to one.

The reviewable signal

An integration event that applies to some injections in a sequence and not others is the single most reviewable thing in chromatographic data. It is visible without any special tooling, it has an obvious question attached to it, and the answer is either a characterised difference between those injections or it is nothing. Reviewers should look for it first; analysts should expect it to be looked for.

4.4 Building one that holds

Develop against worst case, not typical case. The chromatograms used to set the parameters should include the highest expected impurity loading, aged column, and the extremes of whatever the method's robustness study varied.
Set the baseline construction explicitly, including which skim, per Module 2.
Set rejects deliberately. An area reject exists to suppress noise, not to suppress small peaks. Its value should be tied to the reporting threshold, and the relationship should be written down.
Characterise every timed event with the evidence that justifies it — normally a blank or a placebo chromatogram demonstrating the feature it addresses.
Qualify the finished method by applying it unaltered to a representative sequence and confirming that no chromatogram requires intervention.
Lock and version it, and treat subsequent changes as changes to the analytical procedure.

At the bench

If you find yourself adjusting integration on a routine sample, the useful reflex is not “how do I integrate this?” but “why did the processing method not handle this, and has something changed?” A processing method that suddenly needs help is usually reporting a real change — an ageing column, a new impurity, a shifted retention. That is information, and adjusting the integration discards it.

▸

Sandbox — audit trail panel. Every limit move and mode change is logged with a sequence number and the resulting figure. Watch what a record of your own decision-making looks like when someone else reads it.

Check yourself

What technical feature distinguishes a legitimate timed integration event from an illegitimate one?

None. They are the same tools. The distinction is entirely in when the decision was made and on what evidence: set during development against representative chromatograms and applied to the whole sequence, versus applied after seeing a result, to one injection. This is why the justification and its timing — not the event itself — are what has to be recorded.

A sequence of twenty injections has a manual integration event on injections 7 and 13 only. What is the first question?

What is different about injections 7 and 13? There are only two possible answers. Either there is a characterised difference in those chromatograms, in which case it should be documented and the processing method probably should have handled it — or there is not, in which case the events were applied in response to the results. The selective pattern is what makes the question unavoidable.

How should an area reject value be set?

By reference to the reporting threshold, with the relationship written down. An area reject exists to stop noise being integrated as peaks. Set higher than it needs to be, it silently deletes small peaks — potentially reportable ones — and leaves no trace on the chromatogram that anything was suppressed.

Sources for this module

FDA Data Integrity and Compliance With Drug CGMP: Questions and Answers (Final, December 2018) — for the control and change expectations that attach to the processing method; treated fully in Module 7.

Module 5

Impurities: thresholds, response factors, and what integration decides

In an assay, an integration error moves a result. In a related-substances method, it can determine whether a result exists at all.

By the end of this module you should be able to

State the ICH reporting, identification and qualification thresholds and how they vary with daily dose.
Use the correct current terminology for the reporting threshold.
Explain what a relative response factor is, when one is needed, and how it is applied.
Describe how an integration decision interacts with a reporting threshold.

5.1 The three thresholds

ICH sets three thresholds for impurities, and they do different jobs. Below the reporting threshold an impurity need not be reported at all. Above the identification threshold its structure must be established. Above the qualification threshold its biological safety must be established.Requirement

Drug substances — ICH Q3A(R2), Attachment 1

Maximum daily dose	Reporting threshold	Identification threshold	Qualification threshold
≤ 2 g/day	0.05%	0.10% or 1.0 mg/day intake, whichever is lower	0.15% or 1.0 mg/day intake, whichever is lower
> 2 g/day	0.03%	0.05%	0.05%

Drug products — ICH Q3B(R2), degradation products

Threshold	Maximum daily dose	Value
Reporting	≤ 1 g	0.1%
Reporting	> 1 g	0.05%
Identification	< 1 mg	1.0% or 5 µg TDI, whichever is lower
	1 mg – 10 mg	0.5% or 20 µg TDI, whichever is lower
	> 10 mg – 2 g	0.2% or 2 mg TDI, whichever is lower
	> 2 g	0.10%
Qualification	< 10 mg	1.0% or 50 µg TDI, whichever is lower
	10 mg – 100 mg	0.5% or 200 µg TDI, whichever is lower
	> 100 mg – 2 g	0.2% or 3 mg TDI, whichever is lower
	> 2 g	0.15%

Percentages are of the drug substance; TDI is total daily intake. Higher reporting thresholds require scientific justification, and lower thresholds may be appropriate for unusually toxic impurities.

5.2 A terminology change worth catching

Check your SOPs for this

USP <621> has replaced the term “disregard limit” with “reporting threshold”. Procedures still written around a disregard limit are using superseded language. This is a small thing that reads badly in an inspection, because it suggests the procedure has not been reviewed against the current chapter.

5.3 How the number is calculated

Approach	What it does	Where it is vulnerable
Area normalisation	Each impurity expressed as a percentage of total integrated peak area	Assumes every component responds equally; every integration decision anywhere in the chromatogram changes the denominator
External standard	Impurity quantified against a standard of itself, or of the API at a diluted concentration	Requires the standard; response factor differences must be addressed explicitly
Internal standard	Ratio to a compound added to every sample	Internal standard peak is itself integrated, and its own integration errors propagate

Area normalisation is common in related-substances methods and deserves one specific caution: because the denominator is the total integrated area, an integration decision on the main peak changes every impurity percentage in the chromatogram, including ones nobody was looking at.

5.4 Relative response factor

Detector response is compound-specific. A UV detector responds to chromophore, so two compounds at identical concentration can give very different areas. The relative response factor is the ratio of an impurity's response factor to that of the reference compound, normally the API:

RRF = (response factor of impurity) / (response factor of API)

and the corrected area is the observed area divided by the RRF. An impurity with RRF 0.5 gives half the area of the API at the same concentration, so its area must be doubled to express it correctly.

RRF is determined experimentally with an authentic standard of the impurity, most reliably by comparing the slopes of linear calibration curves for the impurity and the API over the relevant range, rather than from a single concentration.Practice

A widely used convention treats an RRF between roughly 0.8 and 1.2 as unity, on the basis that the correction is smaller than the method's uncertainty.Practice It is a convention, not a requirement. Whatever cut-off a laboratory uses, it should be stated in the method with its justification, and the uncorrected values should remain retrievable.

At the bench

Where an RRF is applied, the same integration decision now has a leveraged effect: an area error is multiplied by 1/RRF before it reaches the report. For an impurity with RRF 0.25, a 10% integration error becomes a 10% error on a number that is already four times the raw area. The impurities with the smallest response factors are the ones where integration care matters most.

5.5 Where integration meets the threshold

This is the point at which everything in Modules 1 to 4 stops being technical and starts being consequential. An impurity present at 0.106% in a product dosed below 1 g/day is above the 0.1% reporting threshold and belongs on the certificate of analysis. The same impurity, integrated with limits a few hundredths of a minute tighter, reports at 0.098% and does not.

Nothing about that chromatogram looks wrong afterwards. No rule was broken and no baseline was obviously forced. The result simply ceases to exist, and the quality unit never sees it. That asymmetry — a small, invisible adjustment producing a binary change in what is reported — is the whole reason integration practice attracts the regulatory attention it does.

▸

Sandbox, scenario 3 — 0.1% at the reporting threshold. A well-resolved impurity on a flat baseline, genuinely present at 0.106%. Bringing both limits in by six hundredths of a minute reports 0.098%. Lifting the baseline by 5% of the peak height — a nudge you can barely see — reports 0.086%.

Check yourself

A product is dosed at 500 mg/day. A degradation product is measured at 0.14%. What follows?

At a maximum daily dose below 1 g the Q3B(R2) reporting threshold is 0.1%, so it must be reported. The identification threshold for a dose in the range above 10 mg to 2 g is 0.2% or 2 mg TDI, whichever is lower — 0.14% is below 0.2%, so identification is not triggered on that basis. The qualification threshold in the 100 mg to 2 g band is 0.2% or 3 mg TDI, whichever is lower, so qualification is likewise not triggered. It is reported but neither identified nor qualified on these numbers alone.

An impurity has an RRF of 0.20. Why does integration accuracy matter more for this peak than for one with an RRF of 1.0?

Because the correction multiplies the error. The corrected area is the observed area divided by the RRF, so at RRF 0.20 the observed area is multiplied by five before it reaches the report. The impurity is also intrinsically small in the chromatogram — low response for a given amount — so it is more likely to sit near the noise and near a threshold, where baseline placement dominates.

Why does an integration change on the main peak affect impurities that were not touched?

Under area normalisation every impurity is expressed as a percentage of the total integrated peak area. The main peak is almost all of that denominator. Change it and every impurity percentage in the chromatogram changes, including ones nobody looked at. This is also why a reviewer should check the main peak's integration even when the finding of interest is a trace impurity.

Sources for this module

ICH Q3A(R2) Impurities in New Drug Substances, Step 4, 25 October 2006, Attachment 1.
ICH Q3B(R2) Impurities in New Drug Products, Step 4, 2 June 2006, thresholds for degradation products.
USP General Chapter <621> Chromatography — replacement of “disregard limit” with “reporting thresholds”.
The 0.8–1.2 RRF convention is industry practice and appears in no compendium.

Module 6

System suitability, and the pressure it relieves

System suitability is the gate that decides whether you have a result to report. Its connection to integration practice is behavioural as much as technical: marginal chromatography is where integration discipline fails first.

By the end of this module you should be able to

State the system suitability descriptors in USP <621> and the one numeric limit the chapter sets by default.
Calculate a symmetry factor and a maximum permitted %RSD from the chapter's own definitions.
Describe the limits of permitted adjustment before a procedure becomes a new one.
Explain why marginal system suitability reliably precedes integration findings.

6.1 What system suitability is for

System suitability establishes that the whole system — instrument, column, mobile phase, reagents, and analyst — was performing adequately at the time the samples were run. It is a gate, not a formality. If it fails, there is no reportable result, and no integration technique creates one.

6.2 The descriptors

Descriptor	What USP <621> provides
Resolution (R_S)	Formula only: R_S = 1.18(t_R2 − t_R1) / (W_h1 + W_h2). No universal numeric limit; acceptance criteria belong to the monograph or method
Symmetry factor (T)	The chapter's one hard numeric default: unless otherwise stated, the symmetry factor of the peak used for quantitation is 0.8–1.8Requirement
Repeatability (%RSD)	Maximum permitted %RSD of peak response for an assay, from the chapter's Table 2 — see below
Theoretical plates (N)	Formula only: N = 5.54(t_R/W_h)². Criteria are monograph-specified
System sensitivity	Expressed as signal-to-noise: LOQ corresponds to S/N = 10, LOD to S/N = 3

Symmetry factor

The symmetry (tailing) factor is constructed at 5% of peak height:

T = W_0.05 / 2f

where W_0.05 is the full width of the peak at 5% of its height and f is the distance from the leading edge to a perpendicular dropped from the apex, measured at that same height.

Constructing the symmetry factor. A horizontal line is drawn at 5% of peak height. W is the full width there and f is the distance from the leading edge to the perpendicular dropped from the apex. T = W / 2f.

A perfectly symmetrical peak gives T = 1.0. The peak above gives T = 2.19, outside the 0.8–1.8 default range: on this descriptor the run does not qualify.

Repeatability

USP <621> Table 2 gives the maximum permitted relative standard deviation of peak response for an assay, as a function of the number of replicate injections and of B, where B is the upper limit of content given in the monograph minus 100%.Requirement

B	n = 3	n = 4	n = 5	n = 6
2.0%	0.41	0.59	0.73	0.85
2.5%	0.52	0.74	0.92	1.06
3.0%	0.62	0.89	1.10	1.27

Note what this does. The permitted variability is tied to how wide the specification is: a tighter content limit demands a more repeatable system. It is not a fixed 2% that some laboratories still quote from memory.

6.3 Adjustment, and the line to a new procedure

USP <621> permits defined adjustments to chromatographic conditions, provided system suitability is still met. Beyond them, the procedure is a new one and requires validation. Selected values for liquid chromatography:Requirement

Condition	Isocratic	Gradient
Mobile phase minor components	±30% relative or ±10% absolute, whichever is smaller	Adjustment more restricted; changing column dimensions requires the gradient to be adjusted
Flow rate	±50%	±50%
Column temperature	±10 °C	±5 °C
Column L/d_p ratio	−25% to +50%	—
Retention time of the principal peak	—	±15%

The full table in the chapter is longer than this extract and should be consulted directly rather than paraphrased into an SOP.

6.4 The behavioural link

Now the part that matters most, and that no vendor tutorial will tell you.

Consider the position an analyst is in when the symmetry factor is 1.9, the impurity result is 0.11% against a 0.10% limit, and the sequence finished at two in the morning. Re-running costs a day and a column. The chromatography cannot be fixed at that hour. The only lever within reach is the integration.

Every significant enforcement case in this area begins from that position. The integration is rarely the root cause; it is the place where a chromatographic problem gets converted into a reportable-looking number. This is why a laboratory that is serious about integration integrity invests in chromatography that comfortably passes, rather than in tighter rules about integrating chromatography that barely does.

At the bench

When system suitability is marginal, the finding to record is the marginal system suitability — not a rescued result. A failed or borderline suitability injection is not a problem to be worked around; it is the measurement telling you that the run does not support a conclusion. Recording that costs a day. Not recording it has, repeatedly, cost firms their ability to ship.

▸

Sandbox, scenario 6 — marginal system suitability. The symmetry factor reads 2.19, outside the limit, while every baseline construction lands within about 10% of truth. The integration is fine. The run still does not qualify — which is exactly what makes the scenario dangerous.

Check yourself

A monograph gives a content limit of 98.0–102.0%. Five replicate injections of the standard give an RSD of 0.9%. Does the system pass on repeatability?

No. B is the upper limit minus 100%, so B = 2.0%. From Table 2 at B = 2.0% and n = 5, the maximum permitted %RSD is 0.73. An RSD of 0.9% exceeds it. Note that a laboratory working from a remembered ‘2% rule’ would have passed this run.

A peak has W_0.05 = 0.42 min and f = 0.12 min. Calculate T and state the outcome.

T = W_0.05 / 2f = 0.42 / 0.24 = 1.75. That is inside the 0.8–1.8 default range, so it passes — but only just. A peak at 1.75 is worth treating as a trend to watch rather than a pass to file, because the conditions that produce it usually get worse with column age.

Why does this module argue that investing in better chromatography does more for integration integrity than tightening integration rules?

Because the failure mode is situational. Integration discipline breaks down when the chromatography is marginal, the result is borderline, and integration is the only lever available. Rules written for that moment are being asked to hold against real pressure. Chromatography that comfortably passes removes the moment altogether, and a rule that never has to be tested is the only kind that reliably holds.

Sources for this module

USP General Chapter <621> Chromatography — symmetry factor default range 0.8–1.8; Table 2 maximum permitted %RSD; resolution and plate count formulae; signal-to-noise definitions of LOQ and LOD; permitted adjustments to conditions.
Consult the chapter directly for the complete adjustments table; the extract above is partial.