Two analysts can process the same raw data file and report different area percents without either of them doing anything wrong. The number is not read off the detector; it is built from the detector trace by a chain of decisions, most of them made by default settings nobody looked at. Meyer put figures on the largest of those decisions. Peak pairs of equal width were calculated — two Gaussians, two exponentially modified Gaussians, and an exponentially modified Gaussian followed by a Gaussian — across area ratios from 1:1 to 10:1, tailing factors from 1.10 to 2.0 and resolution between 1.5 and 0.75, and each pair was integrated by the vertical drop method. Relative area errors reached up to plus or minus 40% of the true area within the pairs investigated, and in every case the small peak was more impaired than the large one, with the impairment depending on the size ratio [1].
That asymmetry is the practical problem in one sentence: the peak whose area you care least about is the one integration treats best. What follows is how an area is actually produced — peak detection, baseline construction, the choice between drop line and skim, and the record that has to accompany a hand adjustment. What the resulting percentage means, whether a share of total absorbance is a share of mass and what unequal response factors do to that assumption, is a separate question handled in the library's guide to HPLC purity analysis.
Before any baseline: deciding a peak exists
A data system does not begin by integrating. It begins by classifying. The most basic detection algorithm is slope sensitivity: the software monitors the slope of the baseline over small time intervals and compares successive intervals against each other to detect a change [2]. Peak width is its companion setting, and in a chromatography data system it is as indispensable as slope sensitivity in identifying and integrating peaks [2]. Between them they decide where integration starts, where it stops, and whether a feature is a peak at all or a piece of baseline.
These are not cosmetic settings. A review of uncertainty sources in gas and liquid chromatography reports that an earlier study had compared threshold, area reject and minimum peak width and found the threshold — the minimum peak height the integrator will detect — the most influential of the three; in the study built on that result, as the threshold was raised the number of peaks detected in a complex mixture fell, and the change was not linear. In the same work the significance of chromatographic factors in a robustness design varied with the threshold value, the authors concluding that the effects of column type and mobile phase composition depended critically on it [3]. Read that carefully: an integration parameter changed which chromatographic variables appeared to matter.
The same review collects the failure modes. Baseline noise blurs the foot of a peak and makes the beginning and end hard to locate; noise at the top of a peak can make the integrator assign a valley there, splitting one peak into two halves integrated separately; a slope sensitivity set too large causes the slope of the peak to be detected late and lost early; and tailing makes the end of a peak harder to identify. Where the parameters are set so that integration begins or stops well inside or outside the peak, the resulting error is systematic rather than random [3].
The baseline is constructed, not observed
There is no baseline underneath a peak. There is an interpolation between the two points integration chose, and everything above that line is called analyte. This is why baseline decisions dominate small peaks: the interpolated line displaces a roughly fixed slab of signal, and the smaller the area above it, the larger that displacement is in relative terms.
Noise sets the floor. The relative error in a peak area is inversely proportional to the signal-to-noise ratio, and the component of that error due to noise is proportional to the width of the peak and inversely proportional to the number of data points recorded across it [3]. Two of those three terms work against a late-eluting peptide band on a shallow gradient, which is broad and usually sits where the background is drifting.
The trace being integrated has also already been filtered before you see it. Wahab and colleagues, running fast separations on a column of some 38,000 plates, showed that sampling frequency and detector response time both affect retention time, noise amplitude, peak shape and width in a complex fashion, that the common wisdom of sampling twenty points per peak can be inadequate for high-efficiency columns, and that a manufacturer's data bunching, integration and response-time conditions may be integrally coupled to the sampling frequency in a manner neither transparent to nor controllable by the user [4]. The uncertainty review makes the matching point from the other side: percentage error in the area measurement increases with the sampling interval, but if the interval is too small, unwanted baseline noise will be detected and measured as peaks [3].
Baseline behaviour is worth measuring rather than assuming. In capillary electrophoresis of proteins — a different technique and a different analyte class from peptide reversed-phase work — Gu and colleagues used computer simulation alongside targeted experiments to separate two confounded effects, and found the electrophoretic baseline disturbance had a more pronounced impact on the purity data than the migration time shift did [5]. The numbers do not transfer to an HPLC gradient, but the diagnostic order does: when a percentage moves between runs, interrogate the baseline before the retention axis.
Drop line, valley-to-valley, tangential skim
Once two peaks share a baseline the software has to divide the shared signal, and there are three common constructions.
The drop line, or perpendicular drop, runs a vertical from the valley between the peaks down to the interpolated baseline and assigns everything on each side to the nearer peak. It is exact only when the two profiles are symmetric and identically shaped, because only then does the valley sit above the point where the two true profiles cross. Meyer's calculated pairs are a map of what departures from that cost, and the error grows with tailing and with size ratio [1].
Valley-to-valley integration lifts the baseline itself, running it from valley to valley instead of across the whole cluster. That removes a rising background, and it also removes real analyte signal sitting beneath the valley. Blanc and colleagues, working through an annotated trace in which the under-reported and over-reported regions are shaded so the area at stake is visible, judged a hybrid of valley-to-valley and a common baseline the most accurate of the constructions shown [2]. That is one illustrative electropherogram, not a general theorem.
Tangential skim exists for the case the drop line handles worst, a small peak riding on the flank of a large one. The function applies a tangent line at the inflection point on the shoulder peak, and the area demarcated by that line, the skimmed area, is then subtracted from the total area under the parent peak [2]. Skimming moves area from the rider to the parent; the drop line moves area the other way. Neither is neutral, and choosing between them changes the impurity number, not the main peak's.
Papas and Tougas quantified both on simulated unresolved peaks built from the exponentially modified Gaussian model. Area inaccuracies for the tangent skim and the perpendicular drop exceeded 50% in many cases, and were much greater than those calculated from ideal Gaussian profiles [6]. Skew is what does the damage, so any estimate of integration error derived from symmetric peaks is optimistic for real chromatography. They then built multiple linear regression models that predict the relative error of either method from chromatographic descriptors, so that the method with the smaller forecast error can be selected deliberately, and contrasted those boundaries against the criteria used in data system algorithms [6]. The choice between skim and drop is a quantitative decision with a predictable answer, not a question of which picture looks tidier.
Where it matters: a small peak beside a large one
Put those error magnitudes where they actually land. The pairs Meyer calculated stopped at the size ratios named above, while impurity work routinely runs at a hundred or a thousand to one — outside the range those calculations covered, and on the far side of the trend they showed [1]. A misplaced boundary costs the main peak a rounding difference. It costs the impurity the same absolute area, expressed as a large fraction of a small number, which is the number the experiment exists to produce.
Precision degrades in the same direction. The uncertainty review summarises an interlaboratory study of caffeine solutions from 1 down to 0.05 µg/mL in which, at those low concentrations, the precision of peak height measurements was better than the precision of peak area measurements, with relative reproducibility becoming worse as concentration decreased [3]. Taken with the inverse dependence of area error on signal-to-noise already noted, this is the argument against treating a peak near the reporting threshold as carrying the same relative uncertainty as the main band. It does not license switching to height mid-method; it licenses stating which measure was used and holding it fixed.
Shoulders, riders and pairs that never resolve
A shoulder is a peak the detection algorithm declined to admit. Deciding how many components sit under one envelope is a different problem from integrating them, and derivatives are the standard instrument: high-order derivatives provide information for assessing the number of underlying compounds under a given peak cluster, and a smoothing step is essential before taking them, because noise is amplified when differences are calculated. Vivó-Truyols and colleagues automated the Savitzky-Golay window choice with the Durbin-Watson criterion and produced peak parameter estimates and ranges to hand to a subsequent deconvolution step [7].
The smoothing window is itself a defensible number rather than a matter of taste. One method compares the autocorrelation of the fitting residuals, the differences between the input signal and the smoothed signal, against the autocorrelation of the instrument noise, the lag-one autocorrelation of that noise having been measured first on a blank signal; the window size giving the closest match is taken as optimal, and the approach proved robust across NMR, chromatography and mass spectrometry data [8].
Where a pair genuinely will not resolve, curve fitting replaces geometry. The companion paper to that derivative work sets out interlinked algorithms for peak deconvolution by non-linear regression, with a selectivity-based criterion choosing the approach so that heavily overlapped clusters do not settle into suboptimal solutions while well-resolved ones stay efficient, plus a routine fitting the original signal and its second derivative together for intermediate coelution; the peak model is a Gaussian with a polynomial standard deviation whose complexity can be modulated to improve the fit, modified to account for baseline outside the peak region, with the optimal model complexity selected automatically through error propagation theory [9]. The caveat is structural. A fitted area is only as good as the peak shape imposed on it, so reporting one means reporting the model as well.
None of this recovers information the separation never generated. Yoshida and colleagues note that impurity analysis of synthetic cyclic peptides relies on reversed-phase HPLC, which often provides inaccurate results because of the abundance of chemically similar impurities, and screened three polymer-based HILIC columns and seven mobile phases against four nine-residue cyclic peptides; comparing the elution profiles with reversed phase revealed mutually orthogonal behaviour [10]. An area percent is a statement about one separation.
It is also not a statement about identity. De Spiegeleer and colleagues profiled nominally identical synthetic peptide material obtained from five different sources, using liquid chromatography with photodiode array and fluorescence detection and liquid chromatography with electrospray ionisation mass spectrometry, and found that one of the products examined was in reality a totally different peptide [11]. Integration returns the area of whatever eluted, correctly and uninformatively.
Manual integration, and why it has to be written down
Manual integration is not something to pretend away. Blanc and colleagues recommend that it be kept to a minimum but remain permissible where required, and that where it is used it be described as well as possible; integration parameters must be captured, including slope sensitivity, smoothing factors and timed events such as peak start and stop, and integration settings must be documented, justified and reviewed whether they were applied automatically or by hand [2]. The reason is stated plainly in the same review: the essence of integration relies upon the skill and judgment of the operator, which makes a training programme for analysts, data reviewers and approvers part of the method rather than an administrative extra [2].
Automated integration does not remove that judgment; it standardises it and hides it. In routine targeted LC-MS/MS, mistakes such as accidentally selecting a nearby matrix peak or drawing an incorrect baseline are relatively common, and the authors note that especially where results are to be used for enforcement, a time-consuming manual review of each integrated peak is still required to obtain reliable results. Alewijn and colleagues built a review tool that compares incoming integrations against a database of examples with correct integrations supplied by analysts and checks all available ion transitions for the target compound; approximately 85% of chromatograms were handled automatically, leaving 15% for conventional review, with accuracy and repeatability equivalent to conventional human review [12]. For a laboratory without such a tool the useful reading is the residual: in that routine setting, roughly one chromatogram in seven was a case where the automatic answer should not have been allowed to stand unexamined.
Reproducibility, between people and between versions
Integration variance is measurable, and it surfaces in interlaboratory data. Siskos and colleagues ran a common targeted protocol covering 189 metabolites across six laboratories; for twenty typical biological samples the median interlaboratory coefficient of variation was 7.6%, with 85% of metabolites showing a median interlaboratory CV below 20%. Among the likely sources of irreproducibility the authors named the near limit-of-detection abundance of some metabolites and the degree of manual review and optimisation of peak integration in the LC-MS/MS data after acquisition [13]. Those analytes were small molecules rather than peptides, but the mechanism is generic: wherever integration depends on human intervention, the reported number is exposed to that judgment.
Software is the other axis. Perfect alignment of algorithms across different software is unattainable because each is programmed independently, even where the objectives of the functions are analogous [2]. A consequence follows that the sources do not state and that is best treated as a working assumption: reprocessing archived data on a different system, or a later version of the same one, can move a reported area when nothing about the sample has changed, so a re-integration is a result to be verified against the original rather than a formality.
The practical minimum is to fix the parameters in the method rather than in an analyst's habit, to record slope sensitivity, threshold, peak width, smoothing and every timed event alongside the number, and to check that the number survives a deliberate perturbation of those settings. An area percent that moves when the threshold moves is telling you something about the integration, not about the sample.