CUSUM Control Chart: Formula, Decision Interval and GMP Trending

A CUSUM Control Chart is a sequential statistical process control chart that accumulates deviations from a reference mean to detect sustained shifts in a process average. In statistical process control (SPC), it retains evidence across successive observations rather than treating each result in isolation. Related terms include cusum chart, cumulative sum control chart and tabular cusum; the last refers to a numerical implementation.

Statistical control and conformance to specification answer different questions. NIST’s distinction between control limits and specifications separates consistent process output from product suitability. A stable process can produce out-of-specification results, while an in-specification result can occur during a process shift. Neither a CUSUM signal nor its absence is, by itself, a product acceptance decision.

This article covers mean-monitoring CUSUM for measured variables, such as mass or temperature. An ordinary cumulative production total adds quantities produced, not deviations from a reference mean. Attribute CUSUM designs monitor outcomes such as defect counts or classifications. Retrospective change-point tests search an existing dataset for a change location; they are not the sequential monitoring method described here.

Choose CUSUM when the monitoring question is whether a process mean has moved slightly and persistently—not simply whether the latest result is unusual. In its statistical guidance, NIST’s CUSUM discussion identifies an advantage over Shewhart charts for detecting mean shifts of 2 sigma or less. That advantage depends on chart design and assumptions; faster detection is not assured for every dataset.

Suitable engineering applications include sustained changes in fill mass, tablet weight, concentration or device dimensions. For example, observations from a fill weight checker can support monitoring when they represent a comparable stream. Select one interpretable statistic, such as individual net mass or subgroup mean, and one coherent process population. Separate products, equipment streams or operating modes when their baselines differ.

• Baseline: assess stability before estimating the reference mean and variation. A convenient nominal target is not automatically a defensible in-control reference mean. Justify baseline coverage rather than prescribing a universal minimum sample count.
• Measurement: confirm that resolution and measurement variability allow the intended shift to be distinguished from measurement noise.
• Sampling: choose a cadence that can reveal the change before the operational response becomes too late. Use rational subgrouping so each subgroup represents comparable short-term conditions.
• Dependence: check the independence assumptions used in performance assessment. Autocorrelation can alter signal behaviour; closely spaced readings are not necessarily independent evidence.

Retain dispersion monitoring within the wider SPC plan. A mean chart does not replace it: spread can increase without a corresponding mean shift. NIST’s variables-chart discussion explicitly stresses examining standard deviations when assessing control. Adopt CUSUM only when the selected statistic, baseline and sampling plan answer a defined monitoring question.

Define the statistic and its scale

Let x_t be the plotted observation or subgroup statistic at step t, mu0 its reference mean, and sigma_x its in-control standard deviation. These must describe the same statistic: an individual measurement and a subgroup mean generally have different standard deviations.

Starting with S_0 = 0, the unstandardized cumulative deviation is S_t = S_(t−1) + x_t − mu0. This unrestricted sum can be positive or negative; it is not the reset-to-zero statistic used for tabular monitoring.

Calculate the two tabular statistics

Standardize each observation and apply these recurrences in time order:

• z_t = (x_t − mu0) / sigma_x
• Cplus_0 = Cminus_0 = 0
• Cplus_t = max(0, Cplus_(t−1) + z_t − k)
• Cminus_t = max(0, Cminus_(t−1) − z_t − k)

Both stored statistics are nonnegative. Cplus accumulates evidence of an upward shift; Cminus accumulates evidence of a downward shift. A negative candidate value becomes zero rather than carrying forward.

Here, k is the reference allowance subtracted at each update, and h is the decision interval. Signal when Cplus_t > h or Cminus_t > h. Equality does not signal: a value exactly equal to h continues without a threshold signal. Document that strict comparison explicitly.

NIST’s tabular treatment permits original-data or sigma units. For original-unit calculations, use K = k × sigma_x and H = h × sigma_x; replace z_t with x_t − mu0 and k with K in the recurrences, then compare the resulting statistics with H.

Select parameters, then validate their use

For a normal mean-shift design with constant known variance, the conventional choice is k = delta / 2, where delta is the shift of interest in sigma_x units. For means of equal-size subgroups containing n independent observations, sigma_x = sigma_individual / square root of n. This CUSUM k is not the control-limit multiplier called k in some Shewhart notation.

Treat h as a performance-design choice, not a universal setting. Selection defines intended sensitivity; validation checks calculation correctness and performance under the intended monitoring conditions. Test zero-floor updates, equality at h and equivalent results across units. Do not transfer settings directly from an EWMA control chart: its update rule differs.

NIST’s V-mask explanation describes a graphical overlay on cumulative sums, positioned at the latest point so historical points can be compared against its arms. Tabular monitoring instead updates two stored statistics and evaluates their thresholds; it does not require drawing or inspecting a mask.

Both cases are hypothetical calculations, not manufacturing results or validated chart settings. Start each cumulative sum at zero and use a strict threshold convention: signal only when the relevant cumulative sum exceeds H, not when it equals H. Keep the assumed baseline fixed throughout each sequence.

Case 1: tablet-weight subgroup means

Assume a target of 500 mg, an individual tablet-weight standard deviation of 4 mg, independent observations and a constant subgroup size n = 4. The standard deviation of the plotted subgroup means is therefore sigma_x = 4 / √4 = 2 mg. Treat these baseline values as known for this example.

With standardized k = 0.5 and h = 4, conversion to measurement units gives K = 0.5 × 2 = 1 mg and H = 4 × 2 = 8 mg. Calculate z = (mean − 500) / 2. For each row, add mean − 500 − 1 mg to Cplus and 500 − mean − 1 mg to Cminus, flooring each result at zero.

At sample 2, the upper calculation is max(0, 1 + 503 − 500 − 1) = 3 mg. Sample 5 reaches 8 mg but does not exceed H. Sample 6 produces the first signal at 9 mg. Every lower increment is negative, so Cminus remains zero throughout.

Compare only with a basic three-sigma Shewhart mean chart using the same known baseline and no supplementary run rules. Its limits are 500 ± 3 × 2 mg: 494 and 506 mg. None of these six means crosses either limit. This constructed sequence demonstrates how repeated modest deviations accumulate; it does not measure average detection delay or establish universal CUSUM superiority.

Investigation question: did feeder adjustments or tooling changes occur near the start of the elevated means? Compare event timestamps with the sequence and use a focused gemba review to examine operating context. These are hypotheses to investigate, not causes established by the chart.

Case 2: lower food fill-mass signals

Assume individual food fill-mass measurements with a target of 100 g, independent observations and sigma_x = 1 g. With k = 0.5 and h = 4, K = 0.5 g and H = 4 g. Each lower increment is 100 − mass − 0.5 g; add it to the previous Cminus and apply the zero floor.

At sample 2, Cminus = max(0, 0.5 − 1.5) = 0 g. The recursion discards a negative running value; this mathematical zero floor is not an operator-authorized restart or a change of baseline. Three subsequent lower measurements build Cminus to 4.5 g, giving the first signal at sample 5 because 4.5 > 4 g.

Investigation question: what do equipment conditions and measurement-system evidence show around samples 3–5? Examine filler behavior, scale checks and tare records, including net-versus-gross reconciliation. Equipment drift and measurement error remain competing hypotheses until the investigation establishes a cause.

Select CUSUM for a defined monitoring objective, with justified assumptions and a response procedure—not as a proxy for meeting every GMP framework. Keep the legal scope assessment separate from the statistical case for using the chart.

Guidance context: FDA’s process-validation guidance identifies continued process verification as Stage 3. This places continued process verification within the validation lifecycle; it is not an instruction to select CUSUM.

In its analytical-methodology discussion, Section VII of FDA’s process-validation guidance recommends scientifically sound methods—specific, sensitive and accurate—and reliable results. That recommendation concerns measurement methods; a CUSUM calculation does not itself validate the method producing its inputs.

Engineering practice: Record the chart-selection rationale in the CPV plan, including the monitoring question, assumptions and assigned response. Present this as the site's technical justification, not a universal regulatory prescription.

Software support: Treat calculated signals as review inputs, not compliance certification or automatic disposition. Keep approval and release decisions with authorised people.

A practical CUSUM data model separates measured facts from statistical settings and review decisions. Connect every observation to its event timestamp, receipt timestamp, measurement unit, sample size, batch, equipment and originating record. Event time identifies when measurement occurred; receipt time identifies when it reached the monitoring system. Define calculation order explicitly rather than assuming arrival order represents process sequence.

Mark each input as an individual measurement or subgroup summary. For summaries, retain the statistic, subgroup membership and underlying measurement references through MES–historian integration. Sample size alone cannot reconstruct a subgroup.

An ISA-95-oriented integration convention can locate equipment within site, area, line and unit, with product and recipe-version links supplied by production context tagging. This is a practical hierarchy, not a claim of conformity to a particular ISA-95 clause. Together, these five entities let a reviewer identify the monitored population and reproduce a historical result without substituting today's baseline or settings.

In standard sequential-monitoring practice, in-control average run length (ARL) is the expected number of plotted observations before a signal under a specified in-control model. Shifted-process ARL describes that expectation under a specified change; state the shift size, onset and chart starting state. These averages are expectations, not assurances. Observations are not necessarily batches, and conversion to elapsed time depends on sampling frequency and interruptions.

For fixed k, starting conditions and model, increasing h generally reduces nuisance signals but delays detection; decreasing h reverses that trade-off. Compare credible shift sizes and directions, not one attractive ARL value. Balance the consequences of undetected drift against investigation capacity and alarm workload.

Use simulation or independently checked analytical calculations representing the intended sample statistic, parameter estimation, sampling plan and signal rules. Include baseline uncertainty and plausible departures from assumptions, such as serial correlation. Report uncertainty in estimated performance and detection-delay percentiles alongside averages. A single-trial false-alarm probability does not describe the long-run behaviour of a sequential rule.

Create a candidate-comparison table before selection. The template below identifies entries to complete, not performance results. Label any illustrative values as assumptions, not measured outcomes; connect workload estimates to the planned out-of-trend review.

For CUSUM, preserve four distinct records: original measurements, derived values, result corrections and chart configuration changes. Link corrections to the original result rather than silently replacing it. An audit trail should identify who changed what, when, why and which evaluations were affected. These are implementation controls supporting review and reproducibility—not automatic proof of regulatory compliance.

Spreadsheets need particular attention: overwritten formulas, hidden rows, manual sorting, inconsistent units and undocumented file versions can alter the accumulated result or obscure its origin. Use controlled access, protect calculation cells and perform independent test calculations against known inputs and expected outputs. During review, confirm the effective calculation version, observation order, included rows and units; do not rely on the displayed chart alone.

Assess changes to targets, sigma estimates, subgroup sizes or chart logic before routine use. Document their effect on calculated values and signal interpretation. Preserve the old configuration and its results alongside the new version, clearly distinguishing any retrospective recalculation from the evaluation used for the original decision. The operational restart decision belongs in the integration workflow.

Check 21 CFR Part 11 and EU Annex 11 applicability separately against their current texts. This section addresses chart-data integrity, not electronic-signature, retention or validation obligations.

A proposed, vendor-neutral workflow acquires observations from MES, LIMS or historians, evaluates them against the chart version effective for those observations, and routes signals to the responsible role. Connect the review to continued process verification (CPV) and the relevant batch record or device-history context, retaining observation identifiers and chart-version references.

Distinguish three states: a signal is a statistical trigger; a reviewed process concern is a documented assessment that warrants follow-up; a formal investigation or CAPA record is a separate quality-system decision. The site procedure should assign ownership, response times and escalation criteria according to risk. A threshold crossing alone neither proves a defect nor automatically creates CAPA. MES–QMS integration should preserve links between these states and record the review disposition.

For interface exceptions, identify duplicates and quarantine observations with unresolved measurement units or missing metadata from calculation. Apply an approved event-ordering policy rather than assuming arrival order is process order. After accepting late or corrected data, replay affected evaluations from the appropriate prior state. Preserve original and superseding results, their input versions and the reason for replay; clock-drift monitoring can help identify timing discrepancies.

After intervention, document the intervention evidence, post-intervention observations, unresolved concerns and rationale for continuing, resetting or replacing the chart. Obtain authorization from the role designated in the site procedure before implementing that decision. A reset should not erase unresolved signals. Calculations and routing support decisions; authorized people approve changes and make product-release decisions. Software does not certify site compliance.

Use these diagnostic checks alongside the baseline and integrity controls covered above. They are implementation practices, not additional regulatory requirements.

Use these engineering readiness checks before routine monitoring, then repeat the relevant checks after changes and at scheduled reviews.

1. Approve the configuration. Record the named monitoring objective, process population, statistic, defensible baseline, explicit units, selected k and h, monitored direction and equality rule. Attach the documented performance assessment described under performance and risk. Confirm that the configuration version and effective date match the intended observations, using the data-design controls above.
2. Test before use. Specify expected outputs before execution: accumulator values, signal status and exception disposition where applicable. Include both worked examples, a sequence returning an accumulator to zero, equality at h, missing or duplicate inputs, corrected observations, ordered replay and an authorized restart. Apply the history and change controls covered under data integrity; resolve discrepancies before acceptance.
3. Confirm operational ownership. Name the responsible reviewer and confirm access to source context, the escalation route and investigation linkage described under signal handling. Verify continuity after intervention through the earlier CPV integration controls. Schedule periodic reassessment of assumptions using the performance-and-risk assessment, recording decisions and any resulting configuration changes.