Bioanalytical method validation meets AI-augmented analysis.

The 4-6-15 / 4-6-20 acceptance core that anchors every bioanalytical method validation since Shah et al. (1992) does not specify how the peak got integrated. The acceptance is at the result level — concentration values within ±15% of nominal, ±20% at the LLOQ. The integration that produced the value has historically been treated as instrument-software-internal: deterministic, vendor-qualified, beneath the validation layer. That assumption holds for traditional algorithms (ApexTrack, Genesis, MarkerView in their classical configurations). It does not hold cleanly for the newer AI-assist features that those same vendors have shipped from 2022 onward.

/ 01What ICH M10 says, in v1.

ICH M10 (Step 4 endorsed 24 May 2022; PMDA via PSEHB/PED Notification 4 December 2024; EMA 21 January 2023; Health Canada 1 September 2023; Swissmedic 1 January 2024; ANVISA via RDC 742/2022) does not contain a section on AI/ML peak integration. Section 3.2.3 (Matrix Effect) and section 3.3 (Study Sample Analysis, including the M10 note that any system-suitability assessment should follow a predefined plan or SOP) apply generally; section 4 (Full validation) is method-result-focused; the reconstructibility obligation runs through the documentation expectations across the guideline. The reconstructibility obligation is where the AI question quietly lives. If an integration was assisted by a model the analyst could not reproduce by hand or with documented manual integration, what does sufficient detail mean.

Five regulators currently apply this section to AI-assist features case by case. The variance in inspection outcomes through 2024–2025 is largely about how completely the laboratory documented the model behaviour at the time of analysis. Laboratories that documented the algorithm version, the integration mode, the parameter set, and any human override have been passing. Laboratories that produced "the system did it automatically" have been receiving observations.

/ 02What ICH M10 v2 is likely to say (iFeed projection).

ICH has not published an official M10 v2 timeline. Based on iFeed's read of work-plan signals and member-state public commentary, AI/ML peak integration appears to be in scope at narrow band. Industry-projected Step 2b ~Q4 2026; industry-projected Step 4 endorsement ~Q3 2028 (these are iFeed projections, not ICH-confirmed dates). The narrow scope, if confirmed: peak detection, integration boundary identification, calibration-curve fit selection. Likely out of scope at v2: generative authoring of validation reports, AI-generated CAPA, AI-driven ISR sample selection (the last one may move into a separate workstream around 2029).

The likely v2 obligation set, based on draft circulating language and member-state public commentary:

• Algorithm version recorded contemporaneously to each integration
• Training data summary documented at method-validation time and at every revalidation
• Reference-test-set performance demonstrated before clinical use
• Manual override pathway preserved; override frequency reportable as a method performance indicator
• Change control on algorithm updates, with revalidation triggers explicit

/ 03The audit-trail question.

21 CFR Part 11 and EU Annex 11 both require contemporaneous, attributable, complete audit trail capture for any GxP-relevant action. When a peak is integrated by a deterministic algorithm, the algorithm's behaviour is implicit in the software version. When a peak is integrated with AI assistance, several elements need to be in the trail that historically did not have to be:

Algorithm version and parameters.

Not just the software version — the model version, training-set version, and any tuneable parameters at the time of integration. Vendor instrument software in 2026 is moving toward exposing this; some vendors are still treating it as internal. The laboratory cannot wait for the vendor; the laboratory needs to capture what it can capture today and document the gap where the vendor does not yet expose detail.

Confidence or quality flag.

If the algorithm produces an integration confidence score or quality flag, that flag is part of the result. ALCOA+ "complete" requires that the flag is preserved with the integration, not discarded after the analyst's decision. The flag is what allows post-hoc audit of integration quality across a study. Discarding it eliminates an audit channel.

Override history.

If the analyst overrode the algorithm's integration, the audit trail records the algorithm's proposed value, the analyst's overridden value, the reason, the time, and the user identity. This is straightforward in the abstract; it requires that the software supports it, and not all 2026 vendor configurations do. Laboratories operating in this gap should be documenting the limitation in their validation report and addressing it at the next instrument refresh.

The integration is now an assisted decision, not an automatic one. Assisted decisions have an audit trail. Automatic decisions get folded into the software-version record. Treating the assisted decision as automatic is the documentation gap that produces the 483 observation.

/ 04What an inspector checks now.

Field experience from inspections through 2025 across FDA OGD, EMA via national competent authority, and PMDA shows a consistent line of inquiry on AI-assisted bioanalytical work:

• What algorithm version was in effect for this study? Show me the SOP and the contemporaneous record
• How was this algorithm validated for use in this method? Show me the reference test set and the performance data
• What is the override rate for this study? Show me the trend
• How did you handle the previous algorithm version? Show me the change-control record
• What is your manual integration capability if the algorithm fails? Demonstrate it

The expectation is not that the laboratory has solved every question. The expectation is that the laboratory knows the questions are being asked, has a documented position on each, and can show the audit trail behind the position. Laboratories that have updated SOPs through 2025 to address these questions are inspection-ready. Laboratories operating on pre-AI SOPs are exposed.

/ 05Where this lands in the validation document.

Laboratories asking where in the validation report to put the AI-assist documentation should look at section 5 of their M10-aligned validation parameters: a sub-section on data processing, naming the algorithm, version, training reference, performance, and override-handling SOP. The M10 v2 obligation will likely sit here. Laboratories that are creating the section now will be ready when the obligation lands.