A disproportionality run finished at 03:14 on a Sunday. By the time the safety physician opened the dashboard on Monday, the EBGM engine had surfaced forty-one drug-event combinations above threshold for a marketed monoclonal antibody, and seven of them were new since the last quarterly run. One of those seven was a cluster of immune-mediated colitis reports with an EB05, the lower bound of the 90 percent credible interval, of 2.9 and twenty-two cases, eleven of them serious. The pharmacovigilance lead's instinct was correct and dangerous at the same time: the instinct to treat the alert as a finding. A statistical signal is not a finding. It is a question the database is asking, and the entire defensibility of what happens next depends on the team understanding that the alert is the cheapest, most automatable part of the work, and the medical validation that converts an alert into a confirmed safety signal is the part no model owns. This lesson designs the end-to-end AI-integrated workflow that runs from the disproportionality pipeline through AI-assisted triage, medical assessment drafting, signal management plan documentation, and the downstream Risk Management Plan or REMS update, and it does so the way a GVP inspector and an FDA-EMA Guiding Principles reviewer would expect to see it built: with the human accountability gate placed exactly where the judgment lives.

The Statistical Signal Is an Alert, Not a Finding

Every disproportionality method in routine pharmacovigilance is a measure of observed-versus-expected reporting, and none of them establishes causality. The Proportional Reporting Ratio (PRR) compares the proportion of reports for an event among a drug's reports against the proportion of that event across the rest of the database, with the conventional signal threshold of PRR at least 2, a chi-square of at least 4, and at least 3 cases. The Reporting Odds Ratio (ROR) reframes the same two-by-two contingency table as an odds ratio and is preferred where logistic-regression covariate adjustment is wanted. The Bayesian methods, the Multi-item Gamma Poisson Shrinker that produces the Empirical Bayes Geometric Mean (EBGM) used in FDA's FAERS analyses, and the Bayesian Confidence Propagation Neural Network behind the Information Component (IC) used in the WHO VigiBase, apply shrinkage so that a two-case coincidence does not masquerade as a strong association the way a raw PRR can. The shrinkage is the entire point: in a database of fifteen million reports, a frequentist measure on a sparse cell is unstable, and the EB05 lower bound is what a defensible workflow thresholds on, not the point estimate.

The mental model that has to survive every downstream step is this: disproportionality is a hypothesis generator operating on spontaneous reports, and spontaneous reports carry every bias the literature has documented for forty years. Reporting is stimulated by publicity, by litigation, by a Dear Healthcare Professional letter, by the Weber effect in the first two years post-launch, and by indication-driven confounding where the disease itself produces the event. A strong EBGM for a drug-event pair can be an artifact of a competitor's recall, a journal article three weeks earlier, or a label change that made reporters more attentive. The model that ranks the alerts does not know any of this, and the safety physician who treats the EB05 as evidence rather than as a prompt to investigate has skipped the only step that distinguishes pharmacovigilance from data mining. The FDA-EMA Guiding Principle of fitness for purpose is the relevant constraint here: a disproportionality engine is fit for the purpose of generating hypotheses and is categorically unfit for the purpose of confirming them.

The Disproportionality Pipeline as a Validated Component

The first stage of the end-to-end workflow is the quantitative pipeline, and it is the part of the system most people wrongly assume is the AI. It usually is not a large language model at all. It is a deterministic statistical engine, often an Oracle Empirica Signal deployment or an equivalent computation inside ArisGlobal LifeSphere or an in-house pipeline, that ingests the safety database extract, builds the contingency tables stratified by the chosen variables, and computes PRR, ROR, EBGM, and IC at the MedDRA Preferred Term and sometimes High Level Term level. Because the computation is deterministic and the algorithms are published, this stage validates cleanly under a GAMP 5 Category 4 configured-product posture: you qualify the installation, you run an operational qualification against a reference dataset with known expected scores, and you confirm the engine reproduces the published EBGM for a validation dataset to a defined tolerance. The Performance Qualification confirms that on your actual data, the stratification variables, the subgrouping by age and sex and region, and the threshold logic behave as the signal management SOP specifies.

The decisions that make this stage defensible are mostly upstream of the math. The stratification choice matters because failure to stratify by a strong confounder, indication being the classic one, produces both false signals and masked true ones. The selection of the comparator, whether the full database, a therapeutic-class subset, or a custom reference, changes every score, and that choice must be documented and stable across runs so that a quarter-on-quarter trend is a trend in the data and not an artifact of a changed denominator. The de-duplication and case-versioning logic determines whether the same patient appears three times and inflates a cell. None of this is AI judgment; all of it is signal-management-plan judgment that must be frozen, version-controlled, and auditable, because an inspector who finds that the comparator changed silently between quarters will discount every trend the team presents. The output of this stage is a ranked list of drug-event combinations with their scores, their case counts, and their threshold-crossing flags, and that list is the input to the part of the workflow where AI earns its place.

AI-Assisted Signal Triage: Ranking, Not Deciding

Forty-one threshold-crossing combinations is a triage problem, and triage is where a well-built AI layer compresses days of manual prioritization into a structured worklist a physician can move through in an hour. The triage layer, the kind built into ArisGlobal LifeSphere NavaX or assembled as a retrieval-augmented workflow over the case series, does three things that are genuinely useful and one thing that is genuinely dangerous if misconfigured. It clusters the cases behind each alert by clinical pattern so that the physician sees that nineteen of the twenty-two colitis reports describe a consistent onset window and biopsy-confirmed pattern while three are confounded by a reported concomitant checkpoint inhibitor. It cross-references each alert against the current label, the company core data sheet, and the existing RMP so that the physician immediately knows whether the event is already listed, already an identified risk, or genuinely new. It pulls the recent literature and prior periodic-report mentions so that an alert driven by a single publicized case series is flagged as plausibly notoriety-biased. Those three functions are extraction, retrieval, and clustering, and they are the modalities where AI is strong and where the output is checkable against named sources.

The dangerous function is prioritization scoring, where the layer assigns each alert a triage priority. This is useful as a sort order and indefensible as a decision. The design rule that makes the triage layer pass a GVP inspection is that the AI produces a recommended priority with its reasoning surfaced and its source cases linked, and a named human assigns the actual disposition. The distinction is not bureaucratic. A priority score that down-ranks a true signal because the case narratives are sparse, which is exactly the situation with a newly emerging rare event, would bury the one alert that matters most. The validated workflow therefore configures the triage layer so that every threshold-crossing alert is reviewed by a human regardless of its AI priority, the AI score orders the queue but never closes an alert, and the rationale the AI generated is captured as a draft input to the medical assessment rather than as a conclusion. This is the human-AI handoff at its most consequential: the model accelerates the reading, and the safety physician owns the disposition of every alert under their qualification.

From Alert to Validated Signal: The Medical Assessment

Signal validation under GVP Module IX is the act of examining the evidence supporting a detected signal to determine whether it constitutes a confirmed safety signal that warrants further action, and it is a structured clinical judgment that the AI can draft and only a qualified person can make. The medical assessment that documents this is the load-bearing artifact of the entire workflow, and a well-designed AI step produces a first draft that follows the signal-validation template: a description of the signal and how it was detected, the strength of the disproportionality with the EB05 and case count stated, a clinical review of the case series that separates well-documented cases from confounded or unassessable ones, a biological-plausibility section, a review of the background incidence in the indicated population, and an assessment against the existing label and RMP. The AI is exceptional at assembling this scaffold from the case series and the reference documents because the structure is stable and the source material is in the context window. It is exceptional at the wrong thing too: it will write a confident causality conclusion in the same even tone whether the evidence supports it or not.

The verification discipline mirrors the claim-decomposition logic that governs all regulated AI drafting. Every quantitative claim in the assessment, the EB05, the number of serious cases, the median time to onset, the number of positive dechallenge or rechallenge events, must reconcile to the case series and the disproportionality output, because a drafted assessment that states fourteen serious cases when the line listing holds eleven is a record that will not survive a PRAC question. The causality reasoning is where the named author cannot delegate: the WHO-UMC categorization, the weighing of temporal relationship, dechallenge, rechallenge, confounding, and alternative explanation, is exactly the integrative judgment that the FDA-EMA principles reserve for human accountability, and it is the same class of judgment as the benefit-risk integration that medical writing keeps out of the model's hands. The output of this stage is a signal validation memo with an explicit recommendation, refute, keep under monitoring, or confirm as a signal requiring action, signed by the safety physician and feeding the signal management plan.

Documenting the Signal Management Plan and the Decision Trail

A confirmed signal does not end the workflow; it opens a tracked lifecycle that the signal management plan documents from validation through assessment, recommendation, action, and closure. The plan is a controlled document in the safety system, often the signal management module of Oracle Argus or ArisGlobal LifeSphere, and the AI's role here is to keep the documentation complete and consistent rather than to make any of the decisions. A good integration drafts the signal-tracking entry, populates the cross-references to the validation memo and the case series, generates the timeline against the GVP-expected milestones, and produces the assessment report that will go to the safety governance committee. The value is that the same de-duplicated, source-linked case set flows through every artifact, so the case count in the validation memo, the signal management plan, and the eventual periodic report cannot drift apart the way they do when three writers assemble three documents by hand from three database pulls.

The trail that makes this defensible is the audit trail the FDA-EMA accountability and documentation principles require and that 21 CFR Part 11 and EU Annex 11 enforce. For every AI-drafted artifact in the chain, the record must capture the model and version, the prompt and system prompt, the exact source documents loaded, the timestamp, and the named human who reviewed and accepted each output, linked to the document version under change control. The signal management plan additionally records the decision rationale at each governance gate, because the regulator's question is rarely whether a signal was detected and almost always why a particular disposition was reached and by whom. A signal that was kept under monitoring rather than confirmed must show the evidence that supported watchful waiting, and a signal that was refuted must show why the disproportionality was judged an artifact. The AI can assemble all of this documentation; the decisions inside it belong to the safety physician and the signal management governance committee, and the audit trail must make that ownership unmistakable.

Propagating the Decision Into the RMP and REMS Update

When a signal is confirmed as a new important identified risk, an important potential risk, or missing information, the consequence is a Risk Management Plan update in the EU framework and potentially a REMS modification in the US framework, and this is where the workflow earns its description as end-to-end rather than as a clever triage tool. The RMP is a structured GVP-format document, and the relevant parts move in concert: the safety specification in Part II must add or recategorize the risk, the pharmacovigilance plan in Part III must state whether routine pharmacovigilance suffices or whether additional activity such as a PASS is warranted, and the risk-minimization measures in Part V must state whether routine measures suffice or whether additional measures such as educational materials or a controlled-access programme are needed. An AI workflow that has carried the confirmed signal, its evidence, and its categorization through the earlier stages can draft these updates with the categorization and the supporting evidence already reconciled, which is precisely the cross-document consistency that manual RMP updates routinely fail.

The US REMS pathway carries a different but parallel logic, because a REMS exists only where risk-minimization beyond labeling is necessary to ensure benefits outweigh risks, and a confirmed signal may justify adding an element such as a Medication Guide, a communication plan, or Elements to Assure Safe Use. The AI can draft the REMS modification narrative and the supporting assessment, but the determination that a REMS element is warranted is a benefit-risk judgment that sits with the sponsor's safety governance and ultimately with the agency, never with the model. The discipline that closes the loop is traceability in both directions: the RMP or REMS change must trace back through the signal management plan to the validation memo to the specific case series and disproportionality run that started it, and forward to the label change, the communication, and the next periodic report that will reference the new risk. A workflow that maintains that bidirectional trace is one where an inspector can start at the RMP and walk back to the 03:14 Sunday run, and find a named human accountable at every gate. That trace, not the speed of the triage, is what makes the AI integration defensible.

Designing the Handoffs and the Failure Modes That Hide in Them

The end-to-end design is a sequence of human-AI handoffs, and each handoff has a characteristic failure mode that the workflow spec must name and gate against. At the pipeline-to-triage handoff, the failure is a silently changed comparator or stratification that makes the ranked list non-comparable to the prior quarter, gated by freezing and version-controlling the signal-detection configuration. At the triage-to-assessment handoff, the failure is an AI priority score that buries a sparse-but-real emerging signal, gated by reviewing every threshold-crossing alert regardless of score. At the assessment handoff, the failure is a fluent causality conclusion the evidence does not support, gated by reserving the WHO-UMC and benefit-risk judgment for the named safety physician. At the documentation handoff, the failure is case-count or categorization drift across artifacts, gated by flowing one source-linked case set through every document. At the RMP and REMS handoff, the failure is a risk recategorized in one part of the RMP but not the others, gated by cross-part consistency checking before sign-off.

The reason this matters is that an inspector does not audit the AI; the inspector audits the workflow and the records it produced, and a workflow whose handoffs are unspecified is one where accountability is ambiguous exactly where the regulator looks hardest. The IQ/OQ/PQ mindset applies to the integration as a whole: the intended use is hypothesis-to-action signal management, the fitness-for-purpose statement scopes the AI to extraction, retrieval, clustering, and drafting, and the acceptance criteria require source-linked outputs, full audit capture, and named human disposition at every gate. Built this way, the AI does not weaken the pharmacovigilance system; it strengthens it, because it removes the manual transcription and reconciliation errors that are themselves a documented source of safety-data integrity findings, while leaving every act of medical judgment, where the architecture has nothing to offer, in the hands of the people the regulation holds accountable.

Key Takeaways

• A disproportionality score is a hypothesis, not a finding, and the EB05 lower bound is what a defensible workflow thresholds on, not the point estimate. PRR, ROR, EBGM, and IC measure observed-versus-expected reporting on spontaneous data that carries notoriety, Weber, and indication-confounding bias, so the alert is the cheapest part of the work and the medical validation under GVP Module IX is the part no model owns.
• The quantitative pipeline is a deterministic engine that validates as a GAMP 5 Category 4 component, and its defensibility lives upstream of the math. Stratification, comparator selection, and de-duplication must be frozen, version-controlled, and stable across quarters, because a silently changed denominator turns a real trend into an artifact an inspector will discount.
• The AI triage layer should rank and surface but never close an alert. Clustering, label and RMP cross-referencing, and literature retrieval are strong, checkable modalities, but a priority score can bury a sparse-but-real emerging signal, so every threshold-crossing alert gets human disposition regardless of its AI rank.
• The signal validation memo can be AI-drafted but the causality conclusion cannot be AI-decided. Every quantitative claim must reconcile to the case series and the disproportionality output, and the WHO-UMC categorization and benefit-risk judgment are reserved for the named safety physician under the FDA-EMA accountability principle.
• The RMP and REMS update must trace bidirectionally back to the originating run and forward to the label and periodic report. One source-linked case set flowing through validation, signal management plan, and risk-document update is what lets an inspector start at the RMP and walk back to the 03:14 Sunday alert, finding a named human accountable at every handoff.