A pharmacovigilance operations lead in a global safety department watches a spontaneous report arrive: a call-center intake describing a hospitalization in a patient on a marketed biologic, captured as a few lines of unstructured text and a handful of structured fields. From that fragment, a regulated chain has to execute within a hard clock: the case must be assessed for seriousness and expectedness, a structured case narrative must be drafted, the events and the drug must be coded in MedDRA, an E2B(R3) XML message must be generated, and that message must be transmitted to the FDA Electronic Submissions Gateway, to EudraVigilance, and to the PMDA, with the expedited fifteen-day path running on its own countdown for serious unexpected cases. Each step is a candidate for AI integration and each step carries a failure mode that an inspector reads as a Good Vigilance Practice deficiency. The regulatory floor moved in 2026: E2B(R3) became the FDA-mandated standard for IND safety reports on 1 April 2026, and postmarketing transmission moves to the ESG NextGen environment from 1 October 2026, so the workflow this lesson designs is not optional modernization but the new compliance baseline. This lesson builds the end-to-end ICSR workflow with seriousness assessment, MedDRA coding fidelity, E2B(R3) structural validity, and the fifteen-day clock as the structural controls, and with the medical-judgment boundaries that AI does not cross drawn explicitly.

What the ICSR Lifecycle Actually Is, and Where the Clock Starts

An Individual Case Safety Report is the regulated unit of pharmacovigilance, the structured record of a single patient's suspected adverse reaction to a medicinal product, and its lifecycle runs from the moment of first awareness through assessment, coding, narrative, and electronic transmission to the health authorities that require it. The lifecycle is governed by ICH E2A through E2D for the safety concepts, by ICH E2B(R3) for the transmission format, and by the regional rules of FDA, the EMA through EudraVigilance, and the PMDA for what must be reported and when. The single most consequential fact about the lifecycle is that the regulatory clock starts at Day Zero, the day any person in the organization first becomes aware of a case that meets the minimum criteria, a patient, a reporter, a suspect product, and an event, and for a serious unexpected case the expedited fifteen-calendar-day clock to transmission begins at that moment regardless of how incomplete the initial information is. The workflow is therefore not only a quality problem but a timing problem, and an AI integration that accelerates the work without protecting the clock has solved the wrong half of the task.

This is why the workflow treats the seriousness and expectedness assessment as the first structural control, because that assessment determines which clock the case is on. A case is serious if it meets one of the regulatory seriousness criteria, death, life-threatening, hospitalization or prolongation of it, persistent or significant disability, congenital anomaly, or an other medically important condition, and a case is unexpected if the reaction is not listed in the applicable reference safety information. The combination of serious and unexpected puts the case on the expedited fifteen-day path, and the entire downstream urgency of the workflow flows from that determination, which means a mis-assessment that calls a serious case non-serious does not merely mis-classify a record, it silently removes the case from the clock that legally governs it. The workflow exists to make that determination explicit, traceable, and human-owned, because seriousness and expectedness are the points where AI assists but does not decide.

Intake, Triage, and the Seriousness-Assessment Boundary

The workflow begins at intake, where the case arrives from a spontaneous source, a call center, a patient portal, a healthcare-professional report, or a literature hit, as a mix of unstructured narrative text and partial structured fields, and the first task is to extract the case elements into a normalized structure: the patient demographics, the suspect and concomitant products, the reported events, the dates, the reporter, and the outcome. AI is strong at this extraction, reading free text and populating the structured intake fields, identifying the suspect product and the reported events, and surfacing the candidate seriousness criteria for the assessor to confirm, and this extraction is a genuine acceleration over manual transcription of high-volume intake. The triage layer then routes the case by provisional priority, flagging the candidates for the expedited path so that the fifteen-day clock is visible from the first hour, and an AI triage that surfaces a likely-serious case immediately is protecting the clock rather than merely sorting a queue.

The boundary the workflow draws hardest is at the seriousness and expectedness determination, because that judgment is a medical and regulatory decision that the model can render plausibly and wrongly with severe consequences. A model can read a narrative describing an event that meets the hospitalization criterion and, because the narrative phrased it ambiguously, score it non-serious, and that single misjudgment removes a reportable case from its expedited clock invisibly. The workflow therefore positions the AI as a surfacer of the candidate seriousness criteria and the candidate listedness against the reference safety information, with the determination itself made and signed by a qualified safety assessor, and with the assessor's decision and its rationale recorded against the AI's suggestion. The expectedness assessment carries the same boundary: the model can compare the reaction term against the reference safety information and propose listed or unlisted, but the listed-versus-unlisted call that drives reportability is a human determination, because an unlisted reaction wrongly called listed is a case that should have been expedited and was not.

Narrative Drafting and the MedDRA Coding-Fidelity Problem

With the case structured and assessed, the workflow drafts the case narrative, the Section E.i clinical account that tells the story of the case in a consistent, regulator-expected sequence: the patient's relevant history, the suspect product and dosing, the chronology of the reaction, the treatment and outcome, and the reporter's and the company's causality. AI is well suited to narrative drafting because the narrative is a structured retelling of the captured fields in a stable format, and a model that drafts a clean, chronologically ordered narrative from the structured case is a real time saving over composing each one by hand at the volume a global safety database generates. The discipline is that the narrative is a faithful retelling of the captured and assessed case, not a regeneration that introduces facts the case does not contain, so the workflow constrains the narrative to the structured source and validates that every clinical fact in the prose traces to a captured field, because a narrative that invents a dose, a date, or a concomitant medication is a fabricated clinical record.

The coding step is where fidelity matters most, because the events and the medical history are coded in MedDRA, and the coding choice, the selection of the Lowest Level Term that best represents the verbatim reported event and rolls up to the correct Preferred Term, determines how the case aggregates in signal detection and how it reports to the authorities. AI assists coding by proposing the candidate LLTs for each verbatim term, and a model trained on coding patterns is good at the common mappings, but the failure mode is precise and consequential: a model can select an LLT that is fluent and close but wrong, mapping a verbatim event to a Preferred Term that under- or over-states the medical concept, and that miscoding propagates silently into every aggregate analysis the case feeds. The workflow therefore treats the proposed LLT as a suggestion that a qualified coder confirms against the MedDRA hierarchy and the verbatim term, records the verbatim-to-LLT-to-PT chain for audit, and flags any code the model proposes with low confidence or against an ambiguous verbatim for mandatory human review. Coding is where a small, plausible error becomes a systematic distortion, and the workflow makes the verbatim-to-code mapping explicit and confirmed rather than accepted.

E2B(R3) Generation, Structural Validity, and the Transmission Layer

The assessed, narrated, and coded case is rendered into an E2B(R3)-compliant XML message, the ICH-standard structured representation that carries the case to the health authorities, and this is the step where the workflow's output must satisfy a machine validator before it can satisfy a human one. E2B(R3) is the HL7-based successor to E2B(R2), with a richer, more granular data structure and stricter conformance rules, and a message that fails the structural validation, a missing mandatory field, a malformed date, a code that violates the controlled vocabulary, an identifier that does not conform, is rejected at the gateway before it reaches a reviewer. AI assists the E2B(R3) generation by mapping the structured case fields into the correct message elements, but the controlling reality is that the message must pass conformance validation against the E2B(R3) schema and the regional business rules, so the workflow runs the generated message through structural validation as a hard gate and treats a validation failure as a blocking defect rather than a warning. A structurally invalid message is not a draft to be filed and fixed later; it does not transmit at all.

The transmission layer routes the validated message to the right destination by the right channel: to the FDA Electronic Submissions Gateway, moving to the ESG NextGen environment for postmarketing transmission from 1 October 2026, to EudraVigilance for the EU, and to the PMDA for Japan, each with its own acknowledgment handshake that confirms receipt. The fifteen-day clock governs the expedited cases through this entire chain, which means the workflow must track elapsed time from Day Zero against the transmission deadline and surface any expedited case approaching its limit, because a case that is correctly assessed, narrated, coded, and generated but transmitted on Day Sixteen is a reportable-late deficiency regardless of the quality of its content. The acknowledgment from each gateway is itself part of the record, because a message generated and sent but not acknowledged has not completed transmission, and the workflow captures the acknowledgment as the evidence that the regulatory obligation was met within the clock.

Follow-Up, Nullification, and the Case-Version Problem

A spontaneous case is rarely complete at first awareness, and the lifecycle does not end at the initial transmission, because follow-up information arrives that can change the seriousness, the events, the outcome, or the causality, and each material change generates a follow-up version of the ICSR that must itself be assessed, re-coded if needed, regenerated as an E2B(R3) message with the correct case version and sender identifiers, and retransmitted, sometimes restarting an expedited clock when the new information first makes a case serious. This versioning is where an AI workflow can introduce a subtle and dangerous error, because the model can regenerate a follow-up narrative or message that silently overwrites a fact that was correct in the initial version, or fails to carry forward the case identifiers that link the follow-up to its predecessor, and a follow-up that does not correctly reference its parent case is a broken record at the gateway. The workflow therefore treats every follow-up as a versioned change against the prior case state, reconciling what changed and what must persist, and it preserves the case-identifier chain so that the regulatory history of the case remains intact across versions.

Nullification carries the same discipline in a higher-stakes form, because a case determined to be invalid, a duplicate, or reported in error must be nullified through a formal E2B(R3) nullification message rather than simply deleted, and the nullification is itself a regulated transmission with its own conformance and acknowledgment requirements. An AI that proposes a duplicate match is useful for surfacing candidate duplicates against the safety database, but the determination to nullify a transmitted case is a regulatory decision a named assessor owns, because a wrongly nullified case removes a real safety record from the authorities' view, and a missed duplicate inflates the case count that feeds signal detection. The workflow positions the model as a surfacer of candidate duplicates and follow-up changes, with the nullification and the follow-up materiality determinations made and signed by a qualified assessor, so that the case-version history is both accurate and human-owned across the full life of the case.

The Validation Gate, the Fifteen-Day Clock, and What the Inspector Asks

The validation gate that defines this workflow runs a sequence of reconciliations across the case lifecycle. The seriousness-and-expectedness reconciliation confirms that the determination on the case was made and signed by a qualified assessor and that the case is on the correct clock, flagging any case whose AI-surfaced criteria and human determination disagree without a recorded rationale. The narrative-fidelity reconciliation confirms that every clinical fact in the E.i narrative traces to a captured field, rejecting any narrative that introduces an unsourced fact. The coding reconciliation confirms that every MedDRA code traces from its verbatim term through a confirmed LLT to its Preferred Term, with low-confidence and ambiguous codes routed to human review. The structural reconciliation confirms that the E2B(R3) message passes schema and business-rule conformance, and the timeliness reconciliation confirms that every expedited case transmitted within fifteen days of Day Zero and that the gateway acknowledgment was received and recorded.

The Level 3 deliverable is the validated workflow and the audit trail that make the ICSR lifecycle defensible to a Good Vigilance Practice inspection, not the narrative prose alone. The validation spec states the intended use, the production of E2B(R3)-compliant ICSRs from spontaneous intake through transmission, the fitness-for-purpose statement aligned to the FDA-EMA principle, and the acceptance criteria, which are the seriousness-determination ownership, the narrative-fidelity pass, the coding-traceability pass, the E2B(R3) conformance pass, and the timeliness pass against the fifteen-day clock. The audit trail captures, for each case, the model and version, the system prompt identity, the temperature, the timestamp at each step, Day Zero, the AI-surfaced suggestions and the human determinations against them, the verbatim-to-code chains, the conformance-validation result, the transmission timestamps, and the gateway acknowledgments. The human handoff is unambiguous: the model extracts, drafts, proposes codes, and generates the message, but a named safety assessor owns the seriousness and expectedness determination, a named coder confirms the MedDRA mapping, and a named author signs the case, with the QPPV accountable for the overall pharmacovigilance system. When an inspector asks the sponsor to show that a given expedited case was correctly assessed, coded, and transmitted within fifteen days, the audit trail already holds Day Zero, the human determination, the coding chain, the conformance result, and the acknowledgment. The model accelerates the lifecycle; the named assessor owns the seriousness, the named coder owns the term, and the named author certifies the case.

Key Takeaways

• The ICSR lifecycle is a timing problem as much as a quality problem, and the clock starts at Day Zero. For a serious unexpected case the expedited fifteen-calendar-day clock begins the moment any person in the organization first becomes aware of a case meeting the minimum criteria, so an AI integration that accelerates the work without protecting the clock has solved the wrong half of the task.
• Seriousness and expectedness are AI-surfaced but human-determined. The model surfaces the candidate seriousness criteria and listedness against the reference safety information, but a qualified safety assessor makes and signs the determination, because a serious case wrongly called non-serious is silently removed from the clock that legally governs it.
• Narrative fidelity and MedDRA coding fidelity are the content controls. Every clinical fact in the E.i narrative must trace to a captured field, and every MedDRA code must trace from its verbatim term through a confirmed Lowest Level Term to its Preferred Term, because a fabricated narrative fact is a false clinical record and a miscoded event is a systematic distortion of every aggregate it feeds.
• E2B(R3) conformance is a hard machine gate, and 2026 is the cutover year. E2B(R3) became FDA-mandated for IND safety reports on 1 April 2026 and moves to the ESG NextGen environment for postmarketing from 1 October 2026; a message that fails schema or business-rule validation is rejected at the gateway and does not transmit, so structural validity is a blocking defect, not a warning.
• The audit trail and the gateway acknowledgment make the lifecycle defensible to a GVP inspection. Day Zero, the human seriousness determination, the verbatim-to-code chains, the conformance result, the transmission timestamps, and the acknowledgment are the record that answers an inspector asking whether an expedited case was correctly assessed, coded, and transmitted within fifteen days.