PRODUCTION · 66 SPRINTS · 350+ ENDPOINTS

The signal is in the narrative. Now you have time to read it.

SafetySignal AI automates case intake, MedDRA and WHODrug coding, three-algorithm causality assessment, disproportionality signal detection and E2B(R3) submission — in seconds rather than days. It runs on SafetySignal-PV-7B, a 7.6-billion-parameter model fine-tuned on real adverse event data, and every output it produces carries the evidence behind it.

Request a demo Run a case through the pipeline
THE PROBLEM

Drug safety is the only function measured on volume and judged on what it missed.

A pharmacovigilance team processes what arrives. Spontaneous reports, literature, clinical studies, regulatory feeds — and every one of them carries a clock. Fifteen days for an expedited report. The team is staffed to the volume, which means it is staffed to process, not to notice.

The signal is rarely in the coded fields. It is in the narrative — the phrasing the reporter used, the sequence of events, the clinical detail a physician would flag and a structured field cannot hold. And it is frequently not in any single case at all. It is in a pattern across cases that no individual reviewer sees, because no individual reviewer reads all of them.

So the work compounds in the wrong direction. Manual MedDRA coding runs roughly twenty-one dollars a case. Causality assessment waits for a qualified reviewer. Aggregate reporting consumes the same scientists who should be running signal evaluation. And the one outcome that matters — catching an unlisted event early enough to act — depends on attention that has already been spent.

General-purpose assistants make this worse rather than better. In pharmacovigilance a confident wrong answer is the most expensive output available: it consumes the scarce resource, qualified reviewer attention, and returns nothing defensible to an inspector.

$21Industry cost of coding a single case by hand
15 daysExpedited reporting clock on an unlisted serious event
$2M+Annual run cost of a legacy safety database with no AI causality
0MedDRA auto-coding rate
five-strategy NLP, no reviewer touch
3sCausality assessment per case
Naranjo, WHO-UMC and Kramer in parallel
20msSubmission-ready E2B(R3) XML
FDA and EMA formats, all required fields
0Modelled return on investment
3.4-month payback against legacy licence
CORE CAPABILITIES
FOUR PILLARS

Four pillars covering the full safety lifecycle — intake to submission, and the governance underneath both.

Each one replaces a manual workflow rather than adding a dashboard on top of it. Every output is traceable to the case data that produced it, the model version that generated it, and the reviewer who accepted it.

PILLAR 01

Intelligent case intake & extraction

Structured and unstructured ingestion, coded and enriched before a human opens it.

FAERS bulk import, EMA feeds, manual upload and REST API intake in one pipeline. The extraction layer reads unstructured source — narrative text, scanned medical records, literature abstracts — and resolves it into a structured ICSR with patient demographics, suspect and concomitant drugs, event terms, dates and outcomes. MedDRA coding runs through five strategies in sequence: exact match, normalised match, partial match, synonym expansion, and PV-7B semantic coding for the residue. WHODrug resolution handles trade names, salts and combination products.

  • 94% auto-coding with no reviewer touch
  • Zero re-keying from narrative to structured ICSR
  • Completeness scoring at intake, graded A–D
PILLAR 02

Autonomous triage & causality assessment

Seriousness, expectedness and causality resolved at intake, with the reasoning attached.

Every case is scored for seriousness against the ICH criteria, assessed for listedness against current product labelling, and routed by priority into the assessment queue. Causality runs three established algorithms in parallel — Naranjo, WHO-UMC and Kramer — and reports the consensus with the divergence exposed rather than averaged away. Where the three disagree, that disagreement is the finding, and the case escalates to a reviewer with the conflicting logic laid out.

  • Three-algorithm consensus in under 3 seconds
  • Automated listedness flags unlisted AEs for 15-day expedited reporting
  • 80% agreement with a frontier general model, and the divergence is where domain training earns its keep
PILLAR 03

Proactive signal detection & risk management

Four disproportionality methods triangulating, plus the narrative signal the maths cannot see.

PRR, ROR, IC/BCPNN and EBGM/GPS run simultaneously across the drug–event space, so a signal confirmed by multiple independent methods separates itself from statistical noise. A composite priority score then combines disproportionality strength with causality confidence, clinical severity, case volume and novelty — producing a ranked queue rather than an alert list. Week-over-week emergence tracking across the portfolio classifies every pair as new, rising, declining or stable, and PubMed surveillance surfaces literature-only signals that have not yet appeared in spontaneous data.

  • Multi-method confirmation across PRR, ROR, IC and EBGM
  • Ranked queue, not an undifferentiated alert backlog
  • Literature-only signals caught before they reach FAERS
PILLAR 04

Audit-ready compliance & governance

Human-in-the-loop by construction, with an evidence trail built for inspection.

No regulatory-consequential output leaves the platform on the model's authority alone. Every AI-generated assessment is a recommendation carrying its supporting evidence, presented to a qualified reviewer who accepts, amends or rejects it — and that decision, the reviewer identity, the timestamp and the model version are written to an immutable record. Submission artefacts are generated to format: E2B(R3) ICSRs, CIOMS I forms, PSUR and PBRER packages, ICH-compliant narratives, and BRAT benefit-risk assessments aligned to ICH E2C(R2).

  • 21 CFR Part 11 electronic records and signatures
  • Full audit trail with model version and reviewer decision pinned per case
  • Submission-ready in FDA and EMA formats, not export-and-fix
THE PLATFORM
WHAT SHIPS

Nine production modules, not a roadmap.

Sixty-six production sprints and more than 350 API endpoints. Each module is independently addressable, so a team can adopt causality assessment first and signal prioritisation six months later without a re-platform.

AI causality assessment

Naranjo, WHO-UMC and Kramer consensus on SafetySignal-PV-7B v3.

Signal prioritisation scoring

Composite of PRR, ROR, IC and EBGM with causality, severity, volume and novelty.

Benefit-risk assessment

BRAT evaluation aligned to ICH E2C(R2), generated from case data in under 50ms.

Labelling gap detection

Automated listedness against current labelling, flagging unlisted AEs for expedited reporting.

E2B(R3) XML generator

Submission-ready ICSRs with every required field, in FDA and EMA formats.

Literature surveillance

PubMed portfolio monitoring, cross-referenced against the case database.

Signal trending

Week-over-week emergence across 344+ drug–event pairs with change classification.

ICH narrative generation

E2B-compliant case narratives from structured data, with completeness grading.

Regulatory intelligence

Compliance status, deadline alerts, signal queue and risk heatmap from 130+ CIOMS XIV tables.

WORKED EXAMPLE
INTAKE TO SUBMISSION

One case, from narrative to submission-ready XML.

A spontaneous report arrives as free text. Watch it get coded, assessed, scored against the portfolio, checked against labelling, and emitted as E2B(R3) — with the reviewer gate held open where a human decision is required.

safetysignal · case pipelineidle
$ awaiting case intake
Illustrative sequence built on real platform behaviour. Not connected to a live tenant.
SIGNAL SCIENCE
TRIANGULATION

One method finds candidates. Four methods find signals.

Every disproportionality method has a known failure mode. PRR is unstable at low counts. ROR inflates in sparse strata. Bayesian shrinkage in IC and EBGM corrects for that but can suppress a genuine early signal. Running one method and tuning a threshold is how a safety database produces either a backlog or a blind spot.

SafetySignal runs all four simultaneously and treats agreement as the evidence. A pair confirmed by three or four independent methods is a different object from a pair flagged by one, and the queue is ordered accordingly.

PRR

Proportional reporting ratio

The frequentist baseline. Fast, transparent, and unstable below roughly three cases — which is exactly where early signals live.

ROR

Reporting odds ratio

Handles stratification and covariate adjustment better than PRR, at the cost of inflation in sparse cells.

IC

Information component, BCPNN

Bayesian shrinkage toward the null. Conservative by design, which makes a positive IC meaningful.

EBGM

Empirical Bayes geometric mean

Gamma-Poisson shrinkage with credible intervals. The method regulators most often recognise in a submission.

PORTFOLIO SURVEILLANCE · METHOD AGREEMENT344 DRUG–EVENT PAIRS
Four methods agree — escalate Three agree — evaluate Two agree — monitor Single method — noise until proven otherwise
ECOSYSTEM
ORACLE ARGUS

It augments the system of record. It does not ask you to replace it.

Nobody rips out a validated safety database. Argus is qualified, documented, inspected, and wired into every downstream regulatory process the company owns. Replacing it is not a migration project — it is a revalidation programme with regulatory exposure attached, run by a team that is already oversubscribed.

So SafetySignal connects into the Oracle pharmacovigilance ecosystem and enriches what is already there. Argus holds the record. SafetySignal supplies the intelligence layer above it.

ConnectionMechanismWhat it carriesStatus
Argus SafetyREST API and E2B messagingBidirectional case exchange. Argus cases enriched with generated causality assessments and ICH narratives.LIVE
Argus MartDirect warehouse readAggregate safety data and case listings into the regulatory intelligence dashboard.LIVE
Empirica TopicsSOAP web serviceSignal detection results imported and augmented with composite priority scoring and benefit-risk assessment.LIVE
Clinical OneSAE reconciliationSerious adverse event exchange between clinical and safety.READY
OCITerraform and KubernetesValidated manifests for GxP-compliant deployment alongside Oracle Cloud Infrastructure.READY
The addressable population is not sponsors willing to migrate. It is everyone who already owns the incumbent.
ARCHITECTURE
& TRUST

The model recommends. A qualified person decides. The record proves it.

Human-in-the-loop is not a setting in SafetySignal. It is the shape of the system. No regulatory-consequential output — a causality determination, a listedness call, an expedited-reporting decision, a submission artefact — leaves the platform on the model's authority alone.

Every AI-generated assessment arrives as a recommendation carrying the evidence that produced it: the source text it read, the algorithm outputs it combined, the disproportionality statistics it weighed, and the confidence it assigns. A qualified reviewer accepts, amends or rejects. That decision, the reviewer identity, the timestamp, the model version and the configuration in force are written to an immutable record at the moment of the decision.

Which means the platform answers the question an inspector actually asks. Not does your AI work — but who decided, on what basis, and can you show me.

REGULATORY

Standards built to, not mapped to

ICH E2B(R3), ICH E2C(R2), GVP Module VI and IX, CIOMS XIV, FDA FAERS and EMA EudraVigilance submission formats. Getting the science right and the format wrong is still a finding.

VALIDATION

GxP and 21 CFR Part 11

Electronic records and signatures, computer system validation artefacts, IQ/OQ/PQ documentation, and change control that survives an inspection of the platform itself.

SECURITY

HIPAA-ready, SOC 2 aligned

Encryption in transit and at rest, role-based access control, tenant isolation enforced architecturally, and PHI handling designed for named-patient safety data.

DEPLOYMENT

Your tenancy or ours

Validated Terraform and Kubernetes manifests for deployment inside your own cloud tenancy, including OCI, so regulated data never has to transit infrastructure you do not control.

HUMAN-IN-THE-LOOP GATE
01Model produces assessment with supporting evidence
02Confidence and method divergence surfaced, not hidden
03Qualified reviewer accepts, amends or rejects
04Decision, identity, timestamp and model version written immutably
THE CASE

The comparison is not against another vendor. It is against the status quo.

A legacy safety database runs past two million dollars a year and contains no AI causality assessment, no composite signal prioritisation, and no benefit-risk engine. The question a Head of Safety is answering is not which platform is better. It is what the current arrangement costs in scientist hours that should be spent on evaluation.

94%Of MedDRA coding handled without reviewer touch
3.4 moModelled payback period against legacy licence cost
248%Modelled return on investment, first year
3,000+Cases processed in production to date

Return figures are modelled on published legacy licence costs and observed per-case handling time. We will rebuild the model on your case volume, portfolio size and current cost base before any commercial conversation.

Show us a month of cases you have already closed.

The useful demonstration is not on our data. It is on yours, retrospectively, where you already know the answer — and you can judge exactly what the platform would have surfaced, and when.

Request a demo Book a safety audit

Forty minutes. Your case mix, your portfolio, your current cost base.