We asked 24 pharmacovigilance scientists across mid-size pharmaceutical companies one question: where does your signal management time actually go? Not what your SOPs say, not what your team lead estimates — where does it go minute by minute when you track it honestly for two weeks.
The answers were more consistent than we expected, and more uncomfortable than most PV leadership wants to hear. Signal detection is not the bottleneck. The signal detection algorithms — PRR, ROR, EBGM, IC — run in under an hour on most company databases. What takes days is everything that comes after the algorithm finishes.
What We Asked and How We Listened
The 24 participants came from PV departments ranging from 4 scientists to roughly 30, all working on marketed products with active signal management programs. We asked them to keep a rough time log for two weeks and categorize activity against the ICH E2C(R2) signal management steps: detection, validation, analysis, prioritization, action recommendation, and outcome communication.
We're not claiming this is statistically representative of the industry. It's observational. But the patterns that emerged across independent teams, using different safety databases, different signal management software, and different regulatory footprints, were hard to dismiss as coincidence.
The Validation Swamp
Signal validation — confirming that a detected signal represents a genuine data pattern rather than a coding artifact or data quality issue — consumed an average of 31% of total signal management time across our 24 participants. That was the single largest category by a considerable margin.
The core problem is not that validation is conceptually difficult. It's that the work is fragmented. A scientist receives a disproportionality signal on a PT-level adverse event. To validate it, they need to:
- Pull the underlying ICSRs and verify MedDRA coding is consistent across reporters
- Check for notoriety — is this already in the labeling, a known class effect, or documented in a prior signal assessment?
- Assess data quality: incomplete cases, duplicate reports, expedited vs. non-expedited case mix
- Run the same query against the prior period to determine whether this is a new finding or a stable disproportionality that has existed for multiple review periods
None of these steps is hard. All of them require navigating between at least two systems — the safety database, the signal detection output, often a case narrative tool, and sometimes a separate literature surveillance tracker. The context switching alone accounts for roughly a third of validation time, based on what participants described.
The Documentation Parallel Track
The second largest time sink, at 24% on average, was documentation that runs in parallel with the scientific work rather than being a product of it. Signal assessment reports require a specific structure that most teams maintain in Word or a Word-like template, separate from the safety database. Scientists are writing while also working — capturing rationale, populating data tables, pulling case counts again to make sure the numbers in the narrative match the numbers in the appendix.
Several participants described a specific frustration: finishing a signal assessment, sending it for review, and then having to revise the data tables because the database was refreshed with new cases during the 3-day period between their analysis and the medical reviewer's response. This is not a process failure — it's a natural consequence of living data sources and asynchronous review workflows. But it adds roughly a half-day per signal assessment in many cases.
We're not saying documentation should be eliminated or even dramatically reduced — regulatory agencies expect a clear written record of scientific reasoning. What we're noting is that much of the time isn't spent on that reasoning; it's spent ensuring consistency between the reasoning and the data tables that surround it.
Prioritization: Where Time Goes Unexpectedly
We expected prioritization to be a minor time component. Signal prioritization — deciding which detected signals warrant full assessment this period and which can be held — sounds like a meeting or a checklist. In practice, participants averaged 18% of their signal management time on prioritization activities, including the pre-work that makes prioritization meaningful.
The pre-work is the issue. To make a sound prioritization decision on a signal, a scientist needs at minimum: the cumulative case count, the exposure-adjusted reporting rate, whether the signal has been reviewed before and if so what the prior conclusion was, whether there are signals on related PTs in the same MedDRA SOC that might form a cluster, and whether there are any regulatory commitments (REMS, RMP, post-marketing commitment) that create an external obligation to assess this class of event.
Assembling that context for a weekly or biweekly signal review meeting routinely takes 2-3 hours for a portfolio of 15-20 active signals. That time is mostly invisible in workflow metrics because it doesn't appear in the safety database as a completed work unit.
Multi-Drug Interactions: The Growing Edge of the Problem
Something that several participants mentioned unprompted was the increasing complexity of multi-drug combination signals. As polypharmacy grows in the reported case population — FAERS cases reporting 5 or more concomitant drugs are common in elderly populations — single-drug disproportionality analysis produces a growing number of suspect drug confounding issues.
When a signal appears on Drug A, and 60% of the underlying cases also report Drug B or Drug C as concomitant medications, the validation question becomes whether this is truly a Drug A signal or a co-medication interaction effect. Traditional disproportionality analysis is not built to answer that question efficiently. Working through it manually, case-by-case, is where experienced scientists report spending significant unplanned time.
This is precisely the problem that n-drug graph approaches are designed to address — modeling the co-reporting structure across all drugs in a case rather than computing pairwise associations in isolation. Teams we've spoken with describe the shift from single-drug PRR analysis to network-aware methods as reducing the multi-drug validation burden significantly, because the interaction hypothesis either emerges directly from the analysis or can be ruled out with a single graph traversal rather than a manual case review series.
What the Numbers Don't Capture
The time log data has a limitation we want to name directly: it captures activity, not cognitive load. The 12% of time spent on medical literature review doesn't feel the same as the 12% spent on database querying. The literature review requires sustained analytical focus; the database work is procedurally demanding but less cognitively taxing. When scientists describe feeling exhausted and behind at the end of a signal management cycle, they're not necessarily describing a volume problem. They're describing a mismatch between high-value analytical work and the low-value procedural work that keeps interrupting it.
That mismatch is where automation has its clearest value — not in replacing scientific judgment, but in doing the procedural scaffolding that currently consumes judgment-hours. Query execution, cross-system data pull, case count consistency checking, prior signal retrieval — these are candidates for automation not because they're trivial but because making a scientist do them manually is a poor use of the training that put them in the role.
Where PV Teams Are Starting
When we asked participants what they would automate first if they could, the answers clustered around two things: duplicate detection across case sources, and the prior signal retrieval workflow. Both are bounded, well-defined tasks with clear inputs and outputs. Neither requires a scientific judgment call. Both consume a meaningful slice of time every review cycle.
The more ambitious answer — automating multi-drug interaction analysis across the full case portfolio in real time rather than reactively — came from scientists who had seen the problem most acutely in complex-polypharmacy products. The realistic starting point for most teams is smaller and more tactical. But the direction is consistent: reduce procedural overhead so that the scientists doing signal management are spending their hours on signal assessment rather than signal assembly.