safety.viz release review 2026-08-15

What v1.7.0 adds, annotated

Revised since these captures (2026-08-15, your RC review): the time-to-event endpoint dropdown shown below was replaced by flexible multiselect filters over the event data — you compose the endpoint (body system, term, seriousness, severity), and the chart shows time to first qualifying event, censored at end of follow-up. The intro animation is also gone (every frame is now an estimate). The captures below predate that change; the current behaviour is on the live dev demo and summarized in the RC comment.

The gallery learns to answer "how long until…". One new renderer — the Time-to-Event Explorer, the thirteenth in the gallery — brings Kaplan–Meier displays to the safety portfolio: step curves with censoring marks, pointwise confidence bands, and the at-risk table the FDA Safety Tables & Figures guide mandates beneath every time-to-event plot. Each section below is a behaviour, a capture of it running, and a way into the live demo to check it yourself. Every capture was taken on 15 August from the live dev deploy — the build this cut promotes.

1 new renderer — 13 in the gallery 1 250 unit + 252 browser tests estimator cross-validated against survival::survfit promotes dev → main via sv#131

sv#129 · hub#161

The chart family

Experimental

One estimator pass produces everything on screen: the step curves by treatment group, the censoring tick marks along them, the pointwise 95% confidence bands (Greenwood variance, log-log transform — the survival::survfit default family, cross-validated against it), and the No. at risk / Cumulative events strip beneath. Because the table is derived from the same pass as the curve, the two cannot disagree.

Time-to-Event Explorer: cumulative-incidence step curves for Placebo and two Xanomeline arms with censoring tick marks and 95% confidence bands, data-handling notes above the chart, and the No. at risk / Cumulative events strip table beneath. 1 2 3 4
  1. 1 The header states its statistics: pointwise 95% CIs, not simultaneous — and where 1 − KM overreads risk
  2. 2 Censoring is drawn, not hidden: a tick on the curve at every censored exit
  3. 3 The band quantifies what the curve alone would overstate — precision thins as the at-risk set drains
  4. 4 The FDA ST&F strip: at-risk and cumulative-event counts at each gridline, from the same estimator pass

Why it matters

Time-to-first-event is the display safety review reaches for when incidence tables disagree about "how often" because the arms differ in "how long". The demo endpoint is the CDISC Pilot 01 study's actual safety concern — time to first dermatologic event on a dermal patch — and the arms genuinely separate.

The renderer consumes an ADTTE-shaped analysis dataset: the clinical derivation of "first qualifying event" stays upstream with the data owner, where it belongs. It ships Experimental pending review of design decisions D1–D6.

Try it

Open the Time-to-Event Explorer
  1. Read the header line: events, censored, and the CI caveat are stated before the chart.
  2. Follow the blue Placebo curve to its plateau — the ticks are censored participants leaving the risk set.
  3. Check any gridline against the strip table: the curve and the counts always agree.
  4. Then read the evidence page — every TTE-* row traces to a named test.
sv#129 · hub#161

Two orientations, one estimator

The default display is cumulative incidence (1 − KM) — risk accumulating upward, the direction safety questions are asked in. One control flips to event-free probability (KM) — the classic survival curve stepping down. The axis always names the estimator in force, so a screenshot of either view carries its own definition.

The same endpoint flipped to event-free probability: KM curves stepping down from 100%, the y-axis labeled 'Event-free probability (KM)', with the same censor marks, bands and strip table.

The same endpoint, same estimator pass, flipped to event-free probability. Nothing else moves: bands, censor marks and the strip table carry over, and the y-axis label changes to say exactly what is now plotted. The header's caveat is standing, not decorative: 1 − KM can overestimate absolute risk when competing events are present — censoring here is end-of-study regardless of reason, including death, which is precisely that situation — and the clinical guide walks through it.

Why it matters

Reviewers argue about up-versus-down more than any other property of this display. Offering both, with the estimator named on the axis in every state, means the choice is presentational — the statistics cannot silently change under a toggle.

Try it

Open the demo, find Display → Orientation
  1. Flip Orientation to Event-free probability (KM) and watch the axis label follow.
  2. Toggle the Pointwise 95% CI band checkbox off and on.
  3. Filter Treatment Group to a single arm — counts and curves re-derive.
sv#129 · hub#161

The honest wide band

The demo ships three endpoints, and one is deliberately sparse: time to first serious adverse event — three events across 254 participants. The wide, hugging-the-floor band that produces is not a rendering problem. It is the correct display for a rare endpoint, and the renderer draws it rather than prettifying it.

Time to First Serious Adverse Event: 254 participants, 3 events, 251 censored — near-flat curves at the floor of the chart with a wide confidence band and dense censor ticks.

Three events, 251 censored. The curves barely leave the floor, the band is wide and ends where the data end, and the strip table says plainly that almost nobody had the event. This is the display doing its job: a rare-endpoint KM plot that looked confident would be lying. The header counts events and censored up front, so the sparseness is announced before the eye reaches the curve.

Why it matters

Safety endpoints are often rare by construction. A renderer that only looks good on rich endpoints teaches its readers to over-trust thin data; this one was shipped with the sparse case as a first-class demo endpoint precisely so the honest behaviour is on display.

Try it

Open the demo, switch the endpoint
  1. Set Endpoint to Time to First Serious Adverse Event.
  2. Read the header: 3 events, 251 censored — stated before the chart.
  3. Switch to Time to First Treatment-Emergent AE for the opposite extreme.
sv#129 · hub#161

Ask the curve

Every step answers questions. Hover an event step and the tooltip gives the day, the estimate with its pointwise CI, and the risk-set arithmetic at that moment. Click it, and the participants behind that step are named — the same drill-down-to-the-person discipline the rest of the gallery holds.

Tooltip on a high-dose event step: Day 30 — Xanomeline High Dose, cumulative incidence 50.2%, pointwise 95% CI 39.3–62.2%, 1 event of 36 at risk, 1 censored here (end of study).

Hover: day 30 on the high-dose arm — the estimate (50.2%), its pointwise CI (39.3–62.2%), and the arithmetic (1 event of 36 at risk; 1 censored here, with the censoring reason). The tooltip reads as a sentence a reviewer could quote.

After clicking the same event step: the footnote beneath the chart reads 'Xanomeline High Dose, day 30: 1 participant with an event — 01-703-1076'.

Click: the step's participants are named beneath the chart — here, day 30's single event is participant 01-703-1076. The aggregate curve stays connected to the people in it.

Why it matters

A KM curve is an aggregate over individual exits, and safety review eventually needs the individuals. Keeping "which participants are behind this step" one click away is what makes the display a working tool rather than a figure.

Try it

Open the demo
  1. Hover any event dot and read the estimate, CI and risk-set line.
  2. Click it — the participant IDs appear beneath the chart.
  3. Click empty chart space to clear the selection.
notes

Reading this page

How the captures were made

  • Playwright drove Chromium against the live dev deploy — the release-candidate build as served — on 2026-08-15. Nothing was staged or mocked; the tooltip and participant-selection captures are real interactions located via the chart's own data registry.
  • Stills are 2× JPEG.
  • The demo data is site/data/adtte.csv, derived from pharmaverseadam adae + adsl (CDISC Pilot 01): three endpoints — first dermatologic event, first serious AE (the sparse case), first any TEAE. The KM estimator is cross-validated against survival::survfit on this data.

Also in the release, not shown here

  • NEWS.md becomes the running release log (sv#125, sv#127) — unreleased work accumulates under an (Upcoming) heading per the program convention.
  • Evidence baselines refreshed for every module on the canonical Linux environment as part of the time-to-event landing.
  • Release prep (sv#130) — dist/safety.viz-1.7.0/ vendored, fixtures repointed.

Where this fits

This page is the review surface for RC sv#131 (dev → main). The release notes live in that PR's body and in NEWS.md, and publish on the v1.7.0 tag when it merges — on your approval, via the attested lane, and not before.

The rest of the candidate

  1. The dev gallery — all thirteen renderers
  2. time-to-event evidence — requirement-keyed test records
  3. time-to-event clinical guide — the 1 − KM reading, stated
  4. API reference — the data contract