Deep dive — chart 3

Chart 3 — eligibility-lane rate per patient group, baseline vs recent

The chart with the operational + regulatory bite. Where score-distribution shifts translate into actual candidate routing. P3 ineligible_exclude rate rose 22% → 75%. Other patient groups stable. That is the exact representative-population signature FDA + EMA reviewers look for.

Eligibility-lane rate per patient group — baseline (left) vs recent (right). P3 ineligible_exclude rate rose sharply while other patient groups stayed stable.

What you are looking at

Grouped bars per patient group (P1 / P2 / P3 / P4 / P5)
Three colors per patient group: Eligible/Enroll (green) / Further Screening (blue) / Ineligible/Exclude (red)
Left cluster: baseline (first 30 days) — the "normal" routing per patient group
Right cluster: recent (last 30 days) — the current routing per patient group

What the chart shows

P1, P2, P4, P5 — lane rates roughly unchanged between baseline and recent.
P3 — ineligible_exclude rate climbs (22% baseline → 75% recent). Eligible_enroll rate drops. Further_screening rate compresses.
The differential exists only for P3. Other patient groups are unaffected.

53 percentage-point differential in exclude-lane routing. One patient group. Silently. Over 45+ days. This is the exact operational pattern the FDA Diversity Plan Guidance, EMA Reflection Paper on AI, and IRB Common Rule equitable-selection criterion are calibrated to detect.

Why aggregate metrics missed it

Because P3 is one patient group among five, the aggregate exclude-lane count barely moved: the P3 climb was largely offset in the total by stable P1/P2/P4/P5 volume, and the mean eligibility score across all screenings drifted only mildly. The CTMS enrollment-velocity view stayed green. Aggregate metrics are inherently blind to patient-group-level differential patterns, and patient-group-level differential is inherently what representative-population review is built to surface.

The signal that fires the A/B/C determination

The sensor's high-severity threshold fires on repeated P3 KL-divergence excursions plus the mean-shift signal on Day 51. With 10 high-severity events on file — well above the 3-event floor — the Snapshot lands on Category C by rule. No matter what any other metric shows.

See the Category-A/B/C decision logic →

The exhibit-quality read

This chart is what an FDA BIMO inspector or EMA scientific-advice rapporteur asks the sponsor for first. It is also what the sponsor's Regulatory Affairs lead should have on file before either lands — because the same chart, in the sponsor's hand, tells a different story:

Same chart, sponsor stance: "we detected this pattern through independent verification on Day X, initiated the following remediation on Day X+N, and here is the timeline of that remediation." The chart becomes evidence of good-faith investigation, not evidence of missed monitoring.

$499 Snapshot. 3 business days.

Same lane-shift chart for your sponsor's actual AI surface + full determination + 3 fix-first items.

Buy $499

Back to pharma sample hub.