clinicians.build · interactive · august 26, 2026

The Drop

One line per American hospital, drawn from how often the sepsis bundle gets started to how often it gets finished. Every line falls. The whole band falls 16.8 points, and it does not matter who owns the hospital or how big it is.

Story: Alexandra Byrne, “AI software to detect hospital drug diversion has a problem: Humans,” STAT News, Aug 25 2026
Data: CMS Provider Data Catalog — Timely and Effective Care – Hospital (SEP_1 and SEV_SEP_3HR), reporting period Jan 1 2024 – Jun 30 2025

A travel nurse at Adventist Health in Bakersfield spent late September 2024 taking patients' IV opioids and charting them as given. The hospital ran machine-learning software built to catch exactly that pattern. It fired. Managers ignored the alerts, auditors found.

Nobody publishes how often diversion alerts get ignored — hospitals aren't required to say they run the software, let alone report when it fails. But CMS does publish, for 3,084 hospitals, what happened after a different clinical trigger fired. Here is all of it, at once.

3,084 hospitals, starting and finishing

Each hairline is one hospital. The navy end is how often the first move landed — the 3-hour severe sepsis bundle: lactate, cultures before antibiotics, antibiotics, fluids. The red end is how often the whole thing got finished and closed out. Watch it fall.

The field
▶ Replay the drop Sort by finish Sort by drop Sort by size Hide small hospitals (n < 30)
Started
79.7%
mean 3-hour bundle
Finished
62.9%
mean SEP_1 composite
The drop
16.8
points, on average
Lines that fall
98.6%
3,042 of 3,084
Drop ≥ 20 pts
1,043
33.8% of hospitals
Every hospital in the file is on screen. Hover anywhere on the field.
Started — 3-hour bundle delivered (SEV_SEP_3HR) Finished — full bundle completed (SEP_1)
What the two ends actually are

Navy — SEV_SEP_3HR: of severe sepsis patients, the share who got the complete 3‑hour package.
Red — SEP_1: the composite. All the 3‑hour elements plus the 6‑hour work — repeat lactate, and for septic shock, vasopressors and a documented reassessment. These are two separate CMS measures with overlapping but not identical denominators (SEP_1 covers severe sepsis and septic shock together). SEP_1 requires the 3‑hour elements, so it is close to nested inside the navy end but not strictly so. Read each line as a strong indicator of escalation loss, not an exact conditional dropout.

The drop is the product

Sepsis detection is a solved problem. The trigger logic is decades old, it is baked into every major EHR, and by CMS's own numbers it works about four times in five. Diversion detection is a solved problem too — at Bakersfield the software fired for weeks.

What both stories have in common is that everything downstream of the firing was unowned and uninstrumented. Somebody has forty alerts in a queue, thirty-eight are noise, and the queue becomes a thing you clear rather than a thing you read.

If your product can be fully ignored and never tell you, you don't have telemetry. You have a dashboard.Builder's Briefing, August 26, 2026

Where this dataset is thin

DenominatorHospitalsStartedFinishedDropSD
Under 30 cases27577.758.619.022.8
30–9986779.863.316.516.5
100–2991,61379.762.916.814.8
300 or more32981.165.615.414.0

If you are building clinical AI: the detector is the easy half. Who receives the alert, what their shift looks like when it lands, what the false-positive rate does to their trust by week six, whether closing it requires a reason, whether anyone audits the closures. None of that is in your model card. All of it decides whether your model does anything at all.

⚠︎ AI-generated · not reviewed by a human · verify against the linked sources before relying on it. Hospital-level values are transcribed from the CMS Timely and Effective Care file; the “drop” framing, the small-sample analysis and the diversion analogy are this page's interpretation, not CMS's.