Clinical decision support & deterioration alerting

A machine-learning sepsis alert deployed across five hospitals of an academic health system, evaluated on 590,736 patients — the largest prospective study of its kind. Its mortality benefit was real but conditional: it accrued only to patients whose alert a provider confirmed within three hours, and the alert alone did nothing. The evaluation was prospective and peer-reviewed, but observational and run by the developer who commercialized the model.

An in-hospital deterioration model running around the clock across 21 hospitals, firing about twelve hours ahead of predicted deterioration. Its NEJM-measured mortality benefit is inseparable from where the alert goes: not to the bedside, but to a dedicated regional tier of critical-care virtual nurse consultants who screen every alert before escalating. The governance here is not a checkbox — it is a staffed, 24/7 subsystem with a payroll.

A deep-learning sepsis detector scoring every emergency-department patient every five minutes, fronted by rapid-response nurses on treatment-bundle timers. Its fault line is an authority split: the nurse who receives the alert is not the physician empowered to act on it. An independent ethnography found the system worked only because nurses did hidden, undervalued repair work to make a risk score actionable across a professional hierarchy.

A proprietary sepsis-prediction model shipped inside a common EHR and switched on across hundreds of hospitals — then externally validated to catch only a third of sepsis cases at roughly 109 alerts per true case, a real-world performance the vendor had not fully examined before selling it, shielded behind a firewall from outside scrutiny. This is the domain's failure arc: deployment at scale ahead of independent validation, alert fatigue, and vendor opacity — corrected only after outside criticism forced a retune.

The payer side of clinical decision support: inside the largest Medicare Advantage insurer's post-acute coverage review, a vendor coordinator completes the nH Predict similar-patient stay estimator while the patient is still hospitalized, and subpoenaed committee minutes show the same testing that cut six to ten minutes from each review arriving together with a rise in the adverse-determination rate - and being approved anyway. The one correction with a measured effect, the appeal, overturned 83.2% of the denials that reached it in 2022; 9.9% of denials did.

A commercial risk score used to ration a scarce care-management program was trained to predict next year's medical cost while its stated purpose was to find the sickest patients — and because unequal access means less is spent on Black patients at equal illness, the model was accurate and racially well calibrated on cost while Black patients at the same score carried 26.3% more chronic conditions. The independent dissection (Science, 2019), the manufacturer's own replication on 3.7 million patients, a same-day New York regulator demand letter, and an 84% reduction in one bias measure from changing only the training label — in an experimental, holdout predictor that was not the tool making enrollment decisions.

A rule-and-text child-abuse alerting system built by the CA-CDS Consortium at UPMC Children's Hospital of Pittsburgh and disseminated onto two other health systems' commercial record platforms roughly doubled triggering at both and measurably raised mandated reporting — but its working end product crosses a boundary its builders cannot see across: a report an individual clinician may not withhold, landing in a state child protection agency's durable record that must respond, screens out more than it screens in nationally, and by statute owes nothing back to the chart the report came from.

The domain's inversion case: a claim-review system that decides nothing about care, because the care has already happened. A deterministic code screen compares the procedure a doctor billed against an in-house list of diagnoses the insurer deems acceptable for it; matches are paid, mismatches queue to a company physician who signs the denial. ProPublica and The Capitol Forum, computing from internal company records, reported over 300,000 payment requests denied through this method in two months of 2022 at an average of 1.2 seconds each. Cigna disputes that characterization and has published no substitute figures. In October 2025 California's managed-care regulator fined the state plan entity $500,000, finding claims denied without physicians conducting clinical reviews first, under a review process that differed from the one it had on file.

The domain's routing case: an algorithm that denies nothing and still governs how much gets denied. The largest delegated prior-authorization vendor scores each request against criteria it writes itself; requests above an operating threshold are approved with no clinical review, and only the physicians below it may issue a denial. ProPublica and The Capitol Forum, working from internal documents and five former employees, reported in October 2024 that the threshold is adjustable and that insiders called it the dial. EviCore and Cigna dispute that characterization, saying the algorithms exist only to accelerate approval of appropriate care. The one enforcement loop that has closed is small and admitted: a Connecticut market-conduct examination of 196 files ended in a February 2024 consent order and a $16,000 fine for utilization-review compliance violations, and it read files rather than thresholds.

IBM sold hospitals on five continents a cancer treatment advisor marketed as machine reading of the medical literature, whose recommendations were in fact computed from a knowledge base of synthetic cases written by a few specialists per cancer type at one New York hospital — a provenance the buyers were not told. Its own internal reviewers recorded examples of unsafe and incorrect recommendations in mid-2017 and global selling continued for years; no published study ever measured a patient outcome; and the arc closed with a private-equity divestiture rather than any safety verdict. Its separately owned twin project at MD Anderson consumed more than 62 million dollars and was benched by a state procurement audit that expressly declined to judge the medicine.

In April 2018 the FDA authorized IDx-DR — since renamed LumineticsCore — the first diagnostic in any field of medicine whose clinical decision is rendered by software alone: a locked classifier that reads two retinal photographs taken by a clinic assistant who has never done ocular imaging, and returns one of two messages — refer to an eye care professional, or rescreen in twelve months — with no clinician interpreting the image or the result. Eight years on this is the atlas's benefit-forward anchor and its cleanest autonomy-boundary-by-design case: the human interpretive check was not eroded here, it was removed on purpose, in public, by a regulator, with the compensating controls named out loud. The measured benefit is real and replicated and the adverse-event docket is empty. The honest catch sits in the compensating control itself — in independent real-world use the image-quality gate that makes the design defensible declined a quarter of patients, disproportionately the older, cataract-prone people a screening programme most needs to reach.

For over a decade the standard equations for estimating kidney function multiplied the result upward for any patient recorded as Black — a coefficient of 1.159 in the 2009 CKD-EPI equation, against a measured median overestimate of 3.7 mL/min/1.73m2 — and a candidate needs an estimate of 20 mL/min or lower to begin accruing kidney waiting time. The US organ network did something this atlas rarely records: it prohibited the race-inclusive calculation in July 2022, then ordered every kidney program to recompute affected candidates' history without the coefficient and backdate the waiting time into the live allocation registry. By the one-year report, 14,701 modifications had been processed at a median of 1.7 years each and all 230 active kidney programs had attested. Then an independent national evaluation found the execution had varied significantly between transplant centers — and the governance response was to tighten, not to close the file.

A free cloud record platform sold the content of a point-of-care pain alert to an opioid manufacturer's marketing department for $959,700, let the sponsor's marketers propose edits to the trigger logic and the treatment-option list, and displayed the result more than approximately 230 million times over two and a half years — while its own analyses, delivered to the sponsor as the reporting it had bought and never to the prescribers being alerted, showed extended-release opioids to be the least effective of the listed options at lowering pain. The alert never named a drug. It did not malfunction. It is the atlas's clearest case of a system working exactly as designed, for a party the clinician could not see.

A detector reads every stroke-code CT angiogram as it leaves the scanner and, on a suspected large-vessel occlusion, pushes an alert with a compressed preview to the neurointerventional team's phones — while the radiologist's read continues beside it, unchanged and holding the diagnostic authority. The FDA authorization that created this device class in 2018 turns on exactly that parallelism, and CMS then scaled the deployment with an add-on payment of up to $1,040 a case. The system denies nothing and blocks nothing; its failure surface is alarm burden, a reading queue reordered at someone else's expense, distal occlusions the detector measurably misses, and an adoption curve that rose and fell with the payment rather than with need.

A behavioral-health claims administrator wrote the coverage criteria it applied to its own members, reissued them every year, and seated its Finance and Affordability representatives on the committees that approved them. After a ten-day bench trial a federal court found the 2011-2017 editions significantly and pervasively more restrictive than generally accepted standards of care, in eight enumerated ways, and found that the financial incentives had in fact infected the guideline development process. It is the atlas's best-documented case in which the rule itself was the defect — and its twelve-year appellate history is part of the record: the class-wide wrongful-denial theory failed, no coverage request was ever reprocessed, and what stands is a judgment about how the criteria were written.