Domain Atlas / Clinical decision support & deterioration alerting
CA-CDS Child Abuse Alerting
The Child Abuse Clinical Decision Support system (CA-CDS), built by a consortium at UPMC Children's Hospital of Pittsburgh, combines a five-item nurse-administered screen for every child under 13, a free-text scan of the chief complaint and nursing assessment, physician orders and discharge diagnoses into a rule-and-text trigger — no learned risk score — that raises a dashboard icon and a pop-up recommending a guideline-aligned physical-abuse order set. Implementation measurably moved volume: identification roughly quadrupled at two general emergency departments (P < .001), and triggering roughly doubled at both disseminated health systems (2.4% to 3.5% of children at the University of Wisconsin on Epic; 1.1% to 1.9% at Northwell Health on Allscripts). Its published performance — sensitivity 96.8%, specificity 98.5%, positive predictive value 26.5% — was measured against the hospital child protection team's chart assessment as reference standard: a concordance with an expert record judgment, not detection of abuse in any child. The documented failure mode is a string match: 'burn' in a 3-month-old's chief complaint is a trigger and so is 'burning up with fever'; 'broke' in a young infant triggers on 'the father speaks broken English'; 70% (33 of 47) of audited overtriggers came from the free-text scan, and 81% (195 of 242) of triggers were judged appropriate on chart review.[3]
What happened
The Child Abuse Clinical Decision Support system (CA-CDS) was built at UPMC Children's Hospital of Pittsburgh — Children's Hospital of Pittsburgh, not Children's Hospital of Philadelphia, a one-letter distinction this record obliges every reader to keep — by the consortium based in the hospital's Division of Informatics around Rachel Berger and colleagues, with funding from the Patient-Centered Outcomes Research Institute. Its parts are deliberately unglamorous: a five-item child abuse screen administered by the triage or primary nurse to every child under 13, an EHR trigger that combines the screen result with a free-text scan of the chief complaint and nursing assessment, specific physician orders and discharge diagnoses, a dashboard icon and practitioner pop-up alert, a physical-abuse-specific order set aligned to the American Academy of Pediatrics evaluation guideline, and an in-record Child Abuse Reporting Form documenting that a report was made. There is no learned risk score anywhere in the build: the trigger is rules and string matching over the hospital's own encounter record. The published performance figures — sensitivity 96.8% (the share of the charts the child protection team called concerning that the trigger caught), specificity 98.5% (the share of the charts the team did not call concerning that it correctly left alone), positive predictive value 26.5% — were measured against the hospital child protection team's chart assessment as the reference standard, which makes them a concordance with an expert record judgment, not a detection statistic about any child; roughly three of four alerts fire on charts the expert reviewers would not have called concerning. The documented failure mode is a string match, not an inference: 'burn' in the chief complaint of a 3-month-old is a trigger, and so is 'burning up with fever'; 'broke' in a young infant's chart triggers on 'the father speaks broken English.' Seventy percent of the overtriggers at the two general emergency departments — where the 242 children who triggered were all under 2 years old, and 81% of triggers were judged appropriate on chart review — came from the free-text scan.
The deployment arc ran from silent-mode testing in the Children's Hospital pediatric emergency department starting October 2014, through a randomised trial there in 2015 on the Cerner platform, live operation at the UPMC Hamot and Mercy general emergency departments from May 2016, a screen rollout across 13 UPMC general emergency departments in 2016, a practitioner barrier survey across 19 in February 2020, and dissemination onto two other health systems' commercial platforms published in 2023: the University of Wisconsin (2 emergency departments, Epic) and Northwell Health (4 emergency departments, Allscripts). The consistent finding is that the automation moves volume. Identification roughly quadrupled at the two general emergency departments (P < .001); triggering rose from 2.4% to 3.5% of children at Wisconsin and 1.1% to 1.9% at Northwell; reports to child protective services at Wisconsin rose from 0.6% to 0.9% of children (P = .03), with 17% of triggered children reported against 0.26% of non-triggered. The system's checks are equally well measured and point the other way. The order set produced 100% full guideline compliance when used at the originating hospital and 96% at Northwell — and it was used 43 times in a seven-month trial, 23 times in an entire implementation period, with 58% of surveyed practitioners reporting never using it. Of 71 practitioners analysed across 19 general emergency departments, 69% did not recognise the dashboard icon, 54% did not know they could view the screen result, 27% could not recall seeing an alert, and 65% were uncertain which tests to order — yet only 4.5% disagreed with the system's recommendations, 75% said it raised awareness, and 72% discussed alerts face-to-face with the child's nurse. Site-to-site acceptance diverged sharply on nominally identical software: 81% of Wisconsin respondents wanted continued use, against 3% at Northwell. No litigation and no regulatory action against the system appears anywhere in this record; the governing legal frame is not AI regulation at all, but the mandatory-reporting statutes of each state under the federal Child Abuse Prevention and Treatment Act.
Those statutes are what make this deployment's end product unlike anything else in this domain. What leaves the hospital is not software output: it is a mandated report, transmitted by an individual clinician discharging a personal statutory duty that no hospital policy can absorb — in 17 states plus the District of Columbia and the Virgin Islands the statutes say so expressly, 12 states bar employers from discouraging a report, and the reporting standard is suspicion, with no burden of proof, with physician-patient privilege among the privileges most commonly denied. The receiving state child protection agency must screen every referral against statutory criteria; roughly 37 states and DC assign risk-tiered response clocks. In federal fiscal year 2023, states screened in 2,107,473 referrals and screened out an estimated 2,292,000 — the second consecutive year screen-outs exceeded screen-ins — on 5,936 specialised intake workers and 21,739 investigation and alternative-response workers completing 66 responses each per year, at a mean first-contact time of 102 hours, with 28 states reporting that time lengthening and citing staff shortages and turnover. And nothing comes back: the statutory scheme specifies what a report must contain, how it is screened, who investigates and which third parties are notified, and establishes no duty to return the screening decision or disposition to the reporting clinician's record. The report is durable on the far side — it joins the agency's record, and prior-involvement and central-registry checks are standard elements of the investigation that follows the next report about the same family. The one deployment-level equity measurement in this record is a null: screening rates did not differ by patient or hospital characteristics across 13 emergency departments. No source publishes trigger, report or compliance rates disaggregated by subpopulation for this deployment. The measured disparity this domain is known for — 22.5% of white versus 52.9% of minority children reported, with skeletal surveys ordered at 8.75 times the adjusted odds for minority toddlers — belongs to a different institution (Children's Hospital of Philadelphia) and era (1994–2000), and is precisely the clinician variability a universal, instrument-driven screen was built to standardise: context for this case, never a measurement of this system.
The sociotechnical reading
This is the atlas's first case whose working end product is compelled INTO an organisation the deploying one has no authority over, with no return leg — and the map draws that as topology rather than as caption. Two organisations, two record systems, one compulsory crossing. Everything the hospital's levers can reach sits on the near side: the trigger specification whose documented defect is a string match only its build team can fix, the order set that is near-perfect when exercised and exercised on a minority of encounters, the awareness gap over the control surface itself, the consult team present at the tertiary hospital and absent from the general-ED path that carries the volume. The far side — the agency's screening, its response queue with a published 102-hour clock, its durable registry — is drawn from national statute and federal collection figures, because that is all the record supplies and all the hospital can see. The boundary discipline here is deliberate and double. First, the receiving organisation is a statutory class, 'the state child protection agency in each jurisdiction,' never a named agency: no source documents which agency received this deployment's reports, and none should be implied. Second, this case terminates exactly where the atlas's child-welfare cases begin. The agency-side deployments documented elsewhere in this atlas — call-screening models, prevention-outreach tools, quality-assurance algorithms operating on reports that have already arrived — govern the far side of precisely this crossing. This case governs what generates and shapes the reports that arrive. The two ends of the mandated report are different systems, run by different institutions, and the atlas keeps them in different files for the same reason the statute keeps them in different organisations.
The honest complications are structural rather than scandalous. The system demonstrably works as specified: triggering doubled, identification quadrupled, reporting rose measurably, and the order set achieves total guideline compliance whenever anyone opens it. But the asymmetry of discretion is the deployment's deepest fact — the clinician has full discretion over the alert (advisory by design, and 58% never used the order set) and none over the report, while the artifact that crosses is only as complete as the chart behind it, and full guideline compliance at the moment of reporting ran between 33% and 89% across sites and periods. A report made without the full workup is simultaneously a correct discharge of the statutory duty and an incomplete artifact, and no check on either side of the boundary reconciles the two: the in-record form documents that a report was made without verifying what it carried, the agency screens what arrived without seeing the chart, and the evaluation team — the only actor with a view toward both organisations — counts reports leaving the hospital and does not follow what the agency did with them. What practitioners asked for was not agency feedback but social work support beside them, named by 54%. The Field Guide lesson is about where governance can and cannot reach: the strongest checks in this network are all hospital-side and all partial, the crossing's compulsory character is statutory rather than configurable, so no hospital lever reaches either it or the receiving side's own operations, and the only actor that could rewire the crossing itself — add a return leg, change the duty, alter what a report must contain — is a legislature that appears on no organisation chart. The boundary holds as always: nothing here computes an outcome for any child or family, the record's own equity measurement is a single null stated as such, and whether this crossing happens at different rates for different groups of children is a question every source in this record leaves unmeasured.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.