Domain Atlas / Child welfare & family services
Allegheny Hello Baby
In Allegheny County's Hello Baby program, the top-tier roughly 5% of newborns by predictive risk score accounted for about 54% of children later removed from the home by age three, at roughly twenty times the removal risk of other newborns (methodology relative risk 22.24, 95% CI 17.50-28.25); the model reported an AUC of about 0.93 on holdout data.[2]
What happened
Allegheny County's Hello Baby program scores every newborn shortly after birth to estimate the child's risk of being removed from the home due to maltreatment by age three, then sorts families into three service tiers — Universal, Family Support, and Priority — with the top tier offered proactive outreach and voluntary prevention services rather than an investigation. Built on the same county administrative records as the Family Screening Tool but pointed at prevention, it launched in September 2020 after two independent ethics reviews (Michael Veale of University College London and Deborah Daro of Chapin Hall), whose public, point-by-point DHS response noted that the county has no institutional review board. Its most-contested feature is passive consent: families are informed at hospital discharge and again by a mailed postcard, and have a 20-day window to opt out before the birth record is scored. On holdout data the model reported an AUC of about 0.93, and the top roughly 5% of newborns by score accounted for about 54% of children later removed from the home, at roughly twenty times the removal risk of other newborns (methodology relative risk 22.24, 95% CI 17.50–28.25). A 2025 external evaluation of birth cohorts 2016–2024, controlling for COVID-19, found the program associated with fewer first investigations and substantiations but no reduction in out-of-home placements — the outcome the model was built to predict.
The sociotechnical reading
Most predictive tools in this Atlas point a score at a decision — screen in, investigate, deny. Hello Baby is the case where the score is deliberately pointed away from the decision: its defining governance artifact is a firewall keeping the newborn risk score out of child-welfare intake and investigation, so the same predictive machinery drives an offer of help instead of a coercive action. That makes its central risks the ones a firewall invites. Function creep is the standing pressure to let investigators see the scores, and a wall is only as strong as the sign-off that re-authorizes it each time someone proposes a new connection. And because the help is delivered by mandated reporters, and the score is computed from cumulative records, the system carries a documented loop in which accepting outreach can generate a new report that re-enters the record and raises a family's future score — help metabolizing into surveillance. The lesson the Atlas draws here is that a benevolent front end does not neutralize a coercive back end when both run on the same records and the same reporters; the governance work is not accuracy but boundary maintenance — consent, provenance, data minimization, and an actively defended wall. (Whether that surveillance loop is empirically large is contested: critics treat it as central, the county argues the measured effect is small.)
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.