Pediatric No-Show Prediction: Social Data Added No Accuracy but Pointed to Fixes

A Scienmag story by Harold Sullivan, dated October 9, 2026 and republished by Bioengineer.org, reports on a Pediatric Research study from The University of Texas Health Science Center at Houston ^[1]. The study tested whether a machine learning model could predict missed pediatric appointments and whether data on families' social circumstances improved those predictions. Adding the social data did not raise accuracy ^[1]. Several social and administrative factors were nonetheless associated with higher odds of a no-show, and the story presents many of them as barriers clinics could act on ^[1].

What the study measured

The analysis covered 7,931 scheduled appointments from April to August 2024 at six general pediatrics clinics in the greater Houston area: one academic center in the Texas Medical Center and five community-based sites ^[1]. Of these, 2,761 (34.8 percent) were no-shows. Most appointments involved Hispanic (43.2 percent) or non-Hispanic Black (39.2 percent) patients, and nearly 80 percent were covered by Medicaid ^[1].

The model used a transformer architecture, a design familiar from language processing, applied to each patient's visit history. Its inputs included prior attendance, phone reminders, confirmation status, appointment details and weather ^[1]. Non-medical drivers of health, meaning families' social circumstances, came from a 13-item screening questionnaire covering food insecurity, transportation, housing instability, financial strain, health literacy and health-harming legal needs ^[1].

With baseline variables, the area under the receiver operating characteristic curve (AUROC) was 0.708 with or without social variables. Adding demographics and insurance raised it to 0.721, again with no gain from social data ^[1]. The area under the precision-recall curve stayed near 0.54 in all four configurations, only modestly above the 34.8 percent no-show prevalence ^[1].

Associations with missed visits

Associations were estimated with generalized estimating equations, a regression method that accounts for repeated appointments by the same patient ^[1]. Unconfirmed appointments had more than twice the odds of a no-show compared with confirmed ones (adjusted odds ratio 2.24), and inactive or pending portal accounts had 31 percent higher odds ^[1]. Transportation needs were associated with 54 percent higher odds and housing instability with 26 percent higher odds after adjustment for demographics, insurance, clinic site and visit type ^[1]. Each added day between booking and visit slightly raised the odds, roughly 10 to 16 percent higher for visits booked 30 days ahead ^[1].

Non-Hispanic Black patients had 68 percent higher odds than non-Hispanic White patients, and Medicaid coverage carried 76 percent higher odds than private insurance ^[1]. Unknown insurance status showed the largest effect, a more than fivefold increase. The authors caution that this label reflects incomplete or unverified information and may be entangled with prior missed visits, because insurance is often verified at the visit ^[1].

Why it matters for scheduling teams

The story calls unconfirmed appointments, unactivated portal accounts and transportation or housing needs the strongest predictors in the cohort. It stresses that a clinic can detect these signals before the visit, although unknown insurance status showed a larger effect ^[1]. The authors warn that answering a high-risk flag by overbooking or offering a less convenient slot could penalize families already facing those barriers and deepen inequities in access ^[1]. They argue that elevated risk should instead prompt supportive outreach, such as transportation assistance, help with portal registration and care coordination referrals ^[1].

What to watch

The open questions are whether the findings hold across seasons and in health systems without an established screening and referral workflow. A further question is whether clinics that adopt such risk scores use them for outreach rather than for scheduling decisions that ration access ^[1].

Limitations

The study covered about five months of one spring-summer period, so seasonal effects from school schedules, holidays and illness waves were not tested ^[1]. Transportation needs were unknown for 28.2 percent of appointments, which may have blunted both the model's use of social data and some measured associations ^[1]. The work took place in one health system with an established screening and referral workflow, and the figures here come from a news report rather than the paper itself ^[1].