Prediction Modeling of Missed Appointments in Safety-Net Physiatry: Visit Characteristics Outperform Patient Risk Scores

Objective. To identify appointment-level and patient-level factors associated with missed (no-show) visits in an outpatient physical medicine and rehabilitation (PM&R) clinic at a public safety-net hospital, to test whether electronic health record (EHR) composite risk scores and neighborhood social vulnerability add predictive value beyond scheduling characteristics, and to develop and internally validate a no-show prediction model from routinely available scheduling data. Design. Retrospective cohort study of all scheduled outpatient PM&R appointments, with development and internal (cross-validated and temporal) validation of a prediction model. Setting. Outpatient PM&R clinic, Arrowhead Regional Medical Center, a county safety-net hospital in the Inland Empire region of Southern California. Participants. 8,927 appointments scheduled for 2,409 adults between February 15, 2022 and February 12, 2026; 6,594 appointments for 1,369 patients were linked to age, EPIC General Risk Score, EPIC Preventive Care Gap Score, and ZIP-code Social Vulnerability Index (SVI). Main outcome measure. No-show at the appointment, defined as a scheduled appointment neither attended nor cancelled; cancellations were retained as a separate status. Results. The no-show rate was 13.1% (1,171 of 8,927); 29.0% of appointments were cancelled. Ten percent of patients accounted for 50% of no-shows. In generalized estimating equation models clustered by patient, no-shows were more likely at new-patient consultations than follow-ups (OR 1.40, 95% CI 1.10-1.78) and less likely at telehealth visits (OR 0.25, 95% CI 0.14-0.44), procedure visits (OR 0.27, 95% CI 0.09-0.77), same-day appointments (OR 0.59, 95% CI 0.35-1.01), and with increasing age (OR 0.88, 95% CI 0.81-0.96 per decade). The EPIC General Risk Score (OR 0.99, 95% CI 0.94-1.04), Preventive Care Gap Score (OR 1.02, 95% CI 0.96-1.09), and both SVI measures were not associated with no-show. A logistic model using only scheduling features discriminated no-shows with a patient-grouped cross-validated area under the ROC curve (AUC) of 0.660 (95% CI 0.636-0.681); adding age, EPIC scores, and SVI did not improve discrimination (AUC 0.679 vs 0.679 in the linked cohort). Temporal validation on 2025-2026 appointments gave an AUC of 0.624 (95% CI 0.593-0.658). Sex, race/ethnicity, and completion of SDOH screening were not associated with no-show in the fully linked subset. Conclusions. In safety-net PM&R, missed visits are associated with how a visit is scheduled rather than with EHR-derived clinical risk, neighborhood vulnerability, or demographic characteristics. Automatically generated composite risk scores should not be assumed to identify patients at risk of non-attendance. Scheduling-level levers, in particular telehealth and targeted outreach before new-patient consultations, are the more promising focus for reducing no-shows in this setting.