Artificial Intelligence for Predicting Waiting Times in Specialized Healthcare Regulation: A Machine Learning Study Using Real-World Data from the Brazilian Unified Health System

Long waiting times hinder timely access to specialised care in Brazil’s Unified Health System (SUS), and municipal services need analytical support to find bottlenecks and unequal access. This study evaluated machine learning models for predicting outpatient waiting time, defined as the number of days between a request and the appointment date. It used real-world SUS records from SIMEC / Programa Eficiência in Juazeiro do Norte, Ceará, Brazil. The data were cleaned by removing corrupted dates and impossible waiting times, standardising categories and handling missing values. Temporal, patient and demand features were then derived. XGBoost, Random Forest and LightGBM regression models were trained with cross-validation and hyperparameter tuning. XGBoost performed best (R² = 0.75), followed by Random Forest (0.72) and LightGBM (0.70). The two leading predictors were referral by the regulator (20.0% of feature importance) and service-provider availability (17.6%). Median waits were longer for municipal referral pathways (83–105 days) than for regulator/SESAU referrals (7–13 days). Rheumatology (202 days), paediatric neurology (177 days) and several ophthalmology exams (160–169 days) had the longest median waits. In the fitted model, vacancy reallocation (−26.5 days) and SESAU referral (−18.3 days) were associated with shorter predicted waits, and having few service providers (+7.8 days) with longer ones. These associations are not causal. Machine learning can support waiting-time prediction and bottleneck detection in municipal SUS services. The next steps are integration into SIMEC, prospective evaluation, and monitoring for bias, transparency and equitable access.