Predicting patient no-shows in magnetic resonance imaging appointments using interpretable machine learning

Missed appointments, or no-shows, in Magnetic Resonance Imaging (MRI) procedures lead to underutilised resources and scheduling inefficiencies in healthcare. This study proposes an interpretable machine learning framework to predict patient no-shows using real-world data from a brazilian diagnostic imaging clinic (n = 28991; no-show rate = 8.2%). We evaluated Logistic Regression, Multilayer Perceptron, XGBoost, CatBoost, and LightGBM, incorporating both patient-level and clinic-level features, including historical demand and no-show trends. Time-split cross-validation was used for hyperparameter tuning, selecting models based on AUC-PR, and for adjusting the probability threshold using the Fβ-score with β = 1.5. The LightGBM model was selected and achieved an AUC-PR of 0.2203 and AUC-ROC of 0.6559 on test data, with Precision = 19.68%, Sensitivity = 39.63%, F1-score = 26.30% and overall accuracy = 82.37%. SHAP values were used to interpret feature contributions, revealing that variables such as patient age, sex, distance to clinic, healthcare plan, waiting time, and no-show history were most influential. This framework enables data-driven overbooking strategies and personalised reminders, aligning with operational goals and minimising resource waste. Our approach demonstrates that interpretable models can support clinical decision-making in realistic environments, with potential extensions to other healthcare domains.
Missed appointments in Magnetic Resonance Imaging (MRI) generate significant costs and operational inefficiencies. This study evaluates the application of interpretable machine learning to predict patient no-shows, utilising real-world data from a diagnostic imaging clinic. Rather than seeking perfect prediction accuracy, which is challenging due to the complex nature of patient behaviour, our methodology prioritises operational utility. We calibrated the model to align with specific clinic goals: maximising the detection of potential no-shows (Sensitivity) to minimise wasted slots, even if this results in higher false alarms. By employing the SHAP framework, the model moves beyond “black-box” predictions to reveal the specific drivers of non-attendance. For radiology department managers and clinic administrators, this approach offers a realistic method to guide overbooking strategies and target reminder calls. The results demonstrate that while predicting human behaviour remains difficult, using an interpretable, operationally tuned model outperforms random guessing in identifying missed appointments, providing actionable insights for resource management.