Concept drift poses a critical challenge in clinical decision support systems, where the relationship between patient features and treatment outcomes can shift whenever new treatment guidelines or protocols are introduced. In such non-stationary medical settings, standard distributional metrics may overlook subtle changes in the feature–label relationship, leading to degraded model performance. We address this by proposing a SHAP (SHapley Additive exPlanations)-based drift detection method, which monitors how feature importance distributions evolve across consecutive windows of data. Unlike traditional approaches (e.g. Kolmogorov-Smirnov, Jensen-Shannon, Hellinger), our SHAP-centric method captures hidden concept drifts that do not necessarily manifest in raw feature statistics but significantly alter a model’s decision boundary. We illustrate the effectiveness of this approach using an extended Iris dataset with artificially injected drifts (sudden, gradual, incremental, recurring, and blip), revealing that SHAP-based detection consistently achieves higher F1-scores compared to baseline drift metrics. Moreover, by examining which features exhibit the largest shifts in SHAP values, clinicians, or data scientists can identify and respond to changing clinical guidelines or patient populations more rapidly. These results underscore the value of explainable AI techniques in medical environments, offering both robust drift detection and actionable insights into how evolving practices influence patient outcomes.