Detecting label noise in longitudinal Alzheimer's data with explainable artificial intelligence

Reliable classification of cognitive states in longitudinal Alzheimer's Disease (AD) studies is critical for early diagnosis and intervention. However, inconsistencies in diagnostic labeling, arising from subjective assessments, evolving clinical criteria, and measurement variability, introduce noise that can impact machine learning (ML) model performance. This study explores the potential of explainable artificial intelligence to detect and characterize noisy labels in longitudinal datasets. A predictive model is trained using a Leave-One-Subject-Out validation strategy, ensuring robustness across subjects while enabling individual-level interpretability. By leveraging SHapley Additive exPlanations values, we analyze the temporal variations in feature importance across multiple patient visits, aiming to identify transitions that may reflect either genuine cognitive changes or inconsistencies in labeling. Using statistical thresholds derived from cognitively stable individuals, we propose an approach to flag potential misclassifications while preserving clinical labels. Rather than modifying diagnoses, this framework provides a structured way to highlight cases where diagnostic reassessment may be warranted. By integrating explainability into the assessment of cognitive state transitions, this approach enhances the reliability of longitudinal analyses and supports a more robust use of ML in AD research.