Unifying Interpretability and Explainability for Alzheimer's Disease Progression Prediction

TL;DR

This work addresses predicting Alzheimer's disease progression by unifying interpretability and explainability through an interpretable, mechanistic RL framework (IXRL) that models cognition trajectories over a decade using baseline data. By embedding domain knowledge via differential equations and a two-node brain network (HC and PFC), and applying SHAP for post-hoc explanations, the study evaluates four RL algorithms (TRPO, PPO, DDPG, SAC) and demonstrates that TRPO most accurately captures long-term cognition decline, while explanations reveal limited reliance on amyloid among the decisions. The findings highlight both the promise and limits of RL in healthcare: predictive accuracy can be achieved with domain-guided models, yet pathophysiological biomarkers like amyloid may be underemphasized in decisions, underscoring the need for richer biomarkers and more comprehensive brain models. The IXRL framework advances actionable, transparent AD progression modeling, offering clinicians both global and patient-level insights while pointing to practical improvements for future predictive and explanatory AI in healthcare.

Abstract

Reinforcement learning (RL) has recently shown promise in predicting Alzheimer's disease (AD) progression due to its unique ability to model domain knowledge. However, it is not clear which RL algorithms are well-suited for this task. Furthermore, these methods are not inherently explainable, limiting their applicability in real-world clinical scenarios. Our work addresses these two important questions. Using a causal, interpretable model of AD, we first compare the performance of four contemporary RL algorithms in predicting brain cognition over 10 years using only baseline (year 0) data. We then apply SHAP (SHapley Additive exPlanations) to explain the decisions made by each algorithm in the model. Our approach combines interpretability with explainability to provide insights into the key factors influencing AD progression, offering both global and individual, patient-level analysis. Our findings show that only one of the RL methods is able to satisfactorily model disease progression, but the post-hoc explanations indicate that all methods fail to properly capture the importance of amyloid accumulation, one of the pathological hallmarks of Alzheimer's disease. Our work aims to merge predictive accuracy with transparency, assisting clinicians and researchers in enhancing disease progression modeling for informed healthcare decisions. Code is available at https://github.com/rfali/xrlad.

Figures (16)

• Figure 1: Our proposed framework unifies interpretability with explainability to generate cognition trajectory predictions over 10 years and associated global and local explanations for AD progression.
• Figure 2: Interpretable RL model for AD progression prediction, expressed through DEs based on domain knowledge. The RL agent aims to predict cognition for next time step while balancing cognitive load and energy costs.
• Figure 3: Predictions on MMSE score
• Figure 4: SHAP summary bar plots for all patient predictions by TRPO, PPO, DDPG, and SAC. The set of 6 input features for the 2 brain regions studied, information processing $I_v(t - 1)$, size $X_v(t)$ and amyloid accumulation $D_v(t)$, are ranked by SHAP according to their feature importance score, whereas the colors orange and purple represent their marginal effect on the two outputs/actions (change in cognition $\Delta I_v(t)$).
• Figure 5: SHAP plots for global predictions made by TRPO. (a) and (b): Beeswarm plots for cognition prediction of each region, assigning distinctive colors to sample values (red high, blue low). (c) and (d): Dependence plots capture the partial dependence between the values of a feature (x-axis, $X_v$) and its associated contribution to model prediction (y-axis, SHAP values of $X_v$).