Interpretable Survival Analysis for Alzheimer’s Progression
Original Title: Basic Science and Pathogenesis
Journal: Alzheimer's & dementia : the journal of the Alzheimer's Association
Overview
This research addresses the challenge of predicting the progression of Alzheimer’s disease and related dementias using survival analysis. While deep learning models offer high predictive performance, their complex architectures often obscure the biological factors driving their outputs. To resolve this, the authors introduce the Neural Additive Deep Clustering Survival Machines (NADCSM) framework. This model utilizes data from the Alzheimer’s Disease Neuroimaging Initiative, specifically focusing on AV45 Florbetapir PET imaging, genotyping, and demographic information to track the transition from mild cognitive impairment to early Alzheimer’s disease. The framework models survival times using Weibull distributions, which are well-suited for characterizing time-to-event data in clinical contexts. By integrating Neural Additive Models, the system processes each input feature through independent multilayer perceptrons to generate univariate shape functions. This structure allows the model to quantify the specific contribution of various brain regions to the overall risk profile without sacrificing the capacity to handle non-linear relationships.
Novelty
The technical novelty of this work lies in the integration of Neural Additive Models with Deep Clustering Survival Machines to bridge the gap between performance and transparency. Traditional survival models like DeepCox often lack interpretability, whereas NADCSM provides clear visualizations of how individual brain regions influence disease risk. By using univariate shape functions, the model reveals the functional relationship between regional amyloid burden and survival probability. The results demonstrate that NADCSM achieves a C-index of 0.7772 ± 0.0236, which is comparable to the 0.7789 ± 0.0193 achieved by the more complex DCSM. Furthermore, it significantly outperforms DeepCox in survival curve separation, as evidenced by its LogRank score of 244.1572 ± 24.1248 compared to the DeepCox score of 93.4171 ± 9.7990. This performance indicates that the model maintains high predictive accuracy while providing an interpretable mechanism that identifies the specific impact of regions such as the left fusiform gyrus and the right cerebellum crus.
Potential Clinical / Research Applications
The potential clinical applications for NADCSM are significant for both patient management and clinical trial design. In a clinical setting, physicians could use the generated shape functions to explain to patients how their specific PET imaging results translate into a localized risk of progression, fostering better communication and personalized care plans. For researchers, the model serves as a tool for biomarker discovery by highlighting brain regions that contribute most to the survival distribution. This could refine the selection criteria for clinical trials by identifying individuals at the highest risk of rapid decline based on regional amyloid levels. Additionally, the framework could be adapted to other neurodegenerative conditions where time-to-event analysis is critical, such as Parkinson’s disease or amyotrophic lateral sclerosis. By identifying the most influential regions for different pathologies, the model could help in the development of targeted neuroprotective therapies tailored to the specific anatomical vulnerabilities of each disease subtype.