Interpretable Deep Clustering Survival Machines for Alzheimer's Disease Risk Groups Discovery

Bojian Hou, Zixuan Wen, Jingxuan Bao, R.f Zhang, Boning Tong, Shuhua Yang, Junhao Wen, Yuhan Cui, Jason H. Moore, Andrew J. Saykin, Heng Huang, Paul M. Thompson, Marylyn D. Ritchie, Christos Davatzikos, Li Shen, Alzheimer's Disease Neuroimaging InitiativeAlzheimer 2024 · DOI

TL;DR

This paper introduces Deep Clustering Survival Machines (DCSM), a novel method that simultaneously identifies distinct risk groups for Alzheimer's Disease (AD) progression from Mild Cognitive Impairment (MCI) and predicts individual patient risk. By integrating imaging data with survival analysis, DCSM offers a more nuanced understanding of AD progression than traditional clustering methods.

What Problem Does This Paper Address?

Alzheimer's Disease (AD) is a devastating neurodegenerative disorder. A critical challenge is identifying individuals at high risk of progressing from the Mild Cognitive Impairment (MCI) stage to AD. Early identification of these at-risk groups is crucial for developing targeted treatments and improving drug discovery efforts. Existing methods often cluster patients based on their current state but fail to adequately incorporate the crucial time-to-event information (i.e., when they might convert to AD). This oversight limits their ability to accurately define distinct risk trajectories and understand the underlying biological mechanisms driving progression.

Methods

The core innovation is the Deep Clustering Survival Machines (DCSM) model. DCSM is a hybrid approach that combines discriminative and generative modeling principles. It uses a mixture of expert distributions to model the survival function, allowing for simultaneous clustering of patients into risk groups and prediction of their individual risk of converting to AD. This is achieved through maximum likelihood estimation. The model leverages deep learning for feature extraction from imaging data and integrates this with survival analysis techniques.

Datasets

The study utilizes data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Specifically, positron emission tomography (PET) data was used as features for the real-data analysis. The survival information included whether patients converted from MCI to AD and the time to conversion.

Key Insights and Ideas

  • Bridging Imaging and Survival: DCSM effectively connects neuroimaging data with the temporal aspect of disease progression, a gap often present in prior clustering approaches.
  • Simultaneous Clustering and Prediction: The ability to perform both risk group discovery and individual risk prediction in a single framework is a significant advancement.
  • Biological Validation: The genetic association analysis provides strong biological support for the identified risk groups, suggesting that the clusters are not just statistical artifacts but reflect underlying biological differences.
  • Interpretability: The paper emphasizes the "interpretable" nature of DCSM, suggesting that the identified risk factors and brain regions are meaningful and align with existing knowledge of AD.

Key Results

  • Superior Prediction Performance: DCSM achieved a higher Concordance Index (C-Index) of 0.78±0.04 compared to other baseline methods (e.g., KMeans: 0.69±0.06, Deep Cox: 0.75±0.03, DSM: 0.77±0.03, NSC: 0.72±0.05), indicating better time-to-event prediction accuracy.
  • Distinct Risk Groups: Kaplan-Meier plots (Figure 2(h)) showed a significantly larger difference in survival between the two risk groups identified by DCSM compared to other methods, confirming the distinctiveness of the clusters.
  • Genetic Associations: Targeted genetic association analyses revealed numerous single nucleotide polymorphisms (SNPs) that were significantly different between the high-risk, low-risk, and normal groups, highlighting potential genetic underpinnings of AD risk.
  • Relevant Brain Regions: The brain heat map (Figure 3(b)) identified important brain regions associated with the high-risk group, which were consistent with known brain-AD relationships.

Limitations and Future Work

  • Number of Clusters: The current DCSM model was applied to identify two risk groups. Future work could explore the optimal number of clusters for different datasets or disease stages.
  • Feature Modalities: While PET data was used, future research could integrate other imaging modalities (e.g., MRI) or clinical data to further enhance the model's predictive power and interpretability.
  • Generalizability: Further validation on independent datasets would be beneficial to confirm the generalizability of the DCSM model and its findings.
  • Causal Inference: While genetic associations are identified, exploring causal relationships between genetic factors, imaging biomarkers, and AD progression could be a future direction.

Summary updated

Abstract

Abstract Background Alzheimer's disease (AD) is a complex neurodegenerative disorder that has impacted millions of people worldwide. Identifying different risk groups converting to AD during the mild cognitive impairment (MCI) stage and determining their genetic basis would be immensely valuable for drug discovery and subsequent clinical treatment. Previous studies typically clustered subgroups by unsupervised learning techniques, neglecting the survival information. To address this problem, we propose an interpretable survival analysis method called Deep Clustering Survival Machines (DCSM), which performs clustering and risk prediction simultaneously. Method Our proposed DCSM is a hybrid survival analysis method that integrates the advantages of the discriminative and generative ideas and employ the mixture of expert distributions to fit the final survival function by maximum likelihood estimation. For the real data analysis, we use positron emission tomography (PET) from the ADNI dataset as feature. The survival information includes whether the patients convert from MCI to AD and the corresponding time. Our DCSM model predicts the risk of each patient and clusters the patients into two groups based on their risk simultaneously. Result We first evaluate time‐to‐event prediction performance by Concordance Index which is shown in Table 1 where our model outperforms the other baselines. Besides, we use the LogRank to evaluate the clustering performance. The difference in survival between two risk groups as shown in Kaplan‐Meier plots in Figure 2(h) is significantly larger than other baselines. Furthermore, to biologically validate our clustering findings, we carry out targeted genetic association analyses (Figure 3(a)). There are numerous targeted SNPs discovered when comparing high‐risk group to low‐risk group and normal group. Figure 3(b) illustrates the important brain regions for the high‐risk group. We highlight the top ten important regions which is consistent with the conventional brain‐AD relation findings. Conclusion Our DCSM model builds a good bridge between the imaging modality and survival information. We demonstrate the superiority of the DCSM by applying this approach to cluster patients with MCI into subgroups with different risks of converting to AD. Conventional clustering measurements for survival analysis along with genetic association studies successfully validate the effectiveness of our method and characterize our clustering findings.
BibTeX
@inproceedings{hou2024interpretable,
  author = {Bojian Hou and Zixuan Wen and Jingxuan Bao and R.F Zhang and Boning Tong and Shuhua Yang and Junhao Wen and Yuhan Cui and Jason H. Moore and Andrew J. Saykin and Heng Huang and Paul M. Thompson and Marylyn D. Ritchie and Christos Davatzikos and Li Shen and Alzheimer's Disease Neuroimaging Initiative},
  title = {{Interpretable Deep Clustering Survival Machines for Alzheimer's Disease Risk Groups Discovery}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.092319},
}
Figure