Predicting postmortem neuropathology of Alzheimer’s disease and related dementias using widely accessible clinical data

Yueqi Ren, Babak Shahbaba, Craig E.l. StarkAlzheimer 2024 · DOI

TL;DR

This paper presents a novel framework that leverages widely accessible clinical data to predict postmortem neuropathology of Alzheimer's disease (AD) and related dementias. By combining clustering and classification techniques, the approach demonstrates high accuracy in identifying neuropathologic lesions, offering a cost-effective screening tool for early detection.

What Problem Does This Paper Address?

The rapidly aging global population and the increasing prevalence of Alzheimer's disease (AD) and related dementias necessitate the development of efficient and accessible screening tools. Current diagnostic methods often rely on expensive or invasive procedures. The authors aim to bridge this gap by exploring whether readily available clinical data, such as that collected during routine primary care visits, can accurately predict the underlying neuropathological hallmarks of these diseases, which are typically only confirmed postmortem. This would allow for earlier identification and intervention, potentially improving patient outcomes.

Methods

The researchers employed a two-stage approach. First, they utilized a previously developed supervised encoder (SE) to extract relevant features from potentially noisy clinical data. This encoder is designed to preserve or even enhance important information while filtering out irrelevant noise.

Following feature extraction, the study compared two main strategies for stratifying subjects based on their neuropathology:

  1. Traditional Classification and Clustering: Standard methods were used for classification and clustering.
  2. Novel Bayesian Clustering-Classification Algorithm (Infinite Mixture Classifier - IMC): This Bayesian approach integrates both clustering and classification, aiming to provide a more robust stratification.

A key aspect of their analysis involved examining longitudinal clinical data to identify distinct trajectories of subjects over time, observing how their cluster assignments changed.

Datasets

The data for this study were sourced from the National Alzheimer's Coordinating Center (NACC). This dataset is funded by the National Institute on Aging (NIA)/National Institutes of Health (NIH) Grant U24 AG072122 and includes contributions from NIA-funded Alzheimer's Disease Research Centers (ADRCs). The paper does not introduce a new benchmark dataset but rather utilizes existing, comprehensive clinical and neuropathological data.

Key Insights and Ideas

  • Clinical Data as a Powerful Predictor: The study reinforces the idea that seemingly "low-cost" clinical data, accessible during routine healthcare visits, holds significant predictive power for complex neuropathological conditions like AD.
  • Synergy of Clustering and Classification: Combining clustering (to identify distinct subgroups) with classification (to predict specific outcomes) offers a more comprehensive understanding than either method alone. This hybrid approach mitigates the limitations of purely unsupervised methods.
  • Longitudinal Data Reveals Dynamic Trajectories: Analyzing how individuals' clinical profiles evolve over time reveals distinct trajectories. These trajectories are not just descriptive but are associated with significantly different risks of developing specific neuropathologic lesions. This suggests that disease progression is not monolithic and can be captured through temporal analysis of clinical markers.
  • Generalizability of the Framework: The approach is designed to be broadly applicable, with potential to incorporate emerging data sources like blood biomarkers in the future, enhancing its screening capabilities.

Key Results

  • High Classification Accuracy: The framework achieved "relatively high classification accuracy" in predicting neuropathologic lesions using widely accessible clinical data. While specific accuracy percentages are not detailed in the abstract, the term "high" suggests a promising level of performance.
  • Meaningful Clinical Subgroups: The identified clusters, derived from both the SE's latent features and the IMC, provided meaningful clinical diagnostic information. These subgroups effectively differentiate subjects along the clinical and pathological continuum of AD and related dementias.
  • Distinct Trajectory Subgroups: When longitudinal data was used, distinct subject trajectories were observed. These trajectory subgroups demonstrated significantly different risks of exhibiting specific neuropathologic lesions identified through postmortem examination.

Limitations and Future Work

  • Limitations: The abstract does not explicitly detail limitations acknowledged by the authors. However, the reliance on postmortem neuropathology as the ground truth implies a potential temporal gap between clinical assessment and definitive diagnosis. The "noisy" nature of input features, even after SE processing, suggests that perfect feature extraction might still be a challenge.
  • Future Work: The authors explicitly state that their framework can be "easily extended to include additional data to improve screening for neuropathology using widely accessible data." This includes incorporating emerging technologies like blood biomarkers, which are becoming increasingly accessible. The focus on Medicare annual wellness visits highlights the potential for broad public health screening applications.

Summary updated

Abstract

Abstract Background Faced with a rapidly aging population and the rising prevalence of Alzheimer’s disease (AD) and related dementias, the field needs to urgently consider screening tools that utilize widely accessible data modalities. We have previously shown that lower‐cost data, operationalized as data modalities accessible at primary care visits, can indeed accurately predict AD clinical diagnosis and that clustering these data can provide useful information. Here, we apply a similar approach to predicting histopathological status. Method We first applied our previously‐developed feature extraction method based on a supervised encoder (SE) to transform potentially noisy input features while maintaining or amplifying relevant information. We next performed classification and clustering to stratify subjects by their neuropathology. Here, we compared two traditional classification methods with a novel Bayesian clustering‐classification algorithm called an Infinite Mixture Classifier (IMC). We identified distinct trajectories of subjects based upon changes in cluster assignment over time. Data for this study come from the National Alzheimer’s Coordinating Center, funded by NIA/NIH Grant U24 AG072122 and contributed to by NIA‐funded ADRCs. Result We found that relatively high classification accuracy of neuropathologic lesions was possible using widely accessible, lower cost clinical data. In addition, the supervised clusters, derived from using the SE’s latent features and from the IMC, held meaningful clinical diagnostic information that differentiates subjects along the clinical and pathologic continuum. When clusters were derived using longitudinal clinical data, we further observed distinct trajectories of subjects across time as their cluster assignments changed. These trajectory subgroups have significantly different risk of showcasing each type of neuropathologic lesion obtained from postmortem neuropathology. Conclusion Our framework benefits from the combined strengths of clustering and classification methods while avoiding drawbacks of unsupervised methods. By using lower cost features, which could be obtained at Medicare annual wellness visits, this work is broadly generalizable and has direct implications for screening of neuropathologic lesions of AD and related dementias for the public. As blood biomarkers become more accessible, our framework can be easily extended to include additional data to improve screening for neuropathology using widely accessible data.
BibTeX
@inproceedings{ren2024predicting,
  author = {Yueqi Ren and Babak Shahbaba and Craig E.L. Stark},
  title = {{Predicting postmortem neuropathology of Alzheimer’s disease and related dementias using widely accessible clinical data}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.092688},
}
Figure