Bayesian Event-Based Model for Disease Subtype and Stage Inference

1. Problem Statement

Chronic diseases, particularly neurodegenerative disorders like Alzheimer's Disease (AD), exhibit significant heterogeneity in progression trajectories across patients. While the Subtype and Stage Inference (SuStaIn) model has become the de facto standard for identifying disease subtypes and progression orders from cross-sectional data, it suffers from several critical limitations:

• Robustness: SuStaIn is brittle under model misspecification (e.g., non-Gaussian biomarker distributions, continuous disease stages, or uneven subtype distributions).
• Assumptions: It relies on static estimates of biomarker distribution parameters and assumes uniform priors over disease stages, which often contradicts real-world cohort data (e.g., ADNI) where later stages are underrepresented.
• Performance: In synthetic experiments with known ground truth, SuStaIn often fails to accurately infer ordering, staging, or subtype assignments, particularly when data deviates from ideal Gaussian assumptions.
• Computational Efficiency: SuStaIn is computationally demanding and prone to errors (e.g., singular matrices) in practice.

The authors propose bebms (Bayesian Event-Based Model for Subtyping) to address these issues by embedding the Event-Based Model (EBM) into a principled Bayesian framework.

2. Methodology

bebms models disease progression as a sequence of biomarker events where each biomarker transitions from a "pre-event" (healthy) to a "post-event" (pathological) state.

Key Technical Components:

• Generative Model:

  • Subtypes: The model assumes $T$ distinct subtypes, each defined by a unique sequence of biomarker events.
  • Stages: A patient's stage $k_j$ corresponds to the highest rank of the event that has occurred. Healthy patients are assigned stage $-1$.
  • Distributions: Biomarker measurements are modeled using Gaussian distributions for both healthy ($\phi$) and pathological ($\theta$) states.
  • Priors: The model places Dirichlet priors on subtype probabilities ($\pi_t$) and stage probabilities ($\pi_{k|t}$), allowing the model to learn the distribution of stages and subtypes from the data rather than assuming uniformity.

• Inference Procedure (MCMC):

  • Unlike SuStaIn's Expectation-Maximization (EM) approach, bebms uses a Metropolis-Hastings Markov Chain Monte Carlo (MCMC) sampler.
  • Iterative Estimation: The algorithm iteratively estimates distribution parameters ($\theta, \phi$), stage priors, and subtype priors.
  • Conjugate Priors: It utilizes Normal-Inverse Gamma (NIG) conjugate priors for the mean and variance of biomarker distributions. This allows for analytical updates of parameters based on soft assignments of patients to pre/post-event states, improving stability and reducing sensitivity to initialization.
  • Handling Missing Data: The likelihood calculation naturally excludes missing entries (NaNs), making the model robust to incomplete datasets without requiring imputation.

• Model Selection:

  • The optimal number of subtypes ($T$) is determined via K-fold cross-validation using a Cross-Validation Information Criterion (CVIC), similar to SuStaIn but applied within the Bayesian framework.

• Blind Variant:

  • The authors also introduce bebms (blind), which mirrors SuStaIn's philosophy by using diagnosis labels only for initializing distribution parameters, not for the inference tasks. This allows for a direct comparison of robustness against diagnostic uncertainty.

3. Key Contributions

1. Principled Bayesian Framework: The introduction of bebms, which integrates EBM with Bayesian inference, enabling dynamic learning of stage and subtype distributions rather than relying on static assumptions.
2. Robustness to Misspecification: Demonstrated superior performance over SuStaIn when data violates standard assumptions (non-Gaussian distributions, continuous stages, skewed stage distributions).
3. Efficiency and Stability: The MCMC approach with conjugate priors reduces runtime and eliminates the "IndexError" and "Singular matrix" failures frequently encountered with SuStaIn.
4. Handling Missing Data: Unlike the standard SuStaIn implementation, bebms natively handles missing biomarker values.
5. Real-World Validation: Application to the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, showing results that align more closely with established neuropathological consensus (Murray et al., 2011) than SuStaIn.

4. Experimental Results

Synthetic Data Experiments (1,320 datasets)

The authors tested bebms against SuStaIn (GMM and KDE variants) across 11 experimental conditions varying sample size, healthy ratios, and data generation models (EBM and Sigmoid).

• Ordering Accuracy: bebms reduced ordering error (Kendall's $\tau$ distance) by 27.3% compared to SuStaIn GMM. It remained robust even under non-Gaussian data, whereas SuStaIn GMM performance degraded significantly.
• Subtyping Accuracy: bebms achieved an Adjusted Rand Index (ARI) of 0.25, significantly outperforming SuStaIn GMM (0.16) and KDE (0.13).
• Staging Accuracy: For healthy controls (ground truth stage 0), bebms assigned an average stage of 0.16, whereas SuStaIn GMM assigned 3.03 and KDE assigned 1.45, indicating SuStaIn severely overestimates disease progression in healthy individuals.
• Runtime: bebms was approximately 2-3x faster than SuStaIn (avg. 2.37 mins vs. 6.88 mins for GMM).
• Subtype Count Estimation: bebms accurately estimated the number of subtypes with low Mean Absolute Error (1.08), while SuStaIn KDE tended to overfit (1.92).

Real-World ADNI Dataset

• Model Selection: bebms selected 3 subtypes, while SuStaIn GMM selected 6 (suggesting overfitting).
• Biological Plausibility:
  • bebms Subtype 2 (67.9%): CSF-first progression (Amyloid/Tau $\to$ Neurodegeneration $\to$ Cognition). Matches Typical AD.
  • bebms Subtype 3 (28.7%): Entorhinal/Hippocampal-first. Matches Limbic-predominant AD.
  • bebms Subtype 1 (3.4%): Neocortical/Cognitive-first with delayed hippocampal involvement. Matches Hippocampal-sparing AD.
  • These proportions and sequences align closely with the canonical neuropathological subtypes defined by Murray et al. (2011). In contrast, SuStaIn's 6 subtypes were less distinct and included biologically implausible sequences.

5. Significance and Conclusion

The paper establishes bebms as a more robust, accurate, and computationally efficient alternative to SuStaIn for modeling disease heterogeneity.

• Clinical Impact: By providing more accurate subtype identification and staging, bebms can improve patient stratification in clinical trials and enable precision medicine approaches for Alzheimer's and other chronic diseases.
• Data Efficiency: The model performs well with smaller cohorts (saturation around 300 participants) and is less sensitive to the ratio of healthy controls, making it suitable for rare diseases.
• Future Directions: While bebms excels with cross-sectional data, the authors note that future work should explore longitudinal data integration and further validation on diverse datasets.

In summary, bebms represents a significant methodological advancement in disease progression modeling, offering a principled Bayesian solution that overcomes the brittleness and limitations of current state-of-the-art methods.