Bayesian Nonparametrics for Normative Modelling in Multiple Sclerosis via Modularised Inference

This paper proposes a modularized Bayesian framework combining Bayesian Additive Regression Trees (BART) for flexible, uncertainty-aware normative modeling of Multiple Sclerosis deviations and a SoftBART survival model to propagate this uncertainty, demonstrating superior calibration and prediction accuracy over traditional two-step approaches in large clinical datasets.

The Gist

Imagine you are trying to figure out how much a specific person's health has changed compared to what is "normal" for someone their age and gender. In the world of Multiple Sclerosis (MS), doctors often look at brain scans to spot these changes.

The Problem with the Old Way
Think of the old method like a rigid, straight-line ruler.

1. Too Simple: It tries to draw a straight line through complex, curvy data. Real human biology is messy and full of twists and turns (non-linear effects), but the old ruler can't bend to fit.
2. Ignoring the "Maybe": It takes a single guess (a point estimate) about how sick a person is and treats that guess as absolute fact. It ignores the fact that the measurement itself might be a little fuzzy or uncertain.
3. Bad Adjustments: When trying to account for things that mess up the data (like a blurry scan or a patient's age), it uses clumsy, "make-it-up-as-you-go" fixes.

The New Solution: A Two-Part Team
The authors propose a smarter, two-part team that works together like a specialized construction crew.

• Part 1: The Flexible Architect (The Normative Module)
  Instead of a straight ruler, they use a tool called BART (Bayesian Additive Regression Trees). Imagine this as a team of expert architects who can build a model that bends and twists to perfectly fit the complex shape of the data.

  • They don't just guess; they look at the "population average" (what is normal for everyone) and subtract that from the individual's specific situation.
  • Crucially, they can "erase" the bad parts of the data (like a blurry image) by mathematically averaging them out, so they don't ruin the final score.
  • The Output: Instead of giving a single number, this part produces a whole range of possibilities (a probability distribution), acknowledging that there is some uncertainty in the measurement.

• Part 2: The Careful Foreman (The SoftBART Survival Model)
  This second part takes the work from the Architect and uses it to predict how long a patient might stay healthy or how fast the disease might progress.

  • The Magic Trick: Usually, if you pass a guess from one step to the next, you lose the information about how unsure you were. This new method uses a "cut-posterior" technique. Think of this as a one-way door. The Foreman looks at the Architect's full range of possibilities (the uncertainty) to make a better prediction, but the Foreman's results cannot go back and mess up the Architect's original work. This keeps the two steps honest and separate.

The Results
The team tested this new approach in two ways:

1. Simulations: They created fake, difficult data scenarios to see if the math held up.
2. Real Patients: They applied it to a massive group of over 8,000 people with Multiple Sclerosis.

The Verdict
The new two-part team performed significantly better than the old "plug-in" method. It was:

• Better Calibrated: Its predictions matched reality more closely.
• More Accurate: It predicted outcomes with greater precision.
• Sharper Distinctions: It could better tell the difference between groups of patients over time (like separating those who will progress quickly from those who won't).

In short, by using a flexible, uncertainty-aware system, the researchers created a more reliable way to measure individual deviations in MS patients, leading to clearer insights into how the disease behaves.

Technical Summary

Technical Summary: Bayesian Nonparametrics for Normative Modelling in Multiple Sclerosis via Modularised Inference

Problem Statement

Normative modeling is a critical approach in neuroimaging and clinical research, generating per-subject deviation scores that quantify how an individual differs from a healthy population baseline. These scores are subsequently utilized in downstream analyses to predict clinical outcomes. However, the authors identify two significant limitations in typical pipelines:

1. Inadequate Confounder Handling: Existing methods often rely on ad-hoc or purely linear adjustments for confounding variables (such as image quality or acquisition parameters), failing to capture complex, non-linear relationships and higher-order interactions.
2. Neglect of Uncertainty: Standard pipelines typically pass point estimates of deviation scores directly into downstream models. This "plug-in" approach ignores the uncertainty inherent in the estimation of these scores, potentially leading to biased or overconfident downstream inferences.

Methodology

The paper proposes an integrated, two-module Bayesian framework designed to address these limitations through modularised inference.

1. The Normative Module (Upstream)

• Model Architecture: The framework employs Bayesian Additive Regression Trees (BART) to model the normative relationship. This nonparametric approach allows for the flexible capture of non-linear effects and higher-order interactions between covariates.
• Confounder Adjustment: Instead of simple linear regression, the model marginalizes over image-quality variables via counterfactual averaging. This ensures that the normative baseline is robust to variations in data quality.
• Deviation Definition: A crucial theoretical distinction is made in how individual deviation ($d_i$) is defined. Rather than calculating a simple residual, the authors define deviation as the difference between the individual's expected outcome given their features ($E[Y|X_i, Z_i]$) and the feature-conditional population mean ($\mu(Z_i)$).
  $$d_i = E[Y|X_i, Z_i] - \mu(Z_i)$$
  This formulation ensures the deviation represents a true departure from the expected population norm given the subject's specific characteristics.

2. The Outcome Module (Downstream)

• Model Architecture: A SoftBART survival model is used for the downstream analysis (specifically for time-to-event data in Multiple Sclerosis).
• Uncertainty Propagation: The module ingests the full posterior distribution of the deviation scores from the normative module, rather than a single point estimate.
• Modularised Inference: To prevent feedback loops where the outcome model might distort the normative estimates, the authors utilize a cut-posterior construction. This technique propagates upstream uncertainty into the downstream model while blocking information flow from the outcome back to the normative module.

Key Contributions

• Integrated Framework: The paper introduces a unified Bayesian framework that couples a flexible BART-based normative model with a SoftBART survival model.
• Theoretical Refinement: It redefines individual deviation as a difference in conditional expectations rather than a residual, providing a more rigorous statistical foundation for normative modeling.
• Uncertainty Quantification: By utilizing cut-posterior construction, the method successfully propagates uncertainty from the deviation score estimation to the final survival analysis, a feature often missing in two-step approaches.
• Robust Confounder Control: The use of counterfactual averaging within BART offers a superior alternative to linear adjustments for handling image-quality confounders.

Results

The proposed approach was evaluated through challenging simulations and applied to a large clinical dataset comprising over 8,000 Multiple Sclerosis (MS) patients. The results demonstrate that the integrated modularised approach outperforms traditional two-step plug-in Cox regression models in three key areas:

1. Calibration: The model provides better-calibrated predictions.
2. Prediction Accuracy: It achieves higher accuracy in predicting outcomes.
3. Hazard Separation: It yields improved time-varying hazard separation between patient groups.

Significance and Claims

The paper claims that modularised inference combined with BART-based normative deviations offers a dual advantage: it significantly enhances flexibility in modeling complex data structures and improves uncertainty quantification in downstream clinical analyses. The authors assert that this framework naturally extends to outcomes beyond survival analysis, suggesting a broad applicability for normative modeling in clinical settings where rigorous uncertainty handling is essential. The work positions itself as a solution to the specific methodological gaps of ad-hoc confounder adjustment and the neglect of estimation uncertainty in current normative modeling pipelines.