📊 statistics

Continuous-Time Bayesian Networks with Structured Shrinkage Priors for Modelling Multimorbidity Trajectories in Large-Scale Electronic Health Records

This paper proposes a structured Bayesian continuous-time network framework with order-dependent shrinkage priors to model complex multimorbidity trajectories in large-scale electronic health records, demonstrating through simulations and UK Biobank data that a spike-and-slab prior effectively identifies clinically interpretable disease clusters while controlling false discoveries.

Original authors: Oyebayo R. Olaniran, Soumya S. Paria, Mizanur Khondoker, Alexander J. MacGregor, Alexandra Lewin

Published 2026-07-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Oyebayo R. Olaniran, Soumya S. Paria, Mizanur Khondoker, Alexander J. MacGregor, Alexandra Lewin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Mapping the "Domino Effect" of Illness

Imagine your health as a room full of dominoes. Each domino represents a different long-term condition, like diabetes, high blood pressure, or asthma. Usually, doctors look at these dominoes one by one or just see which ones are standing together at a single moment in time.

But this paper argues that illness is more like a chain reaction. When one domino falls (you get a disease), it might knock over another one later, or it might make it much more likely for a third one to fall. The goal of this research was to build a map that shows exactly how these dominoes knock each other over over time, using massive amounts of patient data.

The Problem: Too Many Possibilities, Not Enough Clarity

The researchers wanted to use data from the UK Biobank (a giant library of health records from 33,000 people) to see how 10 common diseases interact.

The problem is that the number of ways these diseases can interact is mind-bogglingly huge.

The Analogy: Imagine trying to guess the outcome of a game where you roll 10 dice. You have to figure out if rolling a "6" on the first die makes it more likely to roll a "6" on the second die, or if rolling a "6" on the first and a "3" on the second makes a "6" on the third die more likely.
In medical terms, this is called "combinatorial explosion." If you try to calculate every possible combination of diseases interacting, the math becomes impossible to solve, and you end up with too many "false alarms" (thinking diseases are connected when they aren't).

The Solution: A "Smart Filter" for the Math

To solve this, the authors built a new type of statistical model called a Continuous-Time Bayesian Network (CTBN). Think of this model as a super-smart detective that watches the dominoes fall in real-time, rather than just taking a snapshot of the room once a year.

However, the detective needs help deciding which connections are real and which are just random noise. To do this, the team tested four different "filters" (mathematical tools called priors) to see which one was best at finding the true connections while ignoring the fake ones.

They tested:

The "Spike-and-Slab" Filter: This is like a strict bouncer. It has a "spike" that says "zero" (this connection doesn't exist) and a "slab" that says "maybe" (this connection is real). It is very good at making a hard decision: "Yes, this matters" or "No, this is noise."
Three Other Filters (LASSO, Horseshoe, etc.): These are more like gentle nudges. They try to shrink the importance of unlikely connections but rarely say "zero" completely. They are good at estimation but bad at making hard decisions about what to include.

The Experiment: The "Training Gym"

Before looking at real patients, the authors created a simulated gym (a computer simulation). They built a fake world with 5,000 fake patients and a known "truth" about which diseases caused which others. They then ran their four filters through this gym to see which one could correctly identify the true connections.

The Results:

The Winner: The Spike-and-Slab filter was the clear champion. It was the only one that could accurately separate the real disease connections from the noise. It found the right links with very few mistakes.
The Runners-Up: The other three filters were okay at guessing the strength of a connection, but they failed at the most important job: deciding which connections actually exist. They kept almost everything, making the final map too cluttered to be useful.

The Real-World Discovery: Two "Disease Neighborhoods"

Using the winning filter on the real UK Biobank data, the researchers uncovered a clear map of how diseases spread. They found that the 10 diseases didn't just form a messy web; they clustered into two distinct "neighborhoods":

The Cardiometabolic Neighborhood: This group includes Diabetes, High Blood Pressure, and Heart Disease.
- The Finding: These diseases are like a tight-knit family. If you have Diabetes, it significantly increases your chance of getting High Blood Pressure, which in turn increases your chance of Heart Disease. The model showed that having Diabetes makes you twice as likely to develop High Blood Pressure compared to someone without it.
The Inflammatory Neighborhood: This group includes conditions like Rhinitis (hay fever), Eczema, and Chronic Lung Disease.
- The Finding: These are also tightly linked. If you have chronic lung issues, you are much more likely to develop hay fever or eczema, and vice versa.

The "Cross-Talk":
The map also showed that these two neighborhoods don't talk to each other much. There are very few bridges between the "Heart/Diabetes" group and the "Lung/Skin" group. This suggests that the biological reasons for heart disease are quite different from the reasons for skin or lung disease.

The "Double Trouble" Twist

The researchers also looked at what happens when a person has two diseases at once.

The Analogy: Imagine two people pushing a heavy door. You might expect them to push it twice as hard as one person.
The Finding: The model found that for some disease combinations, the "push" was actually less than double. This is called sub-multiplicative antagonism.
Example: If a person already has High Blood Pressure and Heart Disease, adding Diabetes doesn't increase the risk of a third condition as much as you would mathematically expect. It seems the body's "risk pathways" get saturated; once the damage is done by the first two, the third one doesn't add as much extra risk as a simple math equation would predict.

Summary

In simple terms, this paper built a better way to read the "story" of how diseases spread through a population over time.

They created a new math model that handles time and complexity better than old methods.
They tested four different ways to filter out the noise and found that the "Spike-and-Slab" method was the most accurate.
They applied this to real data and found that diseases tend to cluster into two main groups: Heart/Metabolic and Lung/Skin.
They discovered that while having one disease increases the risk of another, having two diseases doesn't always create a "double" risk—it can sometimes be less than expected due to how the body's systems overlap.

This helps doctors understand that treating a patient isn't just about fixing one disease; it's about understanding which "domino" they are standing on and how it might knock over the next one.

Technical Summary: Continuous-Time Bayesian Networks with Structured Shrinkage Priors for Modelling Multimorbidity Trajectories

Problem Statement
Multiple long-term conditions (MLTCs), or multimorbidity, arise through complex, time-dependent interactions among diseases. Existing statistical methods often struggle to jointly model disease progression, multimorbidity networks, and high-dimensional risk factors within a unified framework. Cross-sectional analyses capture co-occurrence but miss temporal dynamics; multi-state models scale poorly to large numbers of interacting conditions; and network-based approaches often lack formal probabilistic models for estimation and uncertainty quantification. Furthermore, applying Continuous-Time Bayesian Networks (CTBNs) to high-dimensional electronic health records (EHRs) faces two specific challenges: the combinatorial explosion of parameters arising from higher-order disease interactions and the need for biologically interpretable sparsity with calibrated uncertainty.

Methodology
The authors propose a structured Bayesian CTBN framework designed for longitudinal EHR data. The core methodological components are:

Model Structure: The joint progression of $M$ inflammatory conditions is modeled as a multivariate stochastic process where each condition evolves as a finite-state continuous-time Markov process. The model assumes an absorbing state (onset only, no remission) for chronic conditions. The instantaneous onset rate (transition intensity) for a condition depends on the current states of its "parent" conditions and exogenous covariates via a log-linear parameterization. This allows for the modeling of main effects and pairwise synergistic interactions.
Likelihood Formulation: The CTBN likelihood is derived and shown to be equivalent to a Poisson Generalized Linear Model (GLM) with a log-offset, where the exposure is the time spent in the at-risk state. This factorization across conditions allows for computationally tractable estimation even with moderate numbers of conditions.
Structured Shrinkage Priors: To address the combinatorial growth of interaction parameters, the authors embed four order-dependent shrinkage priors within the CTBN likelihood. These priors enforce stronger penalization on higher-order interaction terms than on main effects, adhering to the hierarchical principle that complex multi-condition synergies are rare. The four priors compared are:
- Hierarchical Structured (HS) Normal: A scale mixture yielding a Student-t marginal with order-specific variance.
- Bayesian LASSO: A double-exponential marginal with order-dependent rates.
- Regularised Horseshoe: A global-local prior with order-penalized global scale.
- Spike-and-Slab: A two-component mixture (diffuse slab vs. concentrated spike) with order-specific inclusion probabilities.
Variable Selection: For the spike-and-slab prior, Posterior Inclusion Probabilities (PIPs) are derived analytically. For the continuous shrinkage priors (HS, LASSO, Horseshoe), which do not yield explicit PIPs, the authors construct "pseudo-PIPs" based on the posterior shrinkage factor to enable fair comparison of variable selection performance across all methods.

Key Contributions
The paper makes four primary methodological contributions:

Unified Framework: It embeds four distinct order-dependent shrinkage priors into a unified CTBN likelihood, explicitly encoding the biological expectation that complex synergistic pathways are sparse.
Comparative Selection Metric: It derives PIPs for the spike-and-slab and pseudo-PIPs for the continuous priors, placing all four on a common footing for variable selection and network recovery.
Calibrated Simulation: It conducts an extensive simulation study to isolate parameter recovery, variable selection, and predictive behavior across different data-generating scenarios and model specifications (under-, correctly, and over-specified).
Empirical Application: It applies the framework to UK Biobank primary-care records to model the progression of ten prevalent conditions, recovering sparse, interpretable disease networks.

Results

Simulation Findings: Across multiple scenarios, the spike-and-slab prior achieved the best network recovery, variable-selection accuracy, and false-discovery control. It attained True Positive Rates (TPR) of 76–82% for main effects and ~72% for interactions with False Positive Rates (FPR) as low as ~1–4% under correct specification. In contrast, the three continuous shrinkage priors (HS, LASSO, Regularised Horseshoe) failed to perform "hard" variable selection at a standard 0.5 threshold, declaring nearly all coefficients active (TPR $\approx$ FPR $\approx$ 100%) and showing modest selection AUCs (68–71%).
Predictive Performance: All four priors demonstrated nearly identical predictive accuracy (Poisson log-likelihood and Brier scores) across analysis models, confirming that predictive performance is robust to prior specification when the goal is estimation rather than structure recovery.
Empirical Application (UK Biobank): Applying the recommended spike-and-slab model ( $P=2$ $P = 2$ , $\theta=1$ $θ = 1$ ) to 33,558 participants revealed two dominant disease modules:
- A cardiometabolic cluster centered on Diabetes Mellitus (DM), Hypertensive Diseases (HTN), and Ischaemic Heart Diseases (IHD), characterized by dense, bidirectional dependencies (e.g., DM $\to$ HTN with a rate ratio of 2.13).
- An inflammatory cluster linking Rhinitis/Sinusitis (RS), Dermatitis/Eczema (DE), and Chronic Lower Respiratory Diseases (CLRD).
- The model identified sub-multiplicative antagonism in triplets (e.g., {RS, DM} $\to$ HTN), suggesting that joint exposure yields less cumulative risk than the sum of marginal exposures, consistent with overlapping biological pathways.
- Covariate analysis confirmed age as the strongest predictor, with specific inflammatory markers (e.g., GlycA, WBC) and lifestyle factors (smoking, BMI) showing condition-specific associations.

Significance and Claims
The paper claims that the proposed structured Bayesian CTBN framework successfully addresses the dual challenges of high-dimensional parameter spaces and the need for interpretable sparsity in multimorbidity modeling. The authors assert that the spike-and-slab prior is the superior choice for network structure learning, offering the best balance of parameter recovery, false-discovery control, and uncertainty calibration, whereas continuous priors are better suited for pure prediction tasks where structure recovery is not the primary goal.

The application to UK Biobank data demonstrates the framework's ability to recover clinically plausible, directed disease-dependency networks that quantify progression burdens and higher-order interactions. The authors note that while the findings align with established clinical knowledge (e.g., the cardiometabolic cluster), the dynamic, directional nature of the model provides new insights into the temporal ordering and magnitude of disease cascades that cross-sectional analyses cannot capture. The study concludes that extending this framework to the full set of 60 InflAIM conditions represents a substantial advance in population-scale disease-network modeling, with potential translational implications for risk stratification and precision medicine.