LUME-DBN: Full Bayesian Learning of DBNs from Incomplete data in Intensive Care

Basic Information

• Paper ID: 2511.04333
• Title: LUME-DBN: Full Bayesian Learning of DBNs from Incomplete data in Intensive Care
• Authors: Federico Pirola (University of Milano-Bicocca), Fabio Stella (University of Milano-Bicocca), Marco Grzegorczyk (University of Groningen)
• Classification: cs.LG (Machine Learning), cs.AI (Artificial Intelligence)
• Publication Date: November 6, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2511.04333

Abstract

Dynamic Bayesian Networks (DBNs) are increasingly applied in healthcare due to their ability to model complex temporal relationships in patient data while maintaining interpretability—a critical feature for clinical decision-making. However, existing methods for handling missing values in longitudinal clinical datasets primarily derive from static Bayesian network literature and fail to appropriately account for the temporal nature of the data. This gap limits the quantification of temporal uncertainty, which is particularly crucial in intensive care settings where understanding temporal dynamics is essential for model credibility and applicability across different patient populations. This paper proposes a novel Gibbs sampling-based approach for learning DBNs from incomplete data, treating each missing value as an unknown parameter following a Gaussian distribution, enabling principled imputation and uncertainty estimation through full conditional distribution sampling.

Research Background and Motivation

Core Problem

The core problem addressed by this research is how to effectively learn dynamic Bayesian networks in the presence of substantial missing data, particularly in intensive care unit (ICU) applications.

Problem Significance

1. Clinical Urgency: In ICUs, timely and accurate assessment of patient condition evolution is crucial for guiding intervention measures
2. Data Quality Challenges: ICU data frequently suffers from missing values, irregular sampling, and measurement bias
3. Uncertainty Quantification: Traditional methods fail to adequately account for uncertainty introduced by missingness, potentially leading to biased parameter estimates

Limitations of Existing Methods

1. Temporal Blindness of Static Methods: Existing missing data handling methods primarily originate from static Bayesian networks and do not account for temporal properties
2. Insufficiency of Frequentist Approaches: Traditional imputation or frequentist methods may inadequately consider uncertainty introduced by missingness
3. Local Optimality Problem: Structural expectation-maximization (SEM) algorithms and similar methods are prone to converging to local optima

Research Motivation

To develop a fully Bayesian framework capable of simultaneously handling uncertainty in network structure, parameters, and missing values, providing more reliable support for clinical decision-making.

Core Contributions

1. Theoretical Contribution: Derives closed-form solutions for full conditional distributions (FCDs) of missing values in DBNs, proving their tractability
2. Methodological Innovation: Proposes the LUME-DBN algorithm, combining Gibbs sampling for missing data imputation with MCMC structure learning
3. Experimental Validation: Validates the method's effectiveness on simulated and real ICU data, demonstrating superior reconstruction accuracy compared to methods such as MICE
4. Clinical Application: Demonstrates the discovery of meaningful temporal relationships in different ICU types using the PhysioNet 2012 dataset

Methodology Details

Task Definition

Input: Multivariate time series data with missing values $D \in \mathbb{R}^{N \times k \times (T+1)}$, where $N$ is the number of samples, $k$ is the number of variables, and $T+1$ is the number of time points

Output: Posterior distribution samples of DBN structure, parameters, and missing values

Constraints: Assumes first-order Markovian property and absence of instantaneous effects

Model Architecture

DBN Foundational Framework

The DBN is modeled as $k$ independent Bayesian linear regression (BLR) models:

$x_i^t = \beta_0^{(i)} + \sum_{j:(X_j^{t-1} \in \pi(i))} \beta_j^{(i)} x_j^{t-1} + \epsilon_i^t$

where $\pi(i)$ denotes the set of parent nodes of variable $X_i$, and $\epsilon_i^t \sim N(0, \sigma^2_{(i)})$.

Prior Distribution Specification

• Regression coefficients: $\beta^{(i)} \sim N(\mu^{(i)}, \sigma^2_{(i)}\delta^2_{(i)}I)$
• Noise parameters: $\sigma^2_{(i)} \sim \text{Inv-Gamma}(a, b)$
• Uncertainty parameters: $\delta^2_{(i)} \sim \text{Inv-Gamma}(\alpha_\delta, \beta_\delta)$
• Parent set size: $|\pi(i)| \sim \text{Poisson}(\lambda)$

Full Conditional Distribution of Missing Values

For the missing value $x_i^t[MIS]$ of variable $X_i$ at time $t$, its FCD is:

$P(x_i^t[MIS] | \cdot) = N(\mu_*, \sigma^2_*)$

where: $\sigma^2_* = \left(\frac{1}{\sigma^2_{(i)}} + \sum_{j:(X_i^t \in \pi(j))} \frac{(\beta_i^{(j)})^2}{\sigma^2_{(j)}}\right)^{-1}$

$\mu_* = \sigma^2_* \cdot \left(\frac{\mu_i^t}{\sigma^2_{(i)}} + \sum_{j:(X_i^t \in \pi(j))} \frac{\beta_i^{(j)}(x_j^{t+1} - \mu_{{\{-i\}}}^{(j)(t+1)})}{\sigma^2_{(j)}}\right)$

Technical Innovations

1. Unified Imputation Strategy: Designs Gibbs steps for jointly updating missing values across all regression models
2. Closed-Form Solution Derivation: Proves the tractability of missing value FCDs, enabling efficient MCMC inference
3. Temporal Invariance: The FCD structure exhibits temporal invariance relative to DBN parameters, improving computational efficiency
4. Escaping Local Optima: MCMC sampling enables escape from local minima, achieving more accurate network reconstruction

Experimental Setup

Datasets

Simulated Data

• Structure: 10 independent 10-node DBN structures, with each node having at most 5 parents
• Temporal Length: $T \in \{50, 100, 200\}$
• Missing Rates: $\{10\%, 20\%, 30\%, 40\%\}$
• Parameter Settings: Regression coefficients sampled from $\text{Uniform}[0.2, 0.8]$, noise variance $\sigma^2 = 1$

Real Data

• Data Source: PhysioNet 2012 Challenge dataset
• Patient Count: 20,000+ adult ICU patients
• Time Window: First 48 hours of ICU admission
• Variable Count: 11 clinical variables (vital signs, blood indicators, physiological characteristics)
• ICU Grouping: MICU (34 cases), SICU (104 cases), CCU (114 cases), CSRU (62 cases)

Evaluation Metrics

• Structure Reconstruction: Area under the precision-recall curve (AUC-PR)
• Convergence Diagnosis: Potential scale reduction factor (PSRF < 1.1)
• Statistical Significance: Paired t-test

Comparison Methods

• MICE: Multivariate imputation by chained equations
• Temporal MICE: Temporal MICE variant using lagged predictor variables
• Complete Data: Serves as performance upper bound reference

Implementation Details

• Sampling Iterations: 20,000 iterations, with first 5,000 as burn-in
• Missing Value Update Frequency: Updated every 10 iterations ($E_M = 10$)
• Chain Thinning: Retain 1 sample per 5 to reduce autocorrelation
• Prior Parameters: $\lambda = 1$, $\sigma^2_{(i)} = \delta^2_{(i)} = 1$

Experimental Results

Main Results

Simulated Data Results

LUME-DBN significantly outperforms baseline methods across all experimental settings:

1. MICE Performance: Completely fails when missing rates exceed 20%, reflecting its ineffectiveness on temporal data
2. Temporal MICE: Outperforms MICE but remains significantly inferior to LUME-DBN
3. LUME-DBN Advantages: Particularly outstanding at high missing rates, with minimal performance loss compared to complete data in large sample scenarios

Convergence Analysis

• Structure Convergence: Converges within 1.5k iterations across all missing rates
• Missing Value Convergence: Requires 5k iterations at 40% missing rate
• Convergence Stability: Convergence time increases with missing rate, but eventually converges in all cases

ICU Data Analysis

Discovered Clinical Relationships

1. Self-Regulatory Loops: Strong internal connections within pressure parameters (MAP, Sys, Dias) and respiratory variables (FiO2, PaCO2, PaO2, pH)
2. Neurological Interactions: Decreased consciousness level leads to increased heart rate (CCU: GCS → HR)
3. Hemodynamic Effects: Blood pressure strongly influences consciousness level (Medical patients: Dias, MAP → GCS)
4. Temperature Regulation Dynamics: Temperature changes during surgical recovery affect urine output (Temp → Urine)
5. Cardiopulmonary Feedback: Low oxygen levels trigger compensatory heart rate increase (FiO2 → HR)

Standardization Effects

• Local Standardization: Reveals more ICU-specific relationships
• Global Standardization: Network displays more commonalities, but some relationships lack clinical evidence support

Related Work

Missing Data in Static Bayesian Networks

• SEM Algorithm: Hard EM variants effective with limited data but prone to local optima
• MCMC Methods: Recent sampling methods escape local minima, achieving more accurate reconstruction

Dynamic Bayesian Networks

• Existing Methods: Primarily use model-agnostic methods such as MICE for missing data handling
• This Paper's Contribution: First extension of sampling methods to missing data handling in DBNs

Clinical Applications

• Organ Failure Prediction: DBNs used for predicting organ failure trajectories
• Physiological Change Prediction: Predicting physiological changes and mortality risk
• Decision Support: Providing interpretable decision support

Conclusions and Discussion

Main Conclusions

1. Method Effectiveness: LUME-DBN outperforms existing methods in both structure reconstruction and missing value imputation
2. Clinical Relevance: Discovered temporal relationships have clinical significance, supporting safer clinical decision-making
3. Uncertainty Quantification: The fully Bayesian framework provides explicit uncertainty encoding for models, parameters, and missing values

Limitations

1. Computational Complexity: MCMC sampling has high computational costs, requiring parallelization optimization
2. Missing Completely at Random Assumption: Current method only handles MCAR; non-random missing patterns in clinical data require further investigation
3. Sample Size Constraints: Some relationships may lack stability in small sample scenarios
4. Prior Knowledge Integration: Better integration of clinical prior knowledge to guide model inference is needed

Future Directions

1. MNAR Handling: Integrate missing data graph methods to handle non-random missing patterns
2. Non-Homogeneous DBNs: Extend to globally coupled non-homogeneous DBNs to capture non-stationary relationships
3. Mixed Variables: Handle mixed continuous and discrete variable types
4. Real-Time Applications: Develop real-time clinical decision support systems

In-Depth Evaluation

Strengths

1. Theoretical Rigor: Complete derivation of closed-form solutions for missing value FCDs with solid theoretical foundation
2. Methodological Innovation: First application of fully Bayesian methods to missing data learning in DBNs
3. Experimental Sufficiency: Includes simulated and real data validation, covering different missing rates and sample sizes
4. Clinical Relevance: Discovered relationships have clinical significance, validating practical utility
5. Reproducibility: Provides complete algorithm description and open-source code

Weaknesses

1. Computational Efficiency: Lacks detailed computational time analysis and optimization strategies
2. Frequentist Comparison: Missing comparisons with classical frequentist DBN learning methods
3. Parameter Sensitivity: Insufficient sensitivity analysis regarding hyperparameter selection
4. Scalability: Performance on larger-scale networks remains unknown

Impact

1. Academic Contribution: Provides new theoretical framework for missing data handling in DBNs
2. Practical Value: Has significant application prospects in critical domains such as healthcare
3. Method Generalizability: Extensible to other fields requiring temporal sequence missing data handling

Applicable Scenarios

1. Healthcare: ICU monitoring, chronic disease management, clinical trial analysis
2. Finance: Time series risk modeling, market prediction
3. Industry: Equipment health monitoring, quality control
4. Environment: Climate modeling, pollution monitoring

References

The paper cites 42 relevant references, covering important works in Bayesian network learning, missing data handling, medical informatics, and other domains, providing a solid theoretical foundation for the research.

Overall Assessment: This is a high-quality paper with significant methodological innovation, demonstrating both theoretical breakthroughs and practical value. While there is room for improvement in computational efficiency and method comparison, its contributions are sufficient to advance the field.