Bayesian Network structure learning algorithm for highly missing and non imputable data: Application to breast cancer...

Area of Science:

• Computational Statistics
• Machine Learning in Healthcare
• Data Mining

Background:

• Healthcare data frequently exhibits a high proportion of missing values, posing challenges for analysis.
• Traditional methods like imputation or complete case analysis are often unsuitable or lead to significant data loss in clinical settings.
• Existing structure learning algorithms for Bayesian Networks may struggle with substantial missing data.

Purpose of the Study:

• To develop a novel algorithm for learning Bayesian Network (BN) graphs that can handle datasets with missing data without resorting to imputation or complete case analysis.
• To provide a robust method for extracting insights from complex healthcare datasets where data integrity is compromised.
• To evaluate the performance of the proposed algorithm against existing structure learning methods.

Main Methods:

• The proposed algorithm employs a strategy of local bootstrap learning on complete sub-datasets.
• These locally learned models are then aggregated and optimized to form the final Bayesian Network graph.
• This approach bypasses the need for direct imputation or discarding incomplete records.

Main Results:

• The developed learning method demonstrates competitive performance when compared to other established structure learning algorithms.
• The algorithm's effectiveness is consistent across various missing data mechanisms.
• It successfully learns Bayesian Network structures even when imputation and complete case analysis are not feasible.

Conclusions:

• The proposed algorithm offers a valuable alternative for learning Bayesian Networks from incomplete healthcare data.
• It effectively addresses the limitations of imputation and complete case analysis, preserving data utility.
• This method enhances the applicability of Bayesian Networks in real-world clinical data analysis.