A Bayesian derived network of breast pathology co-occurrence

Area of Science:

• Computational pathology
• Machine learning in medicine
• Biostatistics

Background:

• Understanding co-occurrence patterns of breast pathologies is crucial for accurate diagnosis and treatment planning.
• Existing methods for analyzing pathology data can be labor-intensive and may not capture complex interrelationships.
• A data-driven approach is needed to model and predict concurrent breast pathologies.

Purpose of the Study:

• To validate and verify a machine-learning based Bayesian network for modeling breast pathology co-occurrence.
• To assess the robustness and accuracy of the developed network using real-world pathology data.
• To explore the potential applications of the network in clinical and research settings.

Main Methods:

• A Bayesian network was constructed using 1631 breast pathology reports, detailing the presence or absence of 29 common pathologies.
• Network robustness was tested by randomly excluding 25%, 50%, and 75% of the dataset and evaluating structural stability.
• Network validation involved cross-referencing identified co-occurrences with existing breast pathology literature and expert opinions.

Main Results:

• The Bayesian network comprises 25 diagnosis nodes and 40 arcs, representing predicted co-occurrences of breast pathologies.
• The network structure demonstrated significant robustness, with 81% of the original network remaining intact after removing 75% of the data.
• A high degree of validation was achieved, with 95% of identified co-occurrences supported by published literature or expert consensus.

Conclusions:

• The developed Bayesian network is a robust and validated tool for understanding breast pathology co-occurrence.
• The network effectively predicts multiple concurrent pathologies from single observations, offering intuitive pattern exploration.
• This approach has potential for broader application in breast pathology clinical practice and breast cancer research.