Exploiting Domain Knowledge as Causal Independencies in Modeling Gestational Diabetes.

Area of Science:

• Medical Informatics
• Biostatistics
• Machine Learning in Healthcare

Background:

• Gestational diabetes mellitus (GDM) poses significant risks to maternal and infant health.
• Accurate and interpretable models are crucial for understanding GDM risk factors in clinical settings.
• Existing models may lack transparency, hindering clinical adoption and feature importance analysis.

Purpose of the Study:

• To develop a domain expert-guided, interpretable, and explainable probabilistic model for gestational diabetes.
• To leverage causal independence principles for robust GDM modeling.
• To validate the model's efficacy and identify critical predictive features in a clinical study.

Main Methods:

• Constructed a probabilistic model utilizing the causal independence (Noisy-Or) framework.
• Selected features guided by domain expertise for GDM prediction.
• Validated the model's performance and feature importance on a clinical study dataset.

Main Results:

• The developed probabilistic model demonstrated efficacy in modeling gestational diabetes within the clinical study.
• The model successfully highlighted the importance of specific features in predicting GDM.
• The causal independence approach proved effective for creating an interpretable GDM model.

Conclusions:

• Domain expert-guided probabilistic models based on causal independence offer an interpretable and explainable approach to gestational diabetes.
• The identified features are crucial for understanding and potentially mitigating GDM risk.
• This methodology enhances the clinical utility of predictive models by providing clear insights into feature contributions.