SEANN: A domain-informed neural network for epidemiological insights

Area of Science:

• Epidemiology
• Machine Learning
• Artificial Intelligence

Background:

• Traditional statistical models are common in epidemiology for analyzing predictor-outcome associations.
• Non-parametric machine learning, like deep neural networks (DNNs), offers advanced capabilities but is data-limited.
• Explainable AI (XAI) tools can enhance the interpretability of these complex models.

Purpose of the Study:

• To address data limitations in epidemiological machine learning.
• To develop a novel approach for informed deep neural networks (DNNs).
• To integrate domain-specific knowledge, specifically Pooled Effect Sizes (PES), into DNN training.

Main Methods:

• Introduction of SEANN, a novel approach for informed DNNs.
• Leveraging Pooled Effect Sizes (PES) from meta-analysis studies as domain knowledge.
• Integrating PES into the training loss function of DNNs.

Main Results:

• Demonstrated significant improvements in predictive generalization via controlled simulations.
• Showcased enhanced epidemiological plausibility of learned relationships.
• Compared SEANN against a domain-knowledge agnostic neural network, highlighting SEANN's superiority.

Conclusions:

• SEANN effectively integrates domain knowledge (PES) into DNNs for epidemiological research.
• The approach overcomes data scarcity challenges, improving model performance.
• SEANN offers a promising direction for advancing AI applications in epidemiology.