Variational Disentanglement for Rare Event Modeling.

Area of Science:

• Machine Learning
• Healthcare Informatics
• Biostatistics

Background:

• Healthcare data and machine learning advances offer opportunities for better clinical decision support.
• Healthcare risk prediction often faces challenges with imbalanced datasets where target conditions are rare.
• Imbalanced classification is a common problem in healthcare and other fields.

Purpose of the Study:

• To propose a novel variational disentanglement approach for semi-parametric learning from rare events in imbalanced classification.
• To develop a robust prediction model by integrating generalized additive models and isotonic neural networks.
• To address the challenge of learning effectively from low-prevalence events in large datasets.

Main Methods:

• A variational disentanglement approach is employed to learn from rare events.
• An extreme-distribution behavior is leveraged in a latent space to extract information from low-prevalence events.
• A robust prediction arm combines generalized additive models and isotonic neural networks.

Main Results:

• The proposed approach demonstrates superior performance on synthetic datasets.
• The method shows effectiveness on diverse real-world datasets, including COVID-19 mortality prediction.
• Outperforms existing alternative methods in imbalanced classification tasks.

Conclusions:

• The variational disentanglement approach is effective for learning from rare events in imbalanced classification.
• The developed method offers improved risk prediction capabilities in healthcare settings.
• This technique holds promise for enhancing clinical decision support systems dealing with rare conditions.