De-Biased Disentanglement Learning for Pulmonary Embolism Survival Prediction on Multimodal Data

Area of Science:

• Healthcare AI
• Medical Informatics
• Health Disparities Research

Background:

• Health disparities disproportionately affect marginalized populations, impacting healthcare equity.
• Artificial intelligence (AI) in healthcare can exacerbate or alleviate these disparities based on model bias.
• Existing AI algorithms show bias in medical implementation, affecting conditions like pulmonary embolism (PE) prognosis.

Purpose of the Study:

• To address bias challenges in healthcare AI algorithms stemming from population disparities.
• To propose a holistic framework for reducing bias through complementary aggregation.
• To mitigate inequities in medical implementation, particularly for PE prognosis.

Main Methods:

• Exploration of diverse biases within healthcare systems.
• Development of de-biasing deep survival prediction models.
• Leveraging a framework to disentangle identifiable information from multimodal data (images, text, clinical variables).

Main Results:

• A proposed framework to mitigate potential biases within multimodal datasets.
• Demonstration of richer survival-related characteristics compared to traditional methods.
• Achieved bias-complementary predicted results, enhancing survival analysis robustness.

Conclusions:

• The developed framework improves fairness and accuracy in healthcare AI systems.
• Benefits patients, clinicians, and researchers through more robust survival analysis.
• Offers a pathway to more equitable healthcare delivery by addressing AI bias.