A deep state-space analysis framework for cancer patient latent state estimation and classification from EHR...

Area of Science:

• Artificial Intelligence in Medicine
• Computational Biology
• Health Informatics

Background:

• Deep learning excels at short-term disease prediction but struggles with explainable long-term disease progression analysis.
• Estimating gradual disease changes from patient data remains a significant challenge in medical AI.
• Existing methods lack the ability to visualize and interpret temporal shifts in latent patient health states.

Purpose of the Study:

• To introduce a novel "deep state-space analysis framework" for modeling and visualizing long-term disease progression.
• To enable the estimation and interpretation of temporal changes in latent patient states using electronic health records (EHRs).
• To identify key factors associated with poor prognosis and medication effectiveness in chronic diseases and cancer.

Main Methods:

• Utilized a deep state-space analysis framework to process sequentially obtained electronic health records (EHRs).
• Estimated and visualized temporal changes in latent patient states related to disease progression.
• Employed clustering of latent states to categorize disease severity and identified prognostic factors.

Main Results:

• The framework successfully captured clinical status and continuous temporal changes in 12,695 cancer patients.
• Anemia was identified as a significant poor prognostic factor during state transitions in cancer patients.
• Immune cell abnormalities were confirmed as poor prognostic factors for patients treated with specific cancer medications (Nivolumab, Osimertinib, Afatinib).

Conclusions:

• The deep state-space analysis framework provides a novel method for understanding and visualizing long-term disease progression.
• This approach enhances prognostic evaluation and supports personalized, long-term treatment strategy development.
• The findings demonstrate the potential of explainable AI in healthcare for improving patient outcomes and clinical decision-making.