PGCASurv: A Prior-Guided Cross-Attention Framework for Dynamic Survival Model with Longitudinal Data.

Area of Science:

• Biostatistics
• Machine Learning in Healthcare
• Clinical Informatics

Background:

• Traditional survival analysis often provides static, baseline risk scores.
• Static scores limit tracking of patient condition changes during long-term follow-up.
• Dynamic risk assessment is crucial for personalized medicine.

Purpose of the Study:

• To develop a dynamic survival analysis method for evolving patient risk.
• To improve patient risk characterization during long-term follow-up.
• To enable earlier identification of high-risk patients and personalized treatment plans.

Main Methods:

• Utilized a deep neural network to learn risk modifications over time.
• Incorporated a cross-attention mechanism for updating longitudinal data.
• Trained and evaluated the model on five diverse clinical datasets.

Main Results:

• The proposed dynamic method demonstrated superior accuracy and reliability in risk discrimination.
• Achieved statistically competitive performance in predicting survival probabilities.
• Identified key risk factors consistent with existing medical evidence through interpretability analysis.

Conclusions:

• Dynamic risk assessment, combining population priors and individual adjustments, shows significant potential.
• The method aids clinicians in earlier high-risk patient identification.
• Facilitates the formulation of more individualized follow-up and treatment strategies.