Deep Semisupervised Multitask Learning Model and Its Interpretability for Survival Analysis.

Area of Science:

• Medical Statistics
• Bioinformatics
• Machine Learning in Medicine

Background:

• Survival analysis is crucial in medicine for treatment planning and drug development.
• Existing methods often rely on assumptions violated by real-world clinical data.
• There is a need for robust survival analysis techniques that accommodate complex data structures.

Purpose of the Study:

• To propose a novel semisupervised multitask learning (SSMTL) deep learning method for survival analysis.
• To address limitations of traditional methods by avoiding strong assumptions on stochastic processes.
• To provide an effective deep learning approach for survival analysis with complex clinical data, including competing risks.

Main Methods:

• Developed a semisupervised multitask learning (SSMTL) framework for survival analysis.
• Transformed survival analysis into a multitask problem incorporating semisupervised learning and multipoint survival probability prediction.
• Utilized semisupervised loss and ranking loss to handle censored data and non-increasing survival probability trends, modeling outcomes directly without assumptions.

Main Results:

• The SSMTL method demonstrated superior prediction performance compared to previous models in survival analysis, with and without competing risks.
• Effectively modeled the distribution of survival times and covariate-outcome relationships directly.
• Successfully identified the importance of prognostic factors and visualized their time-dependent and nonlinear effects on survival outcomes.

Conclusions:

• The proposed SSMTL deep learning method offers a significant advancement for survival analysis with complex clinical data.
• It provides an effective, assumption-free approach that improves prediction accuracy and interpretability.
• This work highlights the potential of deep learning for structured medical data analysis and clinical decision-making.