IDNetwork: A deep illness-death network based on multi-state event history process for disease prognostication

Area of Science:

• Biostatistics
• Machine Learning in Healthcare
• Survival Analysis

Background:

• Multi-state models, particularly the illness-death model, track disease progression through healthy, intermediate, and death states.
• Current methods use Cox proportional hazard (P.H.) models, which assume linear relationships between covariates and transition risks.
• Clinical data often exhibits complex, non-linear relationships between patient characteristics and disease progression risks.

Purpose of the Study:

• To develop a novel neural network architecture, the illness-death network (IDNetwork), to model disease progression.
• To overcome the limitations of the linear Cox P.H. model in capturing complex covariate effects.
• To improve the accuracy of treatment adaptation based on disease evolution.

Main Methods:

• Proposed a multi-task neural network architecture, IDNetwork, for illness-death processes.
• IDNetwork utilizes fully connected subnetworks to learn transition probabilities, relaxing the linear assumption of Cox P.H. models.
• Evaluated IDNetwork's performance through simulations and on real-world clinical data.

Main Results:

• IDNetwork demonstrated significantly improved predictive performance compared to state-of-the-art methods.
• The model showed added value across simulated datasets, colon cancer clinical trials, and a breast cancer dataset.
• The neural network approach effectively captured non-linear covariate-risk relationships.

Conclusions:

• IDNetwork offers a powerful, flexible alternative to traditional Cox P.H. models for illness-death processes.
• The improved predictive accuracy can lead to more effective, personalized treatment strategies.
• This approach holds promise for advancing precision medicine in oncology and other fields.