Survival prediction models: an introduction to discrete-time modeling

Area of Science:

• Biostatistics
• Machine Learning
• Medical Informatics

Background:

• Time-to-event prediction models are crucial for clinical decisions.
• Censoring in data requires specialized modeling techniques.
• Discrete-time survival models offer an overlooked yet effective approach.

Purpose of the Study:

• To explore discrete-time survival models as an alternative to continuous-time models.
• To compare the predictive performance of discrete-time models with existing methods.
• To provide a guide and R code for implementing discrete-time survival prediction.

Main Methods:

• Applied discrete-time survival models to person-period data.
• Utilized binary classification algorithms within the discrete-time framework.
• Compared model predictions using time-dependent metrics accounting for censoring.

Main Results:

• Some discrete-time models outperformed the continuous-time Cox proportional hazards model.
• Random survival forests showed improved performance but were sometimes surpassed by discrete-time methods.
• The relative performance of classification methods varied across datasets.

Conclusions:

• Discrete-time methods provide a valuable approach for survival prediction, especially with machine learning integration.
• These models can handle continuous time-to-event data and multiple predictors.
• The study provides resources for wider adoption in medical research.