Temporally consistent survival prediction for non-uniform longitudinal data.

Area of Science:

• Biostatistics
• Machine Learning
• Health Informatics

Background:

• Traditional survival prediction models rely on single-time point covariate data.
• Longitudinal datasets often feature patient covariates recorded at non-uniform time intervals.
• Existing dynamic survival prediction algorithms may not fully exploit this temporal irregularity.

Purpose of the Study:

• To develop a novel survival prediction model capable of training on longitudinal datasets with non-uniform time intervals.
• To leverage the temporal structure inherent in multi-time point patient covariate data.
• To provide a more accurate survival prediction method for irregularly sampled data.

Main Methods:

• Proposed Temporally Consistent Multi-Task Logistic Regression (TC-MTLR) algorithm.
• Incorporated distributional reinforcement learning concepts for survival outcome modeling.
• Evaluated TC-MTLR against standard and dynamic survival prediction algorithms on diverse short and long longitudinal datasets.

Main Results:

• TC-MTLR demonstrated top performance in Concordance Index (C-Index) and Mean Average Error (MAE-Uncensored) on short datasets.
• On long datasets, TC-MTLR achieved competitive C-Index performance.
• TC-MTLR outperformed other methods in Pseudo-Observable MAE (MAE-PO) and achieved top performance in MAE-Uncensored and Integrated Brier Score (IBS) on long datasets.

Conclusions:

• TC-MTLR effectively utilizes the non-uniform temporal structure of longitudinal data.
• The proposed method offers a competitive and often superior alternative to existing survival prediction models.
• TC-MTLR enhances the predictive accuracy for events like death or hospital readmission using complex longitudinal patient data.