CliTsSurv: Asymmetry-Aware Survival Prediction from Irregular Longitudinal Clinical Data

Survival prediction from longitudinal clinical records is challenged by heterogeneous feature availability, irregular observation intervals, and imbalanced event distributions with heavy censoring. These interacting feature, temporal, and event asymmetries are only partly addressed by approaches that reduce histories to fixed covariates, place observations on regular grids, or optimize a single survival objective. We propose CliTsSurv, an integrated framework that learns separate representations of longitudinal and static inputs, encodes actual elapsed time at multiple scales, and combines latent-space augmentation with a multi-task objective comprising negative log-likelihood, ranking loss, and Brier loss. We evaluated CliTsSurv on a retrospective breast cancer cohort of 7509 patients with 92,820 longitudinal records and up to 171 months of follow-up. Across 20 repeated experiments using random patient-level splits and independent model initializations, CliTsSurv-Transformer achieved an IPCW C-index of 0.970, a Harrell C-index of 0.965, and an integrated Brier score of 0.032, outperforming the strongest baseline under the common evaluation protocol. Ablation and sensitivity analyses examined the contributions and limitations of the representation learning, temporal encoding, and augmentation components and identified elapsed-time encoding as the largest contributor. These findings support coordinated modeling of feature, temporal, and event asymmetries for survival prediction from irregular clinical records.