Source Code for SIMMT: A Simultaneity-Aware Multi-Task Attention-Based Survival Model

This repository contains the official implementation of SIMMT (Simultaneity-Aware Multi-Task Attention-Based Survival Model), developed to model two temporally dependent clinical events within a discrete-time survival framework. The code supports joint prediction of implant failure and conditional reimplantation while explicitly accounting for event dependency, same-day (zero-gap) occurrences, right-censoring, and temporal coherence. The implementation includes a shared feature encoder, task-specific attention branches, event-masking logic for conditional prediction, a simultaneity alignment loss for same-day events, and monotonicity regularization applied to the derived cumulative incidence. Training and evaluation procedures follow the methodology described in the accompanying manuscript, including discrete-time likelihood optimization and censoring-adjusted performance metrics (C-index, IPCW-AUC, IPCW-BS). The repository provides scripts for data preprocessing, model training, cross-validated evaluation, and result reproduction. All experiments were implemented in Python using TensorFlow/Keras, with additional dependencies for survival analysis and metric computation. This code is released for research and academic use to support transparency, reproducibility, and methodological extension. No patient-level data are included; users must supply their own appropriately formatted datasets.