Supporting Data and Reproducibility Code for QIMCA-Net: A Leakage-Controlled Multimodal Self-Attention Framework for 90-Day Functional Outcome Prediction After Ischaemic Stroke

Version 2.0 — Reviewer-revision update. This repository contains non-identifying derived numerical data and reproducibility materials supporting the revised manuscript “QIMCA-Net: A Leakage-Controlled Multimodal Self-Attention Framework for 90-Day Functional Outcome Prediction After Ischaemic Stroke.”Version 2.0 extends the original supporting materials with reviewer-requested reproducibility and robustness analyses. The deposited materials include the original supporting-data archive and analysis notebook retained for provenance; the revised reviewer-analysis pipeline; repeated model-training and matched-fusion outputs; out-of-fold probability data; smoothed calibration-curve data; and exploratory decision-curve data.The revision analyses address parameter-matched comparison of residual self-attention with direct concatenation, repeated patient-level cross-validation, stronger structured-data comparators, perturbation and attention analyses, centre-related robustness, calibration, uncertainty analysis, and additional reproducibility checks. These analyses are secondary robustness evaluations and do not replace the frozen primary QIMCA-Net out-of-fold results.An implementation-level audit performed during revision established that the primary QIMCA-Net architecture uses three modality-specific 128-dimensional tokens processed by two residual self-attention blocks, followed by mean pooling and a prediction head. It does not contain directed cross-attention or a learnable gating layer. Historical Version 1 filenames are retained only for provenance and may preserve terminology used before this implementation-level correction.The original ISLES’24 medical images and source clinical data are not redistributed. They remain available through the official ISLES’24 repository under its applicable licence and access conditions: DOI 10.5281/zenodo.16731717.Complete execution of the deposited notebooks requires lawful access to the original ISLES’24 dataset and appropriate computing resources.