Derived Data, Phikon Embeddings, and Analysis Code for "Calibration-Aware Patient-Level Prediction of Source-Defined MSIMUT Status in Colorectal Cancer from Whole-Slide Histopathology Images Using Foundation-Model Embeddings"

This repository contains the non-identifying derived numerical data, patient-specific Phikon patch-embedding arrays, pooled patient-level representations, analysis outputs, reproducibility materials, and computational notebooks associated with the study “Calibration-Aware Patient-Level Prediction of Source-Defined MSIMUT Status in Colorectal Cancer from Whole-Slide Histopathology Images Using Foundation-Model Embeddings.” The analyzed cohort comprised 360 patients, including 65 source-defined MSIMUT-category cases and 295 MSS cases. The repository includes the complete patient-specific Phikon patch-embedding arrays for the included cases, the final 1,536-dimensional mean-plus-max patient-level representations, class labels, fixed cross-validation split definitions, documented random seeds and analysis configurations, out-of-sample model predictions, bootstrap outputs, leakage-safe nested probability calibration results, threshold and decision-curve analyses, pooling comparisons, patch-subsampling robustness results, clinicopathological and site-related sensitivity analyses, representation-level SHAP outputs, revision-stage reanalysis outputs, and computational notebooks used for analysis, auditing, and figure generation. The histopathology image archive directly analyzed in the study was the publicly available “TCGA COAD MSI vs MSS Prediction (JPG)” colorectal dataset hosted on Kaggle. This archive is derived from the formalin-fixed, paraffin-embedded histopathology dataset published by Kather et al. on Zenodo (DOI: 10.5281/zenodo.2530835), which in turn was generated from diagnostic whole-slide images originating from The Cancer Genome Atlas (TCGA). The original histopathology image files are not redistributed in this repository. Only non-identifying derived numerical representations, patient-specific patch-embedding arrays, analytical outputs, documentation, reproducibility-audit materials, and computational code generated in the present study are deposited here. The original Kaggle, Kather et al./Zenodo, and TCGA/GDC source materials remain subject to their respective original licenses, data-access policies, and terms of use; the license assigned to this Zenodo record applies only to the original derived materials deposited by the present authors. Patch-level feature extraction was performed using the original Phikon pathology foundation model through the repository identifier owkin/phikon. A commit-specific Phikon model revision was not explicitly pinned during the original embedding-generation run and could not be verified retrospectively; this limitation is documented transparently in the accompanying reproducibility-audit materials. No new participants were recruited, no biological specimens were collected, and no directly identifiable patient information was accessed or generated in this study. The analysis represents a secondary computational study of publicly available, de-identified/coded histopathology data. Full data provenance, file descriptions, analysis structure, validation summaries, and reproducibility information are provided in the accompanying README, validation summary, download manifest, and reproducibility-audit documentation.