SepsisTensor v1: A Harmonized Multi-Cohort Transcriptomic Resource for Mortality Prediction

SepsisTensor is a harmonized cross-platform transcriptomic tensor resource constructed from publicly available Gene Expression Omnibus (GEO) cohorts for reproducible machine learning research in sepsis mortality prediction. This release integrates heterogeneous whole-blood and PBMC transcriptomic cohorts spanning multiple hospitals, geographic regions, and sequencing technologies, including both microarray and RNA-seq platforms. The dataset was generated through a fully reproducible preprocessing and harmonization pipeline involving: - Automated GEO cohort discovery- Metadata harvesting and mortality-label harmonization- Cohort-specific sepsis patient filtering- Probe-to-gene and Ensembl-to-HUGO translation- Cross-platform gene-space standardization- Within-cohort Z-score normalization- Empirical Bayes ComBat batch correction with protected biological covariates- Statistical validation of batch-effect neutralization and biological signal preservation The resulting tensors are optimized for downstream machine learning, survival modeling, explainable AI (XAI), and translational bioinformatics workflows. Importantly, this repository does not redistribute raw GEO payloads (e.g., FASTQ files, CEL files, SOFT archives, or platform annotation dumps). Instead, it provides curated derived transcriptomic matrices, harmonized metadata, reproducible preprocessing scripts, and ML-ready integrated tensors generated from publicly accessible GEO resources. Property Value Integrated Cohorts 7 Total Patients 1636 Non-Survivors 367 Survivors 1269 Common intersecting genes 7964 Platforms Microarray + RNA-seq Endpoint Mortality Batch correction Empirical Bayes ComBat