AIBRIEF
MS MLB (Multiple Sclerosis Machine Learning Benchmark) runs every new blood RNA model through the same standardized pipeline that blocks training data from leaking into the test set and keeps an untouched evaluation set aside, and its authors flag
MS-MLB, a public benchmark released on arXiv, gives researchers a single, leakage-controlled pipeline for testing AI models that classify multiple sclerosis from whole-blood RNA. The framework is built on the public GSE17048 dataset (GSE17048), frames the task as multiple sclerosis versus healthy controls, and ships with a documented external-submission pathway (GitHub).
Every submitted model runs nested cross-validation, an untouched stratified holdout, bootstrap confidence intervals, ROC and precision-recall analysis, and Brier calibration, then gets summarized by a composite the authors call the MS Research Score. In the authors' own run, Gradient Boosting ranked first with a score of 93.83, AUC-ROC 0.989, sensitivity 0.950, specificity 0.778, F1 0.927, and Brier 0.050 (arXiv:2608.05196).
Specificity at 0.778 trails sensitivity at 0.950 by roughly seventeen points, a gap the authors surface in the paper and that any future submission will have to beat on the same ruler.
The authors state the MS Research Score is for research comparison only and has not been clinically validated, and that a blood-RNA classifier cannot replace clinical diagnosis (README). MS-MLB is the first open benchmark of its kind on GSE17048 with an external submission harness (external_models.py). The next milestone is whether outside groups start submitting.