KneeAI Reproducibility Package for Uncertainty-Aware Knee Osteoarthritis Severity Assessment

This repository provides the reproducibility package for the study “Uncertainty-Aware Clinical Decision Support for Knee Osteoarthritis Severity Assessment from Radiographs Using a Hybrid KL 5-to-3 Strategy.” It includes manifests, split-integrity audits, notebooks, figures, tables, statistical outputs, and documentation for an uncertainty-aware knee osteoarthritis (KOA) severity-assessment research prototype. The analyzed data were derived from the public Kaggle Knee Osteoarthritis Dataset with Severity Grading. The final manifest contains 8,260 radiographs from 4,130 corrected identifiers, split at the corrected-identifier/laterality-pair level into 5,282 training, 1,322 validation, and 1,656 independent test images, with no corrected-identifier overlap. This should not be interpreted as definitive patient-level partitioning because curated clinical patient identifiers and metadata were not available in the public mirror. The model uses EfficientNetB3 with internal KL 0–4 supervision and post-hoc KL 5-to-3 aggregation: KL-0/1 = Non-OA, KL-2/3 = Mild–Moderate OA, and KL-4 = Severe OA. The software is a research prototype only, not a clinically validated diagnostic device.