Oncology AI Threshold Evaluation Tool Released

Existing metrics such as AUC, ICC, and MAE measure overall model performance but fail to reflect reliability at the specific cutoff that determines patient treatment in real clinical settings.

oncothresh is a Python library developed to fill this gap, precisely evaluating performance at specific clinical thresholds.

Key Features and Characteristics:

• Threshold-based metrics: Calculates Sensitivity, Specificity, PPV, NPV, and more at specific thresholds.
• Statistical analysis: Supports bootstrap confidence intervals, threshold-sensitivity curves, boundary-weighted calibration, decision curve analysis (DCA) net benefit, and more.
• Application areas: Optimized for pathology AI tasks requiring the conversion of continuous outputs to binary decisions, such as Tumor cellularity, Ki-67, TMB, and PD-L1 scoring.

Additionally, the oncothresh-web dashboard is provided, allowing users to upload CSV files for analysis without coding. It can be run locally via Docker Compose, and analysis results can be downloaded as PDF reports.