Open-source Python library + no-code web dashboard for evaluating oncology AI models at clinical decision thresholds. [P]

Most classification metrics for oncology AI models (AUC, ICC, MAE) measure global agreement. They don't answer the question that actually matters at the point of care: how reliable is this model at the exact cutoff that decides whether a patient gets flagged, biopsied, or treated?

I built oncothresh to evaluate models at a specific clinical threshold rather than in aggregate: sensitivity/specificity/PPV/NPV at the cutoff, bootstrap confidence intervals, threshold-sensitivity curves, boundary-weighted calibration, decision-curve net benefit, and number-needed-to-test. It's a small, dependency-light Python library (numpy/scipy/scikit-learn/pydantic) built for tasks like tumor cellularity, Ki-67, TMB, and PD-L1 scoring, where a continuous model output gets collapsed into a yes/no clinical decision at a fixed cutoff.

Pathology-specific benchmarks like PathBench and PathBench-MIL evaluate foundation models globally but don't evaluate at predefined clinical thresholds with uncertainty quantification, which is the gap this fills.

There's also a companion web dashboard (oncothresh-web) for people who want the same analysis without writing code: upload a CSV of predictions and labels, pick a threshold, get the full set of charts plus a downloadable PDF report. docker compose up and it's running locally, no cloud dependency.

- Library: github.com/omkaradhali/oncothresh

- Dashboard: github.com/omkaradhali/oncothresh-web

Still v0.1, so I'd genuinely welcome feedback: use cases I haven't considered, edge cases in the DCA/calibration math, or places the API doesn't fit how people actually work with threshold-based models.