Physician-led. Evidence-minded.

Clinical AI,
tested like
it matters.

CalibRAItion is a physician-led research and evaluation lab exploring how medical AI performs, where it fails, and what it takes to trust it.

Explore our research Meet the team

IDH prediction · Results explorer

Performance, in perspective.

8 models

1,393 patients186 IDH-mutant · Repeated nested 5-fold CV

MetricHigher is better

0.800.850.900.951.00

AUROC · zoomed axis

Point estimate whiskers show 95% CI

Inspect model

Hybrid0.961 AUROC95% CI 0.946 to 0.974

Hybrid has the highest AUROC. Its paired AUROC interval versus Ensemble + age includes zero.

Internal evaluation · Independent validation needed

Explore the evidence

Behind the result.

Select model

Hybrid · 95% CI
Metric	Estimate (95% CI)
AUROC	0.961 (0.946 to 0.974)
Average precision	0.866 (0.830 to 0.902)
Brier score	0.042 (0.035 to 0.049)
ECE	0.016 (0.004 to 0.020)
Sensitivity	0.866 (0.812 to 0.914)
Specificity	0.922 (0.907 to 0.937)
F1 score	0.730 (0.691 to 0.771)

High-sensitivity operating point

90% target selected in training folds.

SENSITIVITY: 0.898 (0.855 to 0.941)
SPECIFICITY: 0.878 (0.859 to 0.897)
PPV: 0.532 (0.495 to 0.575)
F1: 0.668 (0.633 to 0.706)

How to read these results

Paired differences are bootstrap means, so they may differ from subtracting the displayed model scores.

Ensemble members and the original threshold were selected on these out-of-fold results; ensemble-based estimates are optimistic. Leave-one-site-out refits the combination only, using pooled-site ensemble logits. Age carries much of the signal.

These are internal results, not evidence of clinical readiness. Features + age ECE is 0.009 with a supplied interval of 0.000–0.008; the inconsistency is retained for review.

Repeated nested cross-validation in 1,393 patients. Estimates and intervals come from the supplied results; no curve coordinates are inferred.

Research, with the
questions left in.

From medical images to the evidence we rely on, these are the questions shaping our work.

Medical imaging
Confidence-gated radiomics
Combining radiomic machine learning and deep learning for survival prediction in paediatric high-grade glioma.
Scanner robustness
Beyond a change of scanner
Exploring whether the structure of a tumour can offer more stable imaging features.
Multimodal screening
Reading between the data
Bringing information in ultrasound reports into conversation with biochemical data.
Research methods
Examining the evidence pipeline
How reliable are the tools that prepare the evidence for a systematic review?

Read the methods, evidence and limitations together.

All research

A promising result
is a starting point.

Understanding a model means asking what happens beyond a single score. These are the questions we come back to.

Start with the clinical question.
What decision would this model inform? The context gives the evaluation its purpose.
Look closely at the evidence.
Methods, datasets and assumptions shape what a finding can tell us and what it cannot.
Make uncertainty visible.
Where does confidence hold up? Where should a model abstain? Limitations belong alongside results.