Clinical Metrics Calculator
Turn a confusion matrix into the numbers reviewers ask for, at your real prevalence.
Runs in your browser. Nothing you enter is uploaded.
Confusion matrix
9,100 cases, 1.1% positive in this dataset.
In this dataset (1.1% positive) PPV is 9.1%. At 0.10% prevalence it becomes 0.9%. Per 1,000 people tested: 1 true positives against 100 false positives, so most positives would be wrong.
If 1,000 people are tested at 0.10% prevalence
A positive result means a 0.9% chance of really having the condition. A negative result means a 99.99% chance of really not having it.
| Sensitivity (recall)Of everyone who has the condition, the share the test finds. | 90.0%82.6%–94.5% |
| SpecificityOf everyone without it, the share correctly cleared. | 90.0%89.4%–90.6% |
| PPV (precision)Of those who test positive, the share who really have it. | 9.1%7.5%–11.0% |
| NPVOf those who test negative, the share who really don’t have it. | 99.9%99.8%–99.9% |
| AccuracyShare of all predictions that were right. Misleading when classes are imbalanced. | 90.0%89.4%–90.6% |
| Balanced accuracyMean of sensitivity and specificity; fairer on imbalanced data. | 90.0% |
| F1 scoreHarmonic mean of precision and recall. Ignores true negatives. | 0.165 |
| Matthews correlationCorrelation between prediction and truth, −1 to 1. The most robust single number on imbalanced data. | 0.268 |
| Youden’s JSensitivity + specificity − 1. Used to pick a threshold. | 0.800 |
| Positive likelihood ratioHow much a positive result raises the odds. Above 10 is strong. | 9.000 |
| Negative likelihood ratioHow much a negative result lowers the odds. Below 0.1 is strong. | 0.111 |
Go deeper: Which clinical metric matters for which task, and why calibration is not discrimination — AI for Biotech, Ch. 16, Table 16.2 (Tripathy, 2026) Chapter 16: AI in Clinical Research and Real-World Evidence
Why prevalence changes everything
Sensitivity and specificity are properties of the test. PPV and NPV are not: they depend on how common the condition is in the people being tested. A model that looks excellent on a balanced research dataset can produce mostly false alarms in a screening population, which is the single most common misreading of clinical AI results.
Confidence intervals are Wilson score intervals, which behave better than the normal approximation at small counts and at proportions near 0 or 1.