# Out-of-distribution detection (Mahalanobis guard)

Source: https://onco.cc/terms/ood-detection/  
OnCo record `ood-detection` (Term). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

Out-of-distribution detection flags an input that does not look like anything the model was trained on, so the model can refuse to predict instead of guessing; the Mahalanobis distance from the training cloud is the simplest such guard.

## Summary

Anomaly detection identifies rare items that deviate significantly from the majority of the data (Wikipedia). The Mahalanobis distance measures how far a point lies from a distribution, accounting for its covariance (Wikipedia); computed in PCA space against the training cohort it gives a threshold beyond which a sample (a new platform, a different tissue, a corrupted file) is declared off-manifold and the prediction withheld. Passing the guard is necessary, not sufficient: an in-distribution sample can still yield an unstable prediction.

## Fields

- Kind: Term
- Last checked: 2026-09-24
- Also known as: out-of-distribution; OOD; OOD detection; out-of-distribution detection; OOD guard; Mahalanobis distance; Mahalanobis guard; ManifoldGuard; off-manifold input; refused prediction
- Tags: cansim-terms

## Notes

- Listed in the CanSim terms map 1.0.0 (docs/onco/terms.json, generated 2026-09-24), CC BY 4.0, attribution: CanSim project, an open, public-data-first cancer foundation-model programme; CanSim page path /terms/ood-detection.

## Sources

- Wikipedia: https://en.wikipedia.org/wiki/Anomaly_detection
- Wikipedia: Mahalanobis distance: https://en.wikipedia.org/wiki/Mahalanobis_distance
- Wikipedia: https://en.wikipedia.org/wiki/Anomaly_detection

## Connected records

- terms: [Bootstrap resampling](https://onco.cc/terms/bootstrap/), [Cancer AI vocabulary (CanSim terms map)](https://onco.cc/terms/cancer-ai-vocabulary/), [Conformal prediction](https://onco.cc/terms/conformal-prediction/), [Domain shift and domain adaptation (cell line to patient)](https://onco.cc/terms/domain-adaptation/), [Principal component analysis (PCA) as a feature compressor](https://onco.cc/terms/pca/), [Uncertainty quantification and confidence gates](https://onco.cc/terms/uncertainty-quantification/)

---
JSON: https://onco.cc/api/v1/entities/ood-detection.json