Single-sample outlier detection and conformal claim-gating (Module D)

Outlier detection and conformal anomaly-scoring primitives introduced in the single-sample scoring bank (Module D). These methods flag out-of-distribution samples before a biomarker claim is reported, providing either distribution-free FPR guarantees (conformal approach) or robust compositional distance metrics (Mahalanobis / isolation forest).

Methods provided:

• fit_compositional_mahalanobis / apply_compositional_mahalanobis: MCD-robust Mahalanobis distance on ILR or rCLR log-ratio coordinates.

• fit_conformal_anomaly / os_conformal_anomaly: conformal p-value via k-NN distance against a held-out calibration partition.

• fit_isolation_forest_logratio / apply_isolation_forest_logratio: pure-R isolation forest on rCLR-transformed inputs.

References

Filzmoser P, Hron K, Reimann C. (2009) Principal component analysis for compositional data with outliers. Environmetrics 20: 621-632.

Rousseeuw PJ, Van Driessen K. (1999) A Fast Algorithm for the Minimum Covariance Determinant Estimator. Technometrics 41(3): 212-223.

Vovk V, Gammerman A, Shafer G. (2005) Algorithmic Learning in a Random World. Springer.

Liu FT, Ting KM, Zhou Z-H. (2008) Isolation Forest. ICDM.