Select Rule-Out Features by Specificity at a Sensitivity Target

Source: R/ruleout_features.R

Ranks each feature by the specificity it achieves as a single-marker classifier when its threshold is set to capture at least target_sensitivity of cases. Both directions are tried (high values call positive, or low values call positive) and the better one is kept, so the score is direction-agnostic. With multiple cohorts the per-cohort specificities are combined by cohort_aggregation (the default "min" rewards features that hold up in every cohort) within each direction, and only then is the better direction kept. The direction is therefore shared by every cohort: a feature that is high in cases in one cohort and low in another cannot score well in both, which matters for ratio transforms, where a direction flip cancels rather than adds signal.

select_ruleout_features(
  x_list,
  y_list,
  n_features = 50L,
  target_sensitivity = 0.9,
  cohort_aggregation = c("min", "mean"),
  assay = NULL
)

Arguments

x_list

A matrix-like object, SummarizedExperiment, or list of such objects. Rows represent samples and columns represent features.

y_list

A binary response aligned with x_list. Provide a list when x_list is a list.

n_features

Maximum number of features to return (default 50).

target_sensitivity

Minimum fraction of cases a feature's threshold must capture (default 0.90). Must lie in (0, 1].

cohort_aggregation

How per-cohort specificities are combined into a single ranking score: "min" (default, worst cohort) or "mean".

assay

For SummarizedExperiment inputs, the assay name or index to extract prior to analysis.

Value

Character vector of feature identifiers ordered by decreasing specificity at the sensitivity target. A feature whose score is undefined in any cohort (no non-missing case values there) is dropped.

Details

This differs from AUC-based ranking (e.g. select_discriminative_features()) in that only the high-sensitivity tail of the ROC curve counts: a feature whose cases are mostly well separated but with a minority buried inside the control distribution has a good AUC yet a poor rule-out specificity. It is the univariate, per-feature analogue of the "specificity_at_sensitivity" objective used to score fitted panels during optimisation.

Thresholds are observed case values (order statistics): with n non-missing cases the smallest k = ceiling(n * target_sensitivity) cases that still meet the target are called positive, so the positive side always contains at least target_sensitivity of the cases and never more than needed. Missing values are ignored when locating thresholds and computing specificities. Every cohort must contain at least one case and one control.