When the Threshold Is the Bias: Auditing Sex-Blind Cardiac Decision Cut-offs on Real-World Data

Algorithmic fairness research concentrates on model outputs, yet many clinical decisions reduce to a fixed threshold applied to a measurement: a patient is flagged as elevated when $x > c$. When the guideline-defined cut-off is sex-specific ($c_F \neq c_M$) but a single unisex value $c$ is applied in practice, the decision rule itself becomes a source of inequity, independent of any model. High-sensitivity cardiac troponin is a paradigm case. Using routine real-world data from a university hospital ($n = 7{,}866$; $4{,}192$ men, $3{,}674$ women) accessed via FHIR, we confirm a single applied cut-off $c = 45$ ng/L with no sex distinction in the reference range, while women's troponin distribution runs systematically lower than men's (median $5$ vs $7$ ng/L). Reclassifying each patient under assay-consistent sex-specific cut-offs, $$\text{elevated}i = \mathbb{1}!\left[, x_i > c{s_i} ,\right], \qquad c_F = 39.6,; c_M = 58.5\ \text{ng/L},$$ versus the unisex rule $\mathbb{1}[x_i > c]$ reveals a directional disparity: women under-detected by ${\approx},1.4%$ and men over-flagged by ${\approx},2.7%$. Modest but structural, a unisex threshold on sex-different distributions disadvantages women. We frame sex-blind decision thresholds as an under-examined, model-independent axis of fairness, auditable directly from routine data and correctable by sex-specific reference-interval recalibration. ($s_i$ denotes patient $i$'s sex and $c_{s_i}$ the corresponding sex-specific cut-off; $\mathbb{1}[\cdot]$ is the indicator function.)