Path-specific effects for pulse-oximetry guided decisions in critical care

TL;DR

This work tackles race-based biases in pulse oximetry and their influence on ICU decision-making by introducing a path-specific causal framework that isolates the effect of race mediated through oximetry discrepancies on invasive ventilation. It develops a doubly robust VDE estimator and a self-normalized variant with finite-sample guarantees, and validates them on synthetic, semi-synthetic, and real ICU data from MIMIC-IV and eICU. The real-data analysis finds minimal impact of oximetry bias on ventilation initiation but notable effects on ventilation duration, with variation across datasets, and demonstrates the importance of accounting for multiple mediators beyond SpO$_2$-SaO$_2$ discrepancy. Collectively, the methodology provides a practical pipeline for assessing fairness in critical-care decisions and underscores the necessity of causal methods to robustly quantify disparities in healthcare.

Abstract

Identifying and measuring biases associated with sensitive attributes is a crucial consideration in healthcare to prevent treatment disparities. One prominent issue is inaccurate pulse oximeter readings, which tend to overestimate oxygen saturation for dark-skinned patients and misrepresent supplemental oxygen needs. Most existing research has revealed statistical disparities linking device measurement errors to patient outcomes in intensive care units (ICUs) without causal formalization. This study causally investigates how racial discrepancies in oximetry measurements affect invasive ventilation in ICU settings. We employ a causal inference-based approach using path-specific effects to isolate the impact of bias by race on clinical decision-making. To estimate these effects, we leverage a doubly robust estimator, propose its self-normalized variant for improved sample efficiency, and provide novel finite-sample guarantees. Our methodology is validated on semi-synthetic data and applied to two large real-world health datasets: MIMIC-IV and eICU. Contrary to prior work, our analysis reveals minimal impact of racial discrepancies on invasive ventilation rates. However, path-specific effects mediated by oxygen saturation disparity are more pronounced on ventilation duration, and the severity differs by dataset. Our work provides a novel pipeline for investigating potential disparities in clinical decision-making and, more importantly, highlights the necessity of causal methods to robustly assess fairness in healthcare.

Figures (14)

• Figure 1: Distribution of $\text{SpO}_2$ and $\text{SaO}_2$ in eICU data. Samples with hidden hypoxemia are colored red, following Matos2024BOLDBA.
• Figure 1: Causal diagrams for the modified standard fairness model with two mediators $W$ and $V$. For any other associations between the mediators, the VDE is not identifiable.
• Figure 2: Convergence of the relative error of causal queries using the canonical (left column) and self-normalized (right column) doubly robust estimator on a continuous synthetic setting (top row) and semi-synthetic eICU data (bottom row). The plots show the mean and 95% confidence interval over 100 bootstraps iterations. The estimates using self-normalization exhibit significantly lower variance.
• Figure 3: Distribution of samples in the eICU and MIMIC-IV datasets, showing (a) number of patients by race and ventilation status and (b) invasive ventilation duration, stratified by race. The baseline duration of ventilation is higher for eICU patients, and is associated with greater pre-admission severity (see Appendix \ref{['app:data-viz']}).
• Figure 4: Average causal fairness measures across 500 bootstraps using the self-normalized estimator for the (a) rate and (b) duration of invasive ventilation on eICU and MIMIC-IV data. Colors represent different effects and the numerical label for each bar indicates the mean across all bootstraps. Error bars show $95\%$ confidence intervals. Positive values indicate a higher rate or longer duration of invasive ventilation for Black patients relative to White patients.

Theorems & Definitions (9)

• Proposition 1: Identifiability Miles2017OnSE
• Definition 1: Nuisance Parameters Miles2017OnSE
• Lemma 1: Parametrization
• Lemma 2: Double Robustness Property
• Definition 2: Doubly Robust VDE Estimator
• Theorem 2: Finite Sample Analysis
• Corollary \ref{thm:distributional-finite-sample-error}: Asymptotic error
• Lemma 3: Helper Lemma
• proof