Research map / Probabilistic and scientific modeling / Treatment effects

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

NeurIPS 2025 · Probabilistic and scientific modeling › Treatment effects · approach: Latent-variable causal inference · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Applies a path-specific causal framework (V-specific direct effect) with a doubly robust and self-normalized estimator to measure racial bias mediated by pulse-oximetry discrepancies on invasive ventilation in eICU and MIMIC-IV.

Problem

Existing work shows statistical associations between pulse oximeter bias and ICU outcomes but does not isolate causal pathways of racial disparity mediated only through oximeter discrepancies, and path-specific estimators lack finite-sample guarantees.

Approach

Defines the V-specific Direct Effect (VDE) as a nested-counterfactual path-specific effect with multiple mediators, derives a doubly robust estimator with a self-normalized importance-weighting variant, and provides finite-sample error bounds and asymptotic normality.

Key results

• On eICU, VDE for invasive ventilation rate is -0.21 percentage points (95% CI -0.23 to -0.19); on MIMIC-IV it is -0.46 percentage points (95% CI -0.56 to -0.36), indicating minimal mediated disparity.
• VDE for ventilation duration is 2.3 hours in eICU (95% CI 2.2 to 2.5) and 7.4 hours in MIMIC-IV (95% CI 6.9 to 7.9) for Black vs White patients.
• The self-normalized estimator shows significantly lower variance than the canonical doubly robust estimator in synthetic and semi-synthetic experiments with small samples.

Measured effects reported in the paper

Surprising findings

• Ignoring other post-admission mediators (NIE*) can flip the direction of the estimated disparity, in contrast to prior work relying on single-mediator analyses.
• Racial discrepancies in oximetry have minimal impact on invasive ventilation rates, contrary to prior work, though they are more pronounced for ventilation duration.

Limitations

• Identifiability relies on a specified causal graph; latent common causes between W and V would make the VDE unidentifiable.
• Sparse SaO2 measurements limit matched discrepancy samples and effective sample size.
• Only the first invasive ventilation event is considered and temporality is not explicitly modeled.

Experimental setup

Datasets: eICU Collaborative Research Database MIMIC-IV

Benchmarks: synthetic linear SCM semi-synthetic eICU eICU real-world MIMIC-IV real-world

Metrics: relative error mean and 95% confidence interval over bootstraps VDE, NDE, NIE, TE effect estimates (percentage points, hours

Baselines: non-self-normalized doubly robust estimator NIE* (mediation ignoring other mediators)

Builds on: doubly robust estimation for path-specific effects natural direct and indirect effects structural causal models

Closest papers by meaning

• Two-Stage Shadow Inclusion Estimation: An IV Approach for Causal Inference under Latent Confounding and Collider Bias (ICML 2024 · similarity 0.80)
• A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference (NeurIPS 2025 · similarity 0.80)
• A Unified Framework for the Transportability of Population-Level Causal Measures (NeurIPS 2025 · similarity 0.79)
• Heterogeneous Treatment Effect in Time-to-Event Outcomes: Harnessing Censored Data with Recursively Imputed Trees (ICML 2025 · similarity 0.79)
• Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML (ICLR 2024 · similarity 0.78)
• Distribution-Free Uncertainty Quantification in Mechanical Ventilation Treatment: A Conformal Deep Q-Learning Framework (AAAI 2025 · similarity 0.78)