Introduction
In intensive care units (ICUs), clinical decisions regarding mechanical ventilation—specifically when to initiate intubation or begin the weaning process—are critical, high-stakes interventions. While machine learning (ML) models offer significant potential to assist clinicians in these decisions, there is a growing concern that these models may inherit and perpetuate societal biases.
Traditional research into algorithmic fairness in healthcare typically focuses on a narrow set of protected attributes, such as age, race, and gender. However, this study argues that Social Determinants of Health (SDOH)—the conditions in which individuals live, work, and age—are equally vital. Factors like economic stability, neighborhood employment rates, and housing security are known drivers of health disparities but are frequently overlooked in algorithmic audits. This paper introduces a new temporal benchmark for ICU respiratory support and reveals how AI models predicting ventilation outcomes are influenced by these broader societal factors.
Methods and Model Architecture
The researchers developed a "joint-fusion" deep learning framework designed to handle the high-frequency, dynamic nature of ICU data. They focused on two primary clinical tasks:
- Prolonged Mechanical Ventilation: Predicting whether a patient will require a ventilator for more than 14 days, based on the first 24 hours of ICU admission.
- Successful Weaning: Predicting if a patient can be successfully removed from a ventilator (defined as no re-intubation or death within 7 days) using a 5-day window of data.
The study evaluated several architectures, including Multilayer Perceptrons (MLP), Temporal Convolutional Networks (TCN), and recurrent models like LSTM, BiLSTM, and GRU. The "Hybrid" versions of these models are particularly noteworthy as they integrate hourly time-series data (vitals, labs, ventilator settings) with static patient data (demographics and comorbidities).
Dataset and SDOH Integration
The study leverages MIMIC-IV v2.2, a comprehensive open-access database of electronic health records. The researchers released a curated benchmark containing hourly records for 50,920 patients, encompassing ventilation parameters, laboratory results, and illness severity scores.
A key innovation of this work is the integration of census tract-level SDOH data. By linking clinical records to patient zip codes, the authors were able to analyze how neighborhood-level socio-economic factors—such as employment rates and proximity to healthcare facilities—correlate with model performance and potential bias.
Fairness Audit and Key Results
The core of the research involved a comprehensive Fairness Audit. The team measured performance gaps (parity, recall, and specificity) across both traditional demographics and eight specific SDOH dimensions.
Model Performance
The GRU (Hybrid) model emerged as the top performer, achieving an Area Under the Receiver Operating Characteristic (AUROC) of 0.778 for predicting prolonged ventilation and 0.776 for successful weaning. These results underscore the importance of using hourly temporal data over simple static snapshots.
Identifying Algorithmic Biases
The audit revealed several significant disparities:
- The Language Barrier: Models showed a performance gap favoring English speakers, suggesting that clinical documentation or treatment patterns for non-English speakers may introduce complexities that the models struggle to capture.
- Elderly Vulnerability: In the weaning task, a significant "recall gap" was observed for elderly patients. The model was less accurate at identifying which elderly patients would successfully wean, representing a dangerous blind spot in clinical decision support.
- SDOH as a Hidden Bias: Even when models appeared balanced across race or gender, they exhibited unfairness based on SDOH. For example, the prolonged ventilation model favored individuals from areas with high employment rates.
Discussion and Future Directions
While this study successfully identifies critical biases, the authors acknowledge several limitations. The SDOH data was linked at the census-tract level rather than the individual level, which serves as a neighborhood proxy but may not reflect every patient's specific circumstances. Furthermore, the current work focuses on the identification of bias; future research must prioritize the development of mitigation strategies and de-biasing algorithms tailored for respiratory care.
Ultimately, this benchmark serves as a foundation for moving toward more equitable AI in the ICU. The goal is to transition from predictive models to real-time Reinforcement Learning (RL) systems that can provide fair, bedside decision support in complex clinical environments.
Summary updated