Bias in vital signs? Machine learning models can learn patients’ race or ethnicity from the values of vital signs alone

TL;DR

Machine learning models can predict a patient's race or ethnicity using only vital signs, even after accounting for known biases, raising concerns about fairness in healthcare.

Contribution

The study demonstrates that racial or ethnic information is embedded in vital signs and can be learned by ML models, even when controlling for known biases.

Findings

ML models achieved an AUC of 0.74 between White and Black patients using only vital signs.

Heart rate, oxygen saturation, and blood pressure showed small but significant differences between racial groups.

The model correctly classified race or ethnicity in two out of three patients, indicating non-random results.

Abstract

To investigate whether machine learning (ML) algorithms can learn racial or ethnic information from the vital signs alone.

A retrospective cohort study of critically ill patients between 2014 and 2015 from the multicentre eICU-CRD critical care database involving 335 intensive care units in 208 US hospitals, containing 200 859 admissions. We extracted 10 763 critical care admissions of patients aged 18 and over, alive during the first 24 hours after admission, with recorded race or ethnicity as well as at least two measurements of heart rate, oxygen saturation, respiratory rate and blood pressure. Pairs of subgroups were matched based on age, gender, admission diagnosis and disease severity. XGBoost, Random Forest and Logistic Regression algorithms were used to predict recorded race or ethnicity based on the values of vital signs.

Models derived from only four vital signs can predict patients’ recorded race or ethnicity with an area under the curve (AUC) of 0.74 (±0.030) between White and Black patients, AUC of 0.74 (±0.030) between Hispanic and Black patients and AUC of 0.67 (±0.072) between Hispanic and White patients, even when controlling for known factors. There were very small, but statistically significant differences between heart rate, oxygen saturation and blood pressure, but not respiration rate and invasively measured oxygen saturation.

ML algorithms can extract racial or ethnicity information from vital signs alone across diverse patient populations, even when controlling for known biases such as pulse oximetry variations and comorbidities. The model correctly classified the race or ethnicity in two out of three patients, indicating that this outcome is not random.

Vital signs embed racial information that can be learnt by ML algorithms, posing a significant risk to equitable clinical decision-making. Mitigating measures might be challenging, considering the fundamental role of vital signs in clinical decision-making.

Figures