Evaluating sociodemographic bias in a deployed machine-learned patient deterioration model

Bias evaluations of machine learning (ML) models often focus on performance in research settings, with limited assessment of downstream bias following clinical deployment. The objective of this study was to evaluate whether CHARTwatch, a real-time ML early warning system for inpatient deterioration, demonstrated algorithmic bias in model performance, or produced disparities in care processes, and outcomes across patient sociodemographic groups.

We evaluated CHARTwatch implementation on the internal medicine service at a large academic hospital. Patient outcomes during the intervention period (November 1, 2020-June 1, 2022) were compared to the control period (November 1, 2016-December 31, 2019) using propensity score overlap weighting. We evaluated differences across key sociodemographic subgroups, including age, sex, homelessness, and neighborhood-level socioeconomic and racialized composition. Outcomes included model performance (sensitivity and specificity), processes of care, and patient outcomes (non-palliative in-hospital death).

Among 12 877 patients (9079 control, 3798 intervention), 13.3% were experiencing homelessness and 36.9% lived in the quintile with the highest neighborhood racialized and newcomer populations. Model sensitivity was 70.1% overall, with no significant variation across subgroups. Model specificity varied by age, <60 years: 93% (95% Confidence Interval [CI] 91-95%), 60-80 years: 90% (95%CI 87-92%), and >80 years: 84% (95%CI 79-88%), P < .001, but not other subgroups. CHARTwatch implementation was associated with an increase in code status documentation among patients experiencing homelessness, without significant differences in other care processes or outcomes.

CHARTwatch model performance and impact were generally consistent across measured sociodemographic subgroups. ML-based clinical decision support tools, and associated standardization of care, may reduce existing inequities, as was observed for code status orders among patients experiencing homelessness. This evaluation provides a framework for future bias assessments of deployed ML-CDS tools.