Ethnicity-based bias in clinical severity scores

In the age of digital health, a major focus lies on quantitative analysis large, high-dimensional patient data sets. The goal is to establish clinical decision making that informed by evidence from this big medical data. A particularly promising application domain and source for task are intensive care units (ICUs), which measure characteristics at high frequency hundreds variables, can benefit enormously computational approaches efficiently summarise wealth give an accurate representation patient's state. As early example development, severity scores risk assessment were introduced decades ago, correlate markers illness relevant outcomes. family these used in ICU estimate death measuring acute physiological disturbances (vital signs, biochemical values, diagnoses) pathological processes (admission type), reserves (age comorbidities), location factors (before admission).1Barlow CJ Pilcher D Severity scoring systems outcome prediction.in: Bersten AD Handy JM Oh's manual. eighth edn. Elsevier, China2019: 1173-1175Google Scholar One earliest was Acute Physiology Chronic Health Evaluation (APACHE) I, dating back 1981, based dataset 805 patients using expert consensus only.2Knaus WA Zimmerman JE Wagner DP Draper EA Lawrence DE APACHE-acute physiology chronic health evaluation: physiologically classification system.Crit Care Med. 1981; 9: 591-597Crossref PubMed Scopus (1583) Google Since then, numerous iterations APACHE score other have emerged, building regression techniques larger datasets. Lancet Digital Health, Rahuldeb Sarkar colleagues3Sarkar R Martin C Mattie H et al.Performance unit across different ethnicities USA: retrospective observational study.Lancet Digit Health. 2021; 3: e241-e249Summary Full Text PDF (9) assessed whether include systematic or implicit ethnicity-based bias. Numerous examples show influenced ethnicity-based, attitudes outside conscious awareness, often referred as biases.4Hall WJ Chapman MV Lee KM al.Implicit racial/ethnic bias among professionals its influence outcomes: review.Am J Public 2015; 105: e60-e76Crossref (830) To address question describe status without bias, evaluated most frequently risk-scoring (APACHE IVa, Oxford Illness Score, Sequential Organ Failure Assessment) two public sets (Medical Information Mart Intensive III database electronic Collaborative Research Database, with 43 832 122 919 admissions, respectively). investigated Black, Asian, Hispanic, White. They reported no discrimination hospital mortality prediction individual when stratifying their predictions according ethnic groups (p>0·01); however, they observed statistically significant difference calibration (p<0·0001) Hispanic Black patients. This result means same standardised ratios (SMRs), rendering use such triaging resource allocation problematic. general, has limited should be applied caution because CI too wide scored incorrectly.5Booth FV Short M Shorr AF al.Application population-based system results frequent misclassification.Crit Care. 2005; 1-8Crossref SMR mostly cohort level; eg, evaluating ICUs over time comparing similar hospitals.1Barlow COVID-19 pandemic, instead scores, diagnostic imaging symptoms recommended pragmatic higher sensitivity treated (rather than cohort) level.6Aziz S Arabi YM Alhazzani W al.Managing surge during crisis: rapid guidelines.Intensive 2020; 46: 1303-1325Crossref (159) Scholar, 7Barros LM Pigoga JL Chea al.Pragmatic recommendations identification triage disease low-and middle-income countries.Am Trop Med Hygiene. (published online Jan 6.)https://doi.org/10.4269/ajtmh.20-1064Crossref (5) study adds concerns showing cohorts not subject model. Building findings colleagues,3Sarkar important directions future research open up. First, causality further investigated; specifically, does health-care manifests itself biased? With Simpson's paradox8Simpson EH interpretation interaction contingency tables.J Royal Stat Soc. 1951; 13: 238-241Google mind, explored accounting lurking could strengthen weaken statistical differences. An confounder length stay, ratio closer 1 staying shorter 24 h, due large number deaths discharges occurring first h ICU.1Barlow colleagues note, median lengths stays significantly between groups. Second, adaptation active area machine learning,9Venkataramani Ravishankar Anamandra Towards continuous healthcare.arXiv. 2018; Dec 4.) (preprint).https://arxiv.org/abs/1812.01281Google hope improve learning enhance usability established systematically effects potential confounders distribution, types observed, gender ratio,10Roessler Schmitt Schoffer O Can we trust standardized ratio? formal evaluation axiomatic requirements.arXiv. Sep 8.) (preprint).https://arxiv.org/abs/2009.03650Google trying apply conclude, detecting—as did study—understanding correcting biases will remain challenge research. We declare competing interests. Performance studyThe differences suggest reflect mortality. Full-Text Open Access