(BSPOT07) What we owe to patients: A systematic literature review into reporting of potential biases in clinical decision support tools that use AI to personalise care in viral hepatitis

Background: Artificial intelligence (AI) can be combined with software-based clinical decision support (CDS) tools to assist or guide in common clinical tasks within viral hepatitis care. These tools may use patient-level data to make specific recommendations or diagnoses for patients. However, there have been instances where AI algorithms have been shown to be biased against certain groups. Healthcare providers and researchers should be aware of any potential biases in the resulting recommendations and/or any mis-matches between training/validation sets and real world populations.

Objectives: This is a systematic literature review (SLR) of papers that describe AI-based CDS systems intended to inform viral hepatitis care. The objective of this review was to examine whether a) the source data for AI model training was referenced, and b) whether a proactive assessment of potential bias/unfairness was conducted relating to age, race, and sex.

Methods: This SLR included searches of articles from Scopus, PubMed, clinicaltrials.gov, Ovid MEDLINE and Web of Science libraries. A SLR platform called Dr. Evidence was used to extract and manage articles. Articles had to be in English, be published between 01/01/1995 and 10/01/2024. The search was performed with a combination of terms including “Clinical decision support”, “CDS”, “tool”, “AI”, and “viral hepatitis” to identify relevant articles. To be included, the main purpose of the CDS had to facilitate viral hepatitis care. The final search was conducted on 22 January 2024.

Results: We identified 45 articles as relevant to the stated objectives. After title and abstract screening by two independent reviewers, 25 studies remained, 20 of which were included for final evaluation. Most CDS tools identified fell into three categories i) prediction of infection status to enhance screening or diagnosis (28.6%), ii) staging of disease (23.8%), and iii) prediction of treatment outcome (19.0%). Less than 2/3 of the studies provided information regarding the sex and age distribution of the training data sets on which AI models were built (65% and 60% respectively). Very few models proactively evaluated potential bias in terms of age, sex and gender (only 3 studies reported any identifiable bias assessment).

Conclusions: This review highlights a clear need to better understand the potential for biased or unfair predictions or recommendations according to sex, age, and race. These data also underline the importance of specifying in which situations and populations the described tools can be safely and fairly used to prevent potential bias towards specific groups. These findings should encourage researchers to consider conducting and reporting bias assessments, helping to make AI algorithms fair and equitable for everyone.