A blinded comparison of patient treatments to therapeutic options presented by an artificial intelligence-based clinical decision-support system

Abstract

Background: Comparison of options from clinical decision-support (CDS) systems and decisions made in practice may be biased towards the treating institution. In this retrospective study, bias was minimized by blinding evaluators to the source of treatment recommendations, either Watson for Oncology® (WFO® ) or treatments patients received at Bumrungrad International Hospital (BIH), a user of WFO® . Methods: Treatments given were compared to therapeutic options provided by WFO® . Treatments that were identical to WFO® “recommended” (green, acceptable) were not evaluated further. Paired treatments were evaluated independently in a blinded fashion by each oncologist before consensus ranking of each pair as either acceptable, acceptable alternatives, or unacceptable treatment. The consensus for each treatment was compared to WFO® , with WFO® “for consideration” (yellow, acceptable alternative), and “not recommended” (red, unacceptable). Chi-squared tests analyzed the association between risk factors and discordant recommendations. Results: Of 228 treatments given to patients with lung, colon, breast and rectal cancers, 174 were identical to WFO® acceptable (green) and not evaluated further; 54 nonidentical pairs were evaluated (Table). Overall, 88.6% of decisions were either the same or viewed as equally acceptable by oncologists; oncologists preferred 3.9% of BIH treatments and 4.4% of WFO treatments. In cases where reasons for discordance were provided, 70% were due to BIH oncologist preference, 20% to patient preference and 10% to WFO treatment availability. We found no association between discordant recommendations and patient age or stage of cancer. Conclusions: This blinded study suggests WFO® 's therapeutic options are at as least as good as (or are an acceptable alternative to) treatments in practice. Blinding evaluators to source of treatment may minimize bias in comparisons of CDS systems and decisions made in practice. (Table Presented) .