Identifying decision criteria in initiating treatment for acute myeloid leukemia in frail elderly by discrete choice modelling using behavioural artificial intelligence technology
Clinical decision-making for older patients newly diagnosed with acute myeloid leukemia (AML) is complex and time-sensitive. With a median diagnostic age above 70 years for patients, who are ineligible for intensive chemotherapy, overall survival (OS) is limited. The introduction of venetoclax combined with hypomethylating agents (HMA+Ven) has improved OS compared to HMA monotherapy but also increased toxicity. Determining which patients may benefit most from HMA, HMA+Ven, or supportive care—and balancing survival benefit with quality of life—remains a major challenge for hematologists.
Artificial intelligence (AI) is increasingly being used to support complex medical decisions. Integrating expert clinical judgment into AI-based approaches could enhance decision support. Discrete choice modelling offers a structured way to quantify expert preferences and can be applied within the Behavioural Artificial Intelligence Technology (BAIT) framework to inform shared decision-making. This study presents preliminary findings from a choice analysis exploring how hematologists weigh different patient and disease characteristics in treatment selection for older AML patients.
Three experienced hematologists first identified parameters likely to influence treatment choice, including clinical frailty, cognitive function, comorbidities, social context, and disease-related prognostic features. Using D-efficient design methodology, 32 hypothetical patient vignettes were developed in which these variables were systematically varied. Participating hematologists from all HOVON echelon centers reviewed the vignettes online and selected one of three treatment options: HMA, HMA+Ven, or supportive care. The resulting data were analyzed using a multinomial logistic regression model within the BAIT framework to estimate the relative influence of each factor on treatment decisions.
As of October 7 th, 2025, 32 of 130 invited hematologists (25%) completed the survey. Participants had diverse experience levels (<5 years: 13%; 5–15 years: 50%; >15 years: 37%) and represented all HOVON echelons. Age and Critical Frailty Index (CRI) were the most influential factors differentiating active treatment from supportive care (figure 1). Higher frailty and age levels significantly (p < 0.01) reduced the likelihood of selecting HMA+Ven compared with HMA monotherapy. Across all criteria, worsening patient characteristics decreased preference for both active treatments. In contrast, when all variables indicated favorable patient conditions, HMA+Ven was most preferred, followed by HMA and then supportive care. Prognostic disease factors and comorbidities contributed minimally to decision variance. Gender, leukocyte count, or comorbidity burden were not significant predictors. No differences were observed between hematologists with moderate (5–15 years) versus long (>15 years) experience (FIgure 2).
Preliminary results highlight age and and frailty as the primary determinants in treatment selection for older AML patients. The application of the BAIT framework effectively captures and quantifies expert reasoning, offering a promising approach to model complex clinical decision-making processes and potentially enhance shared decision-making in hematology.