Development and validation of a machine learning model to predict the daily risk of febrile neutropenia in patients with AML undergoing chemotherapy

Infections are among the main complications in patients with acute myeloid leukemia (AML) undergoing intensive chemotherapy, increasing morbidity and mortality. However, no established tools exist to detect early signs of infection. Here, we develop and validate a machine learning model to predict the daily risk of febrile neutropenia (FN). We conducted a retrospective observational study using registry data from two university hospital districts in Finland. We included adult patients with newly diagnosed AML undergoing intensive chemotherapy (n = 271, 48-hour FN outcomes=540), excluding acute promyelocytic leukemia. The model used a 48-hour prediction window and 1.2 M clinical data points to generate daily FN risk estimates. We built the model using the Extreme Gradient Boosting algorithm, explained variable importance with Shapley Additive Explanations, and evaluated and validated performance using measures of discrimination, calibration, and clinical utility. We show how seven variables predict daily FN risk: white blood cell count, temperature, time since cycle start, C-reactive protein trend, absolute neutrophil count, treatment phase, and prior antibiotic treatment. We achieve an AUROC of 0.91 (95% CI, 0.89–0.93) and a calibration slope of 0.95 (95% CI, 0.73–1.16) in the development cohort’s test set (n = 55, 48-hour FN outcomes=109) and validate performance in the external cohort (n = 136, 48-hour FN outcomes=124, AUROC 0.79 (95% CI, 0.75–0.82), calibration slope 0.65 (95% CI, 0.51–0.79)). While prospective validation is warranted, this study suggests that routinely collected clinical data may be useful for daily prediction of FN and that further development of such screening tools could potentially support personalized infection prevention. Infections can be a life-threatening complication of chemotherapy in patients with acute myeloid leukemia. Here, we developed a screening tool to help physicians detect these infections early. Using registry data from two Finnish university hospital districts, we trained an AI model to predict early infection onset over the next 48 hours. The model required seven variables, including routine laboratory tests, temperature measurements, time since each chemotherapy cycle started, and information on the treatment phase and antibiotic use. Such a screening tool could potentially support more personalized supportive care. However, the results must be confirmed prospectively before clinical adoption.