Machine Modeling as a Predictive Tool for Infection Risk in Patients With CLL
Posted: Monday, March 9, 2020
Infections are the major cause of morbidity and mortality in patients treated for chronic lymphocytic leukemia (CLL), and predictive models that combine treatment and infection as an outcome have not been proposed. Consequently, Rudi Agius, PhD, of Technical University of Denmark, and colleagues developed the CLL Treatment Infection Model (TIM), which identifies patients at risk of infection or who had treatment within 2 years of diagnosis. Their model was able to identify patients at risk of severe infection or CLL treatment within 2 years of diagnosis with a 72% precision rate and a recall rate of 75%. Their results were reported in Nature Communications.
According to the international workshop on CLL guidelines, treatment is not recommended unless cytopenia, symptomatic disease, or short lymphocyte doubling time is present. During this waiting period, severe infections prior to treatment result in a higher 30-day mortality (9.8%) than upon treatment, resulting in worse treatment-free and overall survival outcomes compared with those in patients without severe infection.
The researchers collated an ensemble of 28 machine learning algorithms to model changes in patient histories spanning 7 years prior to CLL diagnosis. They identified 3,720 patients with CLL between January 2004 and July 2017 who met inclusion criteria. During the 2-year predictive window, 572 patients (15.4%) had a severe infection, 398 (10.7%) received treatment, and 103 (2.8%) died; a total of 2,647 patients (71.1%) had no study-related events.
The researchers plan to use their model to select patients for a randomized clinical trial (ClinicalTrials.gov identifier NCT03868722) to investigate whether 3 months of venetoclax and acalabrutinib combination therapy improves the grade ≥ 3 infection-free, treatment-free survival compared with the standard-of-care observation arm of the study. They also plan to assess the predictive performance of this machine learning model as they follow patients predicted to be at low risk for infections.
Disclosure: For full disclosures of the study authors, visit nature.com.