A machine learning-based model may help identify newly diagnosed patients with chronic lymphocytic leukemia (CLL) who are at risk for infection or treatment within 2 years of a diagnosis, according to the results of a study recently reported in Nature Communications.
The model, called the CLL Treatment-Infection Model (CLL-TIM), is an ensemble algorithm made up of 28 machine learning algorithms. The model was created using data from 4149 patients with CLL in the Danish National CLL registry.
CLL-TIM was shown to have a precision of 72% and recall of 75% and was also shown to outperform the “current gold standard” prognostic model for CLL, known as CLL-IPI, which predicts time to first treatment and overall survival.
Compared with CLL-IPI, CLL-TIM also had superior performance at handling missing data, which is common in the real world, and this attribute was externally validated using data from the German phase 3 CLL7 watch and wait cohort (ClinicalTrials.gov Identifier: NCT00275054).
In addition, the researchers showed that training the model on 2 outcomes — infections and early treatment — “synergistically” improved the model predictions (P <.05), potentially explaining why the model outperformed the current gold standard.
The researchers plan to use the model to select patients with CLL who are at high risk of infection and/or early treatment for the phase 2/3 PreVent-ACaLL trial (ClinicalTrials.gov Identifier: NCT03868722). Enrolled patients will be randomly assigned to receive acalabrutinib and venetoclax for 3 months or be observed, allowing investigators to determine whether the combination therapy can “improve the natural history of immune dysfunction due to CLL.”
“To our knowledge, this is the first time a machine learning model will be used for patient selection in a randomized clinical trial,” the study authors wrote.
The model is freely available on CLL-TIM.org.
Agius R, Brieghel C, Andersen MA, et al. Machine learning can identify newly diagnosed patients with CLL at high risk of infection. Nat Commun. 2020;11(1):363.