Prediction of clinical outcome in CLL based on recurrent gene mutations, CLL-IPI variables, and (para)clinical data

Abstract

A highly variable clinical course, immune dysfunction, and a complex genetic blueprint pose challenges for treatment decisions and the management of risk of infection in patients with chronic lymphocytic leukemia (CLL). In recent years, the use of machine learning (ML) technologies has made it possible to attempt to untangle such heterogeneous disease entities. In this study, using 3 classes of variables (international prognostic index for CLL [CLL-IPI] variables, baseline [para]clinical data, and data on recurrent gene mutations), we built ML predictive models to identify the individual risk of 4 clinical outcomes: death, treatment, infection, and the combined outcome of treatment or infection. Using the predictive models, we assessed to what extent the different classes of variables are predictive of the 4 different outcomes, within both a short-term 2-year outlook and a long-term 5-year outlook after CLL diagnosis. By adding the baseline (para)clinical data to CLL-IPI variables, predictive performance was improved, whereas no further improvement was observed when including the data on recurrent genetic mutations. We discovered 2 main clusters of variables predictive of treatment and infection. Further emphasizing the high mortality resulting from infection in CLL, we found a close similarity between variables predictive of infection in the short-term outlook and those predictive of death in the long-term outlook. We conclude that at the time of CLL diagnosis, routine (para)clinical data are more predictive of patient outcome than recurrent mutations. Future studies on modeling genetics and clinical outcome should always consider the inclusion of several (para)clinical data to improve performance.

Graphical abstract

Figure 1.

Schematic representation of the data collection and analysis. (A) Data from different data sources, including baseline tests (), routine laboratory tests (), and recurrent mutations (), were combined to construct a heterogeneous data set. Prediction point was set at 3 months postdiagnosis, and clinical outcomes () were predicted. (B) The clinical outcomes were death (), treatment (), the combined event of treatment or infection (composite), and infection (). (C) Based on the combination of feature sets, 4 models were defined: (1) IPI, which included CLL-IPI score and the CLL-IPI features only; (2) +BL, which included CLL-IPI features, baseline tests, and routine laboratory tests; (3) +MUT, which included CLL-IPI features and recurrent mutations; and (4) ALL, which included all features. (D) Clinical outcomes were predicted in 2- and 5-year outlooks postdiagnosis (except for the first 3 months). (E) The data from different sources were merged to create one data set (). Then, for a specific outcome and outlook, the target values were created and later used in the training/test (). Based on the model, feature extraction was performed (). A stacked ML model consisting of 7 algorithms and a fusion stage based on majority voting was trained and tested. The performance of the models () and the contribution of the features () were estimated to identify the risk factors predictive of each combination of outcome, model, and outlook. tNGS, targeted next-generation sequencing.

Figure 2.

Comparison of the performance of the models in predicting clinical outcomes. Box plots depicting the performance of the 4 models in predicting death, treatment, infection, and the combined outcome of infection or treatment within 2-year (A,C,E,G) and 5-year (B,D,F,H) outlooks postdiagnosis. In each subplot, the box plots show the quartiles, the median, and the outliers of MCC values obtained using fivefold cross-validation on set A. The scatter plots (single marker for each model) demonstrate the MCC value of the holdout validation on set C. Death (A-B), treatment (C-D), treatment or infection (composite) (E-F), and infection (G-H). *P =.05, ***P =.001.

Figure 3.

Identified risk factors predictive of the outcomes. SHAP plots on the full cohort illustrate the contribution of the most important features in predicting the clinical (mean absolute SHAP values >0.01) in 2-year outlook (A-D) and 5-year outlook (E-H). Death (A,E), treatment (B,F), treatment or infection (C,G), and infection (D,H). The cluster of features predictive of different outcomes was detected after sorting the features so that the most important features predictive of treatment appear at the top (red) and the most important features predictive of infection appear at the bottom (blue). The features more predictive of death or the composite outcome and not treatment or infection appear in the middle. ECOG, Eastern Cooperative Oncology Group; Ig, immunoglobulin.

Figure 4.

Hierarchic clustering of risk factors predictive of clinical outcomes in the 2 outlooks. The similarity between each pair of risk factor patterns was calculated using cosine similarity, which is the cosine of the angle between 2 n-dimensional vectors. Then, the risk factor patterns were grouped by performing hierarchic clustering on the computed similarity matrix.

Figure 5.

Event-free survival plots for 2-year outlook on the full cohort. Patients were stratified by the classifier using CLL-IPI variables (TP53 aberrations, IGHV mutational status, B2M level, clinical stage, and age [IPI]) (A,C,E,G) and addition of (para)clinical variables to the CLL-IPI variables (+BL) (B,D,F,H). Overall survival (A-B), treatment-free survival (C-D), treatment and infection–free survival (E-F), and infection-free survival (G-H). HR, hazard ratio.

Figure 6.

Event-free survival plots for 5-year outlook on the full cohort. Patients were stratified by the classifier using CLL-IPI variables (TP53 aberrations, IGHV mutational status, B2M level, clinical stage, and age [IPI]) (A,C,E,G) and addition of (para)clinical variables to the CLL-IPI variables (+BL) (B,D,F,H). Overall survival (A-B), treatment-free survival (C-D), treatment and infection–free survival (E-F), and infection-free survival (G-H). HR, hazard ratio.