Chronic lymphocytic leukemia prognostic models in real life: still a long way off

KEYWORDS:

• CLL
• prognostic models
• accuracy
• pre-treatment biomarkers
• post-treatment biomarkers

1. Introduction

The discovery of genetic lesions for chronic lymphocytic leukemia (CLL) is a chance to generate inclusive models integrating a combination of biomarkers [1–5]. Prognostic models were primarily designed to inform clinical decisions, however, their use in day-by-day practice remains undefined [2,4]. The reasons for this limitation have been extensively reviewed in a recently published Cochrane systematic review [6]. The analysis reveals that the current standards of prognostic model development (i.e., predictor assessment, outcome definition, and the relevant performance measures) were often misapplied. Moreover, it is not surprising that prognostic models typically established in the chemo-immunotherapy period were found to be useless in patients receiving targeted agents [7]. A model should be frequently updated and tailored to a particular population of CLL patients to sustain its predictive effectiveness, despite ongoing changes in the treatment paradigm of the disease [8–10]. Finally, frailty is an important feature not included in prognostic models which significantly affects the outcome of patients diagnosed with CLL. Retrospective studies have shown that in patients with CLL treated with new targeted agents (i.e., ibrutinib), comorbidities impact on prognosis and add discriminatory strength to the CLL-International Prognostic Index (CLL-IPI) [11,12].

2. Accuracy of prognostic models: a frequently unmet issue

The creation of models that accurately predict the clinical outcome of patients with high- and low-risk diseases is a key objective of CLL [9]. Generally, two model-related properties are used to determine the magnitude of a model prediction: discrimination and calibration. Discrimination refers to the ability to differentiate patients at higher risk from those at lower risk for developing an event [8]. Discrimination is typically calculated via the concordance index (i.e., c-statistic). The c-statistics can range from 0 to 1, with one indicating perfect discriminative potential and 0.5 indicating fair chance for the model ‘s predictions. Generally, values between 0.70 and 0.80 imply a successful model. From a clinical point of view, this means that in 20–30% of the cases, the prediction provided by the prognostic models investigated would be inadequate [10]. Of note, when a prognostic model is validated in an independent series, c-statistics will fluctuate across different cohorts [10]. The c-statistics by meta-analysis in patients with early and asymptomatic CLL was 0.70 in the recently proposed International Prognostic Score (IPSE) predicting time to first treatment (TTFT); however, it ranged from 0.66 to 0.75 across different validation cohorts [13].

Calibration measures the extent of absolute risk prediction in different patient populations [8]. In this context, information obtained from external validation studies is essential for the appropriate applicability in the clinical practice of prognostic model [6,8,10]. A significant pre-requisite for calibration analyses is the availability of internal and external validation studies. In CLL published studies, however, these data are frequently underreported [6]. A systematic analysis of the prediction of different CLL models indicates that only 12 (23%) of the 52 eligible studies provided data of externally validated models; in detail, six were developed for overall survival (OS), one for progression-free survival (PFS) and five for treatment-free survival (TFS) [6].

3. Can we improve on accuracy of prognostic models?

One challenging issue is whether additional variables such as complex karyotype data and novel mutations (i.e., NOTCH1, BIRC3, and SF3B1) may improve the prediction power of prognostic models in CLL [9]. Since CLL prognostic scores vary in terms of the number of variables used, we sought for correlation between complexity and discriminating power (i.e. c-statistics). For this purpose only studies having as endpoint OS with at least three validation analyses were selected (i.e. 2007 MDACC, CLL-IPI, B-B,BALL) (Table 1) [1,4,5,7]. Twenty-three CLL cohorts, including 16.251 patients, provided data useful

Table 1. Characteristics of prognostic models in CLL with at least 3 validation studies and overall survival as endpoint

Prognostic model	Variables	Endpoint analyzed	Number of patients included in development and validation studies
2007 MDACC*^(1)	age, ALC, β2 m nodal groups Rai stage gender	Overall survival	7246
CLL-IPI^§(4)	age β2-m IGHV mutational status, Rai/Binet stage TP53 mutational status	Overall survival	6003
B-B°^(5)	IGHV mutational status, FISH cytogenetics del (11q) and/or del (17p)	Overall survival	2696
BALL^^ (7)	β2-m LDH Hemoglobin Time from initiation of last therapy <24 mo.	Overall survival	30

*2007 MDACC, MD Anderson Cancer Center

^§CLL-IPI, Chronic Lymphocytic Leukemia-International Prognostic Index

°B-B, Barcelona-Brno

^BALL, B₂-microglobulin, Anemia, LDH, Last therapy

for this analysis [1,4,5,7,14–21]. A linear regression analysis shows that complexity does not guarantee an increase of model accuracy (R2 = 0.048; P = 0.240) (Figure 1). This means that the optimal choice in clinical practice should rely on models including a few variables keeping in mind that increasing the number of variables does not necessarily translate into an improvement in model prediction.

Figure 1. Linear regression correlating the complexity of prognostic models with their discriminating power (i.e., the c-statistic). Complexity of models reflected the number of variables included. Models considered for the purpose of this analysis had overall survival (OS) as endpoint and at least three validation studies. Models fulfilling these criteria were 2007 MDACC, CLL-IPI, Barcelona-Brno, and BALL score [1,4,5,7]

4. Prognostic models in early CLL

Most CLL patients are diagnosed with asymptomatic disease and are followed with no therapy unless disease progression occurs [22,23]. In these patients, phase 3 studies of early treatment based on fludarabine, cyclophosphamide and rituximab (FCR) or ibrutinib have failed to demonstrate OS benefit [24,25]. Accordingly, international guidelines recommend a policy of active surveillance [26]. Since observation remains the standard of care for asymptomatic early stage patients time to first treatment (TTFT) may be a sensitive endpoint [27,28)].

Different prognostic models have been proposed to assess the TTFT for patients with early-stage CLL; however, none has been universally established in routine clinical practice [4,5,13,14,27–30], five different prognostic indexes were compared in a cohort of patients with Binet A CLL from different institutions [31]. None of the scores was able to predict with absolute precision the evolution of early CLL patients. Efforts to harmonize larger cohorts of patients become crucial to identify a prognostic model useful in clinical management and in the design of early intervention clinical trials.

5. Prognostic models and treatment strategies

Prognostic risk scores are established by necessity on the basis of broad population-level studies with long-term follow-up [1–5]. This means that they are frequently obsolete by the time adequate data for validation has been collected by the model. The case of CLL-IPI, which was developed using data from patients registered in eight phase 3 randomized chemotherapy or chemoimmunotherapy clinical trials conducted between 1997 and 2007, is paradigmatic [4]. Moreover, as prognostic models are heavily dependent on the modality of treatment, it is not surprising that CLL-IPI has not been found to be effective in patients receiving targeted agents [4,7].

In patients with relapsed or refractory CLL treated with targeted agents, Soumerai et al. [7] developed a validated risk score to predict OS. This four-factor prognostic model is applicable, with easily obtained data from the patient treatment history and readily available laboratory tests (i.e., serum β 2-microglobulin [B2-M] ≥5 mg/dL, lactate dehydrogenase [LDH] >upper limit of normal, hemoglobin <110 g/L for women or <120 g/L for men, and time from initiation of last therapy <24 months). Of note, del(17p), an independent prognostic factor in patients with relapsed or refractory CLL treated with ibrutinib, is not included in the Soumerai et al. score [7,32]. In an exploratory study, it was found that patients belonging to Soumerai et al. [7] intermediate/high risk can be subclassified based on del (17p) status [33]. Furthermore, in a recent real-life experience, a simplified version of BALL score which excluded time from initiation of last therapy seems to outperform the original model in predicting prognosis of CLL patients stratified according to 17p deletion status [34]. Finally, a new score was established and validated, including LDH, B2-M, relapsed-refractory status, and TP53 status [35]. Notably, Bruton kinase (BTK) and/or phospholipase C gamma 2 (PLCG2) mutations have been observed more frequently in patients of the high-risk group, implying that the model will capture the subsequent risk of clonal evolution [35]. These recently proposed models are good candidates to assist clinicians in the selection of high-risk patients potentially suitable for novel approaches, however, they may integrate but not replace clinical expertise and medical judgment.

6. Dynamic models and the role of post-treatment biomarkers

Risk stratification largely focuses on pretreatment variables. However, post-treatment biomarkers such as minimal residual disease (MRD) assessment are gaining interest [36–38]. Achieving undetectable MRD after receiving therapy affects PFS and OS [38]. However, MRD assessment is indicated in clinical trials but not in everyday practice [39]. In addition to CLL-IPI, MRD is part of a new continuous individualized risk index (CIRI) aimed at dynamically evaluating individual CLL patients’ outcome probabilities [40]. This approach provides a real-time prognostic assessment throughout the patient’s disease course and has been validated in three CLL trials from the German CLL Study Group (i.e., CLL8, CLL10, and CLL11). CIRI represents a proof-of-concept project potentially useful for therapy selection; however, the complexity of this dynamic score prevents a wide applicability in clinical practice [40].

The first-in-class BTK inhibitor to demonstrate efficacy in CLL is ibrutinib [32]. It is also the first BTK inhibitor to have developed resistance in patients [41]. Of note, in around 80% of CLL patients with acquired resistance to ibrutinib, mutations in the BTK and PLCG2 domains are identified, although it remains uncertain if these mutations are merely associated with or specifically trigger disease relapse [41]. Very recently, a potential resistance mechanism to venetoclax has been reported and correlated to acquired mutation in BCL2 [42]. These mutations have the potential to be used as post-treatment biomarkers and tend to be a strong candidate for integration into models aimed at dynamically assessing the prognosis in patients with CLL treated with pathway inhibitor agents [9].

7. Conclusions

Transferring the prognostic model in the current CLL management is a critical issue and some considerations should be done. Validation of prognostic models is a crucial step toward their implementation in clinical practice. Models should be internally and externally validated to obtain reliable estimates of model performance including assessments of discrimination and calibration. Variation in model performance is commonly observed across different settings when a prognostic model is externally and extensively validated. Therefore, validated prognostic models should be applied while keeping in mind that they cannot replace clinical expertise and medical judgment [9,43]. Finally, a decision-analytic evaluation is important to identify models that aim to improve clinical decision-making.

Of note, machine learning offers novel opportunities in the field of prognostic assessment [44]. In contrast to models that simply add to the impact of individual risk factors, machine learning evaluates the more complex and non-linear mechanisms by which prognostic factors contribute toward composite risk in CLL [44]. A CLL treatment-infection model (CLL-TIM) has been developed using the machine learning methodology. The model can handle heterogeneous data including the high rates of missing data to be expected in the real-world setting with a precision of 72% [45].

Declaration of interest

S Molica has received advisory board honorarium from AbbVie, Jansen, and Astra Zeneca; he has also received speaker bureau honoraria from Gilead. The author has no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed here.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

This paper was not funded.