Chronic lymphocytic leukemia (CLL) is the most frequent leukemia in Western countries.1 In the 2016 revision of the World Health Organization (WHO) classification of lymphoid neoplasms, CLL was described as leukemic lymphocytic lymphoma, only to be distinguishable from small lymphocytic lymphoma (SLL) by its leukemic presentation.2 It is important to distinguish CLL from the different disease entities belonging to mature B-cell lymphoid neoplasms and there are still difficulties in defining some cases.2,3 Despite advances in cytogenetics and molecular biology having led to a significant improvement in the diagnosis of lymphoproliferative diseases, it is necessary to evaluate the blood smear and immunophenotype for the successful differential diagnosis of CLL.3,4 A recent, large harmonization study confirmed that a panel of CD19, CD5, CD20, CD23, kappa and lambda is usually adequate for the diagnosis of CLL,5 however, in borderline cases, more markers are needed to make a diagnosis. In ambiguous immunophenotypes, the markers, such as CD43, CD79b, CD81, CD200 or ROR1, may help to refine the diagnosis.6

Healthcare providers use medical clinical decision algorithms based on a set of decision rules to improve care, diagnose, determine the risk for medical conditions and reduce costs.7 A machine learning approach for immunophenotyping with flow cytometry may reduce the cost of diagnostic procedures of lymphoproliferative disease, particularly CLL. The availability of a clinical decision algorithm for diagnosis of CLL would benefit patients with ambiguous immunophenotypes by providing the opportunity to diagnose CLL with reasonable accuracy. Previous studies proposed a decision-tree approach to improve the flow cytometry utilization in the differential diagnosis of lymphoproliferative diseases.8,9 However, the differential diagnosis of LPD has long been considered a difficult task, as it requires the differential diagnosis between different subtypes of heterogeneous and highly variable diseases.10

This study was designed to create a useful and feasible tool for CLL diagnosis, using immunophenotyping with flow cytometry. A predictive model using the Backward LR was built to estimate the diagnostic value of antigens used in routine CLL immunophenotyping and propose a novel differential diagnosis algorithm for B- cell LPD.

The data set was divided into two groups, the training set and the test set. A total of 302 patients were included, of whom n = 242 (80%) were in the training set and 60 (%20), in the test set. Descriptive statistics for the patients were presented in Table 1 for the training set, test set and total.

The CD5, CD43, CD81, ROR1, CD23, CD79b, FMC7, sIg and CD200 were the variables included the modeling process. The logistic regression analysis with the backward elimination method (Backward LR) for the training set with CD5, CD43, CD81, ROR1, CD23, CD79b, FMC7, sIg and CD200 independent variables and the binary outcome variable CLL (CLL and others) was performed. According to the hierarchical (Backward LR) logistic regression result, the CD5, CD81, CD23, CD200 and FMC7 markers remained in the model after 5 steps (p-values 0.094 for FMC7 and < 0.05 for others). At the end of five steps, the FMC7 variable with a value of p = 0.094 was also removed from the model manually and the model with the results in Table 2 was obtained.

In our model, the decrease in the CD81 intensity and increase in the intensity of other markers (CD5, CD23 and CD200) increased the probability of the diagnosis of CLL, (p < 0.05). The odds ratios for the CD5, C23, CD200 and CD81 variables were 1.088 (1.050 - 1.126), 1.044 (1.012 - 1.077), 1.039 (1.007 - 1.072) and 0.946 (0.921 - 0.970) [95% C.I.], respectively. The results of the logistic regression model for the CLL prediction are given in Table 2.

The classification performance (Model predictions) for both the training and test sets are presented in Table 3. The model was established with 95.7% of CCR, 95.27% of sensitivity, 91.46% of specificity, 96.83% of PPV and 92.59% of NPV for all the data.

The number and percentages of the actual diagnosis class regarding the predictive value of the model and Matutes scores of all patients are given in Table 4. For 50.5% (102 + 9 = 111 people) of 220 patients diagnosed with CLL, the Matutes score was 4 and 5. The model estimation for all these 111 patients was CLL. The Matutes score was 3 for 40% (88 people) of 220 patients diagnosed with CLL and the model estimation for the 87 patients was CLL, with the exception of 1 of these 88 patients. The model estimation for 97.3% (214) of 220 patients diagnosed with CLL, was CLL. The Matutes score was 0, 1 and 2 for 74.5% (18 + 30 + 13 = 61 people) of the 82 patients diagnosed with lymphoproliferative disease other than CLL. The model estimation for 58 patients, with the exception of 3 of these 61 patients, was others (LPD other than CLL). For the 18.3% (15 people) of 82 patients diagnosed with others, the Matutes score was 3 and the model estimation for the 11 patients, with the exception of 4 of these 15 patients, was others. The model prediction for 91.5% (75) of 82 patients diagnosed with lymphoproliferative disease other than CLL, was others (Table 4).

The comparison of the model and the new score proposed by D'Arena et al.12 is in Table 5. For 93.2% (158 + 47 = 205 people) of 220 patients diagnosed with CLL, the D'Arena score was 3 and 4. The model estimation for 202 patients, with the exception of 3 of these 205 patients, was CLL. The model estimation for 97.3% (214) of 220 patients diagnosed with CLL, was CLL. A total of 62.2% (5 + 46 = 51 people) of 82 patients diagnosed with LPD other than CLL had low D'Arena scores, including 0 and 1. The model estimation for 49 patients, with the exception of 2 of these 51 patients, was others. For the 26.8% (22 people) of 82 patients diagnosed with others, the D'Arena score was 2. The model prediction for the 19 patients, with the exception of 3 of these 22 patients, was others. The model prediction for 91.5% (75) of 82 patients diagnosed with lymphoproliferative disease other than CLL was others (Table 5).

Flow cytometry immunophenotyping is an essential tool in the diagnostics of B-cell lymphoproliferative disease, including CLL.2,5,13 The Matutes score, based on the expression of CD5, CD23, FMC7, surface immunoglobulin (SmIg) and CD22/CD79b by leukemic cells, has been used for more than 20 years in the diagnosis of CLL.14,15 However, the differential diagnosis of CLL still remains a challenge due to ambiguous immunophenotypes. In this study, we tried to provide a novel scoring algorithm to overcome the diagnostic challenge of CLL.

A composite immunophenotypic panel of different B-cell markers is used to distinguish CLL from other B-cell LPDs because there is no single marker characteristic of a single disease entity. In the Matutes score system, positivity of CD23 and CD5 and weak or absent expression of CD79b (or CD22), FMC7 and SmIg are converted to a score of 0 - 5 and interpreted as CLL in patients with a score greater than 3.14 In addition, D'Arena et al. defined the simplified score, compared to the classical Matutes score, for the differential diagnosis of CLL—which only requires 4 markers, such as CD5, CD23, CD200 and SmIg.12

In our study, the model was created with the variables CD5, CD23, CD200 and CD81 obtained from the Backward LR method, which was established with CD5, CD43, CD81, ROR1, CD23, CD79b, FMC7, sIg and CD200. We found that a 1-unit increase in the MFI of CD5 increases the risk of CLL 1.088-fold, a 1-unit increase in the MFI of C23, 1.044-fold and a 1-unit increase in the MFI of CD200, 1.039fold. A 1-unit decrease in the MFI of CD81 increases the risk of CLL by 1.057-fold (1 / 0.946). In addition, our model estimation for 97.3% of 220 CLL patients with a Matutes score of 3, 4 and 5 was CLL. However, the model prediction for 91.5% of 82 patients diagnosed with lymphoproliferative disease other than CLL was others. The Matutes score of these 75 patients was 0, 1 and 2. We suggest that our model provides a novel diagnostic algorithm with 97.3% of sensitivity and 91.5% of specificity. Our model prediction for 97.3% (214) of 220 patients diagnosed with CLL was CLL and for 91.5% (75) of 82 patients diagnosed with lymphoproliferative disease other than CLL, others, according to the D'Arena score.

Previous studies incorporated antigens, such as CD200, CD43 and CD81, in diagnostic algorithms to facilitate the separation of CLL from another B-cell LPD. They all reported that quantitative evaluation of these markers may be a feasible manner to better identify B- lymphoproliferative disorders.16–20 Falay et al. recommended including the CD43, CD81, CD200 and ROR1 in the diagnostic tool of CLL, particularly in distinguishing it from MCL.21 We suggest that a new score of 4 markers (CD5, CD23, CD 200 and CD 81) is highly effective in distinguishing CLL from other LPDs.

In most cases complete blood counts, a blood smear and immunophenotyping are enough to establish the diagnosis of CLL. However, ambiguous immunophenotypes are common and remain a diagnostic challenge. Here, we propose a differential diagnosis algorithm with higher sensitivity and specificity than the classic Matutes score in diagnosing CLL, with 95.27% of sensitivity and 91.46% of specificity. Our model provides a novel diagnostic algorithm in which the cases were correctly classified as CLL and others, with a 95.7% correctness rate. We suggested that our novel diagnostic model may greatly contribute to the diagnosis of CLL, particularly in cases with ambiguous immunophenotypes.

All procedures performed in this study were in accordance with the ethical standards of the Institutional Review Board of the Clinical Research Ethics Committee of the Ankara City Hospital and with the 1964 Helsinki declaration and its later amendments. Informed consent was obtained from all individual participants included in the study.

No funding.

Study concepts by Mesude Falay and Zehra Narli Ozdemir; study design by Mesude Falay; data acquisition by Zehra Narli Ozdemir, Ozlem Beyler, Eylem Genç and Ahmet Kursad Gunes; quality control of data and algorithm by Mesude Falay, Funda Ceran and Simten Dagdas; data analysis and interpretation by Mesude Falay and Zehra Narli Ozdemir; statistical analysis by Ayhan Parmaksiz; manuscript preparation by Zehra Narli Ozdemir; manuscript editing by Mesude Falay, Simten Dagdas and Gulsum Ozet, and; manuscript review by Mesude Falay.