1. Introduction

Erythrocytosis is characterized by an abnormally high red cell mass (RCM) and is discernible through elevated levels of hematocrit, hemoglobin, or red blood cell (RBC) count. There are instances when the apparent increase is due to a decrease in plasma volume. True erythrocytosis, though, is when the RCM surpasses 125% of the expected amount for a given age and gender.[1–4] As RCM measurement relies on an isotope dilution method, hemoglobin and hematocrit values are commonly used in clinical practice. World Health Organization (WHO) proposed the diagnostic thresholds for hemoglobin and hematocrit as >16.5 or >0.49 g/dL for men, and >16.0 or >0.48 g/dL for women.[5,6] The majority of erythrocytosis cases are acquired and originate from conditions resulting in low oxygen, such as smoking, living at high altitudes, and certain heart, lung, or kidney diseases. Additionally, some tumors and organ transplants can anomalously increase erythropoietin (EPO) production, leading to erythrocytosis.[1] Medications like diuretics, EPO and androgens also may result in erythrocytosis as a potential side effect.[7]

While less frequent, there are genetic triggers for erythrocytosis too, like polycythemia vera (PV) or the inherited condition, familial erythrocytosis.[8] If erythrocytosis manifests without a clear cause and displays well-known symptoms like headaches, vision issues (including blurring and scotomata), erythromelalgia, itching, spleen discomfort due to splenomegaly, skin vein inflammation, or significant thrombosis or bleeding events, PV becomes a primary suspect. One of the diagnostic approach for PV is the Janus Kinase-2 (JAK2) mutation test (targeting V617F and also exon 12). This mutation appears in about 99% of PV cases.[9] Spotting the JAK2V617F mutation in the lab boasts a sensitivity of 97% and almost 100% specificity, helping distinguish PV from other erythrocytosis causes.

In the 2016 WHO update, the diagnostic criteria for PV saw significant reductions in the hemoglobin and hematocrit threshold values. Consequently, there has been a marked increase in the number of cases where the JAK-2 test is requested, with only 1% to 5% returning positive results.[10] Given this, many researchers search for smarter testing methods to reduce unnecessary tests and hence costs.[11–13] Our objective here is to introduce an artificial intelligence application as a more rational approach for testing JAK2 mutations in case of erythrocytosis.

2. Patients and methods 2.1. Patient selection

All test results were obtained retrospectively and anonymously from samples sent to a tertiary hospital’s genetic laboratory to undergo JAK2V617F mutation testing, adhering to WHO criteria between 2017 and 2023. Gender was noted. Positivity and absolute values in percentage for JAK2 were recorded. Concurrent complete blood count (CBC) results were obtained.

2.2. Genetic testing

JAK2V617F mutation was analyzed by the JAK2 Somatic Mutation Screening Kit (Genmark). The JAK2 Somatic Mutation Screening Kit analyses the JAK2 V617F mutation by ASO-Realtime-PCR technique. Primer Probes include Mutant (FAM) and Endogenous (VIC) control-specific primer probe (TaqMan) sets. The genomic DNA was extracted from peripheral blood samples using preparation methodologies. After purification, the acquired DNA samples were diluted to 20 to 100 ng/μL concentration. In the ASO-Realtime-PCR reaction, the master mix, the primer-probe mix, a negative control, a positive control with 100% mutation concentration, and a calibrator (which has a unique mutation concentration for the used kit), were included with each PCR run. All reactions were carried out in the Lightcycler 480 System (Roche, Penzberg, Germany) according to the manufacturers’ protocols. With fit point analysis of positive control, negative control, and calibrator, the percentage of mutant allele was established. The mutation ratio is calculated by a special Formula using DeltaCT of the patient sample and calibrator. The Kit can detect between 0.5% and 100% mutations in wild-type background DNA with high sensitivity.

2.3. Statistical analysis

Statistical analyses were performed via IBM SPSS v.20 (IBM Corp, Armonk, NY). In the acquired dataset, a frequency analysis was performed. Categorical variables were presented as count (%). For continuous variables, parametric ones were expressed as mean ± SD (min–max) and non-parametric ones as median (25–75 percentile). The normality of the data was checked via Kolmogorov–Smirnov test. Independent samples t test or Mann–Whitney U test were used to compare 2 independent variables. Performance metrics were calculated with confusion matrices. Comparative receiver operating characteristics curve analysis was done with MedCalc program (MedCalc Software Ltd, Ostend, Belgium). A P value was accepted as significant below .05.

2.4. Machine learning approach

The approach for model training and testing was implemented using Python programming language, with the employment of numpy, pandas, sklearn, imblearn, and xgboost libraries. The collected data was randomly divided into training and test datasets (80–20%) ensuring that the mutation frequency was identical in both arms. Due to the low prevalence of mutation in the training set, the Adaptive Synthetic Algorithm was applied to create synthetic data in the training set to address the imbalance.[14] Training was executed using various methods, and the trained models were tested on the test dataset. Performance metrics were obtained. Comparative analysis was done using the model performances derived from the test dataset, decision matrices obtained with the literature’s JAK two-stage algorithm, and JAKPOT rules.

2.5. Ethics

This study was approved by the ethical committee of our institution (2023/1173).

3. Results

In the studied cohort involving 458 cases (Fig. 1), the JAK2V617F mutation was detected in 13.3% (61 cases). The average JAK2V617F mutant allele burden in positive cases was 39.38 ± 29.60% (1.21–100.00%). In our analysis of patients who tested positive for the JAK2 mutation, a significant gender disparity was observed. Male patients represented a higher proportion of those with a positive test result, accounting for 63.9% compared to 36.1% of female patients. However, the probability of mutation positivity was notably higher in females than in males, with 33.8% of women testing positive for the mutation against 9.9% of men, a difference that was statistically significant (P < .001).

Figure 1.:

Flowchart for selection of involved patients.

Examination of CBC parameters exhibited noteworthy disparities between the mutation carriers and non-carriers, as showed in Table 1. Hemoglobin levels, revealed statistically significant differences in the JAK2V617F group when juxtaposed with the wild-type group. Similarly, the hematocrit, RBC count, mean corpuscular volume, mean corpuscular hemoglobin, and mean corpuscular hemoglobin concentration brought to light marked deviations, particularly emphasizing an elevation in RBC counts and hematocrit and a reduction in other cellular indices in the mutation-bearing cases. Moreover, a heightened red cell distribution width and altered white blood cell subpopulations, specifically an elevation in neutrophil and eosinophil counts and a decline in lymphocyte count in the JAK2V617F mutation bearers, were observed. The platelet parameters were also different between the groups, wherein the mutation holders manifested remarkably elevated platelet counts and plateletcrit, possibly signifying an aberration in megakaryopoiesis or platelet production by the JAK2V617F mutation.

Table 1 - Characteristics of the involved cases. All cases JAK2 WT JAK2V617F P * Number of cases, n (%) 458 (100.0) 397 (86.7) 61 (13.3) – Complete blood count results, median (25–75th percentile)  Hemoglobin (g/dL) 17.2 (16.7–17.9) 17.3 (16.8–17.9) 16.9 (16.4–18.1) .046  Hematocrit (%) 50.7 (49.2–52.8) 50.5 (49.2–52.6) 52.3 (49.5–55.3) <.001  Red blood cell count (×106/mm3) 5.79 (5.55–6.08) 5.76 (5.53–6.01) 6.56 (5.90–7.43) <.001  Mean corpuscular volume (fL) 87.7 (84.6–90.9) 88.1 (85.3–91.2) 83.3 (72.1–87.8) <.001  Mean corpuscular hemoglobin (pg) 30.0 (28.7–31.3) 30.2 (29.0–31.4) 26.8 (23.1–29.6) <.001  Mean corpuscular hemoglobin concentration (g/dL) 34.0 (33.3–34.8) 34.2 (33.5–34.9) 32.6 (31.8–33.6) <.001  Red cell distribution width (%) 13.6 (13.1–14.5) 13.5 (13.0–14.1) 17.3 (14.8–19.0) <.001  White blood cell count (×106/L) 8650 (7200–10525) 8300 (7100–10200) 10200 (8850–13400) <.001  Neutrophil count (×106/L) 5300 (4300–6800) 5100 (4150–6400) 7800 (5800–10600) <.001  Lymphocyte count (×106/L) 2200 (1800–2800) 2300 (1800–2900) 2000 (1450–2400) <.001  Monocyte count (×106/L) 700 (500–800) 700 (500–800) 700 (500–800) .941  Eosinophil count (×106/L) 200 (100–300) 200 (100–300) 300 (200–450) <.001  Basophil count (×106/L) 0 (0–100) 0 (0–100) 100 (92–100) <.001  Platelet count (×106/L) 235.5 (194.7–286.0) 227.0 (191.0–266.0) 557.0 (305.5–782.5) <.001  Mean platelet volume (fL) 8.8 (8.1–9.5) 8.8 (8.2–9.6) 8.5 (7.8–9.1) .004  Plateletcrit (%) 0.21 (0.18–0.25) 0.20 (0.17–0.23) 0.48 (0.26–0.60) <.001  Platelet distribution width (%) 17.0 (16.7–17.4) 17.0 (16.6–17.4) 17.5 (17.1–18.0) <.001

JAK2 WT = Janus Kinase-2 wild type.

*Mann–Whitney U test.

Several models were trained with data. There was no statistically significant difference in mutation prevalence between original training and test datasets (13.3% vs 13.0%, respectively; P = .845). For training, hemoglobin, hematocrit, RBC count, red cell distribution width (RDW), neutrophil count, monocyte count, platelet count, and platelet distribution width (PDW) variables were used regarding panmyelosis pattern observed in PV. The performance metrics of several trained models on the test dataset, as presented in Table 2, exhibit diverse efficacies in their predictive capabilities. The random forest (RF) model stands out, demonstrating precision, recall, F1-score, accuracy, and area under the receiver operating characteristic (AUROC) – all reaching a peak score of 100% or 1.0000, suggesting an exemplary fit and predictive proficiency on the tested data. In contrast, the gradient boosting (GB) model, while presenting notably high scores, does not match the absolute precision of Random Forest but still boasts formidable metrics, including a 92% precision, 100% recall, and an AUROC of 1.0000. Logistic regression also delineates respectable performance with 73% precision and a remarkably high AUROC of 0.9948, implying a strong capability to distinguish between the classes. The K-nearest neighbor model showed a balanced precision and recall of 41% and 58%, respectively, and an AUROC of 0.7516, indicating a moderate classification power. Lastly, the Support Vector Machine model emerges as the least proficient among the considered models, with the lowest scores across all metrics: 35% precision, 50% recall, and an AUROC of 0.7042. Each model thus reflects variable strengths and potential areas of improvement in predicting outcomes based on the tested data, with RF and GB surfacing as particularly potent in this instance.

Table 2 - Performance metrics of trained models in test dataset. Precision Recall F1-score Accuracy AUROC Support vector machine 35.0 50.0 41.0 82.0 0.7042 K-nearest neighbor 41.0 58.0 48.0 84.0 0.7516 Logistic regression 73.0 92.0 81.0 95.0 0.9948 Gradient boosting 92.0 100.0 96.0 99.0 1.0000 Random forest 100.0 100.0 100.0 100.0 1.0000

AUROC = area under receiver operating characteristic curve.

When RF and GB models were chosen to be compared to existing Piris-Villaespesa algorithm (PVA) and JAKPOT rule, they showed prominent superiority which was shown in Figure 2 and Table 3. In evaluating the performance of 4 predictive models based on various metrics, a number of intriguing patterns and variabilities emerge. GB and RF demonstrate remarkably high effectiveness, with both boasting a sensitivity of 100%. They exhibit substantial specificity, with GB at 98.8% and RF achieving a perfect 100%, coupled with impressive accuracy figures of 98.9% and 100%, respectively. Both models show a robust AUC, affirming their reliable classification capabilities. On the other end of the spectrum, PVA and JAKPOT, while maintaining 100% sensitivity, falter notably in specificity, gauged at 52.5% and 70%, respectively. The positive predictive value also manifests a wide range, with RF at a peak of 100% and PVA at a modest 24%. Furthermore, while GB, RF, and JAKPOT ensure a 100% negative predictive value, the divergence in other metrics, notably AUC (with GB and RF outperforming PVA and JAKPOT) and accuracy (with PVA lagging at 58.6%). The pairwise comparison of models indicates a statistical significance in the differentiation between model performances, revealing superiority of GB and RF over PVA and JAKPOT.

Table 3 - Comparative ROC analysis. Sens Spec PPV NPV Acc AUC 95% CI GB RF PVA JAKPOT GB 100.0 98.8 92.3 100.0 98.9 0.994 0.949–0.997 – 0.699 0.006 0.048 RF 100.0 100.0 100.0 100.0 100.0 1000 0.960–1000 0.699 – 0.005 0.036 PVA 100.0 52.5 24.0 100.0 58.6 0.762 0.662–0.845 0.006 0.005 – 0.369 JAKPOT 100.0 70.0 33.3 100.0 69.5 0.850 0.760–0.916 0.048 0.036 0.369 –

Bold values are the ones with statistical significance.

Acc = accuracy, AUC = area under curve, CI = confidence interval, GB = gradient boosting, NPV = negative predictive value, PPV = positive predictive value, PVA = Piris-Villaespesa algorithm, RF = random forest, ROC = receiver operating characteristics, Sens = sensitivity, Spec = specificity.

Figure 2.:

Comparative receiver operating characteristic curve.

4. Discussion

Currently, there’s no universally agreed-upon method for screening polycythemia vera (PV) in instances of erythrocytosis, beyond the established WHO 2016 definitions and current guidelines.[3,5,6,9] Even though this approach is fully accepted, especially in the presence of various clinical features and findings such as unexplained thrombocytosis, leukocytosis, splenomegaly, splanchnic thrombosis, and iron deficiency, it still carries a low positive prevalence.[10,15] Hence, several investigators searched for a more rational decision-making algorithms to reduce the number of unnecessary tests.

In 2019, Mahe and colleagues carried out a study aiming to develop and verify a computational method named JAK2-tree.[11] This algorithm uses peripheral blood markers to determine the clinical necessity of testing for the JAK2V617F mutation. The method did not only rely on hemoglobin or hematocrit values; it also took into account factors like platelet and white blood cell counts. To validate its efficiency, 2 data sets were employed: The first was the Copenhagen General Population Study, consisting of 50,363 participants who underwent CBC and JAK2V617F allele ratio tests. The second set comprised historical lab tests from Calgary Lab Services between January 2009 and August 2017, totaling 2989 JAK2 V617F mutation test records. Within the Copenhagen General Population Study, the JAK2-tree exhibited a 91% sensitivity and 99% negative predictive value. It had a 43% specificity, and its positive predictive value was 22%. Six JAK2V617F -positive cases went undetected by this algorithm, and half of them were linked to a malignant diagnosis. In the historical lab data, of the 2989 JAK2V617F tests, 580 were positive, indicating a 19% hit rate. Here, the JAK2-tree displayed a 94% sensitivity, 92% negative predictive value, 17% specificity, and a 21% positive predictive value. Using the JAK2-tree would have overlooked only 1.2% of the positive JAK2V617F mutations but would have reduced the testing frequency by 15%. In summary, the JAK2-tree provides an effective strategy based on CBC markers to streamline JAK2V617F mutation testing.

This approach offers the dual benefit of reducing costs and improving the diagnostic process’s efficiency, not just for PV, but also for essential thrombocytosis and primary myelofibrosis. However, in our perspective on establishing more rational testing criteria for PV, specifically regarding the WHO 2016 cutoffs (Hb levels greater than 16.5 g/dL in men or 16.0 g/dL in women), it does not offer any secondary branches as a helpful tool. Two more relevant studies exist that specifically explore which patients should undergo the JAK test. In both studies, attempts were made to provide rules and algorithms to predict the necessity of the JAK test, starting with patient groups that meet the WHO 2016 criteria.

Piris-Villaespesa et al[12] set out to create and validate a two-step algorithm to screen people based on the prevalence of JAK2V617F, using the WHO 2016 guidelines for increased Hb or Htc levels. The research was bifurcated into 2 stages. The initial phase focused on algorithm formulation, and the subsequent phase on its validation. Blood specimens from subjects were assessed for factors such as Hb, Htc, leukocytes, neutrophils, platelets, MCV, MCH, and RDW. In the initial phase, from 15,366 blood samples, those meeting the WHO 2016 thresholds for Hb or Htc were subjected to JAK2V617F mutation screening. Variables showing high diagnostic precision were integrated into the algorithm. Of these samples, 1271 (8.3%) showed elevated Hb or Htc levels per WHO 2016 guidelines. From this subset, 1001 samples were screened for the JAK2 mutation, with 13 testing positive (1.3%). Consequently, the prevalence of the JAK2V617F mutation in samples fitting the WHO criteria was 0.8% (8/996). Those with the mutation exhibited notable disparities in their hematological parameters. In this dataset, the algorithm was developed. Validation was performed on an independent set of 15,298 samples. Those fitting the algorithm’s requirements were screened for the JAK2 mutation. Among these, 1595 (10.4%) showed elevated Hb or Htc. Of this subset, 501 proceeded to the next phase, and 7 tested positive for JAK2V617F. This indicates a mutation prevalence of 0.04% in the entire set, 0.4% among those with elevated Hb or Htc, and 1.2% in those meeting both algorithm steps. Additionally, a random sampling of 300 samples revealed 1 positive JAK2V617F case, indicating a prevalence below 1/300. The team effectively formulated and authenticated a two-step algorithm that leverages JAK2V617F mutation prevalence to screen individuals showing raised Hb or Htc, as per the WHO 2016 guidelines. This method holds promise as a diagnostic instrument for identifying potential PV patients.

Chin-Yee and his team set out to craft a diagnostic methodology for patients referred due to high hemoglobin levels.[13] They used a prediction rule grounded on CBC and white blood cell differential parameters to steer JAK2 mutation testing. This rule was formulated and confirmed through a significant retrospective clinical group named the JAK2 Prediction Cohort or JAKPOT. The study took place at the London Health Sciences Centre in Southwestern Ontario, Canada, and included adult patients referred from January 1, 2015, to May 12, 2021, because of increased hemoglobin levels who then underwent JAK2 mutation testing. The molecular tests employed methods such as qPCR, SNP allelotyping, and next-generation sequencing. Data about age, gender, and CBC parameters at the mutation testing time were collected and examined. The participants were randomly divided into 2 groups: derivation (n = 616) and validation (n = 285). A binary score, grounded on critical variables, was devised and then assessed and internally validated through logistic regression and nonparametric bootstrapping. From the 901 participants, JAK2 mutation positivity was 11.8% in total, 12.2% in the derivation group, and 10.9% in the validation group. The final model classified high-risk patients based on specific criteria: RBC count exceeding 6.45 × 1012/L, platelets surpassing 350 × 109/L, or neutrophils above 6.2 × 109/L. Those who didn’t fit these standards were labeled as low-risk. The proposed model boasted a 94.7% sensitivity and a 98.8% negative predictive value in the derivation set. For the validation set, both these metrics stood at an impressive 100%. The study observed a remarkably low false-negative rate of 0.4%. By leveraging this predictive approach, there could be a more directed JAK2 mutation testing, potentially slashing the number of tests by over half. Thus, Chin-Yee and his colleagues effectively demonstrated an optimized method for JAK2 mutation testing in patients with high hemoglobin, promoting both efficient testing and resource conservation.

In this study, comparison of the patients with and without the JAK2 mutation revealed an unexpected pattern: those with the mutation had lower hemoglobin (Hb) levels, with a marginal P value of .046. The distribution of Hb levels among the mutation-positive group was nonparametric, while it appeared parametric in females and particularly nonparametric in males. Notably, males predominated the mutation-positive group. This gender disparity suggests that the nonparametric distribution of Hb levels in males may account for the observed differences. The significance of these findings was not replicated with parametric tests, indicating a potential influence of data distribution on the results.

These 2 previous studies stand out due to their rational methodologies and high efficiency values. Our study, taking a similar rational approach, aimed to provide a different perspective on this problem by employing more efficient artificial intelligence methods and successfully achieved this with high efficiency. While we also do not have any information regarding clinical features like in previous studies, the inclusion of different features like monocyte count, RDW and PDW, which would better depict the panmyelosis pattern and cellular morphological changes, has allowed us to overcome the limitations of the conventional one-dimensional approach. Reviewing the results from the predictive models (GB, RF, PVA, and JAKPOT), an important pattern becomes clear that deserves careful scientific consideration, especially when looking at how to improve diagnostic approaches. The sensitivity and negative predictive value of 100% across all models signify a capability to rightly identify all actual positive cases and accurately predict a negative result when it is indeed the case, thereby alleviating false negatives. This attribute emerges as paramount, especially in scenarios necessitating the exclusion of a particular condition or ensuring that no actual positive cases are dismissed prematurely. However, when the objective swivels towards minimizing testing, GB and RF provide excellent positive predictive value, accuracy, and AUROC metrics, underlining their prowess in not only correctly identifying positive cases but also in their overall classification and decision-making competency. The substantial AUROC and accuracy, in particular, highlight the models’ adeptness in balancing sensitivity and specificity, as well as their overall classification accuracy, respectively, which are pivotal in reducing unnecessary testing by ensuring that individuals are classified accurately in the first instance. Thus, while all 4 models find positive cases, GB and RF achieve a reduction in testing, affirming that machine learning models, when adeptly developed and validated, can substantially contribute to streamlining diagnostic pathways and optimizing resource allocation within a clinical context.

Machine learning methods, like ours, are inherently more complex and sophisticated than traditional statistical approaches. The underlying mathematical principles of these algorithms have been long established. The recent advent of sufficient computational power to process extensive datasets and execute complex calculations has enabled the practical application of these theories, marking a significant milestone in the AI revolution. The methodology ensures that each feature is weighted to minimize error rates and optimize overall model performance, contributing to the high accuracy rates observed. Our models outperform others, such as the PVA and JAKPOT, due to their advanced ability to interpret and analyze complex data structures and relationships.

Our AI algorithm, requiring a digital platform due to its mathematical complexity, can be used as online scoring tools accessible via web interfaces post external dataset validation and also integrated directly into laboratory systems, where it can automatically provide scores and recommendations alongside patient lab results, offering real-time, evidence-based guidance without additional effort from healthcare providers. By considering a broader range of factors, our AI model not only surpasses simpler variables in comprehensive analysis but also promises straightforward usability.

The study presents notable strengths and limitations. A significant strength lies in employing a broad patient series, enhancing the robustness of the data. Furthermore, the implementation of artificial intelligence methods enabled the development of an efficacious decision support tool, advancing the analytical capabilities of the study. Additionally, the inclusion of not only conventional variables such as erythrocyte, neutrophil, and platelet counts but also additional variables regarding the panmyelosis picture like monocyte, RDW, and PDW enhanced the rationality of features used. However, the retrospective nature of the study poses inherent limitations on establishing causality. Moreover, the absence of information on patients’ clinical features and EPO levels, along with the omission of an examination of Exon 12, constrains the depth of clinical and genetic insights derivable from the data. A methodological limitation emerged from examining all tests sent by a hematologist to the genetic laboratory, resulting in an imbalanced dataset due to a low positivity rate in these samples. Although we used effective artificial intelligence techniques to overcome this imbalance, still, leveraging larger and more balanced training sets, potentially derived from larger databases, may serve to ameliorate this limitation in future studies.

5. Conclusion

All models show a strong ability to correctly identify and rule out the condition. However, when the goal is to reduce testing and save resources, GB and RF offer a substantial superiority with accurately identifying and classifying conditions with fewer tests. These findings may guide future research and application in varied clinical settings, ensuring effective and resource-efficient diagnostic processes.

Acknowledgements

All authors also agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Author contributions

Conceptualization: Fatos Dilan Koseoglu.

Data curation: Fatos Dilan Koseoglu, Erdem Ugur Alici, Berk Ozyilmaz, Taha Resid Ozdemir.

Formal analysis: Fatos Dilan Koseoglu, Fatma Keklik Karadag, Erdem Ugur Alici.

Investigation: Fatos Dilan Koseoglu, Hale Bulbul, Taha Resid Ozdemir.

Methodology: Fatos Dilan Koseoglu, Erdem Ugur Alici, Berk Ozyilmaz.

Project administration: Fatos Dilan Koseoglu, Berk Ozyilmaz.

Resources: Fatos Dilan Koseoglu, Fatma Keklik Karadag, Hale Bulbul.

Software: Fatos Dilan Koseoglu, Erdem Ugur Alici.

Supervision: Fatos Dilan Koseoglu.

Validation: Fatos Dilan Koseoglu, Erdem Ugur Alici.

Visualization: Fatos Dilan Koseoglu.

Writing – original draft: Fatos Dilan Koseoglu, Fatma Keklik Karadag, Hale Bulbul, Erdem Ugur Alici, Berk Ozyilmaz, Taha Resid Ozdemir.

Writing – review & editing: Fatos Dilan Koseoglu, Fatma Keklik Karadag, Hale Bulbul, Erdem Ugur Alici, Berk Ozyilmaz, Taha Resid Ozdemir.