A potential predictive model based on machine learning and CPD parameters in elderly patients with aplastic anemia and myelodysplastic neoplasms

Abstract

Background

Aplastic anemia (AA) and myelodysplastic neoplasms (MDS) have similar peripheral blood manifestations and are clinically characterized by reduced hematological triad. It is challenging to distinguish and diagnose these two diseases. Hence, utilizing machine learning methods, we employed and validated an algorithm that used cell population data (CPD) parameters to diagnose AA and MDS.

Methods

In this study, CPD parameters were obtained from the Beckman Coulter DxH800 analyzer for 160 individuals diagnosed with AA or MDS through a comprehensive retrospective analysis. The individuals were unselectively assigned to a training cohort (77%) and a testing cohort (23%). Additionally, an external validation cohort consisting of eighty-six elderly patients with AA and MDS from two additional centers was established. The discriminative parameters were carefully analyzed through univariate analysis, and the most predictive variables were selected using least absolute shrinkage and selection operator (LASSO) regression. Six machine learning algorithms were utilized to compare the performance of forecasting AA and MDS patients. The area under the curves (AUCs), calibration curves, decision curves analysis (DCA), and shapley additive explanations (SHAP) plots were employed to interpret and assess the model’s predictive accuracy, clinical utility, and stability.

Results

After the comparative evaluation of various models, the logistic regression model emerged as the most suitable machine learning model for predicting the probability of AA and MDS, which utilized five principal variables (age, MNVLY, SDVLY, MNLALSEGC, and MNCEGC) to accurately estimate the risk of these diseases. The best model delivered an AUC of 0.791 in the testing cohort and had a high specificity (0.850) and positive predictive value (0.818). Furthermore, the calibration curve indicated excellent agreement between actual and predicted probabilities. The DCA curve further supported the clinical utility of our model and offered significant clinical advantages in guiding treatment decisions. Moreover, the model’s performance was consistent in an external validation group, with an AUC of 0.719.

Conclusions

We developed a novel model that effectively distinguished elderly patients with AA and MDS, which had the potential to provide physicians assistance in early diagnosis and the proper treatment for the elderly.

Introduction

Aplastic anemia (AA) is a bone marrow hematopoietic failure disorder, mainly manifested by low bone marrow hematopoiesis, decreased whole blood cells, and anemia [1, 2]. Myelodysplastic neoplasms (MDS) are highly heterogeneous myeloid neoplasms manifested by chronic cytopenias, ineffective and dysplastic hematopoiesis. This leads to a decrease in blood cell counts and morphological dysplasia in one or more blood cells. MDS are called the early stages of leukemia, almost 3 out of 10 patients with MDS progress to acute myeloid leukemia (AML) [3]. Routine blood tests are the most common of the clinical laboratory tests. Although changes in blood cell counts are associated with a number of clinical conditions, an abnormal routine blood test may indicate the presence of a hematologic malignancy [4]. These two bone marrow failure disorders have similar peripheral blood manifestations and are clinically characterized by a reduction in the hematological triad, they have different etiologies and involve a variety of clinical and molecular alterations [5,6,7]. Therefore, how to narrow their boundaries or eventually redefine them altogether poses a major challenge to researchers, and treatment and prognosis of both disorders are greatly influenced by an appropriate differential diagnosis.

The diagnostic process of each disease generates a large amount of laboratory data, making it possible to conduct comprehensive data mining to effectively analyze the diseases. By harnessing the power of big data analysis, we can gain insights into the characteristics, patterns, and trends of these diseases. Machine learning (ML) involves models or algorithms that can allow computers to learn from data and identify individual features of data [8]. As medical data volumes grow, ML can assist doctors to make more accurate diagnoses, predict patient outcomes, and personalize treatment plans. It automates tedious tasks, reducing human labor and enabling doctors to focus on providing better patient care [9]. In hematology, machine learning has been used to improve risk stratification, categorical diagnosis, and prognosis of diseases, as well as mortality prediction and treatment of tumors [10].

For the past few years, with the rapid development of technology, blood analyzers have provided more information in addition to the usual blood image parameters, translating cell morphology and characteristic changes into reportable cell count results and derived study parameters [11]. Hematology analyzer data has been used to predict a series of clinical outcomes, from blood culture results [12], sepsis patients [13, 14], and COVID-19 patients [15]. The application of this novel technology allows reporting of new parameters as well as basic complete blood counts. Leukocytes (neutrophils, monocytes, eosinophils, and lymphocytes) can be classified according to their morphological and functional characteristics using cell population data (CPD). Since CPD is derived from the results obtained from routine blood analysis, it has the advantages of being rapid, economical, and reliable, without the need for additional reagents or processes, and possesses good prospects for clinical application.

In the diseases of the blood system, CPD parameters have shown a great role in various diseases that can cause morphological changes in leukocytes. CPD parameters have shown applicability in the diagnosis and differentiation of various hematological diseases, including multiple myeloma [16], neoplastic hematological diseases [17], thalassemia traits [18], chronic myeloid leukemia [19], with a particular emphasis on its significance in the diagnosis of myelodysplastic syndromes [20,21,22]. Given the challenges associated with the screening and diagnosis of MDS, the application of the CPD parameter is particularly crucial.

Currently, MDS and AA are mainly diagnosed by hematology, cytomorphology, bone marrow examination, and cytogenetics [2, 23]. However, bone marrow examination is an invasive technique that tests the doctor’s skills, and the bone marrow cells may not be observed during the bone marrow biopsy due to inappropriate bone marrow extraction, which brings more pain to the patients. The sensitivity of single immunophenotypic indexes for the differential diagnosis of MDS and AA is too low, which restricts the wide application of FCM in diagnosing myelodysplastic neoplasms, and genetic testing is more costly [24,25,26]. These facts prompted us to look for routine laboratory tests based on which to diagnose AA and MDS. Apart from Wu et al. [27], few studies have reported models based on machine learning to distinguish AA and MDS. However, their model utilized numerous indicators including blood cell count, blood smear, and marrow smear, which may not be practical in clinical settings. Additionally, the model lacked an external validation set and primarily targeted the general population rather than the elderly.

In the current study, we built a model based on CPD parameters and machine learning to distinguish between AA and MDS in elderly patients. Detection of these parameters in these diseases may contribute to early diagnosis and rapid intervention of the disease, which contributed to improving elderly patients’ prognosis. In addition, with further research and optimization, the model was expected to become a powerful tool in clinical practice, and could also provide a reference for other medical-related research.

Methods

Patient involvement

According to the guidelines for the diagnosis and management of adult aplastic anemia [7, 28] as well as the clinicians' diagnoses, we collected 252 patients (age ≥ 50 years) diagnosed with AA and MDS from May 16, 2022 to August 28, 2023, in Zhejiang Provincial Hospital of Chinese Medicine (Hubin). According to the exclusion criteria, 92 elderly patients were excluded, as shown in Fig. 1. Finally, included in the study were a combined total of 89 cases classified as AA and 71 cases classified as MDS. Furthermore, 86 cases from Zhejiang Provincial Hospital of Chinese Medicine (Qiantang and Xixi) were collected as an external validation cohort from May 16, 2022 to August 28, 2023.

Fig. 1

The flow chart demonstrated the participants encompassed within in the study

Diagnostic criteria for aplastic anemia: 1. Blood routine examination: Total blood cells (including reticulocytes) uniformly depressed, and the proportion of lymphocytes increased. Meet at least two of the following three criteria items: HGB < 100 g/L; PLT < 50 × 10⁹/L; Neutrophil rejection Opposition value (ANC) < 1.5 × 10⁹/L. 2. Bone marrow aspiration: Erythropoiesis was reduced or absent. Megakaryocytes and granulocytic cells were markedly reduced or absent. The proportion of non-hematopoietic cells (lymphocytes, reticular cells, plasma cells, mast cells, etc.) increased. 3. Bone marrow biopsy (ilium): The biopsy specimen was hypocellular throughout, with reduced hematopoietic tissue, increased non-hematopoietic cells, no increase in reticulin, and no abnormal cells. 4. Congenital and other acquired and secondary BMF were excluded. The diagnosis of myelodysplastic neoplasms was according to the 5th edition of the World Health Organization Classification of Haematolymphoid Tumours: Myeloid and Histiocytic/Dendritic Neoplasms [29]. Exclusion criteria: Patients with other blood diseases or who had undergone bone marrow transplantation or hematopoietic stem cell therapy, or incomplete data were excluded.

Data collection

Detailed information about these patients’ baseline population characteristics (age, gender, and comorbidities) and CPD parameters were carefully collected from their electronic medical records. Following the completion of the enrollment process, a total of 160 elderly individuals were subjected to random assignment, with 77% being allocated to the training cohort and the remaining 23% to the testing cohort. By establishing a random seed, the current investigation could guarantee the replicability of the stochastic procedure, enabling precise replication of research findings as necessary. The hyperparameters of the best model were chosen using grid search and cross-validated ten times. In ten-fold cross-validation, the dataset was divided into ten equal-sized sections. One of the ten sections was used for testing and the remaining nine sections were used for training. Ten-fold cross-validation was looped ten times throughout the process. Utilizing CPD parameters, six machine learning models were built up in the training cohort and subsequently validated in the testing cohort. The final filtered optimal model was validated using an external validation cohort.

Statistical analysis

All analyses were done through SPSS 26.0 and the platform. Frequencies and percentages were used to indicate qualitative variables, while mean ± standard deviation or median and interquartile range (IQR) were used to indicate quantitative variables. Count data were analyzed using the chi-square test, and measurement data were analyzed using the independent samples t-test or Wilcoxon test.

Through univariate analysis, indicators with significant disparities between AA and MDS groups were screened, and we further utilized the least absolute shrinkage and selection operator (LASSO) regression to pick out the factors that were more relevant to AA and MDS. Via random seeds, 77% of patients were allocated to form the training cohort, whereas the remaining 23% of patients were allocated to the testing cohort. Calibration and Decision plots were utilized to visually evaluate the model, while the area under curve (AUC) was employed for assessing calibration. The intricate feature ranking was interpreted via shapley additive explanations (SHAP) plots. P < 0.05 is considered statistical significance.

Machine learning

The building and training of machine learning models were accomplished through the platform Deepwise & Beckman Coulter DxAI, and the reason for choosing these models was that these are the more common machine learning models. This is a mature platform, publishing a lot of high-score literature. The platform is capable of automatically selecting machine learning models and generating an analysis page online.

Results

Demographic characteristics

The demographic characteristics of elderly patients were summarized in Table 1. This study encompassed 89 (55.625%) classified as AA and 71 (44.375%) classified as MDS. There were 39 males (43.820%) and 50 females (56.180%) in the AA group, while 44 males (61.972%) and 27 females (38.028%) were in the MDS group. As shown in Table 1, the gender difference between the two groups was statistically significant (P = 0.022 < 0.05). The MDS group exhibited a significantly higher mean age compared to the AA group (P < 0.001). The median age of the MDS group was 69.000, as compared to the AA group's median age of 61.000. The most common comorbidity in AA patients was hypertension (29.213%), followed by diabetes (16.854%). Additionally, infectious fever, hypoproteinemia, tumors, and coronary heart disease were present in 12.360%, 5.618%, 5.618%, and 2.247% of AA patients, respectively. The most common comorbidity in MDS patients was hypertension (28.169%), followed by diabetes (15.493%). Additionally, tumors, infectious fever, hypoproteinemia, and coronary heart disease were present in 14.085%, 9.859%, 7.042%, and 4.225% of MDS patients, respectively.

Table 1 Baseline Characteristics of AA and MDS Patients

In the external validation cohort, this study encompassed 57 individuals (66.279%) diagnosed as AA and 29 (33.721%) diagnosed as MDS. Within the AA cohort, 24 patients (42.105%) were male, and 33 (57.895%) were female. The median age of the cohort was 60 years old. The most common comorbidity in patients was hypertension (14.035%), followed by diabetes (8.772%). Additionally, tumors, infectious fever, coronary heart disease, and hypoproteinemia were present in 7.018%, 5.263%, 1.754%, and 1.754% of AA patients, respectively. Within the MDS cohort, 15 patients (51.724%) were male, and 14 (48.276%) were female. The median age of the cohort was 65 years old. The most common comorbidity in patients was hypertension (24.138%), followed by tumors (20.690%). Additionally, diabetes, infectious fever, and hypoproteinemia were present in 13.793%, 3.448%, and 3.448% of AA patients, respectively. There were no patients with coronary heart disease in this group.

Comparison of CPD parameters between AA and MDS patients

The results were shown in Table 2, which indicated that there were significant differences in gender, age, SDVNE, MNCNE, MNMALSNE, SDMALSNE, MNUMALSNE, SDUMALSNE, MNLMALSNE, SDLMALSNE, MNLALSNE, SDAL2NE, MNVLY, SDVLY, SDCLY, SDMALSLY, SDUMALSLY, MNLMALSLY, SDLMALSLY, SDLALSLY, MNAL2LY, SDAL2LY, MNVMO, SDVMO, SDCMO, SDLMALSMO, MNLALSMO, SDAL2MO, SDAL2EGC, MNLALSEGC, SDLMALSEGC, MNLMALSEGC, MNUMALSEGC, SDMALSEGC, MNMALSEGC, SDCEGC, MNCEGC, SDVEGC, MNVEGC, MNLMALSEO, SDUMALSEO, MNUMALSEO, SDMALSEO, MNMALSEO, and MNCEO between the two groups (P < 0.05). This meant that there were some morphological changes in neutrophils, lymphocytes, monocytes, early granulated cells, and eosinophils in the blood of patients with AA and MDS.

Table 2 Comparison of CPD parameters between AA and MDS patients

Screening for optimal predictors by LASSO regression

In the current study, we collected a total of 71 indicators from elderly patients classified as AA and MDS. After excluding non-significant indicators, 45 features were retained for LASSO regression analysis to screen the optimal predictors that have correlated with two diseases. The results via LASSO regression showed that age, MNLMALSNE, MNVLY, SDVLY, SDCLY, SDVMO, MNLALSEGC, SDCEGC, and MNCEGC were considered to be relevant factors with AA and MDS (Fig. 2). Furthermore, using the nine indicators chosen via LASSO regression, the current study examined heat maps of correlations and importance rankings between these indicators.

Fig. 2

Screening the optimal predictors via LASSO regression. A Regression coefficient path plot in LASSO regression. Diverse colored lines indicate that different variables will gradually become zero, and the later they become zero, the more important the indicator. B The cross-validation curve of LASSO regression. The minimum standard is on the left line and the 1-SE standard is on the right line. In the current study, we selected 9 non-zero predictors according to the 1-SE standard. SE, the standard error

AUCs of nine indicators

In Fig. 3, the ROC and AUCs were presented, highlighting the significant differences in diverse indicators between the two groups in forecasting AA and MDS. Among these CPD parameters, MNLMALSNE was the most efficient (AUC = 0.760). SDCLY was then followed (AUC = 0.758).

Fig. 3

The ROC curves of AA and MDS were independently predicted by 9 predictors between the two groups

Feature importance and correlation heatmap of CPD parameters

Upon analyzing the importance of diverse indicators, the current study eventually chose five predictors rooted in the count of elderly individuals afflicted with AA and MDS. The feature importance between the nine filtered indicators was shown in Fig. 4A. The most valuable of these nine indicators was age. Additionally, MNVLY, SDVLY, MNLALSEGC, and MNCEGC were followed, respectively. In turn, the interrelationships among the five indicators were analyzed. It was believed that the correlation between the two indicators < 0.7 would not interfere with each other. As presented in Fig. 4B, age, MNVLY, SDVLY, MNLALSEGC, and MNCEGC exhibited a low correlation, potentially preventing insufficient generalization of the model to new data from other sources.

Fig. 4

A The weight importance of nine filtered indicators. B Heat map of correlation of top five indicators. The correlation degree is from low (blue) to high (red)

Comparative evaluation of six machine learning models

The AUCs of six machine learning models for tenfold cross-validation on the training cohort were presented in Table 3. We focused on the AUC performance of each machine learning model on the validation cohort to determine the optimal model. According to the AUCs in the testing cohort, the highest AUC among the six ML algorithms was achieved by logistic regression (AUC = 0.827). The second-highest AUC was resented by random forest (AUC = 0.787). The third-highest AUC was offered by support vector machines (SVM) (AUC = 0.780). In addition, adaptive boosting (AdaBoost) demonstrated the lowest manifestation (AUC = 0.705) and was excluded. The result showed that the logistic regression model excelled in predicting performance compared to the other five models.

Table 3 Comparative evaluation of six machine learning models for ten-fold resampling-validation

Machine learning models establishment and assessment

Drawing from the data presented in Fig. 5 and Table 4, it was evident that the logistic regression model possessed a robust discriminatory capability in distinguishing aplastic anemia and myelodysplastic neoplasms. The model exhibited an AUC of 0.791 in the testing cohort (Fig. 5B), with specificity and positive predictive value exceeding 80% (Table 4). In addition, Fig. 5C indicated excellent calibration of the model. The calibration curve exhibited a good agreement between the actual probability and the predicted probability. The DCA curve highlighted the clinical benefits of the model, indicating its strong performance in clinical settings (Fig. 5D).

Fig. 5

The performance of six machine learning models. A ROC curve of the training cohort; B ROC curve of the testing cohort; C Calibration curve; (D) Decision curve analysis

Table 4 Evaluation of the optimal logistic regression model for ten-fold cross-validation

Figure 6A depicted the correlation between SHAP values of the five most pertinent features we identified. As Fig. 6B illustrated, the logistic regression model’s interpretation of feature ranking, as determined by the SHAP algorithm, was presented. It was explained that the most powerful features for predicting outcomes of elderly patients were MNLALSEGC, MNVLY, age, SDVLY, and MNCEGC. These characteristics had the greatest impact on predicting patient outcomes and should be considered in the evaluation and treatment of elderly patients. By utilizing SHAP force plots, the study offered a visual representation of the SHAP value of a single indicator, demonstrating its impact on modifying the baseline predicted value, whether positive or negative. Figure 6C and D showed the individual force plots for MDS patients and AA patients, respectively. The figures offered a visual representation of the impact of each feature on modifying the model's predicted value for each patient group. The features that contribute positively, denoted in red, propel the model's score upward, whereas those that contribute negatively, denoted in blue, pull the model's score downward. The length of the arrow offered a visual representation of the magnitude of its impact on the prediction. As the arrow lengthens, the greater the influence on the prediction of MDS.

Fig. 6

Model explainability via the SHAP algorithm. A The horizontal SHAP value represents the influence on the prediction result, and the vertical coordinate is each indicator, the contribution degree is from low (blue) to high (red). B The importance ranking of independent variables. C The SHAP force plot of patients with myelodysplastic neoplasms. D The SHAP force plot of patients with aplastic anemia

External validation

Eighty-six elderly patients were recruited from two additional centers to serve as an external validation cohort. As shown in Fig. 7, the AUC of the logistic regression model was 0.719 when validated using the external validation cohort. This suggested that the model based on CPD parameters had high value in practical applications.

Fig. 7

ROC for the external validation cohort

Discussion

Distinguishing between the various features of aplastic anemia and myelodysplastic neoplasms is critical clinically, as it affects patient drug therapy and outcomes [6]. It has been reported that the risk of progressing to AML patients with MDS was much higher than those with AA [30, 31].

Nowadays, the use of machine learning methods to help clinicians process laboratory results can avoid the influence of empirical differences between clinicians on diagnostic results. Novel leukocyte CPD parameters are emerging as potential markers in diverse clinical settings. These parameters have been initially applied to disease identification, such as in COVID-19 [32] and sepsis [14]. Based on VCS technology, morphological analysis of leukocyte subtypes is performed, cell volume (V) is measured by DC impedance to obtain accurate cell size, and electrical conductivity (C) of internal components of each cell is characterized by radio frequency transmittance. The light scattering (S) beam of cytoplasmic particle size and nuclear structure is measured using a laser [33].

The application of CPD parameters has several advantages in the clinic. These parameters are generated during a routine complete blood count (CBC) analysis, eliminating the need for additional samples. CPD parameters are more objective and accurate than manual difference counts due to the automatic assessment of thousands of white blood cells, making them suitable as an additional marker at a lower cost than other laboratory tests [34, 35]. Therefore, we emphasized the importance of CPD parameters not only for the rapid screening of diseases but also as a simple method for its rapid performance, we could acquire a wealth of valuable data from circulating blood for the description of hematologic diseases.

The present study showed that it was possible to distinguish between AA and MDS using white blood cell population data parameters. Age, MNVLY, SDVLY, MNLALSEGC, and MNCEGC were identified to build up the model. In actual clinical practice, the analysis of a single feature was frequently inadequate to capture the entire nature of the disease. Consequently, our model considered the above five indicators as a whole rather than making diagnostic predictions based on individual features in order to distinguish between AA and MDS. The model, as presented in Fig. 5, demonstrated high discrimination and calibration, indicating a strong performance and higher clinical utility. Furthermore, the model performed effectively in both the testing cohort (AUC = 0.791) and the external validation cohort (AUC = 0.719). These results indicated that the model had significant value in accurately and stably classifying the probability of AA and MDS occurring in elderly patients on an individual basis. Our model had the potential to offer doctors a user-friendly and highly effective tool for discriminating between AA and MDS in clinical practice.

Recently, several machine learning algorithms for predicting MDS have been developed. Park et al. [36] have created a model by using cell population data, the model had an AUC of 0.891. Pozdnyakova et al. [21] have created a model by using CBC parameters, the model had an AUC of 0.860. In our logistic regression model, the AUC in the testing cohort was 0.791 which was lower than their model. However, our model performs better in specificity, achieving a value of 0.850, while their specificities are 0.790 and 0.720, respectively. Additionally, it is worth noting that their models have not undergone further validation using a test cohort or external validation cohort.

Aplastic anemia (AA) affects 7.4 people per million per year, with a higher prevalence in China than in the West [28]. The disease’s occurrence also varies with age, with the highest frequency observed among individuals over the age of 60. The gender ratio is also different: AA patients over 60 years of age are predominantly female (60.000%) [37, 38], which is similar to our study (56.180%). From a biological perspective, it has been observed that elderly individuals with AA often exhibit a higher frequency of mutations that are potentially linked to adverse outcomes. Furthermore, several studies have identified age > 60 as an independent risk factor for mortality in aplastic anemia [39]. However, the prevalence of MDS is estimated to be only as high as 75 per 100,000 over the age of 65 [40]. The immune system undergoes morphological or functional changes with aging, as evidenced by a decrease in autoimmune cells and a higher prevalence of autoantibodies, with AA and MDS being increasingly diagnosed in the elderly [41, 42]. In the present study, age also shows great weight and importance in these indicators (Fig. 4A and Fig. 6B).

As the key link of immune defense, white blood cells will change their shape, internal structure, and function, and their CPD parameters can show this change in sensitivities as an immunoreactive change in the body in a pathological state. In the process of exploring the relationship between CPD parameters and disease, we observed notable disparities in the distribution profiles of CPD parameters among patients with AA and MDS (Table 2).

Since granulocyte dysplasia is a high-visibility feature of MDS, neutrophil-associated parameters have been extensively studied and are commonly utilized to discern granulomatous dysplasia [42, 43]. In addition to age, among the neutrophil-associated CPD parameters, this study observed an increased SDVNE, SDUMALSNE, SDLMALSNE, SDAL2NE, and decreased MNCNE, MNMALSNE, MNUMALSNE, MNLMALSNE, MNLALSNE in MDS patients. Among the lymphocyte-associated CPD parameters, the markedly increased variation in MNVLY, SDVLY, SDCLY, SDMALSLY, SDLMALSLY, SDLALSLY, MNAL2LY, SDAL2LY and decreased in MNLMALSLY in MDS patients. Interestingly, in the two groups, almost all lymphocyte CPD parameters changed, which was also consistent with the two lymphocyte-associated parameters (MNVLY and SDVLY) in the LASSO regression. This finding was consistent with the previous findings [43].

Almost all of the CPD parameters changed in the early granulated cells, which was also consistent with the two CPD parameters in the LASSO regression (MNLALSEGC and MNCEGC). These could be dysplastic features of myelodysplastic neoplasms.

In the MDS group, the degree of heterogeneity of SD measurements was increased, and other studies that have utilized various hematology analyzers to explore CPD in MDS patients have also reported heterogeneity in cellular characteristics [21, 36, 43]. In spite of this, it is unclear what the mechanisms behind this matter are.

However, the current study had some limitations. Firstly, the sample size was relatively small, consisting of only 160 elderly individuals with a diagnosis of AA and MDS. It might lead to biases in the model when generalized. In future research, larger studies with more diverse patient populations and data from multiple centers are needed to further validate the model and assess its performance in real-world clinical settings. Secondly, this model was only validated using Chinese patients. Future studies should include patients from diverse countries and ethnic backgrounds to confirm the generalizability of the model. Additionally, there may be some inevitable bias in clinicians' assessments of disease severity, which could introduce subjective elements. Finally, this research only focused on investigational CPD parameters between AA and MDS patients, without considering other potential biomarkers. In terms of future research directions, we aim to explore ways to optimize the model, such as incorporating new biomarkers (like reticulocytes, bone marrow blast percentage, or other routine blood parameters) or refining the existing algorithm, to improve its accuracy and reliability.

Conclusions

In conclusion, a recognition machine learning model based on CPD parameters was constructed to predict which AA and MDS the patient was. Five filtered indicators were utilized to develop the ML models. The logistic regression model excelled in predicting performance compared to the other five models (XGBoost, AdaBoost, SVM, LightGBM, and random forest). This model exhibited excellent discrimination and calibration, making it well-suited for clinical application. The model may be a powerful tool in scenarios where timely and accurate diagnosis is critical but resources are limited. This could enable early screening for cytopenic patients (AA or MDS) and guide clinical decision-making, especially in lower-level hospitals.