key: cord-0528126-zmqy08fi
authors: Ratul, Ishrak Jahan; Wani, Ummay Habiba; Nishat, Mirza Muntasir; Al-Monsur, Abdullah; Ar-Rafi, Abrar Mohammad; Faisal, Fahim; Kabir, Mohammad Ridwan
title: Survival Prediction of Children Undergoing Hematopoietic Stem Cell Transplantation Using Different Machine Learning Classifiers by Performing Chi-squared Test and Hyper-parameter Optimization: A Retrospective Analysis
date: 2022-01-22
journal: nan
DOI: nan
sha: 12b35867f8f6c264deaab080718fe114b07ff5bb
doc_id: 528126
cord_uid: zmqy08fi

Bone Marrow Transplant, a gradational rescue for a wide range of disorders emanating from the bone marrow, is an efficacious surgical treatment. Several risk factors, such as post-transplant illnesses, new malignancies, and even organ damage, can impair long-term survival. Therefore, technologies like Machine Learning are deployed for investigating the survival prediction of BMT receivers along with the influences that limit their resilience. In this study, an efficient survival classification model is presented in a comprehensive manner, incorporating the Chi-squared feature selection method to address the dimensionality problem and Hyper Parameter Optimization (HPO) to increase accuracy. A synthetic dataset is generated by imputing the missing values, transforming the data using dummy variable encoding, and compressing the dataset from 59 features to the 11 most correlated features using Chi-squared feature selection. The dataset was split into train and test sets at a ratio of 80:20, and the hyperparameters were optimized using Grid Search Cross-Validation. Several supervised ML methods were trained in this regard, like Decision Tree, Random Forest, Logistic Regression, K-Nearest Neighbors, Gradient Boosting Classifier, Ada Boost, and XG Boost. The simulations have been performed for both the default and optimized hyperparameters by using the original and reduced synthetic dataset. After ranking the features using the Chi-squared test, it was observed that the top 11 features with HPO, resulted in the same accuracy of prediction (94.73%) as the entire dataset with default parameters. Moreover, this approach requires less time and resources for predicting the survivability of children undergoing BMT. Hence, the proposed approach may aid in the development of a computer-aided diagnostic system with satisfactory accuracy and minimal computation time by utilizing medical data records.

Even in its most curable forms, cancer kills millions of people every year. According to the cancer statistics of 2020, the estimated death tolls in the USA from colon, pancreatic, lung, breast, and prostate cancers are 53200, 47050, 135720, 42690, and 33330, respectively [1] . When there is no cure, physicians endeavor to extend the lifespan of a cancer patient through surgery, radiation therapy, or chemotherapy as alternative methods of cancer treatment [2] . For various reasons, the high dose of medication during chemotherapy or radiation therapy causes bone marrow damage in patients [2] . Bone Marrow (BM), the delicate, elastic, adipose tissue located inside most skeleton structures, is responsible for creating the red blood cells of human blood [3] [4] . It also contains Hematopoietic Stem Cells (HSC) that are merely immature blood-forming stem cells that are endowed with idiosyncratic properties like self-renewal, and they form populations of progenitor cells through cell division and differentiation [4] [5] [6] . However, the concept of BMT, otherwise known as Hematopoietic Stem Cell Transplant (HSCT), gleans from the postulation of eliminating dysfunctional body parts and replacing them with healthy ones [7] . Although it is a lifesaving treatment, it has potential life-threatening risks [8] . Clinical HSCT commenced in 1957, at a time when the health domain was inadequately fathomed about HSCs, immunological reactions to transplants, and even about the specification of the antigens steering the course of action [9] . HSCT is not a surgery, but rather a specialized treatment for people afflicted by specific cancers or certain medical conditions [10] [11] . The target of such a therapy is transfusing functional BM into a patient, subsequent to their own diseased BM being medicated for exterminating the aberrant cells [11] . The three prime objectives of HSCT are -(a) replacement of deceased stem cells affected by chemotherapy, (b) replacement of diseased marrow that is impotent to synthesize its endemic progenitor cells, and (c)

infusion of allografts to assist in locating and destroying malignant cells [12] .

Healthy BM can be either extracted from the patient (autologous transplant) or conferred on by a volunteer donor (allogeneic transplant). In the case of autologous transplant, stem cells come from any other healthy organ of the patient [3] .

And for an allogeneic transplant, a donor with closely matched Human Leukocyte Antigens (HLAs) is needed [10] . Most of Ishrak et. al. '22 2 arXiv preprint the times siblings having the same parents make for the closest matches, although other close relatives or perhaps an unrelated patron can also be a successful match. There are two ways of collecting donor stem cells for transplant -(a) BM collection, and (b) leukapheresis [13] . When a patient receives highly matched proteins from a donor, the odds of developing a severe adverse reaction, known as Graft-Versus-Host Disease (GVHD), are minimized [3] . Given that a donor cannot be found, cord blood transplants (stem cells collected from the umbilical cord), parent-child, and HLA haplotype mismatched transplants (stem cells collected from a parent, child, or sibling) can be performed [3] . HSCT is broadly adopted for hematopoietic system-acquired and congenital illnesses. According to the Health Resources & Services Administration, almost 23 million people have registered on the donor registry. Besides, the donor registry currently contains approximately 305,000 units of cord blood. The National Cord Blood Inventory (NCBI) provides around 112,000 units, which is reflected in this number, with an additional 4,000 units projected to be available in 2020. The Center for International Blood and Marrow Transplant Research (CIBMTR) registered a total of 9,267 related and unrelated BMTs conducted in the United States in 2018 [14] . According to a survey, undertaken by UPMC Children's Hospital of Pittsburgh, the percentages of patients who survived 100 or more days well after the transplant procedure, the percentage of patients who died of causes other than the underlying disease, and the percentage of patients who survived one or more years after the transplant procedure are 100%, 3%, and 94%, respectively [15] .

To summarize, BMT is a treatment that saves and risks life at the same time. Hence, a lot of data collection and generation are required before the therapy. Classification techniques based on ML can be beneficial for disease prediction in a variety of healthcare situations. It has significant predictive ability for this type of problem and has been extensively used in recent years in a variety of sophisticated healthcare systems [16] [17] [18] [19] [20] . Moreover, it has lately been shown to be incredibly effective in the healthcare arena [21] [22] [23] [24] [25] . In this study, the survival prediction of children who received BMT was thoroughly investigated using seven supervised ML classifiers, such as: Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), K-Nearest Neighbors (KNN), Gradient Boosting Classifier (GBC), Ada Boost (AdB), XG Boost (XGB), and a dataset obtained from the UCI ML repository [26] . The Chi-squared feature ranking technique was deployed after preprocessing the dataset to discover the important factors of survivability [27] . The entire study consists of four experiments, such as -(A) with a full set of features and default hyper-parameters, (B) with a full set of features and Hyper Parameter Optimization (HPO), (C) with a reduced feature dataset (based on the Chi-squared Test) and default hyper-parameters, and (D) with a reduced feature dataset (based on the Chi-squared Test) and Hyper Parameter Optimization (HPO) followed by a rigorous quantitative and qualitative analysis. An overall workflow diagram of this work is depicted in Figure 1 .

The contributions of this study may be summarized as follows -(1) development of a suitable predictive model from raw data, (2) determination of critical factors influencing post-BMT survival, and (3) improvement of the prediction accuracy by reducing dimensionality problems. To the best of our knowledge, this dataset has never been exploited and analyzed in this way before, and it may appear as a significant contribution in helping the healthcare industry to develop a more trustworthy e-healthcare system and create a new horizon in the medical sector.

The following sections of the work include an extensive literature review followed by methodology, where dataset description, chi-squared test, hyperparameter optimization, and overall workflow diagram are depicted comprehensively.

Hence, the simulation results are portrayed for four experimental setups, and corresponding discussions are presented in an elaborate fashion. Lastly, the conclusive remarks are carried out. 

In recent times, ML techniques have been extensively exploited for diagnosis, prognosis, and therapeutics in the healthcare sector. Its applications are not limited to treatment procedures; rather, they are expeditiously gaining traction in a variety of research fields. In a narrative review, Nathan et al. highlighted essential ML concepts for novice readers, discussed the applicability of ML in hematology-related malignancies, and indicated key points for practitioners to consider before evaluating ML studies [28] . Vibhuti et al. also conducted a comparative evaluation of ML methods utilized in the discipline of HSCT, examining the categories of dataflows incorporated, designated ML algorithms used, and therapeutic consequences monitored [29] . On the other hand, patients with Acute Leukemia (AL) undergoing HSCT from unrelated donors exhibit a plethora of variations, even after rigorous genetic matching. To address this, Ljubomir et al. sought to develop an algorithm to predict the five-year survival of AL patients post-allogeneic transplant [30] . Similarly, Brent et al.

trained a Bayesian ML model to predict acute GVHD, including mortality by day 180 [31] . However, with better donor data collection, it is possible to generate a more precise approximation of individual donor availability, as estimating group averages to the distinct donor is an untrustworthy proposition. As a solution to this problem, Adarsh et al. suggested an MLbased technique for estimating the availability of each listed donor and validation of forecasting accuracy [32] . Additionally, Ying Li et al. focused on creating and verifying an ML technique for estimating donor availability, implementing and comparing three ML algorithms [33] . As a result of organized registry establishments and biological data incorporation, data procured from HSCT institutions is becoming highly proliferated and labyrinthine. Consequently, conventional statistical methods are confirmed to be obsolescent. In its provision, Shouval [34] . Similarly, Jan-Niklas et al. explored current ML breakthroughs in the Acute Myeloid Leukemia (AML) diagnosis as a prototype condition encompassing hematologic neoplasms [35] . Further, the goal of the research by Liyan et al. was to shape an ALL (Acute Lymphocytic Leukemia)-relapse detection scheme relying on ML methods [36] . In addition, using Alternating Data Tree (ADTree), Kyoko et al. endeavored to design a model for predicting leukemia recidivism within a year following transplantation [37] . For contemplative and prospective analysis, ADTree was also employed by Yasuyuki et al. to scan databases containing information about adult patients with HSCT in Japan [38] . Daniela et al. also examined the organic phenomena associated with self-regeneration and augmentation of Hormone-Sensitive Prostate Cancer (known as CD34+ cells) in stable conditions and subsequent transplantation [39] . Moreover, a DM analysis involving 28,236 registered adult HSCT receiving patients from the European Group for Blood and Marrow Transplantation's AL registry was done by Shouval et al. to predict 100-day overall and non-relapse mortality, free of leukemia, and 2-year overall survival. The ADTree algorithm was employed to create models using 70% of the data set, and the remaining 30% of the data was utilized to validate them [40] . Moreover, in [41] , a total of 25,923 adult AL cases were studied from the EBMT registry, using an insilico approach.

In this study, an ML stratagem was adopted for eliciting a prediction of the survival rate of patients who had BMT or HSCT. All the previous works augmented the prediction study and related investigation through distinctive strategies; however, all have limitations that need to be overcome. The sole purpose of this research is to investigate whether HPO along with a reduced feature set can provide a reliable outcome using an investigative ML approach and to distinguish the most impactful factors on the children's survival who have received BMTs.

The dataset used in this study, was retrieved from the ML repository at the University of California, Irvine, and the version utilized in this study was extracted from [42] . It covers medical information for children who have been diagnosed with a variety of hematologic diseases and who underwent unmodified allogeneic unrelated donor HSCT. Hence, this dataset comprises 187 occurrences and 37 attributes that contain information about individuals who have been diagnosed with a range of hematologic malignant or benign diseases [26] . Most of the properties contain categorical data, while others contain Boolean and numerical values. The dataset's attributes are listed in Table 1 . Following data extraction, it was subjected to exploratory data analysis using Jupyter Notebook and Python to find out the dataset's properties. The dataset has many categorical attributes and missing values. The statistical data for the dataset's numerical variables are summarized in Table   2 . 

As a type of statistical procedure, Chi-squared tests are used to determine the level of independence between categorical variables. It is also a widely used non-parametric method for parametric and normal distribution testing of nominal data [43] .

This technique is intended for feature tests that are independent of one another. This produces the Chi-squared score, which is used to identify the most highly correlated feature for ML models to predict desired outcomes [44] . The Chi-squared score indicates the degree to which the attributes of a dataset are related. An attribute with a low score, indicates that it has a very low predictive ability for the dataset's desired outcome column. Therefore, by utilizing this information, the most critical features may be identified, and more efficient models may be deployed on large datasets. The Chi-squared statistical test formula can be written as follows -

The parameters that define the architecture of ML models are known as hyper-parameters. Hence, the optimization of hyperparameters has a substantial impact on the formation of ideal models for certain tasks. While training the model, hyperparameters are optimized using validation data from a dataset. Typically, Grid Search Cross Validation and Random Search Cross Validation are two HPO processes that work well for a variety of ML tasks [45] . HPO is critical for determining the optimal performance of any ML model because it establishes the model's core architecture [46] . Moreover, the importance of HPO was discovered by several researchers, and is now widely employed in ML-based prediction [47] . The grid search algorithm evaluates all possible combinations from a given set of hyperparameters, whereas the random search algorithm just attempts some random possible combinations [48] . As a result, even though it takes a bit longer than a random search, the grid search technique yields better results when tuning the hyperparameters of any ML algorithm. Hence, the grid search technique is employed in this study to fine-tune the hyperparameters and achieve better results. Figure 

Following early data analysis, the dataset underwent multiple preprocessing stages before being used in the machine learning models. First, the dataset underwent multiple preprocessing stages before being used in ML models. The missing values of the dataset were filled with mean values for numerical ones and most frequent values for categorical ones. Since categorical data cannot be handled by ML models, the categorical variables were encoded into numerical form. The dummy variable encoding technique was employed for this purpose, and the attributes were turned into boolean attributes that could readily fit into any ML model [49] . Second, the attributes were then normalized using the standard scaling method to avoid bias from the ML models [50] , leaving the dataset with 59 columns after preprocessing. To discover the correlation between attributes, the correlation heatmap is generated using the processed dataset, as depicted in Figure 3 . XG Boost (XGB), were fed and trained on this dataset, and performance metrics were obtained. Moreover, the Chi-squared statistical test is used to determine the most important features, and the test score is represented in Table 3 . Once the chisquared score is calculated, a minimum number of features are determined that can still predict survival reliably, using fewer electronic health records and computational resources. As a result, the top 11 features were chosen empirically from Table   3 and were analyzed for the prediction of the models. The correlation heatmap using these 11 features is shown in Figure   4 .

As mentioned earlier, a total of four distinct experiments, A, B, C, and D, were carried out in this study. 4 RESULTS:

After preprocessing the data, which includes filling in missing values, encoding categorical variables, and normalization, the Chi-squared statistical test is used to determine the attributes' independence. The summary of the test in this preprocessed dataset is shown in table 3. The top attribute in this list is "PLT recovery", followed by "ANC recovery", "time_to_acute_GvHD_III_IV", "survival_time," and so on. 

This experiment was conducted using the processed full-feature dataset with no optimization of hyperparameters. The dataset for this experiment has 58 attributes and 1 objective attribute. Figure 3 shows the correlation heat map for the whole feature dataset. The performance of the ML classifiers was evaluated with default hyper-parameters. The confusion matrices are listed from Table 4 to Table 10 , and the Receiver Operating Characteristics (ROC) curve is shown in Figure 5 . The ROC curve can be used to discover the ideal ML model for a given task, thus removing suboptimal models . Then table 11 

In this experiment, the full set of features of the dataset was utilized and the hyper-parameter optimization was also performed using Grid Search Cross-Validation (GSCV). The training dataset was cross-validated 10-folds using GSCV to determine the optimal hyper-parameters and using which all-other performance metrics were assessed. Seven ML algorithms, mentioned earlier, are investigated in this section, and performance metrics are computed. The tables from 12 to 18 provide the confusion matrices for this experiment, and Figure 6 presents the ROC curve. 

This experiment was conducted based on the results of the Chi-squared test. Following data preprocessing, Chi-squared feature ranking is performed, and the resulting scores are summarized in Table 3 . According to that table, the first eleven features are considered in this experiment since they have a higher Chi-squared score. Hence, ML models were initially trained on the selected training set and subsequently verified on the test set without any HPO. For this experiment, the confusion matrices are shown from Table 20 to Table 26 . However, the ROC curve is presented in Figure 7 . The performance metrics are reported in Table 27 , and it is apparent that the KNN surpasses the rest of the classifiers regarding accuracy, F1score, recall, and the ROC_AUC value. However, when it comes to precision, RF, LR, KNN, GNB, and XGB all do well. 

Like the previous one, experiment D makes use of a reduced feature dataset based on the Chi-squared test, and HPO is performed using GSCV with 10-fold cross validation, and the optimal hyper-parameters are determined and used to evaluate performance metrics. The reduced dataset comprises 11 attributes, and the performances are evaluated based on them. As before, multiple ML methods such as DT, RF, LR, KNN, GBC, AdB, and XGB are deployed. From Table 28 to Table 34, the confusion matrices are shown, and Figure 8 illustrates the ROC curve for these tests. Table 35 illustrates that the DT outperforms all other algorithms in every performance metric. However, in terms of precision, the DT and KNN perform the best altogether. 

As can be seen, the overall study included four experiments with four different approaches. The whole feature dataset was employed in experiment A without HPO, and the maximum accuracy was found to be 94.73%, as were the precision Based on the above four experiments, it is evident that HPO is critical to enhancing the performance of the ML algorithms, and the Chi-squared test plays a significant role in determining the most important feature. The computation time of GSCV using complete and reduced feature datasets is shown in Table 36 and is displayed using a bar plot in Figure 13 . Most of the ML classifiers required less time in the reduced feature dataset without significantly affecting performance, which is an encouraging result of our study. The comparison between the four experiments is shown in Figure 14 in terms of accuracy, precision, recall, F1, and ROC_AUC.

The top five critical attributes established in this study are: "PLT recovery", "ANC recovery", "duration to acute GvHD III IV", "survival time", and "recipient body mass". As can be seen, we found maximum accuracy (0.9473), precision (1), recall

(1), F1 score (0.9545), and AUC (0.9523). 

A Bone Marrow Transplant is a crucial life-saving treatment for a certain type of malignancy. For this reason, early detection of survivability after BMT can play a vital role in the patient's treatment process. Moreover, if healthcare providers have a prior prediction, they can make more informed decisions about treatment options. In this regard, technologies like ML can be useful, since they can be used in situations requiring prediction and can uncover hidden patterns in previous data in order to create an accurate prediction. Nowadays, it is increasingly being employed in every situation that requires prediction. In this study, we developed a Chi-squared feature selection method and an HPO based efficient model for predicting the survival of children who received BMT and identified the most significant parameter for survival after BMT. All four experiments that were conducted yielded satisfactory predictions. The models operate well on a synthetic dataset that has been constructed from the raw dataset via a series of preprocessing phases that reduce the dataset's dimensionality. With the entire feature synthetic dataset, experiment A achieves an accuracy of 94.73%. However, as experiment B optimizes the hyper-parameters using the same dataset as experiment A, it achieves the highest overall performance of all the models.

On the other hand, experiments C and D make use of the 11 most correlated feature dataset based on the Chi-squared test, and experiment D outperforms all performance measures when combined with HPO, achieving high accuracy (94.73%) with less time, data, and resource consumption. In this study, we obtained maximum accuracy (0.9473), precision (1), recall (1), Ishrak et. al. '22 22 arXiv preprint F1 (0.9545), AUC (0.9523), and the top five attributes that influence the survivability rate are "PLT recovery," "ANC recovery,"

"duration to acute GvHD III IV," "survival time," and "recipient body mass." Historically, this dataset has not been evaluated in such a manner before, and it could provide the health sector with a unique perspective. Therefore, this study can make a noteworthy contribution to the development of ML-based healthcare prediction systems in environments where resources are scarce and healthcare practitioners lack more data. Additional research on this topic can be conducted by incorporating deep learning and neural networks. From Bangladesh's perspective, new data can also be gathered from medical records.

Moreover, programs or user interfaces can be developed to make this more useful to healthcare professionals.

The authors do not declare any potential conflict of interest that may alter the outcomes of this study in any manner and approve this version of the manuscript for publication.

The authors did not receive any funding for this study.

Understanding Cancer: Metastasis, Stages of Cancer, and More

bone-marrowstem-cell-transplantation/what-bone-marrow-transplant-stem-cell-transplant

Bone Marrow Transplant: Types, Procedure & Risks

Early decisions in lymphoid development

Regulation of hematopoietic stem cell self-renewal

A History of Bone Marrow Transplantation

Big gains in bone marrow transplant survival since mid-2000s

Bone marrow transplantation 1957-2019

What is a Bone Marrow Transplant? | Be The Match

Bone Marrow Transplantation | Johns Hopkins Medicine

Bone Marrow Transplant: Preparation, Procedure, Risks, and Recovery

Bone marrow transplant: MedlinePlus Medical Encyclopedia

Donation and Transplantation Statistics | Blood Stem Cell

Blood and Marrow Transplant Statistics and Outcomes

Comparing different supervised machine learning algorithms for disease prediction

Prediction of Heart Disease Using a Combination of Machine Learning and Deep Learning

Efficient prediction of cardiovascular disease using machine learning algorithms with relief and lasso feature selection techniques

Machine Learning in Rheumatic Diseases

Breast cancer prediction using an optimal machine learning technique for next generation sequences

Coronavirus disease (COVID-19) cases analysis using machinelearning applications

Artificial intelligence and machine learning-aided drug discovery in central nervous system diseases: Stateof-the-arts and future directions

Explainable machine learning can outperform Cox regression predictions and provide insights in breast cancer survival

Machine learning-based analysis of [18F]DCFPyL PET radiomics for risk stratification in primary prostate cancer

Lung Cancer Prediction Using Stochastic Diffusion Search (SDS) Based Feature Selection and Machine Learning Methods

Machine learning techniques and chi-square feature selection for cancer classification using SAGE gene expression profiles

Machine learning in haematological malignancies

A Systematic Review of Machine Learning Techniques in

Hematopoietic Stem Cell Transplantation (HSCT)

Evaluation of a Machine Learning-Based Prognostic Model for Unrelated Hematopoietic Cell Transplantation Donor Selection

Optimal Donor Selection for Hematopoietic Cell Transplantation Using Bayesian Machine Learning

Machine Learning Approach to Predicting Stem Cell Donor Availability

Predicting the Availability of Hematopoietic Stem Cell Donors Using Machine Learning

Application of machine learning algorithms for clinical predictive modeling: A data-mining approach in SCT

Application of machine learning in the management of acute myeloid leukemia: Current practice and future prospects

Machine learning applications for prediction of relapse in childhood acute lymphoblastic leukemia

Patient-based prediction algorithm of relapse after allo-HSCT for acute Leukemia and its usefulness in the decision-making process using a machine learning approach

Using a machine learning algorithm to predict acute graft-versus-host disease following allogeneic transplantation

Transplantation Induces Profound Changes in the Transcriptional Asset of Hematopoietic Stem Cells: Identification of Specific Signatures Using Machine Learning Techniques

Prediction of allogeneic hematopoietic stem-cell transplantation mortality 100 days after transplantation using a machine learning algorithm: A European group for blood and marrow transplantation acute leukemia working party retrospective data mining stud

Prediction of Hematopoietic Stem Cell Transplantation Related Mortality-Lessons Learned from the In-Silico Approach: A European Society for Blood and Marrow Transplantation Acute Leukemia Working Party Data Mining Study

β-Thalassemia Knowledge Elicitation Using Data Engineering: PCA, Pearson's Chi Square and Machine Learning

Review on feature selection techniques and its impact for effective data classification using UCI machine learning repository dataset

Grid Search, Random Search, Genetic Algorithm: A Big Comparison for NAS

On the Importance of Hyperparameter Optimization for Model-based Reinforcement Learning

Disease prediction via Bayesian hyperparameter optimization and ensemble learning

Grid search in hyperparameter optimization of machine learning models for prediction of HIV/AIDS test results

Understanding and Implementing Machine Learning Models with Dummy Variables with Low Variance

Effectiveness of Normalization Pre-Processing of Big Data to the Machine Learning Performance

RuleKit: A comprehensive suite for rule-based learning

Application of rule induction to discover survival factors of patients after bone marrow transplantation

Survival prognostic factors in patients with acute myeloid leukemia using machine learning techniques

A Clinical Decision Support Tool to Find the Best Initial Intravenous Cyclosporine Regimen in Pediatric Hematopoietic Stem Cell Transplantation

Hematopoietic stem cell transplantation in children and adolescents with GATA2-related myelodysplastic syndrome

Risk factors predicting the survival of pediatric patients with relapsed/refractory non-Hodgkin lymphoma who underwent hematopoietic stem cell transplantation: a retrospective study from the Turkish pediatric bone marrow transplantation registry

Prognostic value and prediction of Bleeding in Patients undergoing Hematopoietic Stem-Cell Transplantation: A Retrospective Study of Chinese Group for Blood and Marrow Transplantation Working Party

Choice of Diffusing Capacity Hemoglobin Correction Equation and Prediction of Mortality and Pulmonary Outcomes in Children Receiving Hematopoietic Stem Cell Transplantation

Combined immune score of lymphocyte to monocyte ratio and immunoglobulin levels predicts treatment-free survival of multiple myeloma patients after autologous stem cell transplant

Health-related quality of life predicts length of hospital stay and survival rates for pediatric patients receiving allogeneic hematopoietic cell transplantation

Diagnosis and severity criteria for sinusoidal obstruction syndrome/veno-occlusive disease in pediatric patients: A new classification from the European society for blood and marrow transplantation

Mass Cytometry for the assessment of immune reconstitution after hematopoietic stem cell transplantation