Next Article in Journal
Home Is Not Always Safe: Pediatric Unintentional Home Injuries in a Tertiary Emergency Department Setting
Previous Article in Journal
Bridging the Gap: The Role of Non-Invasive Respiratory Supports in Weaning from Invasive Mechanical Ventilation
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
JCM
Volume 14
Issue 20
10.3390/jcm14207445
jcm-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Development and Validation of Prognostic Models for Treatment Response of Patients with B-Cell Lymphoma: Standard Statistical and Machine-Learning Approaches

by
Adugnaw Zeleke Alem
1,2,*,
Itismita Mohanty
1,
Nalini Pati
1,3,4,
Cameron Wellard
5,
Eliza Chung
5,
Eliza A. Hawkes
5,6,
Zoe K. McQuilten
5,7,8,
Erica M. Wood
5,8,
Stephen Opat
8,9 and
Theophile Niyonsenga
1
1
Health Research Institute, University of Canberra, Canberra 2617, Australia
2
Department of Epidemiology and Biostatistics, Institute of Public Health, College of Medicine and Health Sciences, University of Gondar, Gondar 196, Ethiopia
3
Department of Haematology, The Canberra Hospital, Canberra 2605, Australia
4
ANU Medical School, Canberra 2605, Australia
5
School of Public Health and Preventive Medicine, Monash University, Melbourne 3004, Australia
6
Olivia Newton-John Cancer Research Institute at Austin Health, Melbourne 3084, Australia
7
Alfred Health, Melbourne 3004, Australia
8
Monash Haematology, Monash Health, Melbourne 3168, Australia
9
Department of Medicine, School of Clinical Sciences at Monash Health, Monash University, Melbourne 3004, Australia
*
Author to whom correspondence should be addressed.
J. Clin. Med. 2025, 14(20), 7445; https://doi.org/10.3390/jcm14207445
Submission received: 18 September 2025 / Revised: 13 October 2025 / Accepted: 20 October 2025 / Published: 21 October 2025
(This article belongs to the Section Oncology)
Browse Figures
Review Reports Versions Notes

Abstract

Background: Achieving a complete response after therapy is an important predictor of long-term survival in lymphoma patients. However, previous predictive models have primarily focused on overall survival (OS) and progression-free survival (PFS), often overlooking treatment response. Predicting the likelihood of complete response before initiating therapy can provide more immediate and actionable insights. Thus, this study aims to develop and validate predictive models for treatment response to first-line therapy in patients with B-cell lymphomas. Methods: The study used 2763 patients from the Lymphoma and Related Diseases Registry (LaRDR). The data were randomly divided into training (n = 2221, 80%) and validation (n = 553, 20%) cohorts. Seven algorithms: logistic regression, K-nearest neighbor, support vector machine, random forest, Naïve Bayes, gradient boosting machine, and extreme gradient boosting were evaluated. Model performance was assessed using discrimination and classification metrics. Additionally, model calibration and clinical utility were evaluated using the Brier score and decision curve analysis, respectively. Results: All models demonstrated comparable performance in the validation cohort, with area under the curve (AUC) values ranging from 0.69 to 0.70. A nomogram incorporating the six variables, including stage, lactate dehydrogenase, performance status, BCL2 expression, anemia, and systemic immune-inflammation index, achieved an AUC of 0.70 (95% CI: 0.65–0.75), outperforming the international prognostic index (IPI: AUC = 0.65), revised IPI (AUC = 0.61), and NCCN-IPI (AUC = 0.63). Decision curve analysis confirmed the nomogram’s superior net benefit over IPI-based systems. Conclusions: While our nomogram demonstrated improved discriminative performance and clinical utility compared to IPI-based systems, further external validation is needed before clinical integration.

1. Introduction

Lymphoma is the most common type of hematological cancer, with non-Hodgkin lymphoma (NHL) accounting for approximately 90% of all lymphoma subtypes [1,2]. In 2020, an estimated 5.44 million people were diagnosed with NHL cases, and approximately 260,000 deaths were attributed to NHL globally [3]. Age-specific incidence rates of NHL are estimated to vary globally, with the most pronounced increasing trends observed in Australia and New Zealand [4]. Moreover, in 2019, NHL resulted in 8,650,352 age-standardized disability-adjusted life years (DALYs) globally [4].
Over the past two decades, the prognosis of lymphoma patients has been significantly improved due to advances in diagnostic tools and targeted therapies, including immunotherapy, and cellular therapies [5,6]. Developing accurate prognostic predictions to categorize patients and inform clinical decisions is essential for enhancing patient outcomes. Currently, the International Prognostic Index (IPI), its updated versions, such as the Revised-IPI (R-IPI) and the National Comprehensive Cancer Network-IPI (NCCN-IPI), are widely used for risk stratification in diffuse large B-cell lymphoma (DLBCL) patients [7,8,9]. Moreover, the Follicular Lymphoma IPI (FLIPI) and the Mantle Cell Lymphoma IPI (M-IPI) are also valuable prognostic tools for risk stratification in follicular lymphoma (FL) and mantle cell lymphoma (MCL) [10,11]. However, these prognostic tools have primarily focused on overall survival (OS) and progression-free survival (PFS) as endpoints [8,9,12,13]. While 5-year OS and PFS are important predominating endpoints for measuring treatment efficacy [14], predicting treatment response before initiating therapy can provide more immediate and actionable insights for effective management and may facilitate the development and evaluation of novel therapies [15,16]. Recent evidence has demonstrated that early treatment response is a validated surrogate endpoint for long-term survival outcomes in lymphoma [16]. This underscores the need for prognostic models that incorporate treatment response as a primary endpoint.
Achieving a complete response (CR) after the course of therapy is an important predictor of long-term survival in lymphoma patients [16,17,18,19]. Although the cure rate of lymphoma patients has improved, patients’ response to therapy varies widely depending on the types of lymphoma and patient characteristics, ranging from progressive disease to CR [5,20]. The IPI scores of 0–1, 2, 3, and 4–5, developed in the pre-rituximab era using CT and bone marrow assessments, correspond to CR rates of 87%, 67%, 55%, and 44%, respectively [7]. However, its predictive accuracy and clinical utility for treatment response have not been thoroughly assessed in the context of modern therapeutic and imaging approaches. Moreover, while revised indices such as R-IPI and NCCN-IPI improve survival prediction, they did not estimate response rates across risk groups [8,9]. Furthermore, the IPI tool fails to capture the wide range of clinical factors and biomarkers. More importantly, addition of molecular abnormality adds significant value to the prognostication of lymphomas [21]; however, most exiting tools do not have this incorporated necessitating further updated tools which again needs to be tested in bigger cohorts. Hence, an updated risk stratification model incorporating routinely collected clinical variables and biomarkers is needed.
Several studies have demonstrated that inflammatory and nutritional indicators are closely related to the prognosis of cancer patients [12,13,22,23,24,25,26,27,28,29]. Systemic inflammation and inadequate diet promote the proliferation of tumor cells, provide nutrition for tumor cells, stimulate cell growth, and disrupt the immune system, which in turn leads to poor prognosis [30]. Body mass index (BMI), serum albumin, and the prognostic nutritional index (PNI) are often used to assess nutritional status in cancer patients. Traditional inflammatory parameters such as the platelet-to-lymphocyte ratio (PLR), neutrophil-to-lymphocyte ratio (NLR), and lymphocyte-to-monocyte ratio (LMR), along with novel indicators like the systemic immune-inflammation index (SII) and systemic inflammation response index (SIRI) are simple measures to assess systemic inflammation. An increasing amount of research indicates that these inflammation and nutritional indicators are independent predictors of lymphoma prognosis [13,26,27,28]. For example, Liu et al., [12] demonstrated that a nomogram incorporating inflammatory-nutritional markers (SII and PNI) exhibited superior discriminative ability compared to the IPI and NCCN-IPI in predicting OS for DLBCL. Moreover, other studies showed that PNI and SII were significantly associated with complete remission rate [26,28]. However, these studies have only drawn associations between these inflammatory-nutritional indicators and treatment response, without demonstrating whether combining these markers with other prognostic factors enhances predictive accuracy and clinical utility. Although inflammation-nutritional indicators are routinely collected, and relatively inexpensive, their prognostic value in predicting treatment response has been limited in patients with lymphoma.
The pretreatment prediction of therapy response is essential for stratifying patients by their likelihood of achieving a CR and the delivery of precise treatment [31]. However, current clinical prediction tools for forecasting treatment response in lymphoma remain limited. Wang et al. developed a nomogram that integrates imaging features with clinico-pathological factors to assess the CR to chemotherapy in patients with gastric DLBCL [20]. However, this model is not applicable for predicting treatment response before starting therapy, as it was constructed based on post-treatment indicators. Therefore, this study aims to develop and validate a novel prognostic model incorporating pretreatment inflammation-nutritional indicators and using machine learning (ML) for treatment response in B-cell NHL. Additionally, the study evaluated the predictive performance of the IPI, R-IPI, and NCCN-IPI in stratifying patients based on treatment response.

2. Methods

2.1. Data Source and Study Population

The study utilized data from the prospective binational Lymphoma and Related Diseases Registry (LaRDR; https://lardr.org/), a multicentre registry established in 2016 across Australia and New Zealand. Adult patients (≥18 years) with a new diagnosis of lymphoma, chronic lymphocytic leukemia (CLL), or related diseases in accordance with the World Health Organization (WHO) classification (WHO-HAEM3 or WHO-HAEM4, depending on the time of registration) were included in the registry [32,33]. The methodology of the LaRDR has been described in detail elsewhere [34]. In this study, patients diagnosed with B-cell NHL, namely, DLCBL, FL, MCL and Burkitt lymphoma (BL), who had been treated with chemotherapy/immunotherapy were included.

2.2. Study Variables Measurement

Treatment response to first-line chemotherapy/immunotherapy was the primary outcome variable of this study. According to the Lugano 2014 criteria [35], treatment response is categorized as complete response (Deauville score 1–3/disappearance of all evidence of disease), partial response (Deauville score reduction from 4–5 to 1–3/decrease in the size of previously abnormal lesions by at least 50%), no response/stable disease (insufficient reduction to qualify for partial response, but also not meeting criteria for progressive disease), and progressive disease (appearance of new lesions or increase in the size of measurable disease by at least 50% of previously involved sites). In this study, treatment response to first-line therapy was dichotomized into CR and incomplete response (partial response, no response/stable disease, or progressive disease).
Pretreatment factors covering sociodemographic characteristics, clinical features, biomarkers, and inflammatory-nutritional indicators, including age, sex, lymphoma subtype, number of extranodal disease sites, ECOG performance status, B-symptoms, presence of bulk disease (>5 cm), lactate dehydrogenase (LDH), albumin, bilirubin, BMI, C-reactive protein, serum β2 microglobulin, creatinine, alkaline phosphatase (ALP), calcium, hemoglobin, white blood cell count, NLR, MLR, PLR, PNI, SII, SIRI, BCL6 expression and BCL2 expression were considered as potential prognostic factors.
Performance status was measured according to the ECOG (Eastern Cooperative Oncology Group) scale based on four criteria which has been found to be highly correlated with survival and may help predictability to tolerate therapy. BMI is categorized based on WHO cutoff points [36]: underweight (BMI < 18.5 kg/m2), normal weight (BMI between 18.5–24.9 kg/m2), overweight (BMI between 25–29.9 kg/m2), and obese (BMI ≥ 30 kg/m2). Patients with anemia and hypoalbuminemia were categorized as per local values and criteria. NLR, MLR and PLR were determined by dividing the absolute counts of neutrophils, monocytes, and platelets by the absolute lymphocyte count, respectively. PNI was expressed as (Albumin (g/L) + 5) × total lymphocyte count  ×  109/L [26]. SII was calculated as neutrophil counts × platelet counts/lymphocyte counts [37]. SIRI was calculated as monocyte count (109/L) × neutrophil count (109/L)/lymphocyte count (109/L) [38].

2.3. Statistical Analysis

Patients characteristics were summarized using frequencies and percentages, according to treatment responses (complete versus incomplete), and the Pearson chi-square (χ2) test was employed. Receiver operating characteristic (ROC) curve analysis was used to determine the optimal predictive cutoff values for quantitative variables, including ALP, creatinine, bilirubin, NLR, PLR, MLR, PNI, SII, and SIRI. Since missingness in the dataset ranged from 0.07% to 62.4% (Supplementary Table S1), complete case analyses were utilised on variables with less than 5% missing values, while variables with 5% to 40% missing values were handled by multiple imputations using the mice R package. Variables with more than 40% missing values, including C-reactive protein (CRP) and beta-2 microglobulin (B2M), were excluded. All data management and statistical analyses were performed using R version 4.4.2.

2.4. Model Development

The data were randomly divided into a training cohort (n = 2221) for tuning model parameters and a validation cohort (n = 553) for predicting model performance metrics (for internal validation) at a 4:1 ratio. The chi-square test was applied to compare the differences in characteristics of patients in the training and validation sets. Six widely used ML algorithms, such as gradient boosting (GBM), K-nearest neighbor (KNN), random forest (RF), support vector machine (SVM), Naïve Bayes (NB), and extreme gradient boosting (XgBoost), along with logistic regression (LR) were employed to predict treatment response of patients with B-cell NHL. The R packages gbm, caret, randomForest, e1071, glmnet, and xgboost were utilized to implement these models. To mitigate the risk of overfitting, a 10-fold cross-validation method was employed in the model training process. In this procedure, the dataset is randomly partitioned into 10 equal folds. Then, the model is trained 10 times, with each iteration using 9 folds for training and the remaining fold for validation and the model performance is averaged across the folds. This approach provides a more reliable estimate of out-of-sample performance compared with a single train test split.

2.5. Feature Selection

To select relevant features and achieve efficient data reduction, we employed two stages of variable selection. First, a Boruta algorithm [39,40] was employed to examine the multivariable relationships among the variables, considering all features relevant to the outcome variable. It is a wrapper method built around Random Forests that identifies all features relevant to an outcome rather than only a minimal optimal subset. It works by creating shadow variables (randomly permuted copies of the original features) and then comparing the importance of each real variable to these shadow features. Then, it classifies variables as confirmed (those that outperform the best shadow feature), tentative (those with intermediate performance), or rejected (those that perform worse than the best random shadow feature). Tentative variables whose importance scores are too close to those of shadow features were further evaluated using importance scores to determine inclusion. Secondly, multivariable LR and RF algorithm were employed for variables retained by the Boruta algorithm. The LR model helps identify potential relationships and key variables influencing the outcome [41]. The RF algorithm aids in feature selection by evaluating the importance of variables. To evaluate variable importance, the mean decrease in the Gini index was used. It is a measure of node impurity used in decision trees and Random Forests. It reflects the probability that a randomly chosen observation from a dataset would be incorrectly classified if it were labelled according to the distribution of classes in that node. A Gini index of 0 indicates perfect purity (all cases in the node belong to a single class), while higher values indicate greater heterogeneity. In Random Forests, variable importance is estimated by averaging the reduction in the Gini index each time a variable is used to split the data across all trees in the forest. Variables that achieve larger decreases in impurity are considered more important for prediction [41,42].

2.6. Class Imbalance Management

In classification models, ML algorithms often achieve high accuracy for the majority class but assign less importance to the minority class. This imbalance can significantly affect the performance of classifiers. To address class imbalance, we used oversampling, undersampling, and two hybrid methods, such as the Random oversampling of examples (ROSE) and Synthetic minority oversampling (SMOTE) techniques in the training dataset. SMOTE and ROSE are a widely used method to address class imbalance in classification tasks. SMOTE generates synthetic examples of the minority class by interpolating between existing minority instances and their nearest neighbours [43]. ROSE generates synthetic balanced samples by drawing new examples from a smoothed bootstrap distribution of both classes, improving classifier performance in binary imbalanced learning [44].

2.7. Model Performance Evaluation

Model discrimination metrics such as the area under the curve (AUC) and classification metrics, including accuracy, sensitivity, specificity, positive and negative predictive values were calculated for each algorithm. The AUC is a discriminating performance indicator that indicates how well a model can distinguish event individuals (i.e., with incomplete response) from non-event individuals (i.e., CR). It ranges from 0 to 1, where values of 0, 0.5 and 1, indicate perfect anti-discrimination, no discrimination, and perfect discrimination, respectively. Brier scores were also used to evaluate the overall agreement between predicted and actual treatment response probabilities, with lower values indicating better calibration and accuracy [45]. In addition, to evaluate and compare developed prediction models in the context of clinical decision-making, a decision curve analysis (DCA) was employed. Based on selected variables, a nomogram for predicting the treatment response of B-cell NHL patients was constructed. This nomogram is a visual tool derived from a statistical model that enables clinicians to estimate the likelihood of a particular clinical outcome. It displays variables separately and assigns to each variable a specific score based on its impact on the probability of the event of interest. By assigning different weights to each risk factor, the nomogram provides a more individualized and accurate risk assessment. The overall score is obtained by summing up the individual variables’ scores [46,47]. Furthermore, the predictive ability of our model, IPI, R-IPI, and NCCN-IPI was compared by AUC, Brier score, and DCA (Figure 1). The R packages pROC, pec, rmda, and rms were used to calculate the AUC, compute the Brier score, perform the DCA, and construct a nomogram, respectively.

3. Results

3.1. Determination of Cut-Off Values for Inflammatory Nutritional Indicators

A total of 2763 patients with B-cell NHL were included in this study. Using ROC curve, the optimal cut-off points for PLR, MLR, NLR, PNI, SII, and SIRI were 274.773, 0.611, 5.123, 40.93, 1686.985, and 3.529, respectively, with corresponding AUC values of 0.555, 0.563, 0.574, 0.572, 0.570, and 0.574 (Supplementary Table S2). According to the ROC cut-off value, creatinine, bilirubin, ALP, NLR, MLR, PLR, PNI, SII and SIRI were divided into a low and high group.

3.2. Background Characteristics

Incomplete response to first-line therapy was more common in patients with adverse clinical features, including stage III/IV, ECOG performance status > 1, elevated LDH, anemia, multiple extranodal sites involvement, and low albumin. Regarding inflammatory nutritional indicators, an incomplete response was more common in patients with low PNI and high SIRI, PLR, NLR, and MLR (Table 1). Additionally, univariable logistic regression identified several factors significantly associated with incomplete response, including advanced stage, poor performance status, elevated LDH, anemia, and inflammatory markers (PNI, SII, SIRI, MLR, PLR, and NLR) (Supplementary Table S3). The training and validation sets were comparable in terms of key characteristics (Supplementary Table S4).

3.3. Features Selection

In the first step, based on the Boruta algorithm, out of 25 attributes, 10 were rejected (red boxplots), 13 were confirmed (green boxplots), and 2 were designated as tentative (yellow boxplots) (Figure 2). Both tentative variables were retained (considered as important variables) based on importance scores (Supplementary Table S5). Of the 15 variables retained by the Boruta algorithm, multivariable LR identified six significant independent prognostic factors: performance status, stage, LDH, BCL2 expression, anemia and SII (Supplementary Table S6). The top six important variables from the RF algorithm, including absolute white cell count, bulk, stage, LDH, performance status and BCL2 expression, were identified to compare with the LR model results (Supplementary Figure S1). The combined methods selected eight variables: six significant variables from the LR model, plus additional two factors (bulk disease and white blood cell counts) from RF ranking.

3.4. Model Development and Performance

In our study, we conducted a comparative analysis of seven algorithms using two distinct predictor sets: variables identified as statistically significant through multivariable LR, and a combination of top six variables based on RF importance with those statistically significant in multivariable LR. Additionally, we employed four methods for managing data imbalance as disproportion in treatment response was encountered, with 75.6% of patients achieving a complete response and 24.4% not achieving a CR. Data balancing improved the performance of only the SVM and RF algorithms, and there was no superior data balancing method, as the performance was consistent across the four data balancing methods (Supplementary Table S7). The inclusion of top variables based on RF importance, alongside those statistically significant in multivariable LR, did not yield a significant enhancement in the predictive performance of the models (Supplementary Figure S2). Consequently, we used six prognostic factors identified as significant through multivariable LR, namely, ECOG performance status, stage, LDH, BCL2 expression, anemia and SII to predict treatment response. The AUCs of all ML in the validation cohort were similar, ranging from 0.69 to 0.70 (Table 2 and Figure 3). The AUC values for all algorithms in both the training and validation sets exhibit minimal differences, indicating that the models were not overfitted and could generalize effectively to unseen data. Moreover, the Brier score ranged from 0.223 for NB to 0.298 for RF, with a score of 1 indicating the poorest calibration and a score of 0 representing perfect calibration (Table 2).

3.5. Nomogram Development

The final six independent prognostic factors were integrated into the nomogram to predict treatment response. Variable scores can be obtained by where the vertical line intersects the point scale at the top of the chart for each variable (Figure 4), and these scores can then be summed up to get a total score. This total score provides predictive measures of treatment response for each patient. The bottom of the series shows the model’s estimated probability of incomplete response after chemotherapy/immunotherapy in patients with B-cell NHL. For example, a patient with performance status (≥2), anemia, elevated LDH, high SII, positive BCL2 expression, and stage III/IV could obtain 57 points, 37 points, 52 points, 70 points, 36 points, and 100 points, respectively, resulting in a total score of 352 points. For patients having all these six risk factors, their probability of having an incomplete response is 80%, compared to 20% for patients without these six risk factors (Figure 4). Generally, the probability of an incomplete (IC) response can be approximated using the formula:
Probability of IC response = 0.2 + 0.00171 × Total points, where 0.0017 is the slope (scaling factor).
The AUCs for the nomogram in the validation cohort was 0.70 (95% CI: 0.65, 0.75), indicating that our model had acceptable discriminating ability. The AUC of the nomogram outperformed the Revised IPI (AUC = 0.61, 95% CI: 0.56, 0.66), NCCN-IPI (AUC = 0.63, 95% CI: 0.57, 0.68), and IPI (AUC = 0.65, 95% CI: 0.60, 0.71). Calibration of the nomogram was superior, with a lower Brier score (0.227) compared to existing indices (Table 2). Furthermore, the decision curve analysis confirmed the clinical utility was better for nomogram (Figure 5).
Based on quantiles, the total score of the nomogram was categorized into four risk groups: low risk (<138 points), low intermediate (138–188 points), high intermediate (188–225 points) and high-risk (≥225 points). In the validation cohort, there were 218, 125, 93, and 117 patients in the low-risk, low intermediate, high intermediate, and high-risk groups, respectively, with corresponding incomplete response rates of 9.1%, 25.6%, 32.3%, and 39.3% (Table 3).
As the IPI, R-IPI, and NCCN-IPI were originally developed for prognostication in DLBCL, we evaluated their performance specifically within the DLBCL cohort to ensure a fair comparison and to further contextualize their utility. In the validation cohort, our unified model demonstrated slightly superior discriminative ability compared to IPI, R-IPI, and NCCN-IPI (AUC: 0.64 vs. 0.59, 0.57, and 0.60, respectively) (Supplementary Figure S3). These findings suggest that our model may offer slightly improved predictive accuracy, even within the subtype for which these existing clinical scores were originally designed.

4. Discussion

Here, we developed and internally validated a nomogram and machine learning algorithms using universally collected clinical features, biomarkers, and inflammatory-nutritional markers data from 33 hospitals to assess the predictability of treatment response. In this study, both LR and ML (RF) identified overlapping key predictors, including stage, LDH, ECOG performance status, and BCL2 expression. Overall, ML algorithms and LR analysis demonstrated comparable predictive ability in predicting treatment response. Notably, the nomogram showed significantly better discriminative ability compared to the R-IPI.
In this study, the comparable predictive performance observed between ML algorithms and LR can be attributed to several factors. First, although categorizing continuous variables are not a statistically recommended practice in predictive modelling [48,49], we categorized them based on predefined cutoff points or ROC curve analysis for simplicity in clinical use. Categorizing continuous variables may have influenced ML model performance. For example, categorization can lead to a loss of information and reduce their ability to capture complex, non-linear relationships, particularly in tree-based methods. Second, the number of predictors used in model development was limited. It is noted that standard regression methods perform well when applied to datasets with relatively few predictor variables and large sample sizes [50]. While no definitive threshold exists for the number of predictors required to enhance ML performance, incorporating more variables may have provided ML algorithms with a more significant advantage over LR.
Although the application of ML in predicting healthcare outcomes has been increasing in recent years, evidence regarding its superiority over traditional regression methods remains inconclusive [51,52,53,54,55]. Consistent with our findings, a registry-based study from the European Society for Blood and Marrow Transplantation reported that LR performed comparably to ML algorithms in predicting hematopoietic stem cell transplantation-related mortality in patients with acute leukemia [51]. Similarly, studies utilizing electronic health records (EHR) in China found no significant difference in the discriminatory ability of ML models and LR for predicting recurrence and mortality in patients with DLBCL [52,53]. Similar performances of LR and ML have also been reported for predicting solid tumors and non-oncologic outcomes, for example, predicting gastric cancer risk [55] and hypertension incidence [56]. In contrast, an EHR-based study from Shanxi Tumor Hospital demonstrated that ML outperformed LR in stratifying recurrence risk among DLBCL patients [54]. Moreover, a systematic review and meta-analysis demonstrated the superiority of ML in predicting overall survival in lung cancer [57] and treatment response in rectal cancer patients [58]. These inconclusive results suggest that further investigation is needed to understand the conditions under which ML models outperform traditional methods.
Given the comparable performance of LR and ML, we developed an LR-based nomogram to aid in incorporating our model into clinical practice. It is important to note that although our nomogram’s AUC of 0.70 appear modest, it exceeds that of well-established prognostic indices, including the IPI (0.65), R-IPI (0.61), and NCCN-IPI (0.63). To our knowledge, only one previous study has developed a nomogram for predicting treatment response in lymphoma [20]. Although the nomogram demonstrated promising discrimination (AUC = 0.957) in gastric DLBCL, its application may be limited by a small sample size (n = 108 patients), a single-centre design, dependence on imaging features not routinely available in practice and reliance on post-treatment indicators, which restrict its use for pre-treatment risk stratification. Other studies have also investigated the prediction of treatment response in lymphoma subtypes, such as primary central nervous system lymphoma and bulky Hodgkin and non-Hodgkin lymphomas, using advanced imaging or radiomic features, with a small sample size from single-centre settings. In these studies, discrimination has ranged from AUC 0.618 to 0.868, with combined radiomics–clinical models often performing better than clinical models alone [59,60]. While such studies suggest potential value in incorporating radiomics, our recent meta-analysis demonstrated that models incorporating radiomic features performed similarly to models based on clinical features in hematological malignancies [61]. In our multilevel meta-analysis of 38 ML models developed for lymphoma outcomes, we observed a pooled AUC of 0.779, which was higher than the current study’s performance, with most studies predicting OS and PFS. Taken together, our nomogram derived from routinely collected pre-treatment variables across 33 hospitals enables early prediction before therapy initiation, offers substantial potential for generalizability, and remains feasible even in resource-limited settings.
Importantly, although the AUC is a fundamental metric for assessing model discrimination, it does not capture a model’s calibration or its clinical usefulness. Model’s clinical utility can be assessed using DCA by evaluating the net benefit across different threshold probabilities [62,63]. DCA demonstrated that the nomogram provided a better net benefit over IPI-based scoring systems across a range of threshold probabilities, showing better clinical utility for guiding treatment decisions. Moreover, the Brier score for our nomogram was lower than that of existing prognostic indices, suggesting better agreement between predicted and observed probabilities, thereby providing reliable risk estimation. In our study, four risk groups were categorized based on total nomogram scores derived from ROC curve analysis rather than assigning equal weight to all factors as used in IPI-based scoring systems. Assigning equal weight to all factors can lead to a loss of discrimination power and inaccurate stratification since not all risk factors have an equal impact on the occurrence of an outcome. Our approach addressed this methodological limitation by allowing for more precise and detailed individualized risk stratification. Additionally, our nomogram included BCL2 expression, anemia, and SII, which were not originally part of the IPI-based scoring systems. This demonstrates the slightly enhanced predictive power and clinical usefulness of our model in stratifying patients.
In contrast to disease-specific prognostic indices such as the IPI for DLBCL, FLIPI for FL, and M-IPI for MCL, our model offers a unified prognostic tool applicable across multiple B-cell lymphoma subtypes. Although these existing indices were developed within subtype-specific cohorts, they primarily rely on general clinical parameters such as LDH, age, performance status, stage, extranodal involvement, and blood counts, which are not disease-specific markers but rather indicators of overall tumor burden and patient condition. Our model includes some of the shared factors found across these existing indices namely LDH, stage, and ECOG performance status and incorporating hemoglobin, which is included in FLIPI but not in IPI or MIPI. We also evaluated white blood cell count, a component of MIPI, as an independent variable, though it did not retain significance in the final model. Importantly, our model extends beyond conventional clinical variables by integrating BCL2 expression, a marker of anti-apoptotic signaling and treatment resistance [64,65] and the SII, a composite biomarker reflecting the host’s immune and inflammatory status [66]. Notably, the lymphoma subtype was not a significant predictor of initial treatment response in our cohort, suggesting that the features included in our model may capture shared biological and clinical characteristics relevant to treatment response across subtypes despite their known heterogeneity. As such, our model may serve as a practical and unified tool to stratify treatment response in various B-cell lymphomas. To further ensure fair comparison and clinical relevance, we propose to evaluate the model’s performance within individual subtypes using larger cohorts and benchmarking it against established disease-specific indices such as the FLIPI, and MIPI.
The prognostic significance of variables incorporated in our nomogram, including stage, LDH, ECOG performance status, BCL2 expression, anemia, and SII, has been well-documented in cancer prognosis [7,8,9,67,68,69,70,71]. Apart from factors included in the IPI, increasingly more evidence suggests that inflammatory-nutritional markers play a significant role in lymphoma prognosis [12,13,26,27,28]. We comprehensively tested various inflammatory-nutritional markers and found that only the SII was a significant independent risk factor for treatment response. The SII, a comprehensive inflammatory biomarker that incorporates neutrophil, platelet, and lymphocyte counts, was developed in 2014 to predict poor outcomes in patients with hepatocellular carcinoma. It is linked to circulating tumor cells and reflects the balance of the body’s inflammatory and immune responses, with higher SII values linked to poorer patient outcomes [66]. Our study observed that high SII were associated with a higher rate of incomplete responses in B-cell lymphoma patients. Consistent with our findings, a prospective study by Waley et al. [28] demonstrated that patients with high SII was significantly associated with low CR rate in patients with DLBCL. The prognostic role of SII in predicting survival outcomes for patients with lymphoma has also been well-established [12,13]. In addition, numerous meta-analyses have confirmed that high SII is associated with worse prognoses in a variety of tumors [72,73], as well as poor cardiovascular outcomes and an increased risk of cardiovascular diseases [74,75].
Moreover, our study demonstrated that patients with pretreatment anemia were more likely to have incomplete responses to treatment. Notably, anemia is a key component of the FLIPI score and has been reaffirmed as a significant prognostic factor in a recent model, the FL Evaluation Index (FLEX) [10,76]. Moreover, studies have shown that pretreatment anemia provides additional predictive value to the IPI, R-IPI, and NCCN-IPI in predicting OS for patients with DLBCL [77,78]. This can be attributed to hypoxia induced by low levels of hemoglobin, which has been shown to contribute to tumor progression and therapy resistance by promoting angiogenesis, inducing genomic mutations, and increasing resistance to apoptosis and the cytotoxic effects of chemo/radiotherapy-generated free radicals [79,80]. Given that anemia is a common hematologic abnormality in cancer patients and can be further induced by chemotherapy [81,82], our study suggests that incorporating pretreatment anemia status into existing prognostic tools for lymphoma patients may enhance their accuracy. This could help tailor treatment strategies more effectively, thereby improving therapy response and overall prognosis.
Furthermore, building on the NCCN-IPI recommendation that the inclusion of biological markers such as BCL2 expression may enhance prognostic accuracy [9], we tested the significance of BCL2 expression and found that positive BCL2 expression was significantly associated with a higher rate of incomplete treatment responses. Similarly, in previous studies, positive BCL2 expression has been associated with an increased risk of recurrence, poor treatment response, and shorter PFS and OS in lymphoma patients [83,84,85,86], as well as poor therapy response in acute leukemia [87]. This finding aligns with the general understanding that overexpression of BCL2 can lead to the survival of abnormal cells that should otherwise undergo apoptosis, thereby contributing to tumor growth and resistance to therapy [64,65]. On the contrary, a recent study has shown that patients with BCL2 dependence in chronic lymphocytic leukemia (CLL) tend to respond favorably to therapy [88]. While BCL2 overexpression generally indicates poor prognosis due to enhanced cell survival and resistance to apoptosis, these findings underscore the complex role of BCL2 in cancer prognosis and warrant further investigation.
This study has some strengths and limitations. To the best of our knowledge, this is the first comprehensive comparison of various machine learning algorithms alongside standard regression techniques, as well as the validation of existing tools for predicting treatment response in B-cell lymphomas. Additionally, the utilization of multicenter data enhances the diversity and representativeness of the patient cohort. Although the model was developed using a multi-institutional dataset, and treatment response did not significantly differ across subtypes, the predominance of DLBCL cases may still influence our unified model’s performance. While we validated the model specifically in DLBCL, comparing its performance with IPI, R-IPI, and NCCN-IPI, further validation in rarer subtypes is needed to confirm broader applicability. Moreover, the study’s reliance on internal validation limits its generalizability, underscoring the need for external validation. In addition, although ROC curve analysis was employed to determine optimal predictive cutoff values for inflammatory-nutritional markers, an advantageous method for evaluating model performance across different thresholds, these cutoff points may not be universally applicable across diverse populations or clinical settings. Therefore, future research should prioritize the validation of these cutoff points to ensure their generalizability and reliability in varied clinical settings.

5. Conclusions

In conclusion, our study developed and internally validated predictive models for treatment response in lymphoma patients, demonstrating that machine learning models and standard regression have comparable performance. Although our nomogram, which incorporates clinical (stage, LDH, anemia and ECOG performance status), inflammatory (SII), and molecular (BCL2 expression) features, demonstrated slightly improved discriminative ability and clinical utility compared to existing tools, the overall discriminatory power remains limited. To advance prognostic accuracy and better reflect the evolving landscape of lymphoma care, future work will focus on integrating additional molecular parameters and transcriptomic signatures. Moreover, as our model is only internally validated, external validation is warranted to confirm its generalizability. This supports its potential use in risk stratification and decision-making for tailored treatment strategies in managing lymphoma patients.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/jcm14207445/s1, Table S1: Proportion of missing data for variables; Table S2: Optimal cutoff points and performance metrics for inflammatory nutritional indicators; Table S3: Univariable logistic regression for treatment response; Table S4: Background characteristics of study participants in the training cohort and validation cohort; Table S5: Prognostic factors retained and rejected by final Boruta feature selection methods; Table S6: Multivariable logistic regression in training data dataset; Table S7: Performance of machine learning algorithms according to different data imbalance management methods; Figure S1: Features ranking based on random forest algorithm. LDH, lactate dehydrogenase; MLR, monocyte-to-lymphocyte ratio; Figure S2: Area under the curve (AUC) for machine learning algorithms with features selected using combined methods in the testing data set, balanced with SMOTE; Figure S3: Area under the curve (AUC) for our model (Nomogram) and existing tools in the testing data set for Diffuse large B-cell lymphoma.

Author Contributions

A.Z.A.: Writing first draft of the manuscript, A.Z.A., I.M., N.P., C.W., E.C., E.A.H., Z.K.M., E.M.W., S.O. and T.N.: Conceptualization, Methodology, Investigation, Data Curation, Data interpretation, manuscript—review and editing. A.Z.A., I.M., N.P., C.W., S.O. and T.N.: Data analysis. I.M., N.P. and T.N.: Supervision. All authors have read and agreed to the published version of the manuscript.

Funding

This study received no funding. However, A.Z.A. is receiving support from the University of Canberra higher degree by research stipend scholarship (https://doi.org/10.82133/C42F-K220).

Institutional Review Board Statement

This study obtained ethics approval from the University of Canberra Human Research Ethics Committee (Project ID: 13709, 6 June 2024). Additionally, approval was obtained from the Steering Committee of the Australia and New Zealand LaRDR.

Informed Consent Statement

Patient consent was waived due to the retrospective nature of the analysis and the use of fully anonymized data, as approved by the Steering Committee of the Australia and New Zealand LaRDR.

Data Availability Statement

The data supporting this study’s results can be obtained from the Lymphoma and Related Diseases Registry (LaRDR) with permission from the Steering Committee, and the request complies with the LaRDR Data Access Policy.

Conflicts of Interest

The authors declare that they have no personal relationships or financial conflicts of interest that could have influenced the work presented in this paper.

Abbreviations

AUCArea Under the Curve
CIConfidence Interval
CRComplete response
GBMGradient Boosting Model
IPIInternational Prognostic Index
LaRDRLymphoma and Related Diseases Registry
LDHLactate Dehydrogenase
LRLogistic Regression
MLMachine Learning
MLRmonocyte-to-lymphocyte ratio
NCCN-IPINational Comprehensive Cancer Network International Prognostic Index
NLRNeutrophil-to-Lymphocyte Ratio
PFSProgression Free survival
PLRPlatelet-to-Lymphocyte Ratio
PNIPrognostic nutrition
RFRandom Forest
R-IPIRevised International Prognostic Index
SIISystemic Immune-Inflammation Index
SIRISystemic Inflammation Response Index
OSOverall Survival

References

  1. Thandra, K.C.; Barsouk, A.; Saginala, K.; Padala, S.A.; Barsouk, A.; Rawla, P. Epidemiology of Non-Hodgkin’s Lymphoma. Med. Sci. 2021, 9, 5. [Google Scholar] [CrossRef]
  2. Cancer Research Institute. Immunotherapy: For Lymphoma. What Makes Immunotherapy for Lymphoma a Promising Treatment? Available online: https://www.cancerresearch.org/immunotherapy-by-cancer-type/lymphoma (accessed on 23 July 2025).
  3. Mafra, A.; Laversanne, M.; Gospodarowicz, M.; Klinger, P.; De Paula Silva, N.; Piñeros, M.; Steliarova-Foucher, E.; Bray, F.; Znaor, A. Global patterns of non-Hodgkin lymphoma in 2020. Int. J. Cancer 2022, 151, 1474–1481. [Google Scholar] [CrossRef] [PubMed]
  4. Chu, Y.; Liu, Y.; Fang, X.; Jiang, Y.; Ding, M.; Ge, X.; Yuan, D.; Lu, K.; Li, P.; Li, Y.; et al. The epidemiological patterns of non-Hodgkin lymphoma: Global estimates of disease burden, risk factors, and temporal trends. Front. Oncol. 2023, 13, 1059914. [Google Scholar] [CrossRef] [PubMed]
  5. Chao, M.P. Treatment challenges in the management of relapsed or refractory non-Hodgkin’s lymphoma—Novel and emerging therapies. Cancer Manag. Res. 2013, 5, 251–269. [Google Scholar] [CrossRef] [PubMed]
  6. Patrício, A.; Costa, R.S.; Henriques, R. On the challenges of predicting treatment response in Hodgkin’s Lymphoma using transcriptomic data. BMC Med. Genom. 2023, 16 (Suppl. S1), 170. [Google Scholar] [CrossRef]
  7. International Non-Hodgkin’s Lymphoma Prognostic Factors Project. A predictive model for aggressive non-Hodgkin’s lymphoma. N. Engl. J. Med. 1993, 329, 987–994. [Google Scholar] [CrossRef]
  8. Sehn, L.H.; Berry, B.; Chhanabhai, M.; Fitzgerald, C.; Gill, K.; Hoskins, P.; Klasa, R.; Savage, K.J.; Shenkier, T.; Sutherland, J. The revised International Prognostic Index (R-IPI) is a better predictor of outcome than the standard IPI for patients with diffuse large B-cell lymphoma treated with R-CHOP. Blood 2007, 109, 1857–1861. [Google Scholar] [CrossRef]
  9. Zhou, Z.; Sehn, L.H.; Rademaker, A.W.; Gordon, L.I.; LaCasce, A.S.; Crosby-Thompson, A.; Vanderplas, A.; Zelenetz, A.D.; Abel, G.A.; Rodriguez, M.A. An enhanced International Prognostic Index (NCCN-IPI) for patients with diffuse large B-cell lymphoma treated in the rituximab era. Blood J. Am. Soc. Hematol. 2014, 123, 837–842. [Google Scholar] [CrossRef]
  10. Solal-Céligny, P.; Roy, P.; Colombat, P.; White, J.; Armitage, J.O.; Arranz-Saez, R.; Au, W.Y.; Bellei, M.; Brice, P.; Caballero, D. Follicular lymphoma international prognostic index. Blood 2004, 104, 1258–1265. [Google Scholar] [CrossRef]
  11. Hoster, E.; Dreyling, M.; Klapper, W.; Gisselbrecht, C.; Van Hoof, A.; Kluin-Nelemans, H.C.; Pfreundschuh, M.; Reiser, M.; Metzner, B.; Einsele, H. A new prognostic index (MIPI) for patients with advanced-stage mantle cell lymphoma. Blood J. Am. Soc. Hematol. 2008, 111, 558–565. [Google Scholar] [CrossRef]
  12. Liu, Y.; Sheng, L.; Hua, H.; Zhou, J.; Zhao, Y.; Wang, B. An externally validated nomogram for predicting the overall survival of patients with diffuse large B-cell lymphoma based on clinical characteristics and systemic inflammatory markers. Technol. Cancer Res. Treat. 2023, 22, 15330338231180785. [Google Scholar] [CrossRef]
  13. Wu, J.; Zhu, H.; Zhang, Q.; Sun, Y.; He, X.; Liao, J.; Liu, Y.; Huang, L. Nomogram based on the systemic immune-inflammation index for predicting the prognosis of diffuse large B-cell lymphoma. Asia Pac. J. Clin. Oncol. 2023, 19, e138–e148. [Google Scholar] [CrossRef]
  14. Bröckelmann, P.; Müller, H.; Fuchs, M.; Gillessen, S.; Eichenauer, D.; Borchmann, S.; Jacob, A.; Behringer, K.; Momotow, J.; Ferdinandus, J. Correlation between progression-free and overall survival in patients with Hodgkin lymphoma: A comprehensive analysis of individual patient data from randomized GHSG trials. Ann. Oncol. 2024, 36, 393–402. [Google Scholar] [CrossRef] [PubMed]
  15. Amin, S.; Bathe, O.F. Response biomarkers: Re-envisioning the approach to tailoring drug therapy for cancer. BMC Cancer 2016, 16, 850. [Google Scholar] [CrossRef] [PubMed]
  16. Bommier, C.; Zucca, E.; Chevret, S.; Conconi, A.; Nowakowski, G.; Maurer, M.J.; Cerhan, J.R.; Thieblemont, C.; Lambert, J. Early complete response as a validated surrogate marker in extranodal marginal zone lymphoma systemic therapy. Blood 2024, 143, 422–428. [Google Scholar] [CrossRef]
  17. Liu, Y.; Sheng, L.; Hua, H.; Zhou, J.; Zhao, Y.; Wang, B. A novel and Validated Inflammation-Based Prognosis Score (IBPS) predicts outcomes in patients with diffuse large B-cell lymphoma. Cancer Manag. Res. 2023, 15, 651–666. [Google Scholar] [CrossRef] [PubMed]
  18. Zhang, X.; Fan, W.; Hu, Y.-Y.; Li, Z.-M.; Xia, Z.-J.; Lin, X.-P.; Zhang, Y.-R.; Liang, P.-Y.; Li, Y.-H. Qualitative visual trichotomous assessment improves the value of fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography in predicting the prognosis of diffuse large B-cell lymphoma. Chin. J. Cancer 2015, 34, 264–271. [Google Scholar] [CrossRef]
  19. Sung, K.H.; Lee, E.H.; Kim, Y.Z. Factors influencing the response to high dose methotrexate-based vincristine and procarbazine combination chemotherapy for primary central nervous system lymphoma. J. Korean Med. Sci. 2011, 26, 551. [Google Scholar] [CrossRef]
  20. Wang, P.; Chen, K.; Wang, J.; Ni, Z.; Shang, N.; Meng, W. A new nomogram for assessing complete response (CR) in gastric diffuse large B-cell lymphoma (DLBCL) patients after chemotherapy. J. Cancer Res. Clin. Oncol. 2023, 149, 9757–9765. [Google Scholar] [CrossRef]
  21. Kos, I.A.; Thurner, L.; Bittenbring, J.T.; Christofyllakis, K.; Kaddu-Mulindwa, D. Advances in lymphoma molecular diagnostics. Diagnostics 2021, 11, 2174. [Google Scholar] [CrossRef]
  22. Huai, Q.; Luo, C.; Song, P.; Bie, F.; Bai, G.; Li, Y.; Liu, Y.; Chen, X.; Zhou, B.; Sun, X. Peripheral blood inflammatory biomarkers dynamics reflect treatment response and predict prognosis in non-small cell lung cancer patients with neoadjuvant immunotherapy. Cancer Sci. 2023, 114, 4484–4498. [Google Scholar] [CrossRef]
  23. Dong, J.; Sun, Q.; Pan, Y.; Lu, N.; Han, X.; Zhou, Q. Pretreatment systemic inflammation response index is predictive of pathological complete response in patients with breast cancer receiving neoadjuvant chemotherapy. BMC Cancer 2021, 21, 700. [Google Scholar] [CrossRef]
  24. Zhao, Y.; Liu, J.; Xiong, Z.; Gu, S.; Xia, X. The Predictive Role of Inflammatory Biomarkers for Treatment Response and Progression-Free Survival in Patients with Hepatocellular Carcinoma Receiving Hepatic Arterial Infusion Chemotherapy with FOLFOX Regimen: A Preliminary Study. J. Hepatocell. Carcinoma 2023, 10, 1037–1049. [Google Scholar] [CrossRef]
  25. Namikawa, T.; Yokota, K.; Tanioka, N.; Fukudome, I.; Iwabu, J.; Munekage, M.; Uemura, S.; Maeda, H.; Kitagawa, H.; Kobayashi, M. Systemic inflammatory response and nutritional biomarkers as predictors of nivolumab efficacy for gastric cancer. Surg. Today 2020, 50, 1486–1495. [Google Scholar] [CrossRef] [PubMed]
  26. Ge, J.; Lei, Y.; Wen, Q.; Zhang, Y.; Kong, X.; Wang, W.; Qian, S.; Hou, H.; Wang, Z.; Wu, S.; et al. The prognostic nutritional index, an independent predictor of overall survival for newly diagnosed follicular lymphoma in China. Front. Nutr. 2022, 9, 981338. [Google Scholar] [CrossRef] [PubMed]
  27. Zuo, J.; Lei, T.; Zhong, S.; Zhou, J.; Liu, R.; Wu, C.; Li, S. C-reactive protein levels, the prognostic nutritional index, and the lactate dehydrogenase-to-lymphocyte ratio are important prognostic factors in primary central nervous system lymphoma: A single-center study of 223 patients. Neurosurg. Rev. 2023, 47, 17. [Google Scholar] [CrossRef] [PubMed]
  28. Waley, A.B.; Haggag, R.; Ahmed Barakat, A.A.-e.; Rashied, H.A.; Bayomy, M.F.; Esawy, M.M.; Abdul-Saboor, A.; Bakry, A. Impact of Prognostic Nutritional Index and Systemic Immune-Inflammation Index on the Clinical Outcome of Diffuse Large B Cell Lymphoma Patients Treated with RCHOP. Zagazig Univ. Med. J. 2024, 30, 3907–3917. [Google Scholar] [CrossRef]
  29. Li, D.; Yuan, X.; Liu, J.; Li, C.; Li, W. Prognostic value of prognostic nutritional index in lung cancer: A meta-analysis. J. Thorac. Dis. 2018, 10, 5298. [Google Scholar] [CrossRef]
  30. Greten, F.R.; Grivennikov, S.I. Inflammation and cancer: Triggers, mechanisms, and consequences. Immunity 2019, 51, 27–41. [Google Scholar] [CrossRef]
  31. Tanaka, M.D.; Geubels, B.M.; Grotenhuis, B.A.; Marijnen, C.A.; Peters, F.P.; Van der Mierden, S.; Maas, M.; Couwenberg, A.M. Validated pretreatment prediction models for response to neoadjuvant therapy in patients with rectal cancer: A systematic review and critical appraisal. Cancers 2023, 15, 3945. [Google Scholar] [CrossRef]
  32. Alaggio, R.; Amador, C.; Anagnostopoulos, I.; Attygalle, A.D.; Araujo, I.B.d.O.; Berti, E.; Bhagat, G.; Borges, A.M.; Boyer, D.; Calaminici, M. The 5th edition of the World Health Organization classification of haematolymphoid tumours: Lymphoid neoplasms. Leukemia 2022, 36, 1720–1748. [Google Scholar] [CrossRef]
  33. Swerdlow, S.H.; Campo, E.; Pileri, S.A.; Harris, N.L.; Stein, H.; Siebert, R.; Advani, R.; Ghielmini, M.; Salles, G.A.; Zelenetz, A.D. The 2016 revision of the World Health Organization classification of lymphoid neoplasms. Blood 2016, 127, 2375–2390. [Google Scholar] [CrossRef]
  34. Lymphoma and Related Diseases Registry Investigators. Improving outcomes for patients with lymphoma: Design and development of the Australian and New Zealand Lymphoma and Related Diseases Registry. BMC Med. Res. Methodol. 2022, 22, 266. [Google Scholar]
  35. Cheson, B.D.; Fisher, R.I.; Barrington, S.F.; Cavalli, F.; Schwartz, L.H.; Zucca, E.; Lister, T.A. Recommendations for initial evaluation, staging, and response assessment of Hodgkin and non-Hodgkin lymphoma: The Lugano classification. J. Clin. Oncol. 2014, 32, 3059–3068. [Google Scholar] [CrossRef] [PubMed]
  36. World Health Organization. Obesity: Preventing and Managing the Global Epidemic. Report of a WHO Consultation; WHO Technical Report Series 894; World Health Organization: Geneva, Switzerland, 2000. [Google Scholar]
  37. Wang, Z.; Zhang, J.; Luo, S.; Zhao, X. Prognostic Significance of Systemic Immune-Inflammation Index in Patients With Diffuse Large B-Cell Lymphoma. Front. Oncol. 2021, 11, 655259. [Google Scholar] [CrossRef] [PubMed]
  38. Chu, Y.; Liu, Y.; Jiang, Y.; Ge, X.; Yuan, D.; Ding, M.; Qu, H.; Liu, F.; Zhou, X.; Wang, X. Prognosis and complications of patients with primary gastrointestinal diffuse large B-cell lymphoma: Development and validation of the systemic inflammation response index-covered score. Cancer Med. 2023, 12, 9570–9582. [Google Scholar] [CrossRef]
  39. Kursa, M.B.; Rudnicki, W.R. Feature selection with the Boruta package. J. Stat. Softw. 2010, 36, 1–13. [Google Scholar] [CrossRef]
  40. Bhalla, D. Select Important Variables Using Boruta Algorithm; TechTarget: Boston, MA, USA, 2017. [Google Scholar]
  41. Neumann, U.; Riemenschneider, M.; Sowa, J.-P.; Baars, T.; Kälsch, J.; Canbay, A.; Heider, D. Compensation of feature selection biases accompanied with improved predictive performance for binary classification by using a novel ensemble feature selection approach. BioData Min. 2016, 9, 36. [Google Scholar] [CrossRef]
  42. Strobl, C.; Boulesteix, A.-L.; Augustin, T. Unbiased split selection for classification trees based on the Gini index. Comput. Stat. Data Anal. 2007, 52, 483–501. [Google Scholar] [CrossRef]
  43. Chawla, N.V.; Bowyer, K.W.; Hall, L.O.; Kegelmeyer, W.P. SMOTE: Synthetic minority over-sampling technique. J. Artif. Intell. Res. 2002, 16, 321–357. [Google Scholar] [CrossRef]
  44. Lunardon, N.; Menardi, G.; Torelli, N. ROSE: A package for binary imbalanced learning. R J. 2014, 6, 79–89. [Google Scholar] [CrossRef]
  45. Steyerberg, E.W.; Vickers, A.J.; Cook, N.R.; Gerds, T.; Gonen, M.; Obuchowski, N.; Pencina, M.J.; Kattan, M.W. Assessing the performance of prediction models: A framework for traditional and novel measures. Epidemiology 2010, 21, 128–138. [Google Scholar] [CrossRef] [PubMed]
  46. Balachandran, V.P.; Gonen, M.; Smith, J.J.; DeMatteo, R.P. Nomograms in oncology: More than meets the eye. Lancet Oncol. 2015, 16, e173–e180. [Google Scholar] [CrossRef] [PubMed]
  47. Lee, W.; Lam, S.-K.; Zhang, Y.; Yang, R.; Cai, J. Review of methodological workflow, interpretation and limitations of nomogram application in cancer study. Radiat. Med. Prot. 2022, 3, 200–207. [Google Scholar] [CrossRef]
  48. Wolff, R.F.; Moons, K.G.; Riley, R.D.; Whiting, P.F.; Westwood, M.; Collins, G.S.; Reitsma, J.B.; Kleijnen, J.; Mallett, S.; Group†, P. PROBAST: A tool to assess the risk of bias and applicability of prediction model studies. Ann. Intern. Med. 2019, 170, 51–58. [Google Scholar] [CrossRef]
  49. Gupta, R.; Day, C.N.; Tobin, W.O.; Crowson, C.S. Understanding the effect of categorization of a continuous predictor with application to neuro-oncology. Neuro-Oncol. Pract. 2022, 9, 87–90. [Google Scholar] [CrossRef]
  50. Austin, P.C.; Harrell, F.E., Jr.; Steyerberg, E.W. Predictive performance of machine and statistical learning methods: Impact of data-generating processes on external validity in the “large N, small p” setting. Stat. Methods Med. Res. 2021, 30, 1465–1483. [Google Scholar] [CrossRef]
  51. Shouval, R.; Labopin, M.; Unger, R.; Giebel, S.; Ciceri, F.; Schmid, C.; Esteve, J.; Baron, F.; Gorin, N.C.; Savani, B. 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. PLoS ONE 2016, 11, e0150637. [Google Scholar] [CrossRef]
  52. Fan, S.; Zhao, Z.; Zhang, Y.; Yu, H.; Zheng, C.; Huang, X.; Yang, Z.; Xing, M.; Lu, Q.; Luo, Y. Probability calibration-based prediction of recurrence rate in patients with diffuse large B-cell lymphoma. BioData Min. 2021, 14, 38. [Google Scholar] [CrossRef]
  53. Fan, S.; Zhao, Z.; Yu, H.; Wang, L.; Zheng, C.; Huang, X.; Yang, Z.; Xing, M.; Lu, Q.; Luo, Y. Applying probability calibration to ensemble methods to predict 2-year mortality in patients with DLBCL. BMC Med. Inform. Decis. Mak. 2021, 21, 14. [Google Scholar] [CrossRef]
  54. Wang, L.; Zhao, Z.; Luo, Y.; Yu, H.; Wu, S.; Ren, X.; Zheng, C.; Huang, X. Classifying 2-year recurrence in patients with dlbcl using clinical variables with imbalanced data and machine learning methods. Comput. Methods Programs Biomed. 2020, 196, 105567. [Google Scholar] [CrossRef]
  55. Huang, R.J.; Kwon, N.S.-E.; Tomizawa, Y.; Choi, A.Y.; Hernandez-Boussard, T.; Hwang, J.H. A comparison of logistic regression against machine learning algorithms for gastric cancer risk prediction within real-world clinical data streams. JCO Clin. Cancer Inform. 2022, 6, e2200039. [Google Scholar] [CrossRef]
  56. Chowdhury, M.Z.I.; Leung, A.A.; Walker, R.L.; Sikdar, K.C.; O’Beirne, M.; Quan, H.; Turin, T.C. A comparison of machine learning algorithms and traditional regression-based statistical modeling for predicting hypertension incidence in a Canadian population. Sci. Rep. 2023, 13, 13. [Google Scholar] [CrossRef]
  57. Didier, A.J.; Nigro, A.; Noori, Z.; Omballi, M.A.; Pappada, S.M.; Hamouda, D.M. Application of machine learning for lung cancer survival prognostication—A systematic review and meta-analysis. Front. Artif. Intell. 2024, 7, 1365777. [Google Scholar] [CrossRef] [PubMed]
  58. He, J.; Wang, S.-X.; Liu, P. Machine learning in predicting pathological complete response to neoadjuvant chemoradiotherapy in rectal cancer using MRI: A systematic review and meta-analysis. Br. J. Radiol. 2024, 97, 1243–1254. [Google Scholar] [CrossRef] [PubMed]
  59. Wang, X.; Zhao, L.; Wang, S.; Zhao, X.; Chen, L.; Sun, X.; Liu, Y.; Liu, J.; Sun, S. Utility of contrast-enhanced MRI radiomics features combined with clinical indicators for predicting induction chemotherapy response in primary central nervous system lymphoma. J. Neuro-Oncol. 2024, 166, 451–460. [Google Scholar] [CrossRef]
  60. Ben Bouallègue, F.; Tabaa, Y.A.; Kafrouni, M.; Cartron, G.; Vauchot, F.; Mariano-Goulart, D. Association between textural and morphological tumor indices on baseline PET-CT and early metabolic response on interim PET-CT in bulky malignant lymphomas. Med. Phys. 2017, 44, 4608–4619. [Google Scholar] [CrossRef]
  61. Alem, A.Z.; Mohanty, I.; Pati, N.; Niyonsenga, T. Prognostic performance of machine learning in predicting haematological cancer outcomes: Systematic review and meta-analysis. Blood Rev. 2025, 28, 101325. [Google Scholar] [CrossRef]
  62. Vickers, A.J.; Elkin, E.B. Decision curve analysis: A novel method for evaluating prediction models. Med. Decis. Mak. 2006, 26, 565–574. [Google Scholar] [CrossRef]
  63. Vickers, A.J.; van Calster, B.; Steyerberg, E.W. A simple, step-by-step guide to interpreting decision curve analysis. Diagn. Progn. Res. 2019, 3, 18. [Google Scholar] [CrossRef]
  64. Qian, S.; Wei, Z.; Yang, W.; Huang, J.; Yang, Y.; Wang, J. The role of BCL-2 family proteins in regulating apoptosis and cancer therapy. Front. Oncol. 2022, 12, 985363. [Google Scholar] [CrossRef] [PubMed]
  65. Sesques, P.; Johnson, N.A. Approach to the diagnosis and treatment of high-grade B-cell lymphomas with MYC and BCL2 and/or BCL6 rearrangements. Blood J. Am. Soc. Hematol. 2017, 129, 280–288. [Google Scholar] [CrossRef] [PubMed]
  66. Hu, B.; Yang, X.-R.; Xu, Y.; Sun, Y.-F.; Sun, C.; Guo, W.; Zhang, X.; Wang, W.-M.; Qiu, S.-J.; Zhou, J. Systemic immune-inflammation index predicts prognosis of patients after curative resection for hepatocellular carcinoma. Clin. Cancer Res. 2014, 20, 6212–6222. [Google Scholar] [CrossRef] [PubMed]
  67. Yao, D.-C.; Ye, B.-K.; Yao, D.-J.; Guo, C.-C. A novel lactate dehydrogenase-based risk score model to predict the prognosis of primary central nervous system germ cell tumor treated with chemoradiotherapy. Clin. Neurol. Neurosurg. 2024, 236, 108081. [Google Scholar] [CrossRef]
  68. Yuan, S.; Xia, Y.; Shen, L.; Ye, L.; Li, L.; Chen, L.; Xie, X.; Lou, H.; Zhang, J. Development of nomograms to predict therapeutic response and prognosis of non-small cell lung cancer patients treated with anti-PD-1 antibody. Cancer Immunol. Immunother. 2021, 70, 533–546. [Google Scholar] [CrossRef]
  69. Peng, R.-R.; Liang, Z.-G.; Chen, K.-H.; Li, L.; Qu, S.; Zhu, X.-D. Nomogram based on lactate dehydrogenase-to-albumin ratio (LAR) and platelet-to-lymphocyte ratio (PLR) for predicting survival in nasopharyngeal carcinoma. J. Inflamm. Res. 2021, 14, 4019–4033. [Google Scholar] [CrossRef]
  70. Wu, Y.; Wu, H.; Lin, M.; Liu, T.; Li, J. Factors associated with immunotherapy respond and survival in advanced non-small cell lung cancer patients. Transl. Oncol. 2022, 15, 101268. [Google Scholar] [CrossRef]
  71. Hammarström, K.; Imam, I.; Mezheyeuski, A.; Ekström, J.; Sjöblom, T.; Glimelius, B. A comprehensive evaluation of associations between routinely collected staging information and the response to (chemo) radiotherapy in rectal cancer. Cancers 2020, 13, 16. [Google Scholar] [CrossRef]
  72. Zhong, J.-H.; Huang, D.-H.; Chen, Z.-Y. Prognostic role of systemic immune-inflammation index in solid tumors: A systematic review and meta-analysis. Oncotarget 2017, 8, 75381. [Google Scholar] [CrossRef]
  73. Kou, J.; Huang, J.; Li, J.; Wu, Z.; Ni, L. Systemic immune-inflammation index predicts prognosis and responsiveness to immunotherapy in cancer patients: A systematic review and meta-analysis. Clin. Exp. Med. 2023, 23, 3895–3905. [Google Scholar] [CrossRef]
  74. Agyeman, K.B.; Shafi, N.; Contreras, R.; Parackal, V.; Shah, D.N.; Gurram, A.; Keetha, N.R.; Ameen, D. The Association Between Systemic Immune-Inflammation Index and Cardiovascular Diseases: An In-depth Umbrella Review of Meta-Analyses with Grade Assessment. Heliyon 2025, 11, e42736. [Google Scholar] [CrossRef]
  75. Wang, H.; Nie, H.; Bu, G.; Tong, X.; Bai, X. Systemic immune-inflammation index (SII) and the risk of all-cause, cardiovascular, and cardio-cerebrovascular mortality in the general population. Eur. J. Med. Res. 2023, 28, 575. [Google Scholar] [CrossRef]
  76. Mir, F.; Mattiello, F.; Grigg, A.; Herold, M.; Hiddemann, W.; Marcus, R.; Seymour, J.F.; Bolen, C.R.; Knapp, A.; Nielsen, T. Follicular Lymphoma Evaluation Index (FLEX): A new clinical prognostic model that is superior to existing risk scores for predicting progression-free survival and early treatment failure after frontline immunochemotherapy. Am. J. Hematol. 2020, 95, 1503–1510. [Google Scholar] [CrossRef]
  77. Clausen, M.R.; Maurer, M.J.; Ulrichsen, S.P.; Larsen, T.S.; Himmelstrup, B.; Rønnov-Jessen, D.; Link, B.K.; Feldman, A.L.; Slager, S.L.; Nowakowski, G.S. Pretreatment hemoglobin adds prognostic information to the NCCN-IPI in patients with diffuse large B-cell lymphoma treated with anthracycline-containing chemotherapy. Clin. Epidemiol. 2019, 11, 987–996. [Google Scholar] [CrossRef]
  78. Troppan, K.T.; Melchardt, T.; Deutsch, A.; Schlick, K.; Stojakovic, T.; Bullock, M.D.; Reitz, D.; Beham-Schmid, C.; Weiss, L.; Neureiter, D. The significance of pretreatment anemia in the era of R-IPI and NCCN-IPI prognostic risk assessment tools: A dual-center study in diffuse large B-cell lymphoma patients. Eur. J. Haematol. 2015, 95, 538–544. [Google Scholar] [CrossRef]
  79. Gaspar, B.L.; Sharma, P.; Das, R. Anemia in malignancies: Pathogenetic and diagnostic considerations. Hematology 2015, 20, 18–25. [Google Scholar] [CrossRef]
  80. Zhuang, Y.; Liu, K.; He, Q.; Gu, X.; Jiang, C.; Wu, J. Hypoxia signaling in cancer: Implications for therapeutic interventions. MedComm 2023, 4, e203. [Google Scholar] [CrossRef]
  81. Bozzini, C.; Busti, F.; Marchi, G.; Vianello, A.; Cerchione, C.; Martinelli, G.; Girelli, D. Anemia in patients receiving anticancer treatments: Focus on novel therapeutic approaches. Front. Oncol. 2024, 14, 1380358. [Google Scholar] [CrossRef]
  82. Cannavale, K.; Xu, H.; Xu, L.; Sattayapiwat, O.; Rodriguez, R.; Bohac, C.; Page, J.; Chao, C. Epidemiology of chemotherapy-induced anemia in patients with non-hodgkin lymphoma. Perm. J. 2019, 23, 18–252. [Google Scholar] [CrossRef]
  83. Gong, Y.; Yan, H.; Yang, Y.; Zhai, B.; Huang, Z.; Zhang, Z. Construction and validation of a novel nomogram for predicting the recurrence of diffuse large B cell lymphoma treated with R-CHOP. Pharmacogenomics Pers. Med. 2023, 16, 291–301. [Google Scholar] [CrossRef]
  84. Li, L.; Zhang, X.; Zhang, T.; Song, Z.; Hu, G.; Li, W.; Li, L.; Qiu, L.; Qian, Z.; Zhou, S. Prognostic Significance of BCL-2 and BCL-6 Expression in MYC-positive DLBCL. Clin. Lymphoma Myeloma Leuk. 2018, 18, e381–e389. [Google Scholar] [CrossRef]
  85. Elhendawy, H.A.; Ibrahiem, A.T.; Elmahdi, H.S.; Omar, A.M. Prognostic Significance of BCL2 Protein in Diffuse Large Cell Lymphoma of Head and Neck; Relation to Response to Chemotherapy. Open J. Pathol. 2020, 10, 76. [Google Scholar] [CrossRef]
  86. Correia, C.; Maurer, M.J.; McDonough, S.J.; Schneider, P.A.; Ross, P.E.; Novak, A.J.; Feldman, A.L.; Cerhan, J.R.; Slager, S.L.; Witzig, T.E. Relationship between BCL2 mutations and follicular lymphoma outcome in the chemoimmunotherapy era. Blood Cancer J. 2023, 13, 81. [Google Scholar] [CrossRef]
  87. Liu, Y.; He, P.; Liu, F.; Shi, L.; Zhu, H.; Cheng, X.; Zhao, J.; Wang, Y.; Zhang, M. Prognostic significance of B-cell lymphoma 2 expression in acute leukemia: A systematic review and meta-analysis. Mol. Clin. Oncol. 2014, 2, 411–414. [Google Scholar] [CrossRef]
  88. Chong, S.J.F.; Lu, J.; Valentin, R.; Lehmberg, T.Z.; Eu, J.Q.; Wang, J.; Zhu, F.; Kong, L.R.; Fernandes, S.M.; Zhang, J. BCL-2 dependence is a favorable predictive marker of response to therapy for chronic lymphocytic leukemia. Mol. Cancer 2025, 24, 62. [Google Scholar] [CrossRef]
Jcm 14 07445 g001
Figure 1. Conceptual framework of the machine learning model development for predicting treatment response of B-cell non-Hodgkin lymphoma. IPI; International Prognostic Index, R-IPI; Revised-IPI, NCCN-IPI; National Comprehensive Cancer Network-IPI, LR; logistic regression, RF; random forest, XgBoost; extreme gradient boosting, KNN; K-nearest neighbour, GBM; gradient boosting, SVM; support vector machine, NB; Naïve Bayes, ROSE; Random oversampling of examples technique, SMOTE; Synthetic minority oversampling technique.
Figure 1. Conceptual framework of the machine learning model development for predicting treatment response of B-cell non-Hodgkin lymphoma. IPI; International Prognostic Index, R-IPI; Revised-IPI, NCCN-IPI; National Comprehensive Cancer Network-IPI, LR; logistic regression, RF; random forest, XgBoost; extreme gradient boosting, KNN; K-nearest neighbour, GBM; gradient boosting, SVM; support vector machine, NB; Naïve Bayes, ROSE; Random oversampling of examples technique, SMOTE; Synthetic minority oversampling technique.
Jcm 14 07445 g001
Jcm 14 07445 g002
Figure 2. Boruta algorithm feature selection. Each bar represents a variable’s importance score relative to shadow features. Green bars indicate confirmed important variables; yellow indicates tentative; and red indicates rejected variables.
Figure 2. Boruta algorithm feature selection. Each bar represents a variable’s importance score relative to shadow features. Green bars indicate confirmed important variables; yellow indicates tentative; and red indicates rejected variables.
Jcm 14 07445 g002
Jcm 14 07445 g003
Figure 3. Area under the curve (AUC) for machine learning algorithms and existing prognostic tools: (A) training data set, (B) Validation set. IPI; International Prognostic Index, R-IPI; Revised-IPI, NCCN-IPI; National Comprehensive Cancer Network-IPI, LR; logistic regression, RF; random forest, XgBoost; extreme gradient boosting, KNN; K-nearest neighbour, GBM; gradient boosting, SVM; support vector machine, NB; Naïve Baye.
Figure 3. Area under the curve (AUC) for machine learning algorithms and existing prognostic tools: (A) training data set, (B) Validation set. IPI; International Prognostic Index, R-IPI; Revised-IPI, NCCN-IPI; National Comprehensive Cancer Network-IPI, LR; logistic regression, RF; random forest, XgBoost; extreme gradient boosting, KNN; K-nearest neighbour, GBM; gradient boosting, SVM; support vector machine, NB; Naïve Baye.
Jcm 14 07445 g003
Jcm 14 07445 g004
Figure 4. Nomogram with personalized working example for predicting incomplete (IC) treatment response in B-cell lymphomas. ECOG_PS, Eastern Cooperative Oncology Group Performance Status, stage; LDH; Lactate Dehydrogenase; SII, Systemic Immune-Inflammation Index.
Figure 4. Nomogram with personalized working example for predicting incomplete (IC) treatment response in B-cell lymphomas. ECOG_PS, Eastern Cooperative Oncology Group Performance Status, stage; LDH; Lactate Dehydrogenase; SII, Systemic Immune-Inflammation Index.
Jcm 14 07445 g004
Jcm 14 07445 g005
Figure 5. Decision curve analysis of nomogram and existing prognostic tools. IPI; International Prognostic Index, R-IPI; Revised-IPI, NCCN-IPI; National Comprehensive Cancer Network-IPI.
Figure 5. Decision curve analysis of nomogram and existing prognostic tools. IPI; International Prognostic Index, R-IPI; Revised-IPI, NCCN-IPI; National Comprehensive Cancer Network-IPI.
Jcm 14 07445 g005
Table 1. Treatment response according to patient characteristics.
Table 1. Treatment response according to patient characteristics.
VariablesTreatment Response, N (%)X2 (p-Value)
CompleteIncomplete
Sex
Male1231 (74.9)412 (25.1)0.93 (0.334)
Female858 (76.6)262 (23.4)
Age
≤60675 (77.3)198 (22.7)1.90 (0.168)
>601414 (74.8)476 (25.2)
BMI
Underweight47 (72.1)19 (28.8)6.91 (0.075)
Normal668 (73.1)246 (26.9)
Overweight726 (76.1)228 (23.9)
Obese648 (78.2)181 (21.8)
Stage
I359 (89.5)42 (10.5)84.89(<0.001)
II330 (83.8)64 (16.2)
III402 (75.4)131 (24.6)
IV998 (69.5)437 (30.5)
I or II689 (86.7)106 (13.3)73.19 (<0.001)
III or IV1400 (71.1)568 (28.9)
Subtype
DLBCL1480 (74.1)518 (25.9)9.63 (0.022)
FL412 (79.1)109 (20.9)
MCL144 (80.0)36 (20.0)
BL53 (82.8)11 (17.2)
ECOG performance status
0 or 11863 (78.0)525 (22.0)54.40 (<0.001)
2–4226 (60.3)149 (39.7)
LDH
Normal1109 (83.1)225 (16.9)78.45 (<0.001)
Elevated980 (68.6)449 (31.4)
B symptoms
Absent1686 (77.1)500 (22.9)12.74 (<0.001)
Present403 (69.8)174 (30.2)
BCL6 expression
Negative636 (73.8)226 (26.2)2.12 (0.14)
Positive1453 (76.4)448 (23.6)
BCL2 expression
Negative758 (78.5)207 (21.5)6.72 (0.009)
Positive1331 (74.0)467 (26.0)
Number of Extranodal sites
≤11419 (77.9)402 (22.1)15.19 (<0.001)
>1670 (71.1)272 (28.9)
Bulk disease
No1392 (78.2)388 (21.8)17.89 (<0.001)
Yes697 (70.9)286 (29.1)
Anemia
No1317 (81.2)304 (18.8)66.90 (<0.001)
Yes772 (67.6)370 (32.4)
Albumin
Low662 (67.5)319 (32.5)53.75 (<0.001)
High1427 (80.1)355 (19.9)
Creatinine
Low (≤95.5)1674 (76.4)516 (23.6)3.75 (0.053)
High (>95.5)415 (72.4)158 (27.6)
Alkaline phosphate
Low (≤83.5)1090 (78.5)298 (21.5)12.61 (<0.001)
High (>83.5)999 (72.7)376 (27.3)
Bilirubin
Low (≤40.5)2052 (75.9)653 (24.1)3.85 (0.050)
High (>40.5)37 (63.8)21 (36.2)
PNI
Low (≤40.93)626 (67.8)297 (32.2)44.90 (<0.001)
High (>40.93)1463 (79.5)377 (20.5)
SII
Low (≤1686.985)1651 (79.1)436 (20.9)55.97 (<0.001)
High (>1686.985)438 (64.8)238 (35.2)
SIRI
Low (≤3.529)1542 (79.2)406 (20.8)44.52 (<0.001)
High (>3.529)547 (67.1)268 (32.9)
MLR
Low (≤0.611)1479 (79.0)394 (21.0)34.99 (<0.001)
High (>0.611)610 (68.5)280 (31.5)
PLR
Low (≤274.773)1561 (78.8)419 (21.2)38.96 (<0.001)
High (>274.773)528 (67.4)255 (32.6)
NLR
Low (≤5.123)1503 (79.3)397 (20.7)43.40 (<0.001)
High (>5.123)586 (67.6)281 (32.4)
IPI risk group
Low625 (89.0)77 (11.0)117.27 (<0.001)
Low intermediate592 (76.4)183 (23.6)
High intermediate513 (70.4)216 (29.6)
High359 (64.5)198 (35.5)
Revised IPI risk group
Low159 (91.4)15 (8.6)87.97 (<0.001)
Intermediate1058 (81.2)245 (18.8)
High872 (67.8)414 (32.2)
NCCN-IPI risk group
Low175 (88.8)22 (11.2)106.35 (<0.001)
Low intermediate919 (82.7)192 (17.3)
High intermediate819 (70.7)340 (29.3)
High176 (59.5)120 (40.5)
MLR, monocyte-to-lymphocyte ratio; NLR, Neutrophil-to-Lymphocyte Ratio; PLR, Platelet-to-Lymphocyte Ratio; PNI, Prognostic nutrition index; SII, Systemic Immune-Inflammation Index; SIRI, Systemic Inflammation Response Index; IPI, International Prognostic Index; NCCN-IPI, National Comprehensive Cancer Network-IPI.
Table 2. Performance of machine learning algorithms and IPI based scoring systems.
Table 2. Performance of machine learning algorithms and IPI based scoring systems.
ModelDiscrimination and Classification MetricsCalibration
AUCAccuracySensitivitySpecificityPPVNPVBrier Score
IPI0.650.600.610.590.310.830.235
R-IPI0.610.600.610.590.310.830.239
NCCN-IPI0.630.550.680.510.300.840.238
LR0.700.620.700.600.350.870.227
RF0.690.620.730.590.350.880.298
XgBoost0.700.610.730.580.340.880.231
KNN0.690.610.760.570.350.890.233
GBM0.690.610.680.600.340.860.232
SVM0.690.620.700.590.340.870.229
NB 0.700.680.580.690.360.850.223
IPI, International Prognostic Index; R-IPI, Revised-IPI; NCCN-IPI, National Comprehensive Cancer Network-IPI; LR, logistic regression; RF, random forest; XgBoost, extreme gradient boosting; KNN, K-nearest neighbor; GBM, gradient boosting; SVM, support vector machine; NB, Naïve Bayes; PPV, positive predictive value; NPV, negative predictive value.
Table 3. Treatment response across risk groups of Models in the validation set.
Table 3. Treatment response across risk groups of Models in the validation set.
ModelRisk Groups (Score)
NomogramLow (<138)LI (138–188)HI (188–225)High (≥225)
Risk factors and scoringStage (I/II = 0; III/IV = 100)
ECOG PS (≤1 = 0; >1 = 57)
BCL2 Expression (Negative = 0; Positive = 36)
SII (Low = 0; High = 70)
Anemia (No = 0; Yes = 37)
LDH (Normal = 0; Elevated = 52)
Frequency (%)218 (39.4)125 (22.6)93 (16.8)117 (21.2)
Treatment responseComplete (%)90.974.467.760.7
Incomplete (%)9.125.632.339.3
IPILow (0–1)LI (2)HI (3)High (4–5)
Risk factors and scoringAge (≤60 = 0; >60 = 1)
Stage (I/II = 0; III/IV = 1)
LDH (Normal = 0; Elevated = 1)
ECOG PS (≤1 = 0; >1 = 1)
Extranodal Sites (≤1 = 0; >1 = 1)
Frequency (%)136 (25.6)167 (30.2)140 (25.3)110 (19.9)
Treatment responseComplete (%)89.778.477.957.3
Incomplete (%)10.321.622.142.7
R-IPIVery good (0)Good (1–2)Poor (3–5)
Risk factors and scoringAge (≤60 = 0; >60 = 1)
Stage (I/II = 0; III/IV = 1)
LDH (Normal = 0; Elevated = 1)
ECOG PS (≤1 = 0; >1 = 1)
Extranodal Sites (≤1 = 0; >1 = 1)
Frequency (%)30 (5.4)273 (49.4)250 (45.2)
Treatment responseComplete (%)93.382.468.8
Incomplete (%)6.717.631.2
NCCN-IPILow (0–1)LI (2–3)HI (4–5)High (6–8)
Risk factors and scoringAge (<40 = 0; 41–60 = 1; 61–75 = 2; >75 = 3)
LDH (≤ULN = 0; >ULN–≤3 × ULN = 1; >3 × ULN = 2)
Stage (I/II = 0; III/IV = 1)
ECOG PS (≤1 = 0; >1 = 1)
Major Extranodal Sites (No = 0; Yes = 1)
Frequency (%)34 (6.1)226 (40.9)232 (42.0)61 (11.0)
Treatment responseComplete (%)91.283.274.155.7
Incomplete (%)8.816.825.944.3
LI; Low Intermediate, HI; High Intermediate, LDH; Lactate Dehydrogenase, ECOG PS; Eastern Cooperative Oncology Group performance status, IPI; International Prognostic Index, R-IPI; Revised-IPI, NCCN-IPI; National Comprehensive Cancer Network-IPI, SII; Systemic Immune-Inflammation Index.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Alem, A.Z.; Mohanty, I.; Pati, N.; Wellard, C.; Chung, E.; Hawkes, E.A.; McQuilten, Z.K.; Wood, E.M.; Opat, S.; Niyonsenga, T. Development and Validation of Prognostic Models for Treatment Response of Patients with B-Cell Lymphoma: Standard Statistical and Machine-Learning Approaches. J. Clin. Med. 2025, 14, 7445. https://doi.org/10.3390/jcm14207445

AMA Style

Alem AZ, Mohanty I, Pati N, Wellard C, Chung E, Hawkes EA, McQuilten ZK, Wood EM, Opat S, Niyonsenga T. Development and Validation of Prognostic Models for Treatment Response of Patients with B-Cell Lymphoma: Standard Statistical and Machine-Learning Approaches. Journal of Clinical Medicine. 2025; 14(20):7445. https://doi.org/10.3390/jcm14207445

Chicago/Turabian Style

Alem, Adugnaw Zeleke, Itismita Mohanty, Nalini Pati, Cameron Wellard, Eliza Chung, Eliza A. Hawkes, Zoe K. McQuilten, Erica M. Wood, Stephen Opat, and Theophile Niyonsenga. 2025. "Development and Validation of Prognostic Models for Treatment Response of Patients with B-Cell Lymphoma: Standard Statistical and Machine-Learning Approaches" Journal of Clinical Medicine 14, no. 20: 7445. https://doi.org/10.3390/jcm14207445

APA Style

Alem, A. Z., Mohanty, I., Pati, N., Wellard, C., Chung, E., Hawkes, E. A., McQuilten, Z. K., Wood, E. M., Opat, S., & Niyonsenga, T. (2025). Development and Validation of Prognostic Models for Treatment Response of Patients with B-Cell Lymphoma: Standard Statistical and Machine-Learning Approaches. Journal of Clinical Medicine, 14(20), 7445. https://doi.org/10.3390/jcm14207445

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop