Mantle cell lymphoma (MCL) is a rare, incurable B-cell malignancy with highly heterogeneous clinical outcomes. The majority of patients relapse after treatment or experience disease progression, yet a clinically meaningful subset maintains an indolent course or responds exceptionally to frontline therapy and achieves durable remission. This heterogeneity makes pretreatment stratification essential for therapeutic decision-making, particularly as targeted and cellular therapies are increasingly explored as frontline alternatives to high-dose chemoimmunotherapy.
Limitations of existing indices: The mantle cell lymphoma international prognostic index (MIPI) and its variants, the biological MIPI (MIPI-b) and simplified MIPI (MIPI-s), were developed using data from patients treated with intensive frontline regimens. These tools incorporate only a handful of clinical variables (age, LDH, ECOG performance status, WBC count, and Ki-67% in the MIPI-b) and do not incorporate molecular or histopathologic features from cytogenetics or next-generation sequencing (NGS). A later iteration, the engineered MIPI (eMIPI), integrated seven clinical features with a clustering algorithm but still excluded cytogenetic and genomic data, leaving a significant portion of disease biology unrepresented.
The case for machine learning: Known prognostic biomarkers in MCL include TP53 mutations, complex karyotype (three or more unrelated cytogenetic abnormalities), blastoid or pleomorphic histology, and a variety of genomic alterations. Studies from the Nordic MCL2 and MCL3 cohorts demonstrated that TP53 mutations independently predicted inferior overall survival (OS) and poor response to intensive chemoimmunotherapy. Prior prognostic models incorporating molecular features have consistently outperformed purely clinical approaches, supporting the rationale for a more comprehensive integrative model.
This 2023 study from MD Anderson Cancer Center represents the largest patient cohort deployed in an integrative prognostic ML model for MCL, using 862 patients diagnosed between 2014 and 2022. The work culminates in two practical tools: the integrative MIPI (iMIPI, 20-feature model) and the integrative simplified MIPI (iMIPI-s, 10-feature model), both deployed as a REST API proof of concept for potential clinical implementation.
The research team retrospectively curated an extensive electronic medical record (EMR) database of 862 patients with MCL treated at MD Anderson between January 1, 2014 and June 30, 2022, with outcome data updated through January 2023. Patients seen only for consultation or second opinion were excluded. Of the 862 patients extracted, 68 were removed due to loss to follow-up without a documented outcome (death, relapse, or disease progression), yielding a final analytic cohort of 794 patients. The training set comprised 595 patients (75%) and the test set 199 patients (25%), with stratification by outcome class to preserve class distribution across splits.
Feature categories and engineering: Baseline features were drawn from three data domains: clinicopathologic measures (demographics, laboratory values, morphology, Ki-67%, ECOG performance status, B-symptoms, BMI, smoking status, alcohol use, gastrointestinal involvement), cytogenetics (FISH, comparative genomic hybridization, conventional karyotyping), and genomics (somatic and germline mutations from targeted sequencing panels). A total of 129 predictive features were assembled after removing zero-variance and near-zero variance (NZV) features. Composite engineered variables included an interaction term for co-occurring TP53 and ATM mutations and an "abnormal karyotype" variable capturing patients with three or more cytogenetic aberrations not formally coded as complex by pathologists. Low-frequency chromosomal abnormalities were consolidated at the chromosome level to prevent NZV removal from eliminating biologically meaningful signals.
Outcome classification: Patients were dichotomized into two groups. "Aggressive" MCL encompassed those who relapsed, progressed after frontline treatment, or died without treatment due to aggressive disease onset. "Indolent/responsive" MCL included patients who did not require treatment or remained in remission after frontline therapy. This binary classification served as the target variable for XGBoost classification models. For survival analyses, progression-free survival (PFS) was defined as time from treatment to relapse, progression, or death, and OS as time from diagnosis to death.
XGBoost algorithm selection and preprocessing: Extreme gradient-boosted (XGBoost) ensemble tree models were selected for their well-documented advantages in oncologic datasets: they handle large amounts of missing data natively through a sparsity-aware split-finding algorithm (avoiding imputation bias), accommodate collinear features by distributing importance across correlated variables, handle class imbalance by up-weighting minority class errors in successive iterations, and capture nonlinear relationships between features and outcomes. Hyperparameter tuning used a grid-based Latin hypercube approach of size 50 over 10-fold cross-validation, with early stopping to prevent overfitting. Dummy variables were created for all nominal predictors, and categories with less than 1% representation were collapsed into an "other" bin.
Four primary XGBoost models were constructed and evaluated: a clinical data-only model, a clinical plus NGS model, a clinical plus cytogenetic model, and an all-features model. Two parsimonious models using the top 10 and 20 features identified by variable importance were also built and designated the iMIPI-s and iMIPI, respectively. The all-features model using all 129 features achieved the best test set performance: AUC ROC of 0.83, accuracy of 0.77, sensitivity of 68.7%, and specificity of 81.9%. The positive predictive value (PPV) was 0.73 and the negative predictive value (NPV) was 0.79.
Parsimonious models: The 20-feature iMIPI achieved AUC of 0.82 on both training and test sets, with sensitivity of 70.0% and specificity of 76.7%. The 10-feature iMIPI-s had AUC of 0.81 on both sets, with sensitivity of 67.0% and specificity of 78.0%. Despite using only 10 features, the iMIPI-s maintained competitive diagnostic performance while offering substantially improved clinical scalability. Both parsimonious models outperformed the traditional MIPI (AUC 0.64) and MIPI-b (AUC 0.66) by 0.15-0.18 AUC units on the test set, with accuracy improvements from 0.60 to 0.74-0.75.
Comparison with other approaches: The Clinical + NGS model performed nearly as well as the all-features model (AUC 0.82, sensitivity 72.3%), but generated more false positives (specificity 78.5% vs. 81.9%). The clinical-only model reached AUC 0.80, demonstrating that clinical variables alone carry substantial prognostic information. Cytogenetics-only (AUC 0.67) and NGS-only (AUC 0.61) models performed substantially worse than combined models, confirming that integrating feature types is essential. A random forest model built on the same top 20 features achieved AUC 0.79 on the test set with sensitivity of only 53%, notably inferior to XGBoost on this dataset.
Complete-case cytogenetic model: To assess the XGBoost sparsity-aware algorithm, a separate model was built using only the 616 patients with complete cytogenetic data. This model performed well on the training set (AUC 0.84) but showed reduced performance on the test set (AUC 0.80, accuracy 0.75) with markedly lower sensitivity (59%). This result indicates that building models with full patient sets containing missing data preserves generalizability better than restricting analyses to complete-case subsets, a finding with practical implications for clinical deployment where molecular data availability is uneven.
Variable importance plots (VIPs) and Shapley additive explanation (SHAP) values were used to identify and interpret the most influential predictors across models. SHAP values, rooted in cooperative game theory, quantify each feature's marginal contribution to the model's output for individual predictions. Unlike traditional regression coefficients, SHAP values accommodate nonlinear relationships and feature interactions, making them particularly informative in complex ensemble models.
Top prognostic features in the all-features model: The highest-ranked variables in the full XGBoost model were lactate dehydrogenase (LDH), Ki-67 proliferation index (%), platelet count, bone marrow involvement percentage, beta-2 microglobulin (B2M), WBC count, age at diagnosis, and TP53 mutational status. Morphology (blastoid or pleomorphic histology vs. classic) also ranked highly, as did modifiable exposure variables including smoking status and alcohol consumption, the latter emerging as an unexpected but consistently impactful predictor that the authors note warrants further investigation alongside genomic and transcriptomic interaction data.
The iMIPI-s 10-feature set: The simplified model retained LDH, Ki-67%, platelet count, bone marrow involvement percentage, hemoglobin, total number of observed somatic mutations, TP53 mutational status, ECOG performance status, B2M, and morphology. The total somatic mutation count, rather than any individual gene mutation, emerged as a consistently important predictor, likely serving as a surrogate measure of tumor mutational burden. This finding is supported by the observation that ATM, the most frequently mutated gene in MCL, did not rank among the top prognostic features in any model, suggesting that ATM mutation frequency does not directly translate to prognostic impact.
Molecular vs. cytogenetic contributions: NGS-derived features contributed more to prognostic predictions than cytogenetic features across models. The complete-case cytogenetic model included only one cytogenetic feature (complex karyotype) among its top variables, suggesting that cytogenetic processes may be partly captured by co-correlated biological markers. As NGS becomes more routinely implemented in clinical practice, cytogenetic testing may become less critical for prognostication, with targeted sequencing panels providing higher-resolution and more impactful information.
Survival analyses extended the binary classification findings into time-to-event modeling. Cox proportional hazards (PH) models were constructed using the top 10 and 20 features identified by XGBoost variable importance. The 20-feature OS model achieved a concordance index (C) of 0.76, and the PFS model reached C = 0.74. The 10-feature OS and PFS models achieved C = 0.75 and 0.71, respectively. These concordance indices are modestly lower than the AUC values from XGBoost classification models, reflecting the inherent challenge of time-to-event censoring compared to binary classification, and demonstrating that XGBoost feature selection can inform robust survival models.
Cox regression findings for PFS: In the PFS model, Ki-67% emerged as a highly significant predictor (HR per unit increase, p less than 0.001). LDH was also significant (HR ~1.24, p less than 0.001). B-symptoms at baseline carried an HR of 1.42 (p less than 0.05). Pleomorphic morphology had an HR of 2.03 (p less than 0.01) relative to blastoid histology, and nonsmoking status was protective (HR 0.69, p less than 0.01). TP53 mutation status reached statistical significance in PFS (HR 1.56, p less than 0.01). Age at diagnosis was a significant predictor (HR ~1.18, p less than 0.05).
Cox regression findings for OS: In the OS model, Ki-67% remained highly significant (HR ~1.63, p less than 0.001). Hemoglobin showed a protective association (HR 0.82, p less than 0.05). ECOG status 3 carried an HR of 6.89 for OS (p less than 0.001), and ECOG 2 had HR 3.10 (p less than 0.05). Pleomorphic morphology reached HR 2.09 (p less than 0.05). Nonsmoking was again protective in OS (HR 0.62, p less than 0.01). Notably, TP53 mutation did not reach statistical significance in the 20-feature OS model despite its clear importance in the classification models, likely reflecting differences between predicting binary outcome class and time-to-event outcomes with varying follow-up durations.
Kaplan-Meier curves: Kaplan-Meier analyses confirmed significant survival differences across categories of ECOG performance status, morphology (classic vs. aggressive histology), TP53 mutational status, B2M level (cutoff 1.8 mcg/mL), bone marrow involvement (cutoff 5%), and platelet count (cutoff 150,000 per microliter). All group differences were statistically significant by log-rank test. Total mutation count above 5 genes did not reach significance for OS or PFS in the Kaplan-Meier analysis, suggesting its prognostic value is better captured by the ensemble ML model than by simple binary threshold stratification.
A multivariate generalized linear model (GLM) with logistic regression was constructed using the same top 20 features identified by XGBoost variable importance, enabling a direct performance comparison between the ML ensemble approach and traditional statistical modeling on identical feature sets. This comparison is methodologically important because clinical adoption of complex ML models requires demonstrating superiority over interpretable statistical methods that clinicians are already familiar with.
GLM performance metrics: The logistic GLM achieved AUC of 0.77 on the test set, accuracy of 0.72, sensitivity of 61%, and specificity of 79%. These results are meaningfully lower than the 20-feature XGBoost iMIPI model, which reached AUC 0.82, accuracy 0.75, sensitivity 70%, and specificity 77% on the same test set. The AUC gap of 0.05 reflects XGBoost's ability to capture nonlinear relationships and feature interactions that a logistic regression model cannot represent without explicit interaction term specification.
GLM coefficient interpretation: The logistic regression results provided interpretable coefficient estimates that confirmed the biological plausibility of the XGBoost feature selection. Ki-67%, total number of somatic mutations, LDH, aggressive morphology, and age at diagnosis were statistically significant predictors of poor (aggressive) MCL outcomes in the GLM. Nonsmoking status was the sole protective predictor reaching significance. Several molecular features included in the model (such as specific gene mutation statuses) did not reach statistical significance individually in the GLM, even though they contributed meaningfully to XGBoost predictions through additive and interaction effects in the ensemble framework.
This discrepancy illustrates a fundamental difference between how XGBoost and logistic regression handle multivariate data: the ensemble model distributes predictive weight across many small contributions from correlated and interacting features, whereas logistic regression evaluates each variable in the context of all others, making it prone to coefficient attenuation when predictors are collinear. The XGBoost framework's ability to handle collinear features by selecting one variable from a correlated pair for each tree in the ensemble prevents the coefficient instability that affects logistic regression in high-dimensional settings.
The authors formally designate their 20-feature parsimonious XGBoost model as the integrative MIPI (iMIPI) and the 10-feature model as the integrative simplified MIPI (iMIPI-s). These designations position the tools as direct successors to the MIPI and MIPI-b within the existing clinical nomenclature framework, making their relationship to current standard tools immediately clear to lymphoma clinicians. The iMIPI incorporates LDH, platelet count, Ki-67%, hemoglobin, bone marrow involvement, B2M, WBC, age at diagnosis, TP53, CARD11, BIRC3, NOTCH1, and ROS1 mutational status, ECOG status, total somatic mutation count, morphology, BMI, B-symptoms, smoking status, and gastrointestinal involvement.
REST API deployment as proof of concept: The parsimonious XGBoost models were versioned and deployed using a REST (representational state transfer) application programming interface (API), demonstrating the feasibility of clinical implementation. Clinicians or researchers can enter any subset of the 10 or 20 model features because the XGBoost algorithm tolerates missing inputs and makes predictions from available data. The API returns a predicted disease class (aggressive vs. indolent/responsive), the probability associated with that prediction based on available data, and model monitoring information. Code and the deidentified dataset are publicly available at the authors' GitHub repository. The REST API is explicitly a proof of concept and has not received FDA approval for clinical use.
Dynamic update potential: A key design principle of the iMIPI framework is that it is intended to be dynamic. As more NGS studies are conducted and new molecular features are validated, the XGBoost model can incorporate updated feature sets and be retrained on expanded cohorts. The authors note that NOTCH1 and ROS1 mutational status, the lowest-ranked molecular features in the 20-feature model, should be reassessed as NGS data accumulates. This reflects a broader shift from static clinical indices (fixed formulas derived from cohort regression) toward adaptive ML-based prognostic tools that improve incrementally with new evidence.
The practical clinical value of the iMIPI lies in its ability to predict which patients will follow an aggressive course at baseline, before any treatment is initiated. High-confidence predictions of aggressive disease could support early enrollment in clinical trials of intensified frontline regimens or novel targeted therapies, while patients predicted as indolent could be spared unnecessary treatment toxicity from high-intensity chemotherapy. Multiple chemotherapy-free induction and maintenance strategies are currently in active trials for MCL, making reliable baseline stratification increasingly consequential.
Missing data and incomplete molecular profiling: The cohort included 794 patients with clinical data, 622 with cytogenetic data, and only 320 with genomic sequencing data, reflecting real-world patterns of molecular testing availability at a single center across an eight-year period. Genomic data were particularly sparse for patients diagnosed in earlier years, creating a potential temporal confounder. Although the XGBoost sparsity-aware algorithm handles missing data without imputation, the authors acknowledge that model performance would likely improve with more complete molecular data across all 862 patients, especially from NGS panels.
Absence of external validation: The study was conducted at a single academic institution without an independent external validation cohort. The authors attempted to address this limitation by random test-set splitting with stratification and 10-fold cross-validation, but these internal resampling procedures cannot fully substitute for external validation in a geographically and institutionally distinct patient population. They note that publicly available MCL datasets lack the granular clinicopathologic and molecular features needed to make a suitable external comparison. External validation from multi-institutional cohorts will be required before the iMIPI can be recommended for routine clinical use.
Binary outcome classification and censored data: The current model provides only a binary prediction (aggressive vs. indolent/responsive), limiting the granularity of risk stratification compared to multi-tier risk scores like the conventional MIPI. The authors acknowledge that patients recently treated at the time of data cutoff may yet relapse, potentially slightly attenuating the model's true predictive power by misclassifying some aggressive cases as indolent at follow-up. XGBoost with censored regression exists but remains technically challenging to implement with stable performance metrics and interpretable variable importance.
Future directions: The authors outline several planned extensions. Integration of RNA sequencing (RNA-seq) data would provide insight into the functional consequences of DNA mutations and enable pathway-level feature extraction. Whole-genome and whole-exome sequencing offer higher-resolution genomic profiling than current targeted panels. Time-series ML models incorporating serial circulating tumor DNA (ctDNA) measurements could capture clonal evolution, minimal residual disease dynamics, and response to therapy longitudinally. The authors also envision incorporating these techniques into precision medicine trial stratification frameworks as multi-omic correlative studies mature.