Lymphoma is one of the most diagnostically complex malignancies in oncology, encompassing dozens of distinct subtypes that each carry different prognoses and require different treatments. The WHO classification system, currently the most widely used framework, categorizes lymphomas into subtypes based on morphology, immunophenotype, and genetic characteristics. Without precise subtyping, treatment decisions are made on faulty foundations, directly harming patient outcomes.
The role of immunohistochemistry: Immunohistochemistry (IHC) has become the workhorse technique for lymphoma differential diagnosis because it is more cost-effective and faster than DNA-based or RNA-based molecular assays. IHC identifies specific proteins expressed on or within tumor cells by applying antibody-stained tissue sections. For example, CD20 is the most widely used pan-B-cell marker, while CD3 is the standard pan-T-cell antigen. These markers, when applied in panels, allow pathologists to distinguish B-cell lymphoma from T-cell lymphoma, and to further subdivide those categories into clinically meaningful variants.
The problem this study addresses: While several prior machine learning algorithms have demonstrated good classification performance on individual subtyping tasks, none had integrated those tasks into a complete, sequential diagnostic workflow that mirrors how pathologists actually work through a differential diagnosis in clinical practice. This study aimed to build exactly that: a streamlined, hierarchical machine learning system that handles the full lymphoma diagnostic cascade using EBER in situ hybridization and IHC data, from distinguishing Hodgkin from non-Hodgkin disease all the way down to classifying DLBCL cell-of-origin subtypes.
The authors also recognized that IHC marker panels are not standardized across institutions. Their analysis confirmed discrepancies between the two cancer centers involved in the study, suggesting that local practice, rather than a universal evidence base, often drives panel selection. Reducing panel size without sacrificing diagnostic accuracy would have direct cost-saving implications at scale.
This is a multicenter observational study drawing from two of China's leading cancer institutions: Sun Yat-sen University Cancer Center in Guangzhou (internal cohort) and Peking University Cancer Hospital and Institute in Beijing (external cohort). Pathological reports diagnosed as lymphoma were retrospectively retrieved and labeled according to the 4th revised WHO classification of Hematopoietic Tumors. Reports with multiple simultaneous lymphoma diagnoses were excluded, ensuring the study captured only primary diagnoses and not recurrences or composite cases.
Cohort structure: The internal cohort (January 2015 to August 2021, N = 8,808) was split into two temporally independent sets: an internal primary-stage cohort (January 2015 to June 2019, N = 4,263) randomly divided into training and internal validation sets, and an internal later-stage cohort (July 2019 to August 2021, N = 4,545) used as a temporally independent validation set. The external cohort (January 2015 to December 2019, N = 6,119) served as a geographically independent validation set. This three-way validation structure, with one internal cross-validation set, one temporal holdout, and one external center, is substantially more rigorous than the single-split validation used by most comparable studies.
Marker extraction: EBER in situ hybridization and IHC marker names and values were extracted from pathological reports using a pattern-based natural language processing (NLP) method. Researchers validated the extraction accuracy and found it reliable. Although hundreds of distinct markers appeared across reports, most were not routinely used. Experienced pathologists therefore designated specific subsets of candidate markers for each of the five classification tasks, creating task-specific filtered datasets of varying sizes (for example, 5,057 patients in the training set for NHL subtype differentiation).
Class distribution: Within the training set for HL vs. NHL classification, 84.81% of cases (N = 1,507 of 1,777) were non-Hodgkin lymphoma, reflecting the real-world epidemiology of the disease. Among NHL subtypes in the training set, B-cell lymphoma comprised 77.67% of cases (N = 2,132 of 2,745), T-cell lymphoma 13.30% (N = 365), and NK-cell lymphoma 9.03% (N = 248). Within B-cell cases, DLBCL was the most prevalent subtype at 45.93% of all B-cell lymphomas (N = 627 of 1,365 training cases).
The system development proceeded in three distinct stages. The first stage identified the optimal machine learning algorithm by training six candidate models on the full panel of markers for each of the five hierarchical classification tasks. The six algorithms tested were logistic regression (LR), support vector machine (SVM), multi-layer perceptron (MLP), decision tree (DT), random forest (RF), and extreme gradient boosting (XGBoost). For tree-based methods (DT, RF, XGBoost), each marker was encoded as a categorical feature with three states: negative, positive, or unknown. For LR, SVM, and MLP, one-hot encoding converted each marker into two binary features.
Five hierarchical classification tasks: The system is structured as a five-step decision cascade. Step 1 classifies all lymphoma patients into NHL or HL using markers CD15, CD5, LCA, and CD30. Step 2 classifies NHL into B-cell, T-cell, or NK-cell lymphoma using CD30, CD5, CD56, CD10, EBERs, CD20, and TIA1. Step 3 classifies B-cell lymphoma into nine variants (DLBCL, FL, MZL, MCL, BL, SLL, PMLBCL, B-LBL/ALL, and others) using CD30, CD5, EBERs, CD20, CD10, TdT, CD23, and CyD1 as the simplified panel. Step 4 classifies T-cell lymphoma into four variants (ALCL, AITL, TLBL, PTCL-U) using CD30, CD10, EBERs, CD56, TIA1, CD5, TdT, ALK, and CXCL-13. Step 5 classifies DLBCL into GCB and non-GCB cell-of-origin subtypes using CD10 and MUM-1 only.
Panel simplification strategy: The second stage simplified the full-panel models to reduce the number of markers required in practice. Markers were ranked by importance, using absolute values of LR coefficients. They were then added to the model in descending order of importance until the AUC reached 99% of the reference AUC from the full model (or 95% for B-cell lymphoma subtypes, given the greater difficulty of that task). Critically, markers already tested at a higher level of the hierarchy were prioritized before introducing new markers, enabling systematic marker reuse across the cascade and minimizing the total number of unique tests ordered.
Performance evaluation: The third stage evaluated each simplified model across all three validation sets on AUC (computed one-vs-rest for multi-class tasks), accuracy, macro-precision, macro-recall, and macro-F1. Economic effectiveness was quantified by comparing the number of markers the system required to reach a final subtype diagnosis with the number actually ordered in real-world clinical practice at each center.
A central finding of the algorithm comparison stage is that logistic regression, the simplest model in the comparison set, outperformed decision trees, random forest, SVM, MLP, and XGBoost when evaluated by mean AUC across all three validation sets. The mean AUCs for LR were 0.967 (internal validation), 0.961 (temporally independent validation), and 0.958 (external validation). This result is counterintuitive given the widespread assumption that more flexible, nonlinear models would outperform linear classifiers on complex multi-class medical data.
Why LR won: The authors' interpretation is that lymphoma pathological subtypes can be meaningfully separated using a linear combination of IHC marker values with varying weights. This is clinically coherent, as the diagnostic logic pathologists use when reading IHC panels is itself largely additive. For example, a positive CD10 and negative MUM-1 pattern pushes toward GCB-DLBCL, while the opposite pattern supports non-GCB classification. A linear model captures this logic naturally. The interpretability advantage of LR is also clinically valuable: the model coefficients for each marker in each task directly reveal which markers drive classification in which direction, providing a mechanistic explanation consistent with known lymphoma biology.
Reference AUCs from full LR models: For the five tasks in order, the Reference AUCs (from full-panel LR models on the training set) were: HL vs. NHL = 0.966, NHL subtypes = 0.995, B-cell lymphoma subtypes = 0.973, T-cell lymphoma subtypes = 0.993, and DLBCL subtypes = 0.983. The B-cell subtype task had the lowest mean AUC across validation sets (0.934), reflecting the genuine diagnostic difficulty of distinguishing morphologically similar B-cell entities such as marginal zone lymphoma and low-grade follicular lymphoma on IHC data alone.
Comparison to prior work: Prior studies on IHC-based DLBCL classification, such as those by Hans and Choi, relied on decision tree algorithms and typically used between two and five antibodies. The LR approach in this study achieved superior discriminatory performance compared to DT. More broadly, the result reinforces an emerging consensus in clinical ML that simple, well-regularized models often generalize better than complex architectures when the feature count is modest relative to sample size.
After simplification, the hierarchical system was evaluated across all five classification tasks using three independent validation sets. For the task of HL vs. NHL, AUCs were 0.961 (100.08% of the full model) in internal validation, 0.956 (99.99%) in temporal independent validation, and 0.961 (100.10%) in external validation. The simplified panel for this first-level task requires only four markers: CD15, CD5, LCA, and CD30. This result is particularly notable because it achieved equivalent discrimination to a model trained on the full candidate panel while using a fraction of the markers.
NHL subtype differentiation: For three-way classification of B-cell, T-cell, and NK-cell lymphoma, AUCs were 0.980 (98.71% of full model) internally, 0.983 (99.00%) temporally, and 0.984 (99.22%) externally, using CD30, CD5, CD56, CD10, EBERs, CD20, and TIA1. These results are robust across all three cohorts and demonstrate that the simplified seven-marker panel is essentially equivalent in discriminatory power to the much larger full candidate panel.
B-cell and T-cell subtype differentiation: The B-cell subtype task (nine classes: DLBCL, FL, MZL, MCL, BL, SLL, PMLBCL, B-LBL/ALL, others) was the most challenging. AUCs were 0.853 (90.67% of full model), 0.884 (92.99%), and 0.869 (95.27%) across the three validation sets. These numbers are lower than the other tasks but were obtained with the relaxed 95% retention threshold applied specifically because of the task's greater inherent difficulty. For T-cell subtypes (four classes: ALCL, AITL, TLBL, PTCL-U), AUCs were 0.971, 0.959, and 0.955, representing 99.45%, 100.84%, and 100.37% of corresponding full model performance. DLBCL cell-of-origin classification using just CD10 and MUM-1 achieved AUCs of 0.960, 0.948, and 0.962 (99.50%, 99.35%, and 98.91% of full model), validating the two-marker panel against the Hans algorithm benchmark.
Overall mean AUC across tasks: Pooling results across all five classification tasks and three validation cohorts, mean AUC values were 0.959 (HL vs. NHL), 0.983 (NHL subtypes), 0.868 (B-cell subtypes), 0.962 (T-cell subtypes), and 0.957 (DLBCL subtypes). These figures represent the performance of the simplified, clinically deployable system, not the unconstrained full-panel models, and they hold up consistently across two temporal horizons and an independent external center.
One of the study's most clinically actionable findings concerns the economic impact of the hierarchical simplified panel system. The researchers compared the average number of IHC markers required by their system to reach a final subtype diagnosis against the number actually ordered in real-world clinical practice at each center. The differences were statistically significant across nearly all comparisons (p less than 0.001).
Internal primary-stage cohort (2015-2019, N = 4,263): The hierarchical system required an average of 11.12 markers to reach a final diagnosis, compared to 13.31 markers actually ordered in clinical practice, representing a saving of 16.45% (p less than 0.001). Savings varied by subtype: HL at the first branch of the hierarchy required only 4 markers (compared to 13.58 in practice, a 70.54% saving), while B-cell NHL subtypes required 12 markers (compared to 13.5, a modest 11.11% saving).
Internal later-stage cohort (2019-2021, N = 4,545): The system required an average of 10.95 markers versus 13.99 actually ordered, a saving of 21.73% (p less than 0.001). Savings ranged from 70.70% for HL to 6.10% for B-NHL subtype differentiation. The larger overall saving in this more recent cohort may reflect increasing panel size in clinical practice over time, amplifying the gap between the optimized system and real-world ordering behavior.
External cohort (2015-2019, N = 6,119): The external cohort showed the smallest mean saving at 3.67% (10.77 system markers vs. 11.18 in practice, p less than 0.001). For some subtype pathways in the external cohort, the system actually required more markers than were ordered clinically, with negative saving percentages of -2.04% (B-NHL subtypes) and -3.17% (T-NHL subtypes). The likely explanation is that pathologists at Peking University Cancer Hospital integrated IHC results from referring hospitals where patients had been treated previously, but those external markers were not captured in the electronic system and therefore not counted in the clinical denominator. This data capture gap artificially deflates the apparent clinical panel sizes at the external center.
Single-ethnicity cohort: Both the internal and external centers are large Chinese national cancer centers, and the entire dataset of 14,927 patients consists exclusively of Chinese patients. Lymphoma subtype prevalence differs between East Asian and Western populations, with some variants (such as NK/T-cell lymphoma) being considerably more common in Asian populations. It is not clear whether the marker importance rankings and simplified panels identified in this study would transfer to European or North American cohorts with different subtype distributions and potentially different clinical marker ordering practices.
Missing data and IHC incompleteness: The study's inclusion criteria required that reports contain at least the minimum markers specified for each task. Reports falling below the minimum were excluded, potentially introducing a selection bias toward cases where the clinical team had already ordered a sufficiently comprehensive panel. High rates of missing marker data across all cohorts could add inaccuracy to the findings if the patterns of missingness are informative (for example, if certain subtypes are more likely to have incomplete panels due to diagnostic certainty at an earlier point in the workup).
IHC-only feature set: The system relies exclusively on EBER in situ hybridization and IHC markers as predictors. For most classification tasks, this is appropriate and clinically pragmatic. However, for B-cell lymphoma subtype classification (the task with the lowest AUC), the authors acknowledge that additional morphological, molecular, clinical, and viral/immune status information would likely improve discrimination. B-cell lymphoma entities such as MZL and low-grade FL are frequently difficult to distinguish on IHC alone even for expert pathologists, and their reliable separation may require FISH or gene sequencing data.
System scope: The hierarchical system was designed for settings where a preliminary diagnosis of lymphoma has already been made. It is not intended as a comprehensive differential diagnosis tool for all lymphoid lesions. Cases where the starting diagnosis is uncertain, or where the primary differential includes reactive lymphoid proliferations, lie outside the system's design scope. This limits its utility to cases that have already passed through at least an initial clinical assessment.
The study's hierarchical design offers advantages that go beyond raw classification accuracy. By mimicking the stepwise human diagnostic workflow, it is structured for clinical integration in a way that an all-in-one classifier is not. At each level, only the markers relevant to that specific decision need to be ordered, and markers already tested at a higher level of the hierarchy are automatically reused at lower levels, preventing redundant testing. The interpretable LR framework means that individual marker coefficients can be inspected, providing pathologists with a transparent explanation of why a particular subtype was assigned.
DLBCL cell-of-origin classification and the Hans algorithm: The two-marker simplified panel for DLBCL subtyping (CD10 and MUM-1) produced results consistent with the Hans algorithm. Prior studies, including those by Hans et al. (Blood, 2004) and Choi et al. (Clinical Cancer Research, 2009), typically used decision tree algorithms and evaluated two to five antibodies including CD10, Bcl-6, FOXP1, GCET1, and MUM-1. The current system matched the Hans algorithm's classification output without requiring Bcl-6, streamlining the panel while preserving accuracy, and did so using a logistic regression framework that is more amenable to probabilistic uncertainty quantification than rule-based decision trees.
Panel standardization: A persistent problem in lymphoma IHC practice is the lack of consensus on which marker panels to use for specific diagnostic tasks. The variation the authors observed between their two centers' real-world panels illustrates this directly. The simplified panels identified in this study could serve as evidence-based recommendations grounded in importance-ranked LR coefficients validated across multiple independent cohorts, providing a more rigorous foundation for panel standardization than informal expert consensus.
Future directions: The authors identify several extensions that could strengthen the system. Incorporating morphological features, clinical data, and molecular markers would address the B-cell subtype task's weakest performance zone. Validation in Western or ethnically diverse cohorts is needed to establish generalizability. Prospective real-world deployment studies, measuring time to diagnosis and cost-per-case alongside classification accuracy, would provide the evidence base needed for regulatory and institutional adoption. Federated learning frameworks could enable multi-institutional model refinement without requiring centralized data sharing, an important consideration for systems drawing on sensitive pathological data.