Thrombosis Risk Prediction in Lymphoma Patients: a Multi-Institutional, Retrospective Model Development and Validation Study

Abstract

Venous thromboembolism (VTE) poses significant risk to cancer patients receiving systemic therapy. The generalizability of pan-cancer models to lymphomas is limited. Currently, there are no reliable risk prediction models for thrombosis in patients with lymphoma. Our objective was to create a risk assessment model (RAM) specifically for lymphomas. We performed a retrospective cohort study to develop Fine and Gray subdistribution hazard model for VTE and pulmonary embolism (PE)/ lower extremity deep vein thrombosis (LE-DVT) respectively in adult lymphoma patients from the Veterans Affairs national healthcare system (VA). External validations were performed at the Harris Health System (HHS) and the MD Anderson Cancer Center (MDACC). Time-dependent c-statistic and calibration curves were used to assess discrimination and fit. There were 10,313 (VA), 854 (HHS), and 1,858 (MDACC) patients in the derivation and validation cohorts with diverse baseline. At 6 months, the VTE incidence was 5.8% (VA), 8.2% (HHS), and 8.8% (MDACC), respectively. The corresponding estimates for PE/LE-DVT were 3.9% (VA), 4.5% (HHS), and 3.7% (MDACC), respectively. The variables in the final RAM included lymphoma histology, body mass index, therapy type, recent hospitalization, history of VTE, history of paralysis/immobilization, and time to treatment initiation. The RAM had c-statistics of 0.68 in the derivation and 0.69 and 0.72 in the two external validation cohorts. The two models achieved a clear differentiation in risk stratification in each cohort. Our findings suggest that easy-to-implement, clinical-based model could be used to predict personalized VTE risk for lymphoma patients.

Graphical Abstract

Introduction

Patients with cancer undergoing systemic treatment are at higher risk of experiencing venous thromboembolism (VTE), especially pulmonary embolism (PE) and lower extremity deep vein thrombosis (LE-DVT), compared with the general population.^{1, 2} VTE significantly elevates risks of morbidity and mortality.³

There have been various attempts to predict and prevent VTE among the general cancer population. However, the generalizability of pan-cancer models is questioned in lymphoma settings since multiple studies indicate that risk factors for VTE in patients with lymphoma are different from risk factors in patients with solid tumors.^4–9 For example, the Khorana Score categorizes all subtypes of lymphomas as intermediate-risk cancer types with c-statistics 0.51, 0.50, and 0.60 in external lymphoma validation cohorts.^{6, 10, 11, 12} These limitations underscore a need for a VTE risk assessment model (RAM) specifically for lymphoma. Antic et al.⁶ proposed the ThroLy Score for lymphoma patients which incorporated lymphoma-specific factors like extranodal disease and mediastinal involvement; however, it did not validate well in external cohorts (c-statistic 0.55 and 0.57).^{11, 13} Two other recently derived risk prediction models still require external validation. Dharmavaram et al.⁶ incorporated only diffuse large B-cell lymphoma (DLBCL) and follicular lymphoma (FL), while TiC-LYMPHO¹¹ was derived based on genetic variables, which could be less convenient to use in general patients. Moreover, most of the current models did not incorporate nuanced histologic subtypes of lymphoma with external validation or death as a competing event for VTE.

The current study built upon recently published cohort studies that combined data from cancer registries and electronic data warehouses (EDW) from multiple institutions in the United States (U.S.) with improved phenotyping for VTE.^{14, 15} Our study aimed to derive and validate a simplified VTE RAM tailored specifically to lymphoma patients based on individual clinical and laboratory attributes to identify those at risk for thromboembolic events.

Methods

Study Design, Population

We developed and externally validated the VTE RAM using retrospective data from three separate U.S. healthcare systems (Supplemental Figure 1). Patients from incident diagnosis of lymphoma 2006–2021 from the national Veterans Affairs healthcare system (VA) formed the derivation cohort. Independent lymphoma patients from the Harris Health System (HHS) 2011–2020 and MD Anderson Cancer Center (MDACC) 2017–2020 formed the validation cohorts. VA is an integrated healthcare network with 171 medical centers and 1,113 outpatient clinics dedicated to veterans. In these integrated systems, patients are highly likely to receive continuous oncologic care with consistent long-term follow-up. HHS is a safety-net health care system featuring two medical centers and 18 outpatient clinics, serving patients of varied racial and ethnic origins in the Houston metropolitan region. MDACC is one of the largest comprehensive cancer centers in the US.

Patient selection followed the same inclusion and exclusion criteria as our previous studies.^{14, 15} In summary, patients were included if they had newly diagnosed lymphoma and underwent first-line systemic therapy within one year of diagnosis. Patients were excluded if they had recent diagnosis of acute VTE within the last 6 months or were prescribed anticoagulant within 30 days before the index date. The index date was the day systemic therapy was initiated, and all covariates were extracted on or before the index date (Supplemental Table 1).

Outcome definition

Overall VTE was defined as a radiologically verified PE, LE-DVT, or proximal or distal deep vein thrombosis in the upper extremity (UE-DVT).¹⁶ We selected the 6-month cumulative risk of VTE as our primary endpoint and PE/LE-DVT as the secondary endpoint.^{17, 18} Outcomes were ascertained using our published, institution-specific ICD-10-CM and natural language processing (NLP) radiology algorithms with a positive predictive value (PPV) of 91% at the VA health care system, 98% at HHS, and 95% at MDACC.^{14, 15}

Risk factor definition

Thirty candidate predictors were initially chosen based on clinical relevance and data availability (Supplemental Table 1). The definitions for most variables were detailed in our previous publications.^{14, 15} Lymphoma subtypes were defined in Supplemental Table 2 and additional covariates added for the current study included laboratory variables (red cell distribution width, LDH, albumin, and calcium).

Statistical Analysis

For the RAM derivation, we mandated the classification of lymphoma subtypes based on the logistic regression odds ratios (ORs) of VTE at 6 months. We used the Least Absolute Shrinkage and Selection Operator (LASSO) regression for parsimonious feature selection beyond lymphoma subtypes.^19–21 We then constructed a Fine and Gray subdistribution hazard model using the final selected variables to predict the cumulative incidence with death as competing events. The native subdistribution hazard models and beta coefficients were saved and shared on GitHub for subsequent external validation. Model discrimination and fit were assessed with time-dependent c-statistics^{14, 22} and calibration plots at 6 months for VTE or PE/LE-DVT. Internal validation was conducted using nonparametric bootstrapping. External validation was conducted by direct application of models estimates in the external cohorts without model refitting. Missing values were imputed via random forest chained imputation²³ for the initial feature selection process. Only complete cases were used for model derivation and validations. To determine risk group assignment, a pre-determined threshold of 7% for VTE at 6 months from the predicted probability was used to stratify high vs. low-risk groups.²⁴ All statistical analyses were performed with R 4.2.2 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Baseline patient characteristics

There were 10,313 (VA), 854 (HHS), and 1,858 (MDACC) patients in the derivation and validation cohorts with unique demographics, disease characteristics, treatment patterns, and comorbidities (Table 1). Patients from VA were older (median age 67), more likely to be male (96.6%), non-Hispanic White (77.8%) and non-Hispanic Black (13.5%), and had a higher proportion of small/chronic lymphocytic lymphoma/leukemia (SLL/CLL) diagnosis (12.3%). Patients from HHS were younger (median age 51), more likely to be Hispanic (57.1%) and Non-Hispanic Black (22.8%), uninsured (79.7%) with prolonged hospitalization (74.2%), and had a higher proportion of diffuse large B cell lymphoma (DLBCL, 40.5%) and classical Hodgkin lymphoma (CHL, 20.3%) requiring cytotoxic chemotherapy (95.0%). Finally, patients from MDACC had a median age of 59, were more likely to be non-Hispanic White (67.3%) and Hispanic (15.6%), insured (88.2% private insurance or Medicare), and more likely to receive targeted monotherapy (25.9%).

Table 1.

Baseline patient characteristics in derivation and validation cohorts

Characteristic	VA N=10,313	HHS N=854	MDACC N=1,858
Age, median (Q1, Q3)	67 (61, 74)	51 (40, 60)	59 (45, 69)
Male sex, N (%)	9,958 (96.6%)	500 (58.5%)	1,099 (59.1%)
Body mass index, median (Q1, Q3)	28.5 (25.1, 32.6)	27.1 (23.2, 30.8)	27.8 (24.4, 32.3)
Race/ethnicity, N (%)
Non-Hispanic White	8,024 (77.8%)	128 (15.0%)	1,251 (67.3%)
Non-Hispanic Black	1,392 (13.5%)	195 (22.8%)	118 (6.4%)
Non-Hispanic Asian Pacific Islander	211 (2.0%)	37 (4.3%)	110 (5.9%)
Hispanic	583 (5.7%)	488 (57.1%)	290 (15.6%)
Unknown	103 (1.0%)	6 (0.7%)	89 (4.8%)
Primary insurance, N (%)  Private/commercial  Medicare  Medicaid  Other/unknown  Uninsured  Tricare/VA	0 0 0 0 0 10,313 (100%)	35 (4.1%) 27 (3.2%) 100 (11.7%) 11 (1.3%) 681 (79.7%) 0	1,053 (57.8%) 564 (30.4%) 20 (1.1%) 132 (7.1%) 63 (3.4%) 26 (1.4%)
Lymphoma histology, N (%)
B-cell prolymphocytic leukemia	18 (0.2%)	2 (0.2%)	0 (0.0%)
Burkitt lymphoma/leukemia	95 (0.9%)	49 (5.7%)	21 (1.1%)
Classical Hodgkin lymphoma	767 (7.4%)	173 (20.3%)	261 (14.0%)
Small/chronic lymphocytic lymphoma/leukemia	1,269 (12.3%)	38 (4.4%)	113 (6.1%)
Cutaneous T cell lymphoma a	258 (2.5%)	2 (0.2%)	25 (1.3%)
Diffuse large B cell lymphoma b	3189 (30.9%)	346 (40.5%)	594 (32.0%)
Follicular lymphoma	1,432 (13.9%)	63 (7.4%)	273 (14.7%)
Hairy cell leukemia	170 (1.6%)	3 (0.4%)	22 (1.2%)
Lymphoplasmacytic lymphoma c	348 (3.4%)	6 (0.7%)	27 (1.5%)
Lymphoid neoplasm, NOS	1,044 (10.1%)	63 (7.4%)	26 (1.4%)
Mantle cell lymphoma	627 (6.1%)	14 (1.6%)	235 (12.6%)
Marginal zone lymphoma	530 (5.1%)	27 (3.2%)	62 (3.3%)
Nodular lymphocyte predominant Hodgkin lymphoma	43 (0.4%)	3 (0.4%)	22 (1.2%)
Primary central nervous system (CNS) large B cell lymphoma	80 (0.8%)	19 (2.2%)	33 (1.8%)
Precursor non-Hodgkin lymphoma d	20 (0.2%)	8 (0.9%)	33 (1.8%)
Systemic T cell lymphoma	423 (4.1%)	38 (4.4%)	111 (6.0%)
Stage, N (%)
I	1241 (12.0%)	103 (12.1%)	227 (12.2%)
II	1123 (10.9%)	137 (16.0%)	294 (15.8%)
III	1753 (17.0%)	126 (14.8%)	239 (12.9%)
IV	4382 (42.5%)	424 (49.6%)	1075 (57.9%)
Unknown/unstageable	1814 (17.6%)	64 (7.5%)	23 (1.2%)
Treatment (first-line), N (%)
Time to first treatment (months), median (Q1, Q3)	1.2 (0.7, 2.4)	1.0 (0.4, 2.0)	1.2 (0.7, 1.9)
Cytotoxic chemotherapy	8,547 (82.9%)	811 (95.0%)	1,336 (71.9%)
Targeted therapy	1,727 (16.7%)	43 (5.0%)	482 (25.9%)
Other therapy	39 (0.4%)	0 (0.0%)	40 (2.2%)
Complete blood count, median (Q1, Q3)
White blood cell count, ×109/L	7.5 (5.6, 10.5)	7.2 (5.3, 10.3)	7.2 (5.5, 9.8)
Hemoglobin, g/dL	12.4 (10.4, 13.9)	10.9 (9.0, 12.5)	12.9 (11.1, 14.2)
Platelet count, ×109/L	203 (145, 272)	249 (179, 342)	232 (171, 303)
Other comorbidities, N (%)
History of VTE lifetime	404 (3.9%)	5 (0.6%)	6 (0.3%)
History of hospitalization last 90d	3,837 (37.2%)	634 (74.2%)	651 (35.0%)

^aIncluding Mycosis fungoides/Sezary syndrome

^bIncluding primary mediastinal large B-cell lymphoma, primary effusion lymphoma, and intravascular large B-cell lymphoma.

^cIncluding Waldenstrom macroglobulinemia.

^dIncluding blastic plasmacytoid dendritic cell neoplasm.

Abbreviations: VA, Veterans Affairs. HHS, Health and Human Services. MDACC, MD Anderson Cancer Center. VTE, Venous Thromboembolism.

The median follow-up durations for continuous VTE assessment—censored if no clinical encounter occurred for more than 90 days—were 21.2 months at VA (IQR, 8.8–44.5), 11.5 months at HHS (IQR, 6.9–20.7), and 8.2 months at MDACC (IQR, 4.8–13.8). At 6 months post-treatment, VTE incidence in the overall cohort was 5.8% at VA, 8.2% at HHS, and 8.8% at MDACC. The corresponding figures for PE/LE-DVT were 3.9%, 4.5%, and 3.7%, respectively. The median follow-up for overall survival assessment were 26.8 months at VA (IQR, 9.4–60.1), 27.8 months at HHS (IQR, 11.3–58.3), and 10.9 months at MDACC (IQR, 4.8–13.8). At 6 months post-treatment, the number of patients who died without VTE were 1,157 (11.2%) at VA, 47 (5.5%) at HHS, and 74 (4.0%) at MDACC.

Model derivation

From an initial pool of 30 candidate predictors, 7 covariates beyond lymphoma histology met the LASSO variable selection threshold in the VA derivation cohort (Table 2, Supplemental Table 1, Supplemental Figure 2). There was a strong association between individual lymphoma histology subtype and VTE risk (Supplemental Table 3). In the final multivariable competing risk models, indolent lymphoma histology subtypes including SLL/CLL, marginal zone lymphoma (MZL), lymphoplasmacytic lymphoma (LPL), and nodular lymphocyte predominant Hodgkin Lymphoma (NLPHL) were associated with the lowest risk of VTE. Clinically heterogeneous histology subtypes including follicular lymphoma (FL), mantle cell lymphoma (MCL), and various other rare histology (B-cell prolymphocytic leukemia, Hairy cell leukemia, cutaneous T cell lymphoma) formed the intermediate cancer risk group (SHR 1.18 [95% CI, 0.97 to 1.43]). Aggressive lymphomas such as DLBCL, CHL, systemic T cell lymphoma, precursor non-Hodgkin lymphoma, and Burkitt lymphoma/leukemia were included in the high-risk group (SHR 1.58 [95% CI, 1.31 to 1.92]). Finally, primary central nervous system lymphoma (PCNSL) formed the very high-risk group (SHR 2.81 [95% CI, 1.70 to 4.63]). Other significant risk factors included pre-treatment BMI ≥35 (SHR, 1.29 [95% CI, 1.11 to 1.50]), administration of cytotoxic chemotherapy vs. targeted monotherapy (e.g. CD20 monoclonal antibody) (SHR, 1.40 [95% CI, 1.15 to 1.72]), recent hospitalization >3 days (SHR, 1.24 [95% CI, 1.09 to 1.41]), history of VTE (SHR, 2.63 [95% CI, 2.14 to 3.23]), history of paralysis or immobility (SHR, 2.63 [95% CI, 1.58 to 4.35]). Time to systemic treatment initiation was an additional protective factor (SHR, 0.96 [95% CI, 0.93 to 0.99]). These covariates maintained comparable magnitudes and significance levels when assessing the PE/LE-DVT outcome instead of the overall VTE (Table 2). The final Fine and Gray competing risk models incorporating these 7 covariates had a derivation c-statistic of 0.68 (95% CI, 0.67–0.69) for overall VTE and 0.68 (95% CI, 0.66–0.71) for PE/LE-DVT. An online calculator of the model is available at https://dynamicapp.shinyapps.io/Lymph-CAT/.

Table 2.

Final lymphoma VTE risk model in the derivation cohort

Risk factors	Total N=10,310a	SHR for VTE (95% CI)	SHR for PE/LE-DVT (95% CI)
Lymphoma histology
SLL/CLL, MZL, LPLb, NLPHL	2189 (21.2%)	1	1
FL, MCL, other histologyc	3548 (34.4%)	1.18 (0.97–1.43)	1.14 (0.90–1.41)
DLBCLd, CHL, Systemic T cell lymphoma, Precursor NHLe, BL	4493 (43.6%)	1.58 (1.31–1.92)	1.51 (1.21–1.88)
PCNSL	80 (0.8%)	2.81 (1.70–4.63)	2.97 (1.68–5.26)
Body mass index (BMI) ≥35	1619 (15.7%)	1.29 (1.11–1.50)	1.36 (1.14–1.62)
Cytotoxic chemotherapy	8546 (82.9%)	1.40 (1.15–1.72)	1.30 (1.03–1.63)
History of VTE	404 (3.9%)	2.63 (2.14–3.23)	2.78 (2.20–3.51)
History of prolonged hospitalization	3837 (37.2%)	1.24 (1.09–1.41)	1.18 (1.02–1.37)
History of paralysis/immobilization	60 (0.6%)	2.63 (1.58–4.35)	2.98 (1.67–5.30)
Time to first treatment (months), median (Q1, Q3)	1.2 (0.7, 2.4)	0.96 (0.93–0.99)	0.95 (0.91–0.99)

^aComplete case analysis

^bIncluding Waldenstrom macroglobulinemia.

^cThis category includes B-cell prolymphocytic leukemia, Cutaneous T cell lymphoma (including Mycosis fungoides/Sezary syndrome), Hairy cell leukemia, Lymphoid neoplasm, NOS.

^dIncluding primary mediastinal large B-cell lymphoma, primary effusion lymphoma, and intravascular large B-cell lymphoma.

^eIncluding blastic plasmacytoid dendritic cell neoplasm.

Abbreviations: SHR, subdistribution hazard ratio. SLL/CLL, Small Lymphocytic Lymphoma/Chronic Lymphocytic Leukemia. MZL, Marginal Zone Lymphoma. LPL, Lymphoplasmacytic Lymphoma. NLPHL, Nodular lymphocyte predominant Hodgkin lymphoma. FL, Follicular Lymphoma. MCL, Mantle Cell Lymphoma. DLBCL, Diffuse large B cell lymphoma. PCNSL, Primary Central Nervous System Lymphoma. CHL, Classic Hodgkin Lymphoma. Precursor NHL, Precursor non-Hodgkin lymphoma. BL, Burkitt lymphoma/leukemia. VTE, Venous Thromboembolism. PE/LE-DVT, Pulmonary Embolism/ Lower Extremity Deep Vein Thrombosis

Covariate distribution and model validation

Due to the inherent differences in the healthcare system for the derivation and validation cohorts, the outcome and covariate distribution in the 3 cohorts were expectedly different (Supplemental Table 4). While the derivation and validation cohorts had similar PE/LE-DVT incidence at 6 months (3.9%, 4.5%, 3.7%), the overall VTE was lower in the VA derivation cohort (5.8%, 8.2%, 8.8%). The cancer histology risk group was also significantly different. The VA derivation cohort had more low-risk subtypes (21.2% vs. 8.7% at HHS and 12.1% at MDACC) and fewer high-risk subtypes (43.6% vs. 71.9% at HHS and 54.9% at VA). Patients in the VA derivation cohort had the highest incidence of a history of VTE. As a result of increased high-risk cancer subtypes, patients from the HHS validation cohort also had the highest hospitalization rate (74.2% vs. 37.2% at VA and 35.0% at MDACC) and need for upfront chemotherapy (95.0% vs. 82.9% at VA and 71.9% at MDACC).

The Fine and Gray models derived in the VA cohort were tested in the two external cohorts. In the HHS cohort, the c-statistics were 0.69 (95% CI, 0.64–0.79) for overall VTE and 0.72 (95% CI, 0.65–0.79) for PE/LE-DVT. In the MDACC cohort, the c-statistics were 0.72 (95% CI, 0.68–0.75) for overall VTE and 0.69 (95% CI, 0.63–0.73) for PE/LE-DVT. The calibration plots for both outcomes correspond closely with the 45-degree line in the derivation dataset (Supplemental Figure 3 A, D). The external model calibrations for PE/LE-DVT appeared adequate without systemic bias (Supplemental Figure 3 E-F). However, there were more observed overall VTE events than predicted in the very high-risk strata (Supplemental Figure 3 B-C).

Stratified analyses and comparisons with other models

When stratified into risk groups using the 7% overall VTE threshold, the cumulative incidence showed promising discrimination between risk groups (Figure 1). In the VA derivation cohort, 27.7% of patients were high-risk with 10.1% VTE, while 72.3% were low-risk with 4.2% VTE at 6 months (Table 3). In the HHS validation cohort, 56.7% were high-risk with 11.7% VTE and 43.3% were low-risk with 3.6% VTE at 6 months. In the MDACC validation cohort, 31.5% were high-risk with 16.8% VTE, while 68.5% were low-risk with 5.1% VTE at 6 months.

Figure 1.

VTE incidence (A-C) and PE/LE-DVT incidence (D-F) by new lymphoma risk assessment model in derivation and validation cohorts

Table 3.

Performance of the novel thrombosis risk assessment model in patients with treated lymphoma

		VTE					PE/LE-DVT
Dataset	Classificationa	6-month incidence		Time-dependent C-statistic (95% CI)			6-month incidence		Time-dependent C-statistic (95% CI)
				New model	Li 2023	Khorana 2008			New model	Li 2023	Khorana 2008
VA derivation	Low risk (n=7458)	4.2%	5.8%	0.68 (0.67–0.69)	0.65 (0.62–0.67)	0.57 (0.54–0.58)	2.7%	3.9%	0.68 (0.66–0.71)	0.65 (0.64–0.66)	0.56 (0.53–0.59)
	High risk (n=2852)	10.1%					7.2%
HHS validation	Low risk (n=370)	3.6%	8.2%	0.69 (0.64–0.79)	0.65 (0.58–0.71)	0.55 (0.48–0.60)	0.8%	4.5%	0.72 (0.65–0.79)	0.64 (0.56–0.72)	0.54 (0.45–0.62)
	High risk (n=484)	11.7%					7.3%
MDACC validation	Low risk (n=1272)	5.1%	8.8%	0.72 (0.68–0.75)	0.67 (0.63–0.70)	0.55 (0.50–0.60)	2.3%	3.7%	0.69 (0.63–0.73)	0.67 (0.62–0.75)	0.56 (0.49–0.62)
	High risk (n=586)	16.8%					6.9%

^aWe dichotomized the risk groups based on a predetermined clinical threshold of 7% overall VTE at 6 months.

Abbreviations: VTE, Venous Thromboembolism. PE/LE-DVT, Pulmonary Embolism/ Lower Extremity Deep Vein Thrombosis. VA, Veterans Affairs. HHS, Health and Human Services.

When compared to previous models, the newly developed RAMs showed improved discrimination and risk stratification for lymphoma patients.^{12, 14} The improvement in c-statistic was 0.03–0.08 compared with the modified Khorana Score,¹⁴ while the improvement was 0.11–0.18 compared with the Khorana Score¹² (Table 3). The prediction performance showed similar improvement for PE/LE-DVT. The time-to-event comparisons of various models are shown in Supplemental Figures 3-4.

Discussion

Using independent datasets from three diverse cancer cohorts in the U.S., we developed and validated a novel clinical RAM for incident VTE in lymphoma patients at the time of systemic therapy initiation. Among numerous clinical and laboratory parameters tested, the final models included lymphoma subtype risk group, treatment regimen, BMI≥35, recent hospitalization, history of VTE, history of paralysis/immobility, and time to systemic treatment as the 7 covariates. The new RAM grouped patients into a high-risk category with a 10.1–16.8% VTE risk (6.9–7.3% for PE/LE-DVT) at 6 months and a low-risk category with a 3.6%−5.1% VTE risk (0.8%−2.7% for PE/LE-DVT). Despite significant differences in patient populations in the derivation and two validation cohorts, the consistent results in discrimination (0.69 to 0.72) indicated the robustness, reliability, and generalizability of the models. The RAM also had better discrimination than most other published risk scores for VTE risk assessment in ambulatory patients with lymphoma.

Our refined RAM for VTE should be interpreted in the context of pan-cancer setting and lymphoma-specific setting. The Khorana score remains the current clinical benchmark for VTE prediction in cancer patients according to several national cancer guidelines.^{25, 26} While it has reasonable performance for solid tumor patients, its discrimination in lymphoma has been sub-optimal because lymphoma is classified into the same intermediate risk group despite the diversity of histologic subtypes ranging from very indolent to very aggressive disease.^{12, 27, 28} Furthermore, risk factors such as leukocyte count generally reflect increased neutrophils in solid tumor malignancies but lymphocytes in certain lymphoma subtypes; therefore, it is unlikely to have the same predictiveness. While we attempted to divide lymphoma into aggressive and indolent subtypes in our previous pan-cancer risk prediction model in the same populations,¹⁴ it was difficult to estimate the attributable risk of each histologic subtype when mixed with a variety of solid tumor malignancies. In the current study, we carefully created 16 histologic subgroups based on the latest SEER lymphoid malignancy classification recommendations²⁹ (Supplemental Table 3) to create a lymphoma-specific RAM. When assessed concurrently with other clinical risk factors, we found that the lymphoma subtype along with history of VTE and history of paralysis/immobilization provided the most appreciable VTE risks in this patient population.

The risk predictors in the current VTE RAM are intuitive and consistent with the previously published literature. The modification of lymphoma subtype risk was concordant but more comprehensive when compared with previous studies.^{6, 8, 14, 30} Dharmavaram et al.⁶ assigned −1.025 for patients with FL, and 0 for DLBCL as the risk scores in their model for the two types of lymphomas. Lund et al.⁸ reported that patients with peripheral T-cell lymphomas (SHR = 2.05, 95% CI: 1.27, 3.30), DLBCL (SHR= 1.79, 95% CI: 1.31, 2.43), or CHL (SHR=1.64, 95% CI: 1.08, 2.48) were in higher risks compared with indolent lymphoma. Similarly, Santi et al. showed that DLBCL patients had a higher risk of experiencing VTE (SHR: 3.42 [1.32–8.84]; p = 0.011) compared to those with FL.³⁰ VTE history was confirmed as prognostic factors in comparable lymphoma studies.^{7, 31–33} Recent hospitalization and history of paralysis/immobility were other well-known predictors for VTE,^{14, 34–36} but they have not been explored in the lymphoma setting. Also, the differential risk associated with chemotherapy versus targeted therapy has not been systematically incorporated into lymphoma-specific RAMs, although several studies have reported the increased VTE risk associated with cytotoxic therapy,^{7, 31, 37, 38} while one study has reported a lower incidence of VTE among patients receiving ibrutinib.³⁹ Finally, concordant with previous studies,^{8, 28, 40, 41} time to treatment initiation was also a key risk factor for VTE, likely a proxy feature associated with the tumor burden and aggressiveness of the disease.

Several aspects of our study deserve consideration. Firstly, our lymphoma-VTE RAM exclusively utilized clinical predictors with broad applicability across diverse clinical settings; however, factors such as ECOG performance status, immunohistochemical, or tumor mutational profiling could not be included in the current study. Secondly, despite the modest discrimination of the novel RAM and reasonable calibration for PE/LE-DVT outcomes in external validation, the RAM underestimated the risk of overall VTE in the predicted higher strata. Given this discrepancy was most likely driven by the under-capture of UE-DVT in the VA derivation cohort, our RAM likely underestimated unique risk factors for UE-DVT. In addition, given that treatment protocols evolve over time and can influence thrombotic risk, the temporal discrepancy between derivation cohort (2006–2021) and validation cohorts (2011–2020, 2017–2020) could contribute to the imperfect calibrations observed in the current study. Furthermore, to ensure uniformity of initial diagnosis and staging, the cohort inclusion mandated patients to have a histologic diagnosis of lymphoma and have started systemic therapy within 1 year of diagnosis. Thus, patients with indolent lymphomas who started initial therapy many years after observation would be excluded from the cohort. Finally, due to the retrospective nature of our cohorts, inherent biases including selection and information bias could be introduced. Since we relied on the cancer registry to determine accurate information on original cancer diagnosis date, cancer subtype (ICD-O-3) and initial staging, the potential delay in data collection and reporting by the cancer registry can lead to challenges of applying the model prospectively. Further informatics work is needed to translate and validate the above information to ICD-10-CM codes used in the electronic health records. While our computable phenotype algorithm combining ICD codes and NLP radiology reports achieved excellent precision and recall in all cohorts to prevent significant misclassification,^{14, 15} the outcome capture was not as robust as prospective cohort studies with independent adjudication.

In summary, we derived and externally validated a novel RAM for VTE in patients with lymphoma initiating systemic therapy. The models contained only clinical predictors that were assessable without specialized biomarker testing. The RAM outperformed other existing pan-cancer and lymphoma-specific RAMs. We are confident that our enhanced models for lymphoma-associated thrombosis can better guide the selection of patients for individualized thromboprophylaxis across a broad spectrum of lymphoma patient groups. We have provided details of our model (Lymph-CAT) derivation and application (https://github.com/Shenglinggithub/Lymph-CAT-calculator) and we encourage additional external validations from researchers and collaborators.

Data Sharing Statement:

Since our datasets are from 3 separate hospital systems, each governed by its own IRB and DUA, data sharing can only be made available with appropriate regulatory approval from each participating site. For original data, please contact ang.li2@bcm.edu