The prognostic utility of a combined PNI-HB immunonutritional index in CD19 CAR T-cell therapy for relapsed/refractory non-Hodgkin lymphoma

Summary

CD19 chimeric antigen receptor (CAR) T cell therapy has improved outcomes in relapsed/refractory non-Hodgkin lymphoma (R/R NHL). However, relapse and treatment related toxicities remain common, highlighting the need for prognostic biomarkers. This study evaluates a composite immunonutritional index, integrating the prognostic nutritional index (PNI) and hemoglobin (HB) in a cohort of 107 patients. The PNI-HB was assessed using Cox proportional hazards models, time-dependent receiver operating characteristic analysis, and bootstrap validation. The PNI-HB score independently predicted inferior progression-free survival (HR = 1.34, 95% CI, 1.08–1.66; p = 0.009) and stratified patients into risk groups with distinct outcomes. The model outperformed IPI, with lower AIC/BIC values (ΔAIC/ΔBIC = 8.97) and improved discrimination. Higher PNI-HB scores were associated with increased risk of cytokine release syndrome (OR = 1.86, p = 0.004). The PNI-HB score is a non-invasive tool for predicting survival and toxicity in patients receiving CD19 CAR T cell therapy.

Graphical abstract

Highlights

• •PNI-HB score predicts survival in R/R NHL after CD19 CAR T cell therapy
• •High PNI-HB scores correlate with significantly shorter PFS and OS
• •The PNI-HB index out-performs single PNI or HB markers in prognosis
• •PNI-HB provides a cost-effective tool for patient risk stratification

Health sciences

Introduction

Chimeric antigen receptor (CAR) T cell therapy targeting CD19 has revolutionized the treatment of relapsed/refractory non-Hodgkin lymphoma (R/R NHL), demonstrating remarkable efficacy in patients who have exhausted conventional therapies.^1,2 Nearly 40% of patients are expected to achieve durable remission, over 60% of those with diffuse large B cell lymphoma (DLBCL) fail to attain long-term disease control.³ Given the treatment responses exhibit considerable interpatient variability associated with CD19 CAR T cell therapy, identifying robust predictive biomarkers is essential to optimize patient selection and improve treatment outcomes. Current clinical investigations have identified multiple independent risk factors predictive of poor prognosis following CD19 CAR T cell therapy, such as TP53 alterations,⁴ elevated pre-treatment Eastern Cooperative Oncology Group (ECOG) performance status,⁵ increased lactate dehydrogenase (LDH) levels,⁶ and high-risk international prognostic index (IPI) scores,⁷ all of which significantly influence patient outcomes.^8,9 The IPI is well-established as an integrative prognostic tool reflecting baseline tumor burden and host factors, there remains a lack of universally adopted nutritional and immune prognostic indices in clinical practice for predicting outcomes after CD19 CAR T cell therapy.

Nutritional and inflammatory immunity status has emerged as a critical determinant of outcomes in cancer patients, influencing both survival and treatment tolerance. In lymphoma, several nutritional and laboratory markers, such as albumin levels (ALBs) have been identified as independent predictors of adverse prognosis.^10,11 Some nutritional assessment tools including the C-reactive protein-to-albumin ratio (CAR), Glasgow prognostic score (GPS) have demonstrated significant prognostic value in multivariate analyses of patients with hematological malignancies.^12,13,14 The prognostic nutritional index (PNI), calculated from serum albumin and lymphocyte count, reflects integrated nutritional and immune status.^15,16 Low PNI has been consistently associated with inferior survival across various lymphoma subtypes, potentially due to impaired immune surveillance and reduced tolerance to therapy. A recent multicenter study of CD5-positive DLBCL patients demonstrated that low PNI was significantly associated with inferior overall survival (OS) (p < 0.05, BCL 2-negative group; p < 0.05, BCL 2-positive group).^17,18 Similarly, hemoglobin (HB) level as a key indicator of inflammatory immunity and nutritional deficiency but may also contribute to tumor hypoxia, promoting aggressive tumor behavior and resistance to treatment.¹⁹ In DLBCL, anemia (HB < 100 g/L) was an international index-independent prognostic factors (HR = 2.4, p = 0.040).²⁰

Although both the PNI and HB levels have been independently associated with outcomes in lymphoma patients undergoing chemotherapy, their combined prognostic utility in CD19 CAR T cell therapy remains unexamined. This study aimed to develop and validate a PNI-HB composite score to enhance prognostic assessment in patients with R/R NHL patients receiving CD19 CAR T cell therapy. We evaluated its ability to predict survival outcomes and toxicity events, including cytokine release syndrome (CRS), and compared its performance with the IPI.

Results

Baseline characteristics and univariate analysis of PFS

Baseline characteristics of the 107 patients were included. The study concluded follow-up on June 30, 2025, with a median observation period of 42.57 months (IQR, 31.97–50.27). 107 patients receiving CAR T cell therapy had a median age of 59 years (IQR, 48.5–67.00), with males comprising 68.22% (n = 73) of the cohort.

Univariate Cox proportional hazards regression analysis was used to evaluate factors associated with PFS following CD19 CAR T cell therapy. The pretreatment clinical characteristics are summarized in Table 1. Our analysis identified HB and the PNI as significant predictors of post-CAR T cell therapy relapse (p < 0.05). Although total cholesterol (TC) also showed statistical significance, all cholesterol values fell within normal limits. Based on clinical relevance, TC was excluded from further predictive modeling. Consequently, HB and PNI were selected for subsequent analysis to determine their prognostic value for PFS.

Table 1.

Baseline characteristics of Cox proportional hazards regression analysis to evaluate factors associated with PFS

Index	N	HR	95% CI	p value
HB g/L, median (IQR)	107	0.98	[0.97,1.00]	0.005
WBC∗10ˆ9/L, median (IQR)	107	0.89	[0.77,1.02]	0.093
PLT∗10ˆ9/L, median (IQR)	107	1.00	[0.99,1.00]	0.247
CRP mg/L, median (IQR)	107	1.01	[1.00,1.02]	0.066
N∗10ˆ9/L, median (IQR)	107	0.84	[0.69,1.02]	0.073
TC mmol/L, median (IQR)	107	0.72	[0.57,0.90]	0.005
Ferritin, n (%)
<300 ng/mL	44 (41.12)	–	–	–
≥300 ng/mL	63 (58.88)	1.20	[0.72,1.99]	0.480
L∗10ˆ9/L, median (IQR)	107	0.81	[0.57,1.16]	0.252
M∗10ˆ9/L, median (IQR)	107	0.58	[0.16,2.07]	0.403
PNI, median (IQR)	107	0.96	[0.93,1.00]	0.030
CAR, median (IQR)	107	1.35	[1.00,1.82]	0.051
CALLY, median (IQR)	107	0.99	[0.97,1.01]	0.244
Gender, n (%)
Female	34 (31.78)	–	–	–
Male	73 (68.22)	1.49	[0.85,2.61]	0.163
Age at enrolment, n (%)
<60 years	55 (51.40)	–	–	–
≥60 years	52 (48.60)	1.09	[0.67,1.79]	0.723
Hans classification, n (%)
GCB	38 (35.51)	–	–	–
N-GCB	69 (64.49)	1.38	[0.80,2.36]	0.246
Double expression, n (%)
No	72 (67.29)	–	–	–
Yes	35 (32.71)	1.67	[1.00,2.76]	0.048
Double/triple-hit, n (%)
No	98 (91.59)	–	–	–
Yes	9 (8.41)	0.85	[0.34,2.11]	0.722
Ann Arbor stage, n (%)
Ⅰ–Ⅱ	11 (10.28)	–	–	–
Ⅲ–Ⅳ	96 (89.72)	2.12	[0.77,5.85]	0.146
ECOG, n (%)
0–1	102 (95.33)	–	–	–
2	5 (4.67)	1.02	[0.32,3.25]	0.974
IPI at enrolment, n (%)
Low risk	12 (11.21)	–	–	–
Low intermediate risk	26 (24.30)	0.7	[0.27,1.77]	0.450
High intermediate risk	45 (42.06)	1.04	[0.45,2.40]	0.920
High risk	24 (22.43)	1.45	[0.60,3.49]	0.410
Bulky disease, n (%)
<10 cm	94 (87.85)	–	–	–
≥10 cm	13 (12.15)	1.16	[0.55,2.44]	0.689
Extra-nodal disease, n (%)
None	16 (14.95)	–	–	–
≥1 organs	91 (85.05)	1.26	[0.60,2.64]	0.545
Prior lines of therapy, n (%)
0∼3	53 (49.53)	–	–	–
≥4 lines	54 (50.47)	1.31	[0.80,2.16]	0.282
TP53, n (%)
WT	56 (52.34)	–	–	–
Altered	43 (40.19)	1.17	[0.69,1.98]	0.568
Missing	8 (7.48)	1.86	[0.81,4.24]	0.141
LDH, n (%)
<250 U/L	60 (56.07)	–	–	–
≥250 U/L	47 (43.93)	1.45	[0.89,2.38]	0.140
ALB, n (%)
<40 g/L	41 (38.32)	–	–	–
≥40 g/L	66 (61.68)	1.38	[0.82,2.32]	0.228
CRP, n (%)
<10 mg/L	33 (30.84)	–	–	–
≥10 mg/L	74 (69.16)	0.63	[0.37,1.05]	0.070
GPS, n (%)
0	67 (62.62)	–	–	–
1	28 (26.17)	1.30	[0.73,2.31]	0.379
2	12 (11.21)	2.07	[1.02,4.19]	0.043
NPS, n (%)
1	5 (4.67)	–	–	–
2	20 (18.69)	0.93	[0.26,3.38]	0.910
3	43 (40.19)	1.05	[0.32,3.46]	0.943
4	39 (36.45)	1.51	[0.45,5.00]	0.503
Prior ASCT, n (%)
No	90 (84.11)	–	–	–
Yes	17 (15.89)	0.80	[0.41,1.57]	0.519
The last treatment response, n (%)
PR	10 (9.35)	–	–	–
SD	24 (22.43)	0.98	[0.35,2.75]	0.969
PD	73 (68.22)	1.38	[0.55,3.48]	0.495
CAR T cell product type, n (%)
41BB	90 (84.11)	–	–	–
CD28	17 (15.89)	0.661	[0.31,1.39]	0.277

HR, hazard ratio; CI, confidence interval. p values are derived from two-sided univariate Cox proportional hazards regression analysis. p < 0.05 was considered statistically significant.

Survival outcomes of PNI and HB three-layer classification

Patients in the PNI-Mal group exhibited substantially higher rates of recurrence and mortality. The one-year recurrence rates were 31.25% in the PNI-Ade group, 53.70% in the PNI-Sub group, and 70.27% in the PNI-Mal group. Corresponding one-year mortality rates were 18.75%, 25.93%, and 62.16%, respectively. Despite a statistically significant difference observed in the OS analysis, the primary endpoint of this study PFS curve did not reach statistical significance and exhibited no clear discriminatory trend (Figure 1A p = 0.124; Figure 1B p = 0.030). The survival analysis stratified by HB levels is presented in Figure 1. Anemia was significantly associated with increased one-year mortality, with rates of 13.33%, 43.40%, and 54.17% in the HB-A, HB-S, and HB-N groups, respectively (p = 0.012; Figure 1D). In contrast, although a trend toward higher recurrence was observed in more anemic groups, with one-year recurrence rates of 40.00%, 58.49%, and 70.83% across the same cohorts, this difference did not reach statistical significance (p = 0.060; Figure 1C).

Figure 1.

Survival outcomes of PNI and HB three-layer classification

Kaplan-Meier survival curves illustrating progression-free survival (PFS) and overall survival (OS). The log rank test was used to compare differences between groups; a p value <0.05 was considered statistically significant. Survival probabilities are presented with 95% confidence intervals (CIs).

(A and B) Kaplan-Meier curves for PFS and OS under PNI three-layer classification.

(C and D) Kaplan-Meier curves for PFS and OS under HB three-layer classification.

Construction of the PNI-HB composite score and its prognostic performance comparison with IPI

Moderate correlation was observed between PNI and HB (Pearson r = 0.46, p < 0.001), to synthesize this information into a more robust predictor, we integrated them into the composite PNI-HB score. Multivariable Cox regression analysis showed that the PNI-HB score remained a significant independent predictor of PFS (HR = 1.27, 95% CI, 1.01–1.59; p = 0.043). In contrast, double expression (p = 0.197) and IPI score (p = 0.488) were not statistically significant in this model. VIF for all variables were low (range, 1.08–1.39), indicating no substantial multicollinearity (Table S2). Patients treated after adjustment for treatment year, the PNI-HB score remained an independent predictor of PFS (HR = 1.27, p = 0.032 Table S3). The resulting cumulative scores were further categorized for subsequent survival analysis. As illustrated in Figure 2, although the PFS analysis did not reach statistical significance (p = 0.114, Figure 2A), a clear separation was observed among the curves for score groups 0, 1–2, and 3–4. We acknowledge that this reflects limited statistical power rather than a lack of biological effect. Similarly, in OS analysis, a statistically significant difference was evident, particularly driven by the group with the highest score of 4 (p = 0.002, Figure 2B). Patients were stratified into low, intermediate, and high-risk groups based on integrated PNI-HB scores of 0, 1–2, and 3–4, respectively (Figures 2C and 2D). The one-year recurrence rates were 14.29%, 50.88%, and 69.77% across these groups (p = 0.035; Figure 2C), the corresponding one-year mortality rates were 14.29%, 26.32%, and 55.81% (p = 0.043; Figure 2D).

Figure 2.

Survival outcomes of PNI-HB classification

Kaplan-Meier survival curves illustrating PFS and OS stratified by the composite PNI-HB score and PNI-HB risk grade. Pooled cumulative survival probabilities are presented with 95% confidence intervals (CIs). The log rank test was used to compare survival differences between groups; a two-sided p value <0.05 was considered statistically significant.

(A and B) Kaplan-Meier curves for PFS and OS under PNI-HB score classification (total score range, 0 to 4).

(C and D) Kaplan-Meier curves for PFS and OS under PNI-HB grade classification (low-risk, score 0; intermediate risk, score 1–2; high-risk, score 3–4).

Time-dependent ROC comparison among PNI, HB and PNI-HB were revealed. For the primary endpoint of PFS, the PNI-HB score achieved AUCs of 0.644, 0.642, 0.678, 0.726, and 0.715 at 12, 24, 36, 48, and 60 months, respectively (Figure 3A). Across nearly all time points, these values were consistently higher than those of PNI alone (AUCs, 0.625, 0.658, 0.673, 0.707, and 0.641) and HB alone (highest AUC 0.689 at 60 months) (Figures 3B and 3C), indicating superior discriminative ability of the composite score for PFS. For OS, the PNI-HB score achieved AUCs of 0.714, 0.654, 0.666, 0.689, and 0.644 at the same time points. These values generally exceeded those of PNI alone (AUCs, 0.678, 0.655, 0.663, 0.678, and 0.605) and HB alone (AUCs, 0.677, 0.600, 0.608, 0.623, and 0.620). Notably, the PNI-HB score showed slightly higher discriminative ability for OS at earlier time points (e.g., 12 months), but overall, the predictive performance for PFS was somewhat stronger than for OS (Figure S4). Furthermore, DCA and CIC demonstrated that the PNI-HB model provided a significant net benefit and high predictive accuracy across a wide range of threshold probabilities, specifically from 0 to 0.72 (Figures 3D and 3E).

Figure 3.

Time-dependent ROC and clinical utility analysis of the PNI-HB score

Time-dependent receiver operating characteristic (ROC) analysis evaluating the predictive accuracy at 12, 24, 36, 48, and 60 months. The areas under the curve (AUCs) values are reported at each time point, demonstrating sustained discriminative performance.

(A) Time-dependent ROC of PNI-HB score.

(B) Time-dependent ROC of PNI classification.

(C) Time-dependent ROC of HB classification.

(D) Decision curve analysis (DCA) illustrating PNI-HB score provides a superior net benefit compared to both benchmark strategies across a wide threshold range of 0–0.72.

(E) Clinical impact curve (CIC) quantifying the clinical efficacy of the model. The high degree of concordance between the two curves within practical threshold ranges confirms the robustness and precision of the PNI-HB score in identifying high-risk individuals.

Using the conventional four-tier IPI classification,²¹ both systems were evaluated for their ability to stratify outcomes. Survival analysis based on IPI scores revealed no significant discriminative trend in PFS. The one-year recurrence rates across IPI categories were 50.00%, 53.85%, 57.78%, and 58.33% (p = 0.272; Figure S1A), the corresponding one-year mortality rates were 33.33%, 23.80%, 42.22%, and 45.83% (p = 0.044; Figure S1B). Although mortality differences reached statistical significance, visual assessment of the survival curves indicated that the low-risk IPI group exhibited unexpectedly poor outcomes, suggesting limited reliability of the IPI system in predicting efficacy after CD19 CAR T cell therapy. The PNI-HB model demonstrated improved prognostic performance compared to the IPI score, with lower AIC (968.27 vs. 977.24) and BIC (973.60 vs. 982.57) values and a higher C index (0.59, 95% CI, 0.52–0.66 vs. 0.57, 95% CI, 0.50–0.64). The differences in AIC and BIC between the two models were 8.97 (Table S4). Although the absolute improvement in C index seems modest, the significant difference in AIC/BIC indicates that the PNI-HB model has a better goodness of fit.

Internal validation and stratified analyses of the PNI-HB score

The internal validity of the PNI-HB score was assessed using bootstrap resampling, which yielded a mean absolute error (MAE) of 0.051 (Figure S2A). In the entire cohort (N = 107), the PNI-HB score significantly predicted poorer PFS (HR = 1.34, 95% CI, 1.08–1.66, p = 0.009). Stratification analyses demonstrated particularly strong associations in male patients (HR = 1.44, 95% CI, 1.08–1.92, p = 0.013), aged ≥60 years (HR = 1.36, 95% CI, 1.01–1.83, p = 0.046), advanced Ann Arbor stage (III–IV) (HR = 1.33, 95% CI, 1.05–1.67, p = 0.017), patients with ECOG 0–1 (HR = 1.31, 95% CI, 1.05–1.63, p = 0.02), TP53-altered cases (HR = 1.49, 95% CI, 1.05–2.11, p = 0.025), elevated ferritin ≥300 ng/mL (HR = 1.47, 95% CI, 1.11–1.96, p = 0.008), elevated LDH ≥250 U/L (HR = 1.53, 95% CI, 1.12–2.09, p = 0.008), and treatment year 2021–2024 (HR = 1.45, 95% CI, 1.05–2.00, p = 0.025). An exploratory RSF analysis ranked PNI and HB highly among baseline predictors supporting their combination. Non-significant trends were observed in females, patients <60 years, early-stage disease, ECOG ≥2, TP53-wild-type, normal CRP, and normal ferritin levels (Table S5; Figure S2B). The non-significance may have been underpowered due to the small sample size. RCS analysis revealed a clear linear dose-response relationship between the PNI-HB score and the log hazard of PFS (Figure S3). The risk of progression increased steadily with rising PNI-HB scores from 0 to 4, with no evidence of non-linearity or plateau effects.

Predictive value of the PNI-HB score for treatment related CRS

Regarding treatment-related toxicities, among the 107 patients, 29 (27.1%) did not develop CRS, 78 (72.9%) experienced CRS, including 74 cases of grade 1 and four cases of grade 2 or higher. ICANS occurred in six patients (5.61%). Univariate logistic regression was performed to assess the association between the PNI-HB score and the occurrence of CRS (Table 2). The PNI-HB score demonstrated significant predictive value for CRS (OR = 1.856, p = 0.004; 95% CI, 1.22–2.82). Subsequently, we performed a chi-square test to evaluate CRS incidence according to PNI-HB-based risk categories (Table 3). In the initial grouping (Group 1), patients were stratified into three groups, with a significant difference in CRS incidence (p = 0.016). Further refinement was performed by consolidating patients into two categories: low-risk (scores 0–1) and high-risk (scores ≥2), which yielded a more robust statistical difference (p = 0.004). In contrast, grouping with low-risk defined as scores 0–2 and high-risk as scores ≥3 showed inferior discriminative ability. Thus, a PNI-HB score ≥2 effectively identified patients at high risk of developing CRS. Moreover, the IPI score showed no significant predictive value for the occurrence of CRS (p = 0.526; Table S6).

Table 2.

Univariate logistic regression to assess the occurrence of CRS

Index	N	OR	95% CI	p value
PNI HB score	107	1.86	[1.22,2.82]	0.004
PNI HB grade	107	2.18	[1.04,4.58]	0.040
HB	107	0.97	[0.95,0.99]	0.015
PNI	107	0.96	[0.92,1.00]	0.053

p < 0.05 was considered statistically significant.

Table 3.

CRS incidence according to PNI-HB-based risk categories

Variable	Total	Overview	Classification	no-CRS (n = 29)	CRS (n = 78)	p value
Grouping 1, n(%)	107 (100)	36 (33.65)	0–1	16 (55.17)	20 (25.64)	0.016
	–	28 (26.17)	2	5 (17.24)	23 (29.49)	–
	–	43 (40.19)	≥3	8 (27.59)	35 (44.87)	–
Grouping 2, n(%)	107 (100)	36 (33.65)	0–1	16 (55.17)	20 (25.64)	0.004
	–	71 (66.36)	≥2	13 (44.83)	58 (74.36)	–
Grouping 3, n(%)	107 (100)	64 (59.81)	0–2	21 (72.41)	43 (55.13)	0.105
	–	43 (40.19)	≥3	8 (27.59)	35 (44.87)	–

p < 0.05 was considered statistically significant.

Discussion

Current research efforts are focused on a variety of factors to develop comprehensive prognostic models.^22,23 There is still a lack of large-scale predictor studies for nutritional indicators and inflammatory immune indicators for R/R NHL patients received CD19 CAR T cell therapy. This study developed and validated an integrated scoring system based on the PNI and HB levels. Results indicate that the PNI-HB score effectively stratifies patients into subgroups with significantly different prognoses, with one-year recurrence rates: 14.29% vs. 50.88% vs. 69.77%; p = 0.035 (Figure 2C). It also demonstrates superior predictive performance compared to the traditional IPI (ΔAIC/ΔBIC = 8.97; C index, 0.59 vs. 0.57, Table S4) and providing independent prediction of CRS occurrence (OR = 1.86, 95% CI, 1.22–2.82; p = 0.004, Table 3). This readily accessible scoring tool offers evidence for precise risk stratification and personalized management strategies prior to CAR T cell therapy.

The prognostic significance of PNI and HB in lymphomas is underpinned by distinct yet interconnected biological pathways, which are particularly pronounced in this population. In patients with NHL, PNI depletion often stems from chemotherapy-induced malnutrition and malabsorption, coupled with nutrient diversion by proliferating tumors. Bone marrow infiltration can impair lymphopoiesis and hematopoiesis, and hepatic involvement may reduce albumin synthesis.²⁴ A low PNI, integrating hypoalbuminemia and lymphopenia, signifies a state of compromised nutrition and impaired immune competence. Specifically, hypoalbuminemia may limit arginine availability, potentially dampening T cell function and the efficacy of CD19 CAR T cell therapy. Lymphopenia directly indicates an attenuated adaptive immune system.^25,26 Low HB in these patients primarily results from bone marrow infiltration-induced hematopoietic failure and inflammation-mediated chronic anemia. The consequent tissue hypoxia activates transcriptional programs in tumor cells that promote aggressiveness and foster an immunosuppressive microenvironment, which can inhibit the proliferation and effector functions of CAR T cell therapy.^27,28,29,30 Both markers capture critical aspects of the host-tumor interaction that are particularly relevant to immunotherapy outcomes.^{17,18,31,32,33,34,35}

The integration of PNI and HB into a composite score demonstrated superior prognostic performance compared to either marker alone. Both PNI and HB showed significant associations with outcomes in univariate analysis, their independent predictive power for the primary endpoint of PFS was limited, as evidenced by the non-significant survival curves in their individual three-tier classifications (Figures 1A and 1C). Across all models, PNI-HB consistently emerged as an independent prognostic factor, with concordant risk gradients observed in group-based analyses. The composite PNI-HB score, successfully stratified patients into low, intermediate, and high-risk groups with significantly distinct PFS (p = 0.035, Figure 2C). This enhanced discriminatory power was confirmed by time-dependent ROC analysis, where the PNI-HB score achieved higher and more sustained AUC values compared to either PNI or HB alone (Figures 3A–3C). Within the practical threshold range, the model’s high concordance between predicted high-risk individuals and actual events suggests it is a reliable tool for guiding individualized treatment strategies in CAR T cell therapy. The synergy between these markers likely stems from their collective capture of synergistic dysregulation in nutritional-inflammatory pathways. For instance, hypoalbuminemia impairs iron transport and utilization, aggravates anemia by inhibiting the production of erythropoietin, lymphopenia and anemia often coexist in a state of chronic inflammation, collectively contributing to a more profoundly immunosuppressive milieu. It is indicating a need for enhanced monitoring by PNI-HB Grade, and strategies such as earlier bridging therapy or combination regimens. RCS analysis further demonstrated a linear dose-response relationship between the continuous PNI-HB score and the risk of disease progression.

The internal validity of the PNI-HB score was assessed via bootstrap resampling (1,000 repetitions). This validation demonstrated excellent model calibration, with predicted risks closely aligning with observed outcomes. In the entire cohort, the PNI-HB score served as a significant predictor of poorer PFS (HR = 1.34, 95% CI, 1.08–1.66; p = 0.009). The model’s prognostic precision was further supported by a low mean absolute error (MAE = 0.051). The broad applicability of the score was evaluated through stratified analyses. A consistent increase in risk (HR > 1) was observed across most predefined subgroups. Statistically significant associations were particularly prominent in key high-risk subgroups, including males, patients aged ≥60 years, patients received therapy through 2021–2024, those with advanced Ann Arbor stage (III–IV), TP53-altered cases, and those with elevated ferritin (≥300 ng/mL) or LDH (≥250 U/L) (all p < 0.05). Notably, a significant association was also detected in the large subgroup of patients with ECOG 0–1. Conversely, non-significant trends (HR > 1) were observed in smaller subgroups, such as females and patients with ECOG ≥2 (n = 5), likely due to limited statistical power.

When compared to the established IPI, our PNI-HB score demonstrated superior performance in predicting survival outcomes in this CAR T cell-specific cohort. The IPI, developed in the chemotherapy era, primarily incorporates tumor burden-related factors but lacks specific metrics for host nutritional and inflammatory immunity status. In our study, the IPI failed to significantly stratify patients by PFS (p = 0.272, Figure S1A), and its predictive power for OS was compromised by an unexpected outcome in the low-risk group. In contrast, the PNI-HB model exhibited a significantly better goodness of fit, as indicated by lower AIC (968.27 vs. 977.24) and BIC (973.60 vs. 982.57) values, with differences (ΔAIC/ΔBIC = 8.97) substantially exceeding the threshold of four considered to represent a meaningful improvement.³⁶ The PNI-HB score also achieved a higher C index (0.59 vs. 0.57) for PFS prediction. In the context of CAR T cell therapy, where patient-specific immune and nutritional statuses are pivotal determinants of treatment response and toxicity. Although the prognostic impact of tumor-intrinsic factors such as DEL appeared attenuated in our CAR T cell cohort, and the effect of PNI-HB remained stable across treatment eras, these findings require further validation in larger and independent cohorts to confirm their generalizability.

Our study shows that the PNI-HB score predicts the development of CRS. The CAR-HEMATOTOX (CAR T cell-related hematotoxicity) score is a model developed to predict hematologic toxicity following CAR T cell therapy in R/R B cell lymphoma. It incorporates HB, ferritin, platelet count, absolute neutrophil count, and C-reactive protein as prognostic indicators and has been widely adopted in clinical settings.³⁷ Similarly, our study also integrated immunological and nutritional parameters to identify predictors of CRS. Patients with higher scores (≥2) were at significantly greater risk of severe CRS, suggesting that baseline inflammatory immunity and nutritional status may prime the immune system for excessive activation upon CAR T cell therapy engagement. The low incidence of ICANS events (n = 6) precluded a meaningful analysis of its association with the PNI-HB score, which should be investigated in larger cohorts.

In conclusion, both the PNI and HB independently predict PFS in patients receiving CD19 CAR T cell therapy. Their integration into a composite PNI-HB score significantly improves the prediction of clinical outcomes, including relapse and mortality. Furthermore, the PNI-HB score also demonstrates predictive utility for treatment-related toxicities and complements the IPI in predicting both survival and toxicity. Further external validation and mechanistic exploration are essential steps toward integrating this index into clinical decision-making and developing targeted interventions for high-risk patients.

Limitations of the study

Despite introducing a potentially valuable composite score, several limitations of this study must be acknowledged. First, the retrospective, single-center design and relatively small sample size may introduce selection bias and limit the generalizability of our findings. The heterogeneity in NHL subtypes and the low frequency of critical toxicity events (e.g., severe CRS) could not be fully accounted for due to the limited sample size. The limited statistical power in certain stratified analyses such as ECOG status likely contributed to wide confidence intervals and non-significant trends, underscoring the need for validation in larger, multi-center cohorts. Second, although the PNI-HB score demonstrated superior prognostic performance over the IPI in several respects, direct comparison remains suboptimal. The precise biological mechanisms underpinning the PNI-HB score are not fully elucidated, and the cutoff values, based on literature, were not optimized specifically for the CAR T cell therapy population. Furthermore, this retrospective study lacked data on serum cytokine dynamics and CAR T cell therapy expansion kinetics, which would help elucidate the biological mechanisms underlying the PNI-HB score association with outcomes. Lastly, the PNI-HB score was developed and validated internally using bootstrap resampling, which lacks an independent, external validation cohort necessitates caution in interpreting its broad applicability.

Subsequent prospective studies will validate the PNI-HB score in larger multi-center cohorts with external validation to reduce bias, identify relevant factors and incorporate previously missing indicators. Furthermore rigorous testing of its incremental prognostic value against the IPI and other modern scoring systems is warranted. Multi-omics analyses will be performed to explore the associations of the PNI-HB score with T cell adaptability systemic inflammation and tumor microenvironment characteristics thereby clarifying the underlying biological mechanisms. All patient-level clinical data are maintained in a secure, institutional repository in full compliance with Chinese data privacy laws, biosafety regulations, and institutional IRB requirements. De-identified analysis datasets are available to qualified academic researchers for non-commercial research purposes upon reasonable request and formal IRB approval, as detailed in the resource availability section.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Aibin Liang (lab7182@tongji.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

The datasets used in the study are available from the lead contact on reasonable request.

The clinical data used in this study are derived from human participants enrolled at Shanghai Tongji Hospital, Tongji University. Public deposition of individual patient-level clinical data in public repositories is strictly restricted by (1) the Personal Information Protection Law of the People’s Republic of China and Biosafety Law of the People’s Republic of China, which prohibit unauthorized public sharing of identifiable personal health data; (2) the approval requirements of the Institutional Review Board (IRB) of Shanghai Tongji Hospital; and (3) the terms of written informed consent signed by all study participants, which do not authorize full public release of individual patient-level medical data to protect participant privacy and prevent re-identification. To ensure the reproducibility of our findings and academic transparency, de-identified participant-level clinical datasets used in the final analysis will be made available upon reasonable request from qualified academic researchers for non-commercial research purposes. All requests will be reviewed and approved by the IRB of Shanghai Tongji Hospital, and data will be shared under a data use agreement to ensure compliance with ethical and legal requirements. Requests for data access may be directed to the lead contact at the email address listed above.

All original code for PNI-HB score calculation and statistical modeling has been deposited at Zenodo and is publicly available without access restrictions, embargoes, or delayed release arrangements as of the date of publication. The DOI can be accessed at https://doi.org/10.5281/zenodo.18657047.

Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request.

Ethics approval and consent to participate

The study protocols were approved by the Ethics Committee of Shanghai Tongji Hospital, Tongji University, and were conducted in compliance with the principles of the Declaration of Helsinki. All patients provided written informed consent.

Acknowledgments

We thank all the patients who participated in this study, as well as clinical research unit colleagues who provided support for this work. This work was supported by funds from the National Key R&D Program of China (2021YFA1100800), National Natural Science Foundation of China (nos. U23A20418 and 82470225), Foundation of Shanghai Shen Kang Hospital Development Center (SHDC2023CRD021), and cultivation project of National Natural Science Foundation of Shanghai Tongji Hospital (GJPY2314, GJPY2332).

Author contributions

Y.L., B.X., and A.L. conceived and designed the study; W.H. and Y.L. collected data; Y.L. and X.C. conducted statistical analysis; Y.L., X.C., and P.L. wrote the manuscript; and all authors contributed to the acquisition or analysis of data. Y.L., X.C., and P.L. have contributed equally to this work.

Declaration of interests

The authors declare no competing financial interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
R software (version 4.3.2)	R Foundation for Statistical Computing	https://www.R-project.org/
Python (version 3.11.x)	Python Software Foundation	https://www.python.org/
GraphPad Prism (version 8.0.1)	GraphPad Software, Inc.	https://www.graphpad.com/
Original code for PNI-HB score calculation and statistical modeling	This study	Zenodo, persistent DOI: https://doi.org/10.5281/zenodo.18657047

Experimental model and study participant details

This study reviewed 107 R/R NHL patients who received CD19 CAR T cell therapy in the Shanghai Tongji Hospital from July 2017 to December 2024. The median age of the cohort was 54 years (range: 17–81). Among them, 62.6% were male and 37.4% were female. In stratified analyses, the PNI-HB score showed a significant prognostic association with PFS in male patients (HR = 1.44, 95% CI: 1.08–1.92, p = 0.013), while a non-significant trend was observed in female patients (HR > 1), which was likely attributed to the limited sample size of the female subgroup (n = 34). All the patients are from China and belong to the yellow race. Race and age did not show significant statistical differences. Clinical data were collected from medical records and electronic patient files, the follow-up period ended on June 30, 2025. Prior to participation, all patients provided written informed consent voluntarily. Patients who were lost to follow-up or had incomplete laboratory data were excluded based on strictly predefined inclusion criteria. This study was conducted following the principles of the Declaration of Helsinki, with protocols approved by the Ethics Committee of Shanghai Tongji Hospital, Tongji University.

Method details

Diagnosis and CAR T cell therapy treatment

The diagnosis of NHL subtypes was established through pathological evaluation in accordance with the 2016 World Health Organization classification.³⁸ Three patients were diagnosed with Follicular Lymphoma (FL), two with Transformed Follicular Lymphoma (tFL), and 102 with DLBCL. The Hans algorithm was applied to determine the cell-of-origin (COO) classification, distinguishing between germinal center B-cell (GCB) and non-germinal center B-cell (non-GCB) subtypes.³⁹ Double-expression lymphoma (DEL), characterized by overexpression of MYC and BCL-2 proteins, were classified based on standard diagnostic criteria.⁴⁰ TP53 alterations were identified either through mutations detected via next-generation sequencing (NGS)⁴¹ or via deletions observed by fluorescence in situ hybridization (FISH) analysis.⁴²

All CAR T cell products utilized in this study were constructed to contain both 4-1BB or CD28 as co-stimulatory domains. 47 cases used commercial CAR T cell products (Axi-cel, Relma-cel), and 60 cases used clinical trial CAR T cell products (NCT02537977, NCT03154775). All CAR T cell infusions were performed during hospitalization. Prior to CAR T cell infusion (designated as day 0), patients received preconditioning with the FC chemotherapy regimen, which consisted of fludarabine (25–30 mg/m²/day) and cyclophosphamide (250–500 mg/m²/day) administered from day −5 to day −3.^43,44 All cases of CRS and immune effector cell-associated neurotoxicity syndrome (ICANS) were managed in accordance with the consensus guidelines from the American Society for Transplantation and Cellular Therapy (ASTCT)⁴⁵ and the 2022 Chinese expert consensus.⁴⁶

Clinical data collection

The baseline characteristics documented the patients' pretreatment disease status in detail prior to CD19 CAR T cell therapy. These parameters included sex, age at enrollment, Hans subtype, lymphoma stage, extranodal involvement, prior lines of therapy, ECOG score, IPI score, bulky disease (defined as a maximum tumor diameter ≥10 cm), Double-expression/hit status and TP53 alterations (mutation or deletion).

Laboratory data were collected from medical records to evaluate overall health and baseline clinical features. The parameters included HB, white blood cell count (WBC), platelet count (PLT), neutrophil count (N), lymphocyte count (L), monocyte count (M), ferritin, LDH, total cholesterol, albumin, and C-reactive protein (CRP). Among them, the normal value of ferritin is defined as < 300 ng/mL, that of LDH is defined as < 250 U/L, of ALB is defined as ≥40 g/L, and that of CRP is defined as < 10 mg/L. These measurements were taken prior to lymphodepletion. Additionally, several integrated indicators related to nutritional status were calculated, including the PNI, CAR, CRP-Albumin-Lymphocyte Index (CALLY), Naples Prognostic Score (NPS), and GPS.

Informed by previous studies, we adopted a refined classification using two additional thresholds (40 and 50) in PNI analysis.^17,18 Thresholds for PNI and HB were based on established literature cutoffs and were not data-driven from this cohort to avoid overfitting. Patients were grouped into three categories and the PNI-HB composite score was derived by assigning point: PNI-Ade (Adequate nutrition, PNI ≥50, 0 point), PNI-Sub (Suboptimal nutrition, 40 ≤ PNI <50, 1 point), and PNI-Mal (Malnutrition, PNI <40, 2 points). In HB analysis, patients were categorized into three groups: HB-A (HB-Anemic, HB < 90 g/L, 2 points), HB-S (HB-Suboptimal, 90 ≤ HB < 120 g/L, 1 point), and HB-N (HB-Normal, HB ≥ 120 g/L, 0 point). Points from both components were summed to yield a total score ranging from 0 to 4 (Table S1).

Treatment responses were assessed according to the Lugano classification using positron emission tomography and contrast-enhanced computed tomography (PET-CT).⁴⁷ Responses were categorized as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD). Progression-Free Survival (PFS) was defined as the time from CAR T cell infusion to disease progression or the last follow-up. OS was calculated from the date of infusion until death or the last follow-up.

Quantification and statistical analysis

Statistical analysis was performed using R software (version 4.3.2), Python (version 3.11.x), and GraphPad Prism (version 8.0.1).⁴⁸ All the code data are listed in the supplementary materials.

Based on variable types, appropriate statistical tests were applied. Continuous variables were first assessed for normality using the Shapiro–Wilk test. Those following a normal distribution (p > 0.05) were compared between groups using the independent samples t test (two groups) or one-way ANOVA (three or more groups). Non-normally distributed variables were compared using the nonparametric Kruskal–Wallis H test. Categorical variables were analyzed using Pearson’s chi-square test; when expected cell frequencies were below 5, Fisher’s exact test was applied.

In the inter-group comparison and therapeutic effect evaluation, PFS and OS were visualized using Kaplan–Meier curves. All the collected data in baseline were evaluated using the univariate and multivariable Cox proportional hazards model to assess the impact of treatment on PFS. The proportional hazards hypothesis of the Cox model was verified by the Schoenfeld residual test, and no significant evidence of violation of this hypothesis was found. Variance inflation factor (VIF) was calculated to assess multicollinearity among variables included in the multivariable models, with a VIF <5 generally indicating acceptable levels of collinearity. Restricted cubic spline (RCS) analysis was applied within the Cox proportional hazards framework to assess potential non-linear associations between the continuous PNI-HB score and the log hazard of PFS, with adjustment for double-expression status.

In the construction and validation of prognostic prediction models. The discriminative ability of the resulting prognostic model was assessed via time-dependent receiver operating characteristic (ROC) curves and corresponding area under the curve (AUC) estimates.⁴⁹ Decision curve analysis (DCA) comparing the clinical net benefit of the PNI-HB model with all-treatment and no-treatment strategies. Clinical impact curve (CIC) visualizing the estimated number of high-risk patients and the number of actual events, the shaded areas represent the 95% confidence interval (CI) based on 500 bootstrap replicates. The final constructed prognostic model was internally validated using the Bootstrap resampling method (1000 repetitions) to correct the model. In each Bootstrap sampling, the model was reconstructed based on the sampled training set and the MAE of the model was calculated. The concordance index (C-index) was used to quantify the model’s ability to correctly rank patients by risk. The Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were used to evaluate and compare the prognostic performance of competing models. Compare the AIC/BIC of the two models. The larger ΔAIC/ΔBIC is (>4), the stronger the evidence that the goodness-of-fit of the model is superior to that of the old model.

Stratified analyses were performed across clinically relevant subgroups to evaluate the consistency of the prognostic score. Formal testing for interaction was performed by introducing multiplicative interaction terms between the treatment strategy and each stratification variable into the Cox proportional hazards models. Data are presented in the form of Hazard Ratio (HR). This study does not use asterisks to mark statistical significance; all between-group comparisons and statistical differences are reported as exact two-sided P-values, with p < 0.05 defined as the threshold for statistically significant difference. A Random Survival Forest (RSF) algorithm was implemented in an exploratory manner to assess variable importance and inform the selection of candidate biomarkers for the composite score.⁵⁰ All 95% CIs are reported for effect sizes (HR, OR, AUC) and survival probabilities, and are clearly defined in the corresponding figure legends.

Additional resources

This study is associated with clinical trials registered at ClinicalTrials.gov: NCT02537977, NCT03154775.

Published: March 6, 2026