Predicting one-year overall survival in patients with AITL using machine learning algorithms: a multicenter study

Abstract

Angioimmunoblastic T-cell lymphoma (AITL) is a life-threatening hematological malignancy. For patients with poor prognosis, especially those with expected survival less than 1 year, the benefits from traditional regimens are extremely limited. Therefore, we aimed to develop an interpretable machine learning (ML)–based model to predict the 1-year overall survival (OS) of AITL patients. A total of 223 patients with AITL treated in 4 centers in China were included. Five ML algorithms were built to predict 1-year outcome based on 16 baseline characteristics. Recursive feature elimination (RFE) method was used to filter for the most important features. The ML models were interpreted and the relevance of the selected features was determined using the Shapley additive explanations (SHAP) method and the local interpretable model–agnostic explanation (LIME) algorithm. Catboost model presented to be the best predictive model (AUC = 0.8277). After RFE screening, 8 variables demonstrated the best performance (AUC = 0.8125). This study demonstrated that the Catboost model with 8 variables could effectively predict the 1-year overall survival of AITL patients.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-18148-x.

Introduction

Angioimmunoblastic T-cell lymphoma (AITL) is recognized as a distinct TFH-derived peripheral T-cell lymphoma entity in the WHO classification¹. AITL accounts for approximately 2% of all non-Hodgkin’s lymphomas and 15–20% of peripheral T-cell lymphomas². AITL is characterized by its poor prognosis and complex clinical presentation¹. In a large international retrospective study of patients treated between 1990 and 2002, 5-year overall survival (OS) for AITL was 32%³. More recently, in an international prospective T-cell Project which enrolled 282 AITL patients between 2006 and 2018, the 5-year OS rate was only 44%⁴. The OS in patients with AITL is significantly influenced by various clinicopathologic factors. AITL presents unique clinical challenges due to its pronounced biological heterogeneity and complex tumor microenvironment interaction^2,5. Therefore, although some patients shared the identical staging or prognostic scores that commonly used to evaluate T cell lymphoma, their clinical outcomes varied considerably. For patients with poor prognosis, especially those with expected survival less than 1 year, the benefits from traditional regimens are extremely limited. Early identification of this high-risk subgroup is critical for redirecting them toward clinical trials targeting novel mechanisms. Therefore, novel models that can better stratify patients are required.

Machine learning (ML) algorithms constitute a crucial component of artificial intelligence, facilitating the analysis of intricate datasets through advanced computational techniques to identify predictive features⁶. As the medical research landscape continues to evolve, ML has emerged as a groundbreaking approach, significantly enhancing the methodologies for data interpretation⁶. The rapid increase in healthcare data, coupled with improvements in computational modeling, has propelled the adoption of ML in clinical settings⁷. ML offers a transformative approach by decoding complex, nonlinear relationships among multidimensional variables—including clinical features, and laboratory data—that collectively dictate AITL outcomes⁸. Explainable AI techniques (SHAP/LIME) further bridge translational gaps by providing clinically interpretable feature attributions, addressing the “black box” limitation of early ML models⁶. ML has demonstrated robust predictive performance in multicenter studies of some non-Hodgkin lymphoma subtypes, such as mantle cell lymphoma and diffuse large B-cell lymphoma^9,10. Leveraging these advances, we developed an interpretable ML framework to predict 1-year OS in AITL, aiming to surpass existing prognostic indices and enable risk-stratified therapeutic decisions. These innovations enable the extraction of relevant patterns from complex data, thereby improving diagnostic accuracy, forecasting patient outcomes, and personalizing treatment strategies¹¹. Hence, the purposes of the present study were to establish ML models with greater discriminant power than established prognostic indices to predict 1-year survival of AITL.

Methods

Patients

This retrospective multicenter study included 252 patients with newly diagnosed AITL between 2012 and 2023 from 4 tertiary hospitals in China, namely, West China Hospital of Sichuan University (n = 167), Fujian Medical University Union Hospital (n = 40), Sichuan Provincial People’s Hospital (n = 25), and Sichuan Cancer Hospital (n = 20). The study was conducted in compliance with the Helsinki Declaration, and this study was approved by the Ethics Committees of 4 hospitals.

Study design

Clinical and laboratory data for the patients were extracted from four hospital databases. A total of 16 baseline variables were analyzed, including age, gender, Ann Arbor staging, B symptoms, extranodal involvement, Eastern Cooperative Oncology Group (ECOG) Performance Status, International Prognostic Index (IPI), rash, edema/serous effusion, hemoglobin (Hb), platelet count (PLT), absolute lymphocyte count (ALC), absolute eosinophil count (AEC), serum albumin (ALB), serum globulin (GLB), and lactate dehydrogenase (LDH). Monthly follow-ups were conducted for each patient, with the primary endpoint of the study being overall survival (OS) at one year. OS was defined as the time from diagnosis until death from any cause or the last known contact for patients who remained alive. Among the 223 patients analyzed, 49 experienced mortality within one year of diagnosis, while 174 survived. Approximately 75% (167) of patients were randomly assigned to the training set, with the remaining 56 patients designated for validation.

Modeling and evaluation methods referred to our previous research¹². In brief, to predict 1-year OS, multiple machine learning models were developed, including logistic regression (LR), random forest (RF), light gradient boosting machine (LGBM), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost). The performance of these models was assessed using the receiver operating characteristic (ROC) curve, with various metrics computed, such as area under the curve (AUC), accuracy, precision, recall, specificity, and F1 score to identify the optimal model. Feature selection was conducted using Recursive Feature Elimination (RFE), which evaluates feature importance and employs cross-validation to determine the optimal number of features. Ultimately, eight key features were identified: age, B symptoms, ECOG performance status, rash, edema/serous effusion, Hb, PLT, and ALB. The Shapley additive explanations (SHAP) method was utilized to interpret model predictions, providing detailed insights into the contribution of each variable. Additionally, local interpretable model-agnostic explanations (LIME) were applied to enhance the understanding of individual predictions through local linear approximations of model behavior.

The code for constructing the CatBoost model in this study has been uploaded to GitHub (https://github.com/tyewu/Predicting-One-Year-Overall-Survival-in-Patients-with-AITL).

Statistical analysis

Patients were classified into two groups according to their 1-year outcome: alive and dead. Binary variables were analyzed using Fisher’s exact or χ² tests, and continuous variables with a normal distribution were assessed via Student’s t-test, reported as medians with standard deviations. Non-normally distributed variables were analyzed using the Mann-Whitney U test, presenting results as medians with interquartile ranges (IQRs). Five machine learning algorithms were developed and validated using Python (version 3.9.13), with statistical significance set at p < 0.05.

Results

Patient characteristics

This multicenter study enrolled 252 treatment-naïve AITL patients from four Chinese institutions between January 2012 and December 2023. Of these, 239 patients had complete baseline characteristic data available. Attrition included 6 patients who voluntarily declined standard chemotherapeutic intervention and 10 cases lost to follow-up. Ultimately, 223 evaluable patients were included in the final analysis, with no missing values observed across all features. (Fig. 1). All analyzed patients received systemic chemotherapy, predominantly anthracycline-based regimens with no significant inter-institutional heterogeneity in therapeutic protocols. Biopsy samples were centrally reviewed, and baseline characteristics are summarized in Table 1. The median age of the cohort was 62 years (range: 24–78 years), with males comprising 60.54% of the population. Notably, over 90% of patients presented with advanced-stage disease. Symptoms such as edema/serous effusion, rash, extranodal involvement, and B symptoms were observed in 48%, 25%, 49%, and 55% of patients, respectively. Out of the 223 patients, 174 survived beyond one year.

Fig. 1.

Patient enrollment flowchart.

Table 1.

Baseline characteristics of AITL patients.

Parameter	Total (n = 223)	Alive(n = 174)	Dead (n = 49)	p-value
Gender				0.950
Male	135 (60.54%)	105 (60.34%)	30 (61.22%)
Female	88 (39.46%)	69 (39.66%)	19 (38.78%)
Age	62.0(54.0, 68.0)	62.0 (53.0, 68.0)	65.0 (60.0, 70.0)	0.005
Edema/Serous effusion				0.010
Yes	107 (47.98%)	75 (43.1%)	32 (65.31%)
No	116 (52.02%)	99 (56.9%)	17 (34.69%)
Rash				0.118
Yes	56 (25.11%)	39 (22.41%)	17 (34.69%)
No	167 (74.89%)	135 (77.59%)	32 (65.31%)
IPI				< 0.001
0–1	44 (19.73%)	39 (22.41%)	5 (10.2%)
2	49 (21.97%)	40 (22.99%)	9 (18.37%)
3	86 (38.57%)	71 (40.8%)	15 (30.61%)
4–5	44 (19.73%)	24 (13.79%)	20 (40.82%)
ECOG				< 0.001
0	93 (41.7%)	85 (48.85%)	8 (16.33%)
1	83 (37.22%)	66 (37.93%)	17 (34.69%)
2–4	47 (21.08%)	23 (13.22%)	24 (48.98%)
Extranodal involvement				0.409
Yes	109 (48.88%)	82 (47.13%)	27 (55.1%)
No	114 (51.12%)	92 (52.87%)	22 (44.9%)
B symptom				< 0.001
Yes	122 (54.71%)	84 (48.28%)	38 (77.55%)
No	101 (45.29%)	90 (51.72%)	11 (22.45%)
Ann Arbor staging				0.067
1–2	21 (9.42%)	20 (11.49%)	1 (2.04%)
3	103 (46.19%)	82 (47.13%)	21 (42.86%)
4	99 (44.39%)	72 (41.38%)	27 (55.1%)
Hb	118.0 (96.0, 131.0)	119.0 (101.25, 131.0)	103.0 (88.0, 126.0)	0.003
PLT	162.0 (126.0, 215.5)	163.0 (136.25, 225.0)	140.0 (74.0, 193.0)	0.002
ALC	0.93 (0.63, 1.34)	0.93 (0.65, 1.38)	0.87 (0.58, 1.16)	0.091
AEC	0.12 (0.05, 0.24)	0.12 (0.05, 0.26)	0.12 (0.03, 0.21)	0.175
ALB	37.4 (32.8, 40.85)	37.8 (33.92, 41.38)	33.5 (28.4, 39.2)	< 0.001
GLB	29.0 (24.95, 35.7)	29.0 (24.92, 35.08)	30.5 (25.6, 38.3)	0.197
LDH	256.0 (196.5, 331.5)	256.0 (193.25, 330.5)	267.0 (212.0, 350.0)	0.262

ECOG, Eastern Cooperative Oncology Group Performance Status; IPI, International Prognostic Index; Hb, hemoglobin; PLT, platelet; ALC, absolute lymphocyte count; AEC, absolute eosinophil count; ALB, serum albumin; GLB, serum globulin; LDH, lactate dehydrogenase.

Predicting 1-year OS

The modeling we conducted involved a thorough set of 16 baseline variables, which included age, gender, Ann Arbor staging, B symptom, extranodal involvement, ECOG, IPI, rash, edema/serous effusion, Hb, PLT, ALC, AEC, ALB, GLB, and LDH. The patients were categorized into two groups: the training set, which was used to construct the model, and the testing set, which was used to evaluate the model’s performance. This categorization was carried out using a random stratification approach with a ratio of 3:1. Supplementary Table 1 displays the clinical characteristics of the two data sets.

Predictive models for 1-year survival were generated using five algorithms: LR, RF, LGBM, XGBoost, and CatBoost. Among these, CatBoost achieved the highest predictive accuracy, with an area under the curve (AUC) of 0.8738 for the training set and 0.8277 for the testing set. The performance metrics, including AUC, accuracy, precision, recall, specificity, and F1 score for all five models across both datasets, are detailed in Table 2. The corresponding ROC curves for each model are illustrated in Fig. 2.

Table 2.

Predictive performance of five ML models.

Model	AUC (Training set)	AUC (Testing set)	Accuracy (Testing set)	Precision (Testing set)	Recall (Testing set)	Specificity (Testing set)	F1 (Testing set)
LR	0.7653	0.7140	0.6429	0.3571	0.8333	0.5909	0.500
Random Forest	0.8653	0.8030	0.8393	0.6154	0.6667	0.8864	0.640
LGBM	0.8287	0.7434	0.7857	0.500	0.750	0.7955	0.600
XGBoost	0.8274	0.7822	0.750	0.450	0.750	0.750	0.5625
CatBoost	0.8738	0.8277	0.8571	0.6667	0.6667	0.9091	0.6667

Fig. 2.

The ROC curves of logistic regression (LR), random forest (RF), light gradient boosting machine (LGBM), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost) models. (A) Training set. (B) Testing set.

The CatBoost model was further refined using the RFE method to identify the most important predictors. In its optimal configuration, the CatBoost model incorporated eight key features: Age, B symptom, ECOG, Rash, Edema/Serous effusion, Hb, PLT, and ALB (Fig. 3). The performance metrics, including AUC, accuracy, precision, recall, specificity, and F1 score for the optimized CatBoost model, are summarized in Table 3. The Decision Curve Analysis (DCA) curves are presented in Supplementary Fig. 1.

Fig. 3.

Applying RFE method to screen for the optimal variables on Catboost model (A). The ROC curves of the optimized Catboost model (B).

Table 3.

Predictive performance of the optimized catboost model.

Model	AUC (Training set)	AUC (Testing set)	Accuracy (Testing set)	Precision (Testing set)	Recall (Testing set)	Specificity (Testing set)	F1 (Testing set)
CatBoost	0.8266	0.8125	0.7857	0.5	0.75	0.7955	0.6

Interpretation and evaluation of machine learning model

We utilized the SHAP methodology to precisely evaluate the quantitative influence of each feature on the predictive results of the model (Fig. 4). Upon analyzing the ranking of features, it was determined that the five most important features were ECOG, B symptom, edema/Serous effusion, rash and PLT. In addition, the LIME technique was utilized to elucidate the impact of various variables of the Catboost model on the prediction outcomes. Two scenarios were chosen at random to analyze the visualized prediction results (Fig. 5).

Fig. 4.

Attribution of 8 characteristics in the enhanced Catboost model using the SHAP algorithm. (A) Overview of SHAP analysis conducted on the dataset. Summary plot showing feature contributions to 1-year mortality prediction. The x-axis represents SHAP values (positive values increase predicted mortality risk). Features are ranked vertically by impact magnitude. Point color indicates feature value: blue (low) to pink (high). Key observations: High ECOG performance status (pink) strongly associates with increased mortality; Presence of B symptoms, edema/serous effusion, or rash (pink) also elevates mortality risk. (B) Ranking of feature importance indicated by SHAP. Figure 4B quantifies feature importance using mean absolute SHAP values, where larger values indicate stronger contributions to model predictions. The analysis identifies ECOG performance status, B symptoms, and edema/serous effusion as dominant prognostic features for 1-year mortality.

Fig. 5.

Outcomes of applying LIME with Catboost to two patients chosen at random. Orange characteristics promote mortality, whereas blue characteristics promote longevity. The upper left corner of the picture displayed the prediction odds for being alive or dead.

Prognostic indices

Through machine learning methods, 8 factors were included in the optimization model finally obtained in this study. Table 4 compared the adverse prognostic factors in our model and commonly used prognostic indicators for AITL, including IPI, Prognostic Index for T-cell lymphoma (PIT), International peripheral T-cell lymphoma Project score (IPTCL) and the Prognostic Index for AITL (PIAI)^13–16. It is worth noting that all the factors included in these prognostic models were also analyzed as part of the 16 baseline features in our study. However, the combination of factors in the existing models did not provide any advantage in our analysis (Supplementary Table 2). Therefore, our study demonstrated that the optimized Catboost model outperformed previous models in predicting 1-year survival in AITL.

Table 4.

Adverse prognostic factors in AITL.

Adverse prognostic factor	IPI	PIT	IPTCLP	PIAI	Optimized Catboost model
Age	+	+	+	+	+
ECOG	+	+	+	+	+
LDH	+	+
Ann Arbor Stage	+			+
Extranodal envolvement	+
Bone marrow involvement		+
B symptoms				+	+
PLT			+	+	+
ALC				+
Rash					+
Edema/Serous effusion					+
Hb					+
ALB					+

Discussion

AITL is a subtype of peripheral T-cell lymphoma as classified by the WHO; however, we have observed significant heterogeneity within this subtype itself. Certain cases of AITL exhibited notably aggressive biological characteristics, resulting in patients having an OS of less than one year. For these patients, traditional treatments offer little advantages, and it is advisable to explore participating in clinical trials or exploring new therapy protocols. Therefore, prompt identification of this high-risk patient subset at diagnosis is clinically imperative to guide risk-adapted therapeutic strategies.

AITL represents a unique clinicopathological entity among lymphomas, defined by its TFH-cell derivation, prominent angiogenic microenvironment, recurrent epigenetic mutations (e.g., TET2, RHOA^G17V, IDH2), and frequent autoimmune phenomena^2,5. These distinctive features contribute to the limited applicability of conventional prognostic indices in AITL. Importantly, the preexisting models overlook AITL-specific biology by omitting features present in over 25% of patients, such as rash and edema. For this investigation, we developed a Catboost model using 16 baseline characteristics. The final model included 8 features: Age, B symptom, ECOG, Rash, Edema/Serous effusion, Hb, PLT, and ALB. This model accurately forecasted the likelihood of AITL patients surviving for one year and performed better than earlier models.

Among the 8 prognostic factors included in our model, Age, B symptom, ECOG, and PLT had already been integrated in earlier models. SHAP analyse demonstrated that Age had the highest quantitative influence, followed by B symptom in second place, ECOG in fourth place, and PLT in fifth place. In contrast to the preexisting models, Edema/Serous effusion, ALB, Rash and Hb were also closed related with outcome in our model. All of these four parameters had been linked to prognosis in prior researches. Associations were observed between the presence of serous effusion or reduced albumin levels and a poor prognosis^17,18. Several studies had regarded rash as one of the prognostic factors associated with shorter survival for AITL^19–21. In another study, Hb was independently associated with shorter OS in AITL²².

The study possesses several limitations that should not be overlooked. Initially, this study was carried out in a retrospective manner and had a restricted sample size. While ML models have shown promise in improving clinical decision-making, they often require careful validation and adaptation to specific clinical contexts to ensure their reliability and applicability²³. Further validation in large cohorts is necessary to provide additional evidence of the effectiveness of the model. Furthermore, the lack of comprehensive data has led to the omission of several potential influential factors, such as C-reactive protein, Beta-2-microglobulin, and specific molecular abnormalities that could affect the prognosis of lymphoma^24–27. In the future, a multimodal model incorporating indicators such as comprehensive clinical features, laboratory data, pathology, imaging and genetic abnormalities to better stratify patients should be established.

Conclusion

The Catboost model developed in our study demonstrated a strong prediction capability for the one-year OS of patients with AITL. Further validation of the model’s predictive accuracy should be conducted using larger cohorts of participants.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

XH and CZ performed the research and wrote the paper, KL, FW and TF analyzed the data, CX designed the research study and revised the paper.

Funding information

This work was supported by grants from the Sichuan science and technology department for key research and development projects (Grant/Award Number: 2019YFS0027), National Natural Science Foundation of China (82000210), and National Key Clinical Specialty Discipline Construction Program (2021-76).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate.

This study was approved by the Ethics Committee of Fujian Medical University Union Hospital, the Ethics Committee of West China Hospital of Sichuan University, the Ethics Committee of Sichuan Cancer Hospital, and the Ethics Committee of Sichuan Provincial People’s Hospital. All the patients included in the study signed the informed consent.