Prognosis Prediction and Treatment Recommendation for Follicular Lymphoma: Enhancing Decision-Making

Backgroud:Despite the abundance of therapeutic options, advanced-stage follicular lymphoma remains incurable. Randomized trials of rituximab maintenance (MR) have shown improved progression-free survival (PFS) in patients with follicular lymphoma, but its effect on overall survival remains inconclusive. Data on R-maintenance in the Chinese population has always been lacking, and there have been few reports on the duration of maintenance cycles and safety.To evaluate the impact of MR on overall survival based on patient and disease characteristics, and to explore certain adverse events, we utilized real-world patient electronic health records (EHR). Our objective was to develop a lymphoma disease model that aligns with EHR data characteristics and meets precision medicine needs. This involved constructing a knowledge analysis and treatment recommendation system to achieve precise risk stratification and diagnosis for lymphoma patients, and to enhance the effectiveness and sequencing of treatment plans.

Methods:

This study retrospectively included clinical data from 444 patients diagnosed with FL grades 1-3A, treated between November 1, 2001, and December 31, 2019, at four centers: Peking University Cancer Hospital, Peking University International Hospital, Inner Mongolia Cancer Hospital, and Tangshan Workers' Hospital. Data preprocessing and screening included handling missing values and performing basic statistical observations, such as plotting curves of progression time with or without R maintenance.The primary endpoint was POD24, while secondary endpoints included progression-free survival (PFS), overall survival (OS), and safety. We used a combination of random forest, uplift model, and catboost methods to predict the risk of tumor progression within 24 months. Features highly correlated with tumor progression were extracted based on model feature importance.

Results:

We conducted an evaluation using ten-fold cross-validation, covering baseline static information, laboratory tests, imaging, and pathological examination features from 444 patients. The model's AUROC for predicting POD24 was 0.859, AUPRC was 0.780, and accuracy was 0.768. Using feature selection via random forest and ranking by Qini importance, we identified that highly important features include PET-CT lesion length, SUVmax, lactate dehydrogenase, and β2-microglobulin, all of which are clinically accessible indicators.

Regarding the model for R maintenance therapy for each patient, we utilized individual treatment effect (ITE) estimation. It was found that patients with SUVmax > 17 had the lowest mean uplift, indicating significant benefit from R maintenance therapy when SUVmax is very high. The youngest patients (<24 years old) had the lowest benefit level, while the oldest patients (≥65 years old) had the highest benefit level, suggesting that R maintenance therapy is more recommended for older patients.

Conclusion：Our model offers evidence-based recommendations and highlights potential risks for FL patients, aiding healthcare professionals in real-world decision-making. This support can enhance patient adherence and improve overall outcomes.

No relevant conflicts of interest to declare.