A machine learning-based model to predict POD24 in follicular lymphoma: a study by the Chinese workshop on follicular lymphoma

Abstract Background Disease progression within 24 months (POD24) significantly impacts overall survival (OS) in patients with follicular lymphoma (FL). This study aimed to develop a robust predictive model, FLIPI-C, using a machine learning approach to identify FL patients at high risk of POD24. Methods A cohort of 1,938 FL patients (FL1-3a) from seventeen centers nationwide in China was randomly divided into training and internal validation sets (2:1 ratio). XGBoost was utilized to construct the POD24-predicting model, which was internally validated in the validation set and externally validated in the GALLIUM cohort. Key predictors of POD24 included lymphocyte-to-monocyte ratio (LMR), lactate dehydrogenase (LDH) > ULN, low hemoglobin (Hb), elevated beta-2 microglobulin (β2-MG), maximum standardized uptake value (SUVmax), and lymph node involvement. The FLIPI-C model assigned 2 points to LMR and 1 point to each of the other variables. Results The FLIPI-C model demonstrated superior accuracy (AUC) for predicting POD24 and 3-year overall survival (OS) in both the internal (AUC POD24: 0.764, OS: 0.700) and external validation cohorts (AUC POD24: 0.703, OS: 0.653), compared to existing models (FLIPI, FLIPI-2, PRIMA-PI, FLEX). Decision curve analysis confirmed the superior net benefits of FLIPI-C. Conclusions Developed using a machine learning approach, the FLIPI-C model offers superior predictive accuracy and utilizes simple, widely available markers. It holds promise for informing treatment decisions and prognostic assessments in clinical practice for FL patients at high risk of POD24.