JMIR Publications

ABSTRACT

Background:

Adult-type diffuse glioma (ADG) is the most common primary malignant tumor of the central nervous system. Its highly invasive nature, marked heterogeneity, and resistance to therapy contribute to a high risk of recurrence and poor prognosis. At present, the lack of reliable prognostic tools poses a significant barrier to the development of individualized treatment strategies.

Objective:

This study aimed to develop an effective prognostic model for ADG by integrating multiple machine learning algorithms, in order to enhance the precision of individualized clinical decision-making.

Methods:

In this retrospective study, 160 newly diagnosed ADG patients who underwent surgical resection and histopathological confirmation at our institution between June 2019 and September 2021 were included. A total of 32 variables, including clinical characteristics, molecular biomarkers, and preoperative hematological indicators, were collected. Overall survival (OS) and progression-free survival (PFS) were defined as the study endpoints. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression, extreme gradient boosting (XGBoost), and random forest algorithms. Kaplan–Meier (KM) survival curves and log-rank tests were used for survival analysis. Multivariate Cox proportional hazards models were constructed to identify independent prognostic factors, and nomograms were developed accordingly. Model performance was evaluated using receiver operating characteristic (ROC) curves, concordance index (C-index), area under the curve (AUC), calibration plots, and KM survival analysis to assess the model’s discriminative ability, calibration, and clinical utility.

Results:

Age, neutrophil percentage-to-albumin ratio (NPAR), and platelet-to-mean platelet volume ratio (PMR) were identified as independent prognostic factors for OS, while age and NPAR were independent predictors for PFS (all P < .001). The prognostic models based on these variables demonstrated good predictive performance, with C-index values of 0.731 and 0.763 for the training and validation cohorts in the OS model, respectively. The PFS model also showed robust performance. AUC values and calibration curves further supported the models' accuracy and stability. Risk stratification analysis revealed clear survival differences between risk groups (all P < .05), indicating strong clinical applicability.

Conclusions:

This study is the first to identify preoperative NPAR as a significant prognostic biomarker for ADG using machine learning approaches. The prognostic model incorporating NPAR, PMR, and age demonstrated favorable predictive performance, offering a novel perspective for accurate risk stratification and personalized treatment in ADG patients.