Variable screening and model construction for prognosis of elderly patients with lower-grade gliomas based on LASSO-Cox regression: a population-based cohort study

Abstract

Background: This study aimed to identify prognostic factors for survival and develop a prognostic nomogram to predict the survival probability of elderly patients with lower-grade gliomas (LGGs).

Methods: Elderly patients with histologically confirmed LGG were recruited from the Surveillance, Epidemiology, and End Results (SEER) database. These individuals were randomly allocated to the training and validation cohorts at a 2:1 ratio. First, Kaplan-Meier survival analysis and subgroup analysis were performed. Second, variable screening of all 13 variables and a comparison of predictive models based on full Cox regression and LASSO-Cox regression analyses were performed, and the key variables in the optimal model were selected to construct prognostic nomograms for OS and CSS. Finally, a risk stratification system and a web-based dynamic nomogram were constructed.

Results: A total of 2307 elderly patients included 1220 males and 1087 females, with a median age of 72 years and a mean age of 73.30 ± 6.22 years. Among them, 520 patients (22.5%) had Grade 2 gliomas, and 1787 (77.5%) had Grade 3 gliomas. Multivariate Cox regression analysis revealed four independent prognostic factors (age, WHO grade, surgery, and chemotherapy) that were used to construct the full Cox model. In addition, LASSO-Cox regression analysis revealed five prognostic factors (age, WHO grade, surgery, radiotherapy, and chemotherapy), and a LASSO model was constructed. A comparison of the two models revealed that the LASSO model with five variables had better predictive performance than the full Cox model with four variables. Ultimately, five key variables based on LASSO-Cox regression were utilized to develop prognostic nomograms for predicting the 1-, 2-, and 5-year OS and CSS rates. The nomograms exhibited relatively good predictive ability and clinical utility. Moreover, the risk stratification system based on the nomograms effectively divided patients into low-risk and high-risk subgroups.

Conclusion: Variable screening based on LASSO-Cox regression was used to determine the optimal prediction model in this study. Prognostic nomograms could serve as practical tools for predicting survival probabilities, categorizing these patients into different mortality risk subgroups, and developing personalized decision-making strategies for elderly patients with LGGs. Moreover, the web-based dynamic nomogram could facilitate its use in the clinic.

Keywords: SEER; cancer-specific survival; elderly; lower-grade glioma; nomogram; overall survival.

Figure 1

The flow diagram of case inclusion and exclusion. LGG, lower-grade glioma; SEER, Surveillance, Epidemiology, and End Results.

Figure 2

K−M analysis determining the impact of variables on OS (months) in the training cohort. Stratified by age (A), WHO grade (B), surgery (C), radiotherapy (D), chemotherapy (E), and adjuvant therapy (F). Chemo, chemotherapy; Radio, radiotherapy.

Figure 3

LASSO regression curves of variables in the training cohort. (A, B) The curve of the regression coefficient versus log (λ) and the partial likelihood deviation versus log (λ) for OS. (C, D) The curve of the regression coefficient versus log (λ) and the partial likelihood deviation versus log (λ) for CSS. Lambda.1se represents the optimal lambda (λ) for screening the variables.

Figure 4

The importance of variables in the prognostic nomogram. (A) The IncMSE (left) and IncNodePurity (right) based on the random forest model emphasized the notable influence of variables on the model’s predictive accuracy. (B, C) SHAP summary plots of variables in the nomogram. The ranking of variables’ importance according to the mean SHAP value in machine learning (B). A wider spread of points in the beeswarm plot indicated a stronger influence of the variables in the model (C).

Figure 5

Prognostic nomogram for OS in the training cohort. GTR, gross total resection; STR, subtotal resection; PR, partial resection; Pr, probability.

Figure 6

Assessment and validation of the prognostic nomogram for OS. ROC curves (A), calibration curves (B), and DCA curves (C) of the nomogram in the training cohort. ROC curves (D), calibration curves (E), and DCA curves (F) of the nomogram in the validation cohort. ROC curve, Receiver operating characteristic curve; DCA, Decision Curve Analysis; FPR, False positive rate; TPR, True positive rate.

Figure 7

K−M analyses were used to determine the impact of the risk stratification system based on the prognostic nomograms on OS and CSS. (A, B) Based on the nomograms on OS (A) and CSS (B) in the training cohort. (C, D) Based on the nomogram on OS (C) and CSS (D) in the validation cohort.

Figure 8

A web-based dynamic nomogram for predicting OS of older patients with LGGs. (A) The nomogram incorporated a panel of independent prognostic factors for the prediction of survival outcomes. (B) A survival plot depicting OS based on different clinical features. (C) The graphical summary included line segments indicating the 95% confidence intervals (CIs) for the estimated survival probability. (D) The numerical summary provided detailed information on the variable features and exact predictive values of OS.