This was a retrospective analysis of data from a single institution. All patients underwent surgery at the Department of Neurosurgery, Acıbadem Mehmet Ali Aydınlar University School of Medicine between September 2011 and February 2019. Only adult patients with supratentorial, hemispheric diffuse IDH1/2-wt gliomas carrying one of the two TERTp hotspot mutations (C228T or C250T), identified on molecular analysis of the solid tumor, and who underwent preoperative diagnostic ¹H-MRS were included. Thalamic, brainstem, cerebellar, spinal, recurrent, and pediatric cases (aged<18 years) were excluded. The final cohort included 57 (IDH-wt, TERTp-mut) adult patients with hemispheric diffuse gliomas (35 men/22 women, mean age: 58 ± 10.47 years, median: 59 years, range: 39–75 years). All the patients were followed up for up to 65 months, and their overall and progression free survival times were recorded. The study protocol was approved by the Institutional Review Board of Acıbadem Mehmet Ali Aydınlar University School of Medicine, and written informed consent was obtained from all patients, who approved the use of their data for research.

The histopathological grade was determined by a single neuropathologist (AED, 12 years of experience) (Hoshide and Jandial, 2016). Histological grading was performed based on the presence of mitosis, vascular endothelial proliferation, necrosis, and IDH and TERTp mutations. One case that had necrosis on magnetic resonance imaging (MRI) but no histopathological examination was considered to have a WHO 2016 grade-4 tumor. The cohort included 45 (79%) patients with histological WHO 2016 grade-4 tumors, eight (14%) patients with WHO 2016 grade-3 tumors, and four (7%) patients with WHO 2016 grade-2 tumors. Molecular testing for IDH1(R132) and IDH2(R140 and R172) and TERTp (C228T and C250T) hotspot mutations was performed in all cases at the time of diagnosis using mini-sequencing and/or Sanger sequencing, as previously described (Akyerli et al., 2018).

The patients underwent MRS 1–7 days before surgery using a clinical 3 T MR scanner (Siemens, Erlangen-Germany) and a standard brain tumor MR imaging protocol that included pre- and post-contrast (gadolinium DTPA) T1-weighted TSE (repetition time [TR] = 500 ms, TE = 10 ms), post-contrast T1-weighted volumetric TurboFLASH (TR = 2000 ms, TE = 4 ms), post-contrast T2-weighted TSE (TR = 5,000 ms, TE = 105 ms), and T2*-weighted gradient-echo echo-planar imaging dynamic susceptibility contrast MRI (TR = 1,500 ms, TE = 30 ms). ¹H-MRS data were acquired from the solid tumor region excluding necrosis, edema, and hemorrhage using a point resolved spectroscopy (PRESS) sequence (TR/TE = 2000/30 ms, 1,024 points, 1,200 Hz, voxel size = 1–8 cm³, number of signal averages = 192, acquisition time = 6.5 min). Hemorrhage and necrosis were determined from pre- and post-contrast T1-weighted MRI scans. Since ¹H-MRS data acquisition close to the air/tissue interface or scalp results in shimming difficulties, magnetic susceptibility artifacts, and lipid contamination, the voxel was placed as far away from these regions as possible. Contrast enhancement and necrosis were determined from the post-contrast T1-weighted MRI by an experienced radiologist with more than 30 years of experience (AD) to evaluate their effects on the survival of IDH-wt, TERTp-mut gliomas. The percentage of the extent of resection was calculated by comparing the pre-and post-operative segmented tumor volumes on axial T2-weighted MRI.

A basis set of 19 metabolites including creatine (Cr), phosphocreatine (PCr), gamma-aminobutyric acid, glutamine (Gln), glutamate (Glu), glycine (Glyc), glycerophosphocholine (GPC), phosphocholine (PCh), glutathione (GSH), 2-hydroxyglutarate (2HG), myo-inositol (mIns), lactate (Lac), N-acetyl aspartate (NAA), N-acetylaspartylglutamic acid (NAAG), and five composite peak concentrations (total choline [tCho = GPC + PCh], total creatine [tCr = Cr + PCr], total NAA [tNAA = NAA + NAAG], glutamine–glutamate complex [Glx], and myoinositol and glycine [mIns + Glyc]) were used to quantify the concentrations of the metabolites in the LCModel spectral fitting program (Provencher, 2001). The basis spectra were simulated using the chemical shifts and coupling constants provided by Govindaraju et al. (2001). In the LCModel control file, which is provided in the Appendix, the ppm window range and some data acquisition parameters (hzppm, echot, nunfill and deltat) were defined. In addition to metabolite concentrations in arbitrary units, metabolite to tCr ratios were also assessed. The basis set also included alanine, aspartate, glucose, scyllo-inositol, and taurine; however, these metabolites could not be quantified in most patients and were not included in further analysis. The signal to noise ratio (SNR) is defined as the ratio of the maximum in the spectrum-minus-baseline over the analysis window to twice the root mean square of residuals in LCModel, while FWHM (full width at half-maximum) is a rough estimate of the linewidths of the in vivo spectrum. Any spectrum with an SNR < 4 and a FWHM of >0.12 ppm was excluded from further analysis. A Cramer–Rao lower bound threshold of >30% was used, and the metabolites estimated with higher CRLB values were assigned “not a number.” The correlation matrix of quantification was created in LCModel, and negative correlations (≤-0.5) between the metabolites, which indicates coupled quantification due to their similar peak patterns, were investigated. Additionally, Pearson correlation coefficients were computed between the metabolites identified as correlated by LCModel. The coupled metabolites were excluded from the Cox multivariate regression and classification analyses.

A Mann–Whitney U test was used to assess the metabolic differences between “IDH-wt, TERTp-mut diffuse gliomas” with and without contrast enhancement and between those with and without radiological necrosis. Bonferroni multiple comparison correction was applied, and a p-value of ≤0.003 was considered as statistically significant for metabolic differences. The effects of contrast enhancement, necrosis, extent of resection, and metabolite concentrations on OS and PFS of “IDH-wt, TERTp-mut diffuse gliomas” were evaluated using Kaplan–Meier survival analysis, followed by a log-rank test. Associations between survival outcomes and different clinical predictors (age and sex), imaging features (the presence of contrast enhancement and necrosis), extent of resection, and metabolite concentrations were assessed using univariate and multivariate Cox regression models. Only the significant univariate parameters were included in the multivariate Cox regression analysis. Classification and regression tree (CART) analysis was used to calculate the predictive power of these variables on OS and PFS. Then, supervised machine learning algorithms were employed to identify IDH-wt, TERTp-mut gliomas with PFS of longer than 6 months and OS of longer than 12 months. A baseline model was developed based on clinical features, including age, contrast enhancement, necrosis and extent of resection for classification of IDH-wt, TERTp-mut gliomas with high PFS or OS. We also created classification models using the features selected between the metabolite concentrations as well as age, extent of resection, contrast enhancement and necrosis. Feature selection with the least absolute shrinkage and selection operator (Lasso) (Tibshirani, 1996), random forest recursive feature elimination (RFRFE) (Granitto et al., 2006), or sequential floating forward selection (SFFS) (Pudil et al., 1994) was applied before machine learning-based classification to eliminate redundant features and prevent overfitting. Fifteen models derived from well-known machine learning algorithms, such as k nearest neighbor (KNN), support vector machines (SVM), and ensemble trees, were used to classify IDH-wt, TERTp-mut gliomas with high PFS or OS based on their metabolic profiles. Five-fold cross-validation was used, and the models were executed 100 times to report the mean performance measures. The synthetic minority oversampling technique (SMOTE) (Chawla et al., 2002) was also employed to overcome the class imbalance problem in our classification. SMOTE was implemented in Python, CART analysis was performed using SPSS statistics 25 (IBM Corp., Armonk, NY), and the remaining computations were performed using MATLAB 2022a (MathWorks, Natick, MA).

Table 1 shows the characteristics of the study participants, including the pathological group and WHO 2016 grade, age, sex, tumor location, contrast enhancement and necrosis. Most tumors occurred in the frontal lobe region (19), followed by the temporal lobe (12), and the insular lobe (10). Gliomatosis cerebri growth pattern were observed only in two patients. Contrast enhancement was observed in 2/12 (17%) patients with WHO 2016 grade-2/3 tumors and 44/45 (98%) patients with WHO 2016 grade-4 tumors. Necrosis occurred in 43/45 (96%) patients with WHO 2016 grade-4 tumors. The median value for extent of resection was 61% [range = 10–100%]. Additionally, the extent of resection was higher than 95% in only seven patients.

The average maximum SNR was 18.36 ± 14.43 and the average FWHM was 0.06 ± 0.02 ppm for the spectra of this patient population. Supplementary Figure S1 shows the mean correlation matrix output of the LCModel averaged over all patients. The correlation coefficient was <−0.5 for Cr and PCr in 39 patients (r = −0.64, p < 0.001) and for Glyc and mIns in 35 patients (r = −0.52, p < 0.001). The Pearson correlation coefficients were computed between the overlapping metabolites, which were −0.82 (p < 0.001) for Cr and PCr and −0.70 (p < 0.001) for Glyc and mIns. 2HG/tCr and NAAG/tCr could not be determined in more than 30% of the patients in each group, and these metabolites were excluded from further analysis. Additionally, Cr-PCr, Glu-Gln, GPC-PCh, and NAA-NAAG were coupled metabolites, so their composite values were included in the Cox multivariate regression analysis and the classification. Figure 1 shows short-TE PRESS MRS data along with the LCModel quantification results of two example gliomas.

Contrast-enhanced tumors had significantly higher Gln/tCr (p = 0.002), GPC/tCr (p = 0.002), tCho/tCr (p < 0.001), and Glx/tCr (p < 0.001) levels than non-contrast-enhanced tumors (Supplementary Figure S2A). In addition, there was a trend for higher Glu/tCr (p = 0.009) and GSH/tCr (p = 0.007) (Supplementary Figure S2A).

Tumors with radiologically determined necrosis had higher Gln/tCr (p < 0.001), Glu/tCr (p < 0.001), GPC/tCr (p = 0.003), GSH/tCr (p < 0.001), Lac/tCr (p < 0.001), tCho (p < 0.001), and Glx (p < 0.001) levels than the other tumors (Supplementary Figure S2B).

Supplementary Figure S3 shows an additional spider plot visualization of the results presented in Supplementary Figure S2.

The anatomical localization was not statistically significantly associated with either PFS or OS by using a log-rank test (p = 0.212 for PFS and p = 0.192 for OS) or univariate Cox regression analysis (p = 0.949 for PFS; p = 0.465 for OS). On the other hand, both contrast enhancement and necrosis were associated with significantly shorter PFS (p = 0.014 and p = 0.012, respectively, log-rank test; Figures 2A,C). However, contrast enhancement and necrosis were not found to be associated with a shorter OS (p = 0.061 and p = 0.057, respectively, log-rank test; Figures 2B,D). The Kaplan–Meier analysis also showed that a higher Glx/tCr measurement was associated with a significantly shorter PFS (p = 0.032) and a significantly shorter OS (p = 0.047; Figures 3A,B). A higher GSH/tCr ratio was also associated with a significantly shorter PFS (p = 0.004; Figures 3C,D). The median PFS and OS of WHO 2016 grade-2/3 tumors were 18 and 22.5 months, while those of WHO 2016 grade-4 tumors were 10 and 19 months, respectively. WHO 2016 grade-2/3 tumors had significantly longer PFS (p = 0.027; log-rank test) than WHO 2016 grade-4 tumors; however, there was no statistically significant difference in OS between the two groups (p = 0.057; log-rank test).

Univariate Cox model analysis using nine metabolic parameters as well as age, sex, contrast enhancement, and radiological necrosis indicated that six parameters were associated with a shorter PFS, which were the contrast enhancement (p = 0.015); presence of necrosis (p = 0.012); and higher levels of Glu/tCr (p = 0.007), GSH/tCr (p = 0.019), tCho/tCr (p = 0.032), and Glx/tCr (p = 0.010) (Table 2). Univariate Cox model analysis also revealed that the presence of necrosis (p = 0.049); and higher levels of Glu/tCr (p = 0.039), and Glx/tCr (p = 0.047) were associated with a significantly shorter OS (Table 2). On the other hand, multivariate Cox model analyses identified no factors that were significantly associated with OS or PFS (Table 3).

CART analysis of “IDH-wt, TERTp-mut diffuse gliomas” indicated that a high GSH/tCr (>0.449), and a high Glx/tCr (>1.474) were associated with a significantly shorter PFS (Figure 4A). Additionally, high Glx/tCr (>2.966) was associated with a significantly shorter OS (Figure 4B). The sensitivity, specificity, and accuracy results along with the classification model, feature selection method, and selected metabolites for classification according to high PFS or OS are presented in Table 4. The models using MRS markers had better performance after oversampling than both the baseline models using only clinical features and the models that incorporated clinical features along with the MRS markers. The highest classification accuracy based on 12 months OS was 70.37%, but the specificity was lower (41.53%), since 67% of the patients had a longer than 12 months OS. Oversampling assisted models to learn better about the patients with longer OS, and following SMOTE, the specificity increased to 76.36% (accuracy = 81.71%, sensitivity = 87.05%). The most useful features for the classification of low and high OS were the extent of Glx/tCr, tCho/tCr, tNAA/tCr and (mIns+Glyc)/tCr. Similarly, for PFS based classification, the accuracy of the machine learning models for identifying the patients having longer than 6 months PFS was not very high (best accuracy = 63.93%, sensitivity = 68.41%, specificity = 54.22% using the MRS markers models). SMOTE also improved the PFS classification results for the minority class, resulting in an accuracy of 77.41% (sensitivity = 75.47%, specificity = 79.48%). Lac/tCr, tNAA/tCr, Glx/tCr and (mIns+Glyc)/tCr were used in this prediction.

The machine learning results for predicting short PFS and OS using the combinations of all different models and three different feature selection algorithms with or without oversampling are provided at Supplementary Tables S1 and S2, respectively. Median AUC for all the models when Lasso was employed as a feature selection algorithm to predict short OS was 0.82 [0.5–1]. On the other hand, the median AUC values were 0.84 [0.52–1] and 0.83 [0.53–1] when SFFS and RFE were employed for predicting short OS, respectively. For the oversampled dataset, the median AUC value was 0.85 [0.67–1] when Lasso was used as a feature selection algorithm. The median AUC values were 0.86 [0.62–1] and 0.85 [0.65–1] when SFFS and RFE were employed for predicting short OS, respectively.

The median AUC for all models, when utilizing Lasso as the feature selection algorithm to predict short PFS, was 0.84 [0.45–1]. On the other hand, when SFFS and RFE were employed to predict short PFS, the median AUC values were 0.88 [0.49–1] and 0.83 [0.54–1], respectively. When examining the oversampled dataset, the median AUC value for predicting short PFS was 0.91 [0.56–1] when utilizing Lasso as the feature selection algorithm. Similarly, when applying SFFS and RFE to predict short overall survival (OS) with the oversampled dataset, the median AUC values were 0.89 [0.57–1] and 0.92 [0.57–1], respectively.