This population-based retrospective cohort study utilized data from SEER 22 Registries, which collectively cover approximately 28% of the U.S. population. Patients diagnosed with diffuse midline glioma (DMG) between 2018 and 2021 were identified using the ICD-O-3 morphology code 9385/3. Although the SEER database labels this code as “Diffuse intrinsic pontine glioma, H3 K27M-mutant,” it was introduced in 2018 following the WHO 2016 CNS classification to represent diffuse midline glioma, H3 K27M-mutant. This terminology lag is common in SEER, as legacy names are retained to ensure compatibility with earlier datasets (e.g., “glioblastoma multiforme, NOS [9440/3]” remains listed despite the removal of “multiforme” in recent WHO editions). Extracted variables included demographic characteristics (age, sex, race, household income), clinical information (primary tumor location, time from diagnosis to treatment), and treatment modalities (extent of surgery, radiotherapy, and chemotherapy).

In addition, an independent institutional cohort comprising 22 pathologically confirmed DMG patients treated with chemoradiotherapy was retrospectively collected for external validation. Clinical and treatment data were extracted following the same variable definitions as the SEER cohort.

Patients from the SEER dataset were randomly assigned to a training cohort (70%) and a validation cohort (30%) for nomogram construction and internal validation, respectively. Continuous variables were summarized as mean ± standard deviation (for normally distributed data) or median [interquartile range, IQR] (for non-normally distributed data), and compared using Student’s t-test or Mann–Whitney U test, as appropriate. Categorical variables were presented as frequencies and percentages, with group comparisons performed using Chi-square or Fisher’s exact test. We used the missRanger package in R, which performs multiple imputation based on random forest algorithms.

Univariate Cox proportional hazards regression was used to identify candidate prognostic factors associated with overall survival (OS), and variables with p < 0.20 were entered into a multivariable Cox model. A bidirectional stepwise selection procedure, guided by the Akaike Information Criterion (AIC), was employed to determine independent predictors (p < 0.05).

Based on the final model, a prognostic nomogram was developed using the rms package in R. Model performance was assessed by discrimination and calibration. Discrimination was evaluated using the concordance index (C-index) and time-dependent receiver operating characteristic (ROC) curves (via the riskRegression package). Calibration was evaluated by plotting the predicted versus observed OS at 6, 12, and 24 months.

A total risk score was calculated for each patient in the SEER cohort based on the nomogram, and the optimal cutoff value was determined using the survival and survminer packages. Patients were classified into high- and low-risk groups accordingly. Kaplan–Meier survival analysis and log-rank tests were used to assess differences in survival between risk strata.

For external validation, the institutional cohort (n = 22) was stratified using the SEER-derived cutoff. Kaplan–Meier curves were generated, and log-rank tests were conducted to evaluate the prognostic separation between risk groups. All statistical tests were two-sided, and p < 0.05 was considered statistically significant.

We identified 409 patients diagnosed with DMG from the SEER database. The median overall survival (OS) was 9 months (IQR: 4–16 months), and the median age was 12 years. Female patients accounted for 53.8% of the cohort. The brainstem was the most common tumor location (49.4%). Radiotherapy and chemotherapy were administered to 74.1% and 51.3% of patients, respectively. The training and validation cohort distributions are shown in Table 1 .

The incidence of DMG reached its peak in children aged 5 to 9 years old and gradually decreased thereafter. There was a significant association between age and tumor location (χ²=24.6, p<0.001): brainstem tumors were predominant in patients under 14 years (70.8%), whereas non-brainstem tumors were more common in adults and the elderly ( Figure 1 ).

Kaplan–Meier analysis ( Figure 2 ) demonstrated that older age (>12 years; p < 0.001), non-brainstem tumor location (p < 0.001), receipt of chemotherapy (p < 0.001), higher household income (≥$100,000; p = 0.038), and shorter diagnosis-to-treatment intervals (p = 0.045) were significantly associated with improved OS. Sex, race, and extent of surgical resection showed no significant correlation with survival (all p > 0.05).

Univariate Cox regression analysis ( Table 2 ) showed protection factors, including age >12 years (HR = 0.565, 95%CI 0.416-0.769; p<0.001), male sex (HR = 0.724, 95%CI 0.535-0.981; p=0.037), non-Hispanic White ethnicity (HR = 0.678, 95%CI 0.476-0.965; p=0.031), and higher income (≥$100,000; HR = 0.661, 95%CI 0.452-0.968; p=0.033). Conversely, the lack of chemotherapy (HR = 1.814, 95%CI 1.338-2.458; p<0.001) or radiotherapy therapy (HR = 1.608, 95%CI 1.129-2.289; p=0.008) independently predicted poorer results.

Multivariate Cox regression ( Figure 3 ) confirmed that age >12 years (HR = 0.602, 95%CI 0.413-0.877; p=0.0083), household income ≥$100,000 (HR = 0.642, 95%CI 0.436-0.947; p=0.0254), and cerebellum tumors location (vs. brainstem: HR = 2.839, 95%CI 1.310-6.152; p=0.0082) as independent prognostic factors. Therapeutically, radiotherapy alone (HR = 0.593, 95%CI 0.378-0.931; p=0.0232) and combined chemoradiotherapy (HR = 0.411, 95%CI 0.270-0.626; p<0.001) significantly improved survival compared with no/unknown treatment, whereas chemotherapy alone showed no benefit (HR = 0.606, 95%CI 0.270-1.725; p=0.3481).

Based on these independent predictors, we developed a nomogram predicting OS at 6, 12, and 24 months ( Figure 4 ). Younger age (≤12 years), lower income, cerebellum location, and lack of treatment (radiotherapy or chemotherapy) corresponded to higher nomogram scores, indicating poorer prognosis.

The nomogram exhibited moderate discrimination in both training (6-month AUC = 0.708; C-index=0.660) and validation cohorts (24-month AUC = 0.725; C-index=0.575) ( Figure 5 ). Calibration curves demonstrated acceptable predictive accuracy. Risk stratification based on optimal cutoff (71.099 points) succeeded in differentiating high- and low-risk groups, demonstrating a significant difference in survival in training (HR = 2.53, p<0.001) and validation cohorts (HR = 1.67, p=0.021).

To assess real-world applicability, we applied the nomogram to an independent institutional cohort of 22 DMG patients. Patients were stratified into high- and low-risk groups using the same cutoff value. Survival curves showed a trend toward poorer outcomes in the high-risk group, although the difference did not reach statistical significance (p = 0.188) ( Figure 6 ). Baseline clinical features of these patients are summarized in Supplementary Table 1 .

As in previous literature, we found a clear age-dependent tumor site distribution: pediatric DMG was primarily located in the brainstem (classic DIPG); versus adults with mainly thalamus or spinal cord involvement (17–21). Consistent with previous findings, this age-dependent survival difference, combined with the anatomical distribution of tumors, suggesting fundamental biological heterogeneity between pediatric and adult DMGs, with age probably being a key driver for the heterogeneity of the molecular behavior of tumors (17, 19, 22–25). The survival time of children with DMG was significantly different, and the prognosis of the patients with thalamic or spinal cord tumor was better than those with brainstem tumors (26). These observations support the role of developmental biology in disease behavior and advocate for location- and age-adapted treatment strategies. These findings are consistent with prior reports identifying both age and tumor location as independent prognostic factors in DMG, supporting their inclusion in our prognostic model.

One of the main methodological advances in our study was the introduction of a refined three-tiered classification of therapy (radiation therapy alone, chemotherapy alone, or combined chemoradiotherapy) to overcome the limitations of conventional binary classifications (for example, radiation versus no radiation, chemotherapy versus no chemotherapy).Our analysis showed that chemotherapy alone did not provide a survival benefit, while the combination treatment significantly improved the outcome — thus correcting the misleading conclusion that only chemotherapy in the binary analysis improved the prognosis. This difference can be explained by the fact that the brain stem and thalamic areas are particularly intact (BBB), which limits the penetration of chemotherapy agents like temozolomide (TMZ), and eventually reduces the effectiveness of monotherapy (27–29). In addition, the methylation of the MGMT promoter is typically absent in DMG, resulting in a poor efficacy of TMZ (27, 30). On the contrary, it has been found that radiation therapy can temporarily destroy the blood-brain barrier, which may increase the delivery efficiency and therapeutic efficacy of chemotherapy drugs at the tumor site (31). Although SEER lacks molecular data, these interpretations are supported by biological plausibility and prior literature. These findings highlight the need for individualized treatment regimens based on tumor location and biological characteristics, ideally guided by future molecular-integrated datasets.

Consistent with prior research, our analysis demonstrated that surgical resection extent—biopsy, partial, or subtotal—did not significantly affect OS in DMG patients (32–34). This result highlights DMG’s diffuse infiltrative nature and frequent involvement of anatomically critical midline structures, significantly limiting surgical efficacy. Although our study did not specifically analyze surgical outcomes stratified by tumor location, the predominance of brainstem lesions (where aggressive resection is rarely feasible) likely contributed to this overall negative result.

These findings support minimally invasive biopsy for molecular characterization as the preferred standard approach, given its vital role in diagnosis, prognosis, and eligibility for targeted therapy trials (13). Prospective analyses involving larger cohorts of non-brainstem DMGs may help clarify whether location-specific surgical approaches could provide selective survival benefits.

Based on our prognostic nomogram, we propose a risk-adapted management approach to improve clinical outcomes. Patients identified as high-risk may be candidates for early enrollment in clinical trials investigating novel targeted therapies (e.g., ONC201 or GD2 CAR-T therapy), which were not captured in the SEER database but represent promising investigational options for DMG (11, 12, 35–40), while low-risk patients could benefit from standard chemoradiation as recommended by current guidelines. This stratification approach aligns with precision medicine principles, facilitating personalized clinical decisions and optimizing therapeutic strategies. Prospective validation of this strategy is crucial for further refinement of individualized treatment algorithms.

To further evaluate the generalizability of our model, we performed external validation using an independent institutional cohort (n=22). Although statistical significance was not achieved in this small dataset (p=0.081), the direction and magnitude of the risk stratification effect mirrored the SEER findings, suggesting a consistent prognostic trend across populations. Given the inherent challenges in assembling large, histologically confirmed DMG cohorts, this level of validation is rare and valuable. Nonetheless, caution should be exercised in interpreting these results due to the limited sample size, and larger multicenter prospective studies are warranted to substantiate the external validity of our model.

This study is constrained by the inherent limitations of retrospective SEER-based analysis, including the lack of molecular and detailed treatment data. While we incorporated an external validation cohort (n=22) to enhance model credibility, statistical power remains limited. Future prospective studies should incorporate comprehensive molecular profiles (e.g., H3K27M status, MGMT methylation) and multicenter validation to refine prognostic models and strengthen clinical utility.

In addition, socioeconomic factors also emerged as a prognostic variable in our model. Although not directly related to tumor biology, previous studies in glioma and other cancers have shown that lower socioeconomic status is associated with limited access to care, reduced treatment adherence, and worse survival outcomes (41–43). This highlights the need to consider not only biological but also socioeconomic factors in DMG management, and future prospective studies should further examine these associations.