Childhood brain tumor prognosis moves past subgroup labels: MANTIS draws 10-year individual survival curves from methylation, reaching c-index 0.80

Risk stratification for childhood brain tumors shifts from grouping patients by subgroup and clinical indicators to giving each patient a 10-year survival probability curve. Previous stratification relied on methylation subgroup, metastatic status and extent of resection, and substantial outcome heterogeneity persisted outside favorable-risk groups, with many patients relapsing despite no established high-risk features. In the average-risk group of the independent SJMB03 trial cohort (M0 with gross total or near-total resection, n = 206, receiving 23.4 Gy standard-dose craniospinal irradiation), MANTIS-MB reached a c-index of 0.80 ± 0.07, a 5-year mean AUROC of 0.85 ± 0.04 and an IBS of 0.08 ± 0.01.

The model groups methylation probes by gene, compresses them into a gene layer, and then predicts survival, cutting input dimensionality while keeping time information in the prediction. Conventional practice treats methylation profiles as binary features or feeds them to generic survival models; MANTIS instead uses multi-task logistic regression, splitting 10 years into intervals and predicting event probabilities jointly so censored follow-up data are used. Training used 820 medulloblastoma patients with follow-up and calibration used 100; the c-index was 0.79 ± 0.05 across the full SJMB03 cohort (n = 305) and 0.76 ± 0.04 across the combined standard-dose ACNS0331 and ACNS0332 cohorts (n = 507).

The same framework is extended to estimate individual treatment effects, used to judge which patients can safely receive reduced craniospinal irradiation dose. In the randomized ACNS0331 trial, low-dose irradiation produced worse outcomes overall, but the difference concentrated in some subgroups, suggesting cohort-level failure does not rule out individual-level suitability. In the randomized ACNS0331 cohort, patients with high predicted treatment effect (ITE > 0.1, n = 66) had significantly worse survival under low-dose irradiation (P = 0.00245), while patients with low predicted effect (ITE < 0.1, n = 104) showed no detected survival disadvantage between low and standard dose (P = 0.683).

The validated scope is retrospective data: the medulloblastoma model was evaluated in three independent trial cohorts, SJMB03, ACNS0331 and ACNS0332, and the reported performance estimates pertain to patients older than 3 years; the ependymoma model was validated in 100 PF-A and PF-B patients treated under standard clinical protocols. The treatment-effect analysis is restricted to patients aged 3 to 7 years with average-risk medulloblastoma randomized in ACNS0331 and covers only craniospinal irradiation dose as the treatment intensity.

The authors state that clinical utility still requires prospective trials, and that clinical interpretability of predictions and patient benefit have not been demonstrated prospectively; treatment-effect modeling currently covers only craniospinal irradiation dose, not chemotherapy intensity, targeted therapy or immunotherapy; the model also does not model late effects such as secondary malignancies. In the SHH subgroup, MANTIS-MB reached a c-index of 0.77 ± 0.07, not outperforming SJMB03 (0.74 ± 0.05, P = 0.21) or the revised clinico-molecular scheme (0.79 ± 0.05, P = 0.65).