The work, led by Jheremy S. Reyes, Constantinos G. Hadjipanayis, L. Dade Lunsford, and Ajay Niranjan at the University of Pittsburgh Medical Center’s Center for Image-Guided Neurosurgery, addresses a population that has long been underserved by predictive modeling. Brain metastases from gastrointestinal primaries, including colorectal, pancreatic, gastric, and biliary cancers, are biologically heterogeneous and historically associated with poorer local control than metastases from lung or breast cancers. Because these tumors are relatively less common in the radiosurgery literature, dose selection for them has been extrapolated from broader metastasis cohorts rather than tailored to their specific behavior. The Pittsburgh team set out to build a model trained specifically on this patient group.

At the heart of the framework is a machine learning architecture known as a mixture of experts, a concept that dates back to foundational work by Jacobs, Jordan, Nowlan, and Hinton in 1991 and was later formalized hierarchically by Jordan and Jacobs. Rather than forcing a single neural network to learn every pattern in the data, a mixture-of-experts model divides the problem among several specialized subnetworks, each of which becomes adept at a particular region of the clinical feature space. A gating network then learns which expert to trust for a given patient, weighting their contributions accordingly. In THINKERS-GI, three such experts operate in parallel, allowing the model to capture distinct subpopulations among GI brain metastasis patients whose tumors and outcomes behave in fundamentally different ways.

The technical machinery beneath the experts is equally deliberate. The framework employs discrete-time competing-risk survival modeling, a statistical approach that estimates the probability of an event, such as local tumor failure, over successive time intervals while explicitly accounting for competing events that can preclude that failure, most notably death from systemic disease. This matters enormously in brain metastases, where many patients have limited life expectancies driven by their systemic cancer, and a model that ignores competing risks will systematically overestimate the benefit of aggressive local therapy. By modeling both local failure and adverse radiation effects in this competing-risk framework, THINKERS-GI can express its outputs as time-dependent probabilities rather than crude binary classifications.

One of the most clinically significant design choices is that the margin dose, the radiation dose prescribed at the edge of the tumor, was built into the model as a queryable input. This transforms the network from a passive outcome predictor into an active decision-support tool. A clinician can ask the model a counterfactual question: what would the predicted probability of local control and of adverse radiation effects be if this lesion received 18 Gy instead of 16 Gy? By sweeping across a range of candidate doses, the framework generates a dose-response profile unique to each lesion, effectively allowing the care team to explore the trade-off between tumor control and toxicity before the first beam is ever delivered.

The study drew on a retrospective single-center cohort of 300 patients treated with Gamma Knife radiosurgery, encompassing 1,233 individual GI brain metastases. Pretreatment clinical and lesion-level variables were used to train the network, and the authors took validation seriously, employing grouped five-fold cross-validation alongside a closed, grouped holdout set. Grouping is a crucial methodological safeguard in this setting: because a single patient can contribute many lesions, randomly splitting lesions across training and test sets would leak patient-specific information and inflate performance estimates. By keeping all lesions from one patient together, the researchers ensured that their reported metrics reflect genuine generalization to unseen patients.

The results were strong by the standards of clinical prediction modeling. In grouped cross-validation, the time-dependent area under the curve ranged from 0.808 to 0.894, with a concordance index of 0.889 plus or minus 0.034, indicating good discrimination between patients who experienced local failure and those who did not. The calibrated twelve-month Brier score, which penalizes both poor discrimination and poor calibration, was a remarkably low 0.019 plus or minus 0.013, and the mean absolute error for predicting the time to local failure was just 1.8 plus or minus 0.7 months. On the more demanding grouped holdout set, performance remained respectable, with a twelve-month AUC of 0.806, a C-index of 0.781, a calibrated Brier score of 0.101, and a mean absolute error for time-to-failure of 1.27 months.

The dose-sweeping experiments offer a glimpse of what routine clinical use might look like. When the framework evaluated candidate doses across the cohort, it selected a dose within two Gy of the dose actually delivered by physicians in 91.8 percent of lesions, a striking signal that the model’s reasoning converges with expert clinical judgment in the vast majority of routine cases. More intriguingly, in the cases where the model suggested deviations, the simulated optimization increased the predicted twelve-month local control rate from 82.3 percent to 84.4 percent while simultaneously reducing the predicted risk of adverse radiation effects from 7.9 percent to 7.6 percent. Improving both endpoints at once is the holy grail of radiation dose optimization, since these outcomes typically pull in opposite directions.

The THINKERS-GI study is not an isolated effort but part of a rapidly expanding research program from the Pittsburgh group, which has applied the same hybrid intelligence philosophy to meningiomas, glioblastomas, vestibular schwannomas, lung cancer brain metastases, melanoma brain metastases, pituitary adenomas, craniopharyngiomas, and post-resection cavities. This disease-specific strategy reflects a growing recognition in oncology AI that a single general-purpose model often performs worse than a family of specialized ones, particularly for rare tumor types with distinctive biology. Related work elsewhere, such as the MoE-Surv framework for rare cancers, has reached similar conclusions, suggesting that mixture-of-experts architectures may become a standard tool for survival prediction in oncology.

Important caveats remain before such tools reach the treatment planning console. The study was retrospective and single-center, and the authors describe their work as internally validated, meaning the model has not yet been tested on patients from other institutions with different treatment protocols and patient populations. External validation, and eventually prospective evaluation, will be essential to establish that the framework’s dose recommendations translate into genuinely better outcomes rather than merely better predictions. The authors also note that no external datasets were generated or analyzed in the current study, underscoring that this is a foundation rather than a finished clinical product. Still, the combination of rigorous competing-risk survival modeling, patient-grouped validation, and queryable dose inputs marks a meaningful step toward a future in which every radiosurgery prescription is informed by an individualized estimate of what the treatment is likely to achieve, and what it is likely to cost the surrounding brain.

Subject of Research: A mixture-of-experts AI framework for predicting outcomes and optimizing radiation doses in gastrointestinal brain metastases treated with Gamma Knife radiosurgery

Article Title: Therapeutic hybrid intelligence with neural and knowledge-based expert reasoning for SRS (THINKERS): a mixture-of-experts framework for gastrointestinal brain metastases

Article References: Reyes, J. S., Hadjipanayis, C. G., Lunsford, L. D., & Niranjan, A. (2026). Therapeutic hybrid intelligence with neural and knowledge-based expert reasoning for SRS (THINKERS): a mixture-of-experts framework for gastrointestinal brain metastases. Journal of Neuro-Oncology, 179(2), Article 68. https://doi.org/10.1007/s11060-026-05781-3

Image Credits: AI Generated

DOI: 10.1007/s11060-026-05781-3

Keywords: brain metastases, gastrointestinal cancer, Gamma Knife radiosurgery, artificial intelligence, mixture of experts, dose optimization, survival modeling, local control, radiation toxicity, predictive modeling, neuro-oncology, competing risks