Dual AI models for gamma knife radiosurgery in craniopharyngioma: prescription dose modeling and outcome risk prediction.
Gamma Knife radiosurgery (GKRS) is an established adjunct for residual or recurrent craniopharyngioma, yet prescription dose selection remains experience-driven and long-term failure risk is difficult to individualize. Machine learning may support more consistent planning and risk-informed follow-up.
To develop and internally validate two complementary AI models for craniopharyngioma GKRS: (1) prescription dose prediction and (2) treated-lesion progression risk prediction.
In this retrospective single-center cohort, we trained a random forest regressor to predict delivered single-fraction margin dose (Gy) from baseline clinical and tumor features and a random forest classifier to estimate the probability of treated-lesion progression using baseline features and delivered dose. Internal validation used cross-validation with discrimination and calibration metrics.
Seventy-two treated tumors were analyzed. Prescription dose prediction showed clinically tight error, with MAE 1.30 Gy and RMSE 1.65 Gy (R² 0.21), indicating the model approximated physician dosing patterns within ~1-2 Gy for most cases. For outcome modeling, risk prediction achieved ROC-AUC 0.75 and PR-AUC 0.582, with reasonable calibration (Brier score 0.112; recalibration slope 0.86, intercept 0.096). Together, the two models enabled simultaneous estimation of an expected prescription dose and an individualized probability of long-term treated-lesion failure, supporting risk stratification beyond dose alone.
A dual-model AI framework for craniopharyngioma GKRS is feasible and provides both dose estimates and individualized long-horizon failure risk predictions, with potential to standardize prescriptions and tailor surveillance intensity.
Not applicable.
To develop and internally validate two complementary AI models for craniopharyngioma GKRS: (1) prescription dose prediction and (2) treated-lesion progression risk prediction.
In this retrospective single-center cohort, we trained a random forest regressor to predict delivered single-fraction margin dose (Gy) from baseline clinical and tumor features and a random forest classifier to estimate the probability of treated-lesion progression using baseline features and delivered dose. Internal validation used cross-validation with discrimination and calibration metrics.
Seventy-two treated tumors were analyzed. Prescription dose prediction showed clinically tight error, with MAE 1.30 Gy and RMSE 1.65 Gy (R² 0.21), indicating the model approximated physician dosing patterns within ~1-2 Gy for most cases. For outcome modeling, risk prediction achieved ROC-AUC 0.75 and PR-AUC 0.582, with reasonable calibration (Brier score 0.112; recalibration slope 0.86, intercept 0.096). Together, the two models enabled simultaneous estimation of an expected prescription dose and an individualized probability of long-term treated-lesion failure, supporting risk stratification beyond dose alone.
A dual-model AI framework for craniopharyngioma GKRS is feasible and provides both dose estimates and individualized long-horizon failure risk predictions, with potential to standardize prescriptions and tailor surveillance intensity.
Not applicable.
Authors
Reyes Reyes, Bouras Bouras, Hadjipanayis Hadjipanayis, Lunsford Lunsford, Niranjan Niranjan
View on Pubmed