Study of 139 tremor patients suggests computer model can forecast radiosurgery results

Improvement after Gamma Knife thalamotomy typically takes several months to appear, so patients often wait a long time to learn whether the treatment worked. The authors say a day-of-treatment forecast could support more personalized counseling and outcome forecasting in functional radiosurgery. Repeated 5-fold cross-validation gave a lower mean R² of 0.72, so the hold-out figure may be somewhat optimistic.

Predicting when benefit would begin worked much less well. The exploratory time-to-benefit model had an R² of 0.31 and an error of 1.78 months.

The team used a random forest model trained on routinely collected data: demographics, clinical history, baseline Fahn-Tolosa-Marín tremor scores and radiosurgical planning variables. Validation combined a patient-level 80/20 hold-out test with repeated 5-fold cross-validation, and performance was measured by R², mean absolute error and root mean squared error.

Cross-validation gave a mean R² of 0.72 (95% CI, 0.64-0.79), a mean MAE of 15.8 percentage points (95% CI, 14.6-17.1) and a mean RMSE of 20.4 percentage points (95% CI, 18.7-22.2). That means predictions typically missed the true percentage improvement by about 16 points.

A separate exploratory model for time to clinical benefit was weaker, with an R² of 0.31, an MAE of 1.78 months and an RMSE of 2.34 months. The authors conclude that predicting tremor improvement on the day of treatment is feasible.