Volumetric Modulated Arc Therapy (VMAT) for glioblastoma often generates complex treatment plans, which can pose challenges for accurate delivery. This study aimed to develop and validate a machine learning framework, the Anatomy-based Complexity Engine for Radiotherapy (ACER), to predict VMAT plan complexity from pretreatment anatomical features before the onset of planning. A cohort of 40 VMAT plans for glioblastoma from a public data archive was analyzed. A custom pipeline was developed to extract three anatomical features from the Planning Target Volume (PTV): volume, surface area, and sphericity. A Modulation Complexity Score (MCS) was calculated for each plan and dichotomized at the median to create ‘Low’ and ‘High’ complexity classes. The performance of three classifiers (Logistic Regression, Random Forest, XGBoost) was rigorously compared using a 70/30 train-test split, with the Area Under the Curve (AUC) as the primary evaluation metric. PTV surface area and sphericity were found to be strongly correlated with the plan MCS (r = 0.72 and r = -0.56, respectively). In the comparative analysis, the simple, interpretable Logistic Regression model achieved the highest discriminative performance with an excellent AUC of 0.87, outperforming both the Random Forest (AUC = 0.83) and XGBoost (AUC = 0.73) models. VMAT plan complexity for glioblastoma is accurately predictable from fundamental anatomical features. The proposed ACER framework, utilizing a simple linear model, provides a rapid and transparent automated tool to identify potentially challenging cases, thereby offering valuable decision support for clinical workflow management and resource allocation.