Machine Learning-based beam delivery time model for Mevion 250i with Hyperscan technology

Purpose: Accurate prediction of beam delivery time (BDT) is essential for operational efficiency, 4D dose calculations, and advanced proton therapy techniques. Despite its importance, no machine-specific BDT model exists for Mevion systems. Methods: We developed the first machine learning-based BDT model for the Mevion S250i Hyperscan system. Institutional machine log files from 11 patients (1120 files) were used to extract features including spot position, energy layer changes, Adaptive Aperture (AA) movements, and spot charge. Inter-pulse time ($Δ$T) was the target variable. A Random Forest model was trained with cross-validation and tested on held-out data. SHAP (Shapley Additive Explanations) analysis was used to quantify feature contributions. Results: The model achieved mean absolute errors (MAE) ranging from 0.9 ms for short intervals (<50 ms) to 222 ms for long delays (>1000 ms). AA movements were the dominant global predictor for $Δ$T > 50 ms, while spot positions and pulse charge influenced short intervals. Energy changes had minor global impact but locally contributed to large $Δ$T values, consistent with range modulator physics. The model was tested in two clinical applications: volumetric repainting and 4D dose recalculation for interplay evaluation. Predicted cumulative delivery times deviated by only -1.7% from machine log data, and dosimetric metrics (D98, D95, V95) remained within intrinsic delivery variability. Conclusions: This study presents the first machine-specific BDT model for the Mevion S250i, accurately capturing temporal dynamics and predictive performance. SHAP analysis provided insight into system behavior, highlighting the roles of AA adjustments, energy switching, and spot positioning. The model supports applications in interplay assessment, 4D dose calculation, and delivery time-based plan optimization.

💡 Research Summary

This paper presents the first machine‑specific beam‑delivery‑time (BDT) model for the Mevion S250i Hyperscan proton therapy system, leveraging a Random Forest regression algorithm and explainable AI techniques. Accurate BDT prediction is crucial for operational efficiency, 4‑D dose calculations, and advanced techniques such as volumetric repainting and proton arc therapy, yet no prior model existed for Mevion machines.

Data acquisition and preprocessing
Machine log files from 11 patients treated in 2025 at the Maastro Proton Therapy Center were collected, yielding 1,120 treatment records and 2,827,772 individual pulses. Each pulse contains over 300 parameters (timestamp, target position, charge, Adaptive Aperture (AA) leaf positions, range‑modulator settings, etc.). Features available before delivery—target charge, spot coordinates, AA positions, and beam energy—were transformed into differences between consecutive pulses (ΔAA, ΔS, ΔE, ΔCharge). Log transformations (e.g., log(1+ΔAA)) were applied to reduce skewness. Additional engineered variables included interaction terms (ΔAA × ΔS, |ΔE| × ΔAA, |ΔE| × ΔS), categorical bins for energy and AA magnitude, binary flags indicating major changes, and a composite movement metric (Total_Movement = ΔAA² + ΔS² + |ΔE|²). The target variable ΔT (inter‑pulse time) was also log‑transformed. A stratified train‑test split based on quantiles of the log‑target ensured balanced representation of short and long intervals (70 % training, 30 % testing).

Model development
A Random Forest regressor was implemented using scikit‑learn 1.6.1. Pre‑processing comprised robust scaling (median removal, IQR scaling), ordinal encoding of discretized categories, and retention of binary features. Hyper‑parameters were tuned via a randomized search with 5‑fold cross‑validation.

Performance evaluation
Model accuracy was assessed with Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), both overall and within four ΔT intervals: 0‑50 ms, 50‑500 ms, 500‑1000 ms, and >1000 ms. Results showed:

• 0‑50 ms: MAE ≈ 0.9 ms
• 50‑500 ms: MAE ≈ 30‑70 ms
• 500‑1000 ms: MAE ≈ 120‑150 ms

• 1000 ms: MAE ≈ 222 ms

Overall MAPE was around 5 %, indicating that the model’s predictions lie well within clinically acceptable tolerances.

Explainable AI (SHAP) analysis
TreeSHAP was applied to 100 test samples spanning the full ΔT range. Global explanations revealed that ΔAA (the magnitude of Adaptive Aperture leaf movement) dominates predictions for intervals longer than 50 ms, reflecting the mechanical and electronic settling time required when the AA system repositions. For the shortest intervals, ΔS (spot‑to‑spot distance) and pulse charge were the primary contributors, consistent with the fact that the synchrocyclotron extracts protons in discrete charge‑limited pulses and the time between pulses is largely governed by spatial travel and charge replenishment. Energy changes (ΔE) contributed modestly overall but produced noticeable spikes in ΔT when large range‑modulator plate swaps occurred, aligning with the physics of the RMS system. Binary flags such as Is_Major_AA_Change and Is_Energy_Change highlighted specific events that cause outlier delays.

Clinical applications
Two use‑cases were examined:

1. Volumetric repainting – A lung plan with five repaintings was simulated. The model’s cumulative delivery time differed from the actual log by only –1.7 %, and dosimetric metrics (D98, D95, V95) remained within the intrinsic variability of the machine, demonstrating that the model can replace log files for time‑based plan evaluation.

2. 4‑D interplay evaluation – By coupling the predicted ΔT sequence with 4‑DCT phases and a synthetic breathing trace, the authors performed 4‑D dose recalculations. The resulting dose distributions showed negligible deviation from those generated with true log data, confirming that the model’s temporal fidelity is sufficient for accurate motion‑interplay studies.

Conclusions and future work
The study delivers a robust, machine‑specific BDT model for the Mevion S250i, achieving sub‑millisecond accuracy for short intervals and acceptable errors for longer pauses. SHAP analysis provides physical insight into the dominant role of Adaptive Aperture adjustments, spot geometry, and energy switching in shaping delivery timing. Limitations include the modest patient cohort (11 patients) and the focus on a single institution, which may restrict generalizability across different clinical workflows or hardware configurations. Future directions involve expanding the dataset across multiple centers, integrating real‑time log streaming for online prediction, and comparing the Random Forest approach with deep learning sequence models (e.g., LSTM, Transformer) to capture potential temporal dependencies beyond the current feature set.

Overall, the presented model bridges a critical gap in proton therapy workflow automation, enabling more efficient treatment planning, accurate 4‑D dose calculations, and informed optimization of delivery‑time‑constrained techniques.