# Machine Learning-Based Beam Delivery Time Model for Mevion S250i With Hyperscan Technology

## TL;DR

This paper introduces a machine learning model to predict beam delivery times for a specific proton therapy system, improving efficiency and accuracy in treatment planning.

## Contribution

The first machine learning-based beam delivery time model for the Mevion S250i Hyperscan system is developed and validated.

## Key findings

- The model achieved mean absolute errors ranging from 0.9 ms for short intervals to 211 ms for long delays.
- Adaptive Aperture movements were the dominant predictor for intervals over 50 ms.
- The model showed strong performance in clinical applications like 4D dose recalculation and volumetric repainting.

## Abstract

Accurate prediction of beam delivery time (BDT) is critical for operational efficiency, 4D dose calculations, and advanced proton therapy techniques. Despite its importance, no machine-specific BDT model exists for Mevion systems.

We developed the first machine learning–based model for the Mevion S250i Hyperscan system. Institutional machine log files from 11 patients (1120 machine log files) were used to extract features describing spot position, energy layer changes, Adaptive Aperture (AA) movements, and spot charge. Inter-pulse time (ΔT) served as the target variable. A Random Forest model was trained with cross-validation and tested on held-out data. SHAP (SHapley Additive exPlanations) analysis quantified feature contributions.

The model achieved mean absolute errors (MAE) ranging from 0.9 ms for short intervals (<50 ms) to 211 ms for long delays (>1000 ms). AA movements were the dominant global predictor for ΔT>50 ms, whereas spot positions and pulse charge dominated short intervals. Energy changes had a minor global influence but locally contributed to a large ΔT, consistent with range modulator physics. The model was tested on two clinically relevant applications: volumetric repainting beam sequence and 4D dose recalculation for interplay evaluation. The predicted cumulative delivery times deviated by only −1.6% from machine log files, and dosimetric metrics (D98, D95, and V95) remained within intrinsic delivery variability.

This study presents the first machine learning–based BDT model for the Mevion S250i system, accurately capturing both predictive performance and machine-specific temporal dynamics. Explainable AI analysis using SHAP provided detailed insights into the operational characteristics of the system, highlighting the contributions of energy layer switching, AA adjustments, and spot position shifts to delivery time. The proposed BDT model demonstrated strong predictive performance across the clinical applications evaluated, supporting its potential use for interplay assessment, 4D dose calculation, and delivery time–based plan optimization.