A Time-Series Approach for Machine Learning-Based Patient-Specific Quality Assurance of Radiosurgery Plans

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning

Background:

• Stereotactic radiosurgery (SRS) for multiple brain metastases offers conformal dose distributions.
• SRS plans require extreme machine parameter modulation, increasing patient-specific quality assurance (PSQA) failure risk.

Purpose of the Study:

• Develop a machine learning (ML) model to predict the PSQA outcome (gamma passing rate, GPR) for SRS plans.
• Improve the efficiency and reliability of PSQA for SRS treatments.

Main Methods:

• Utilized data from 592 patients treated between 2020-2024.
• Extracted 15 plan complexity metrics as input features.
• Employed stratified and time-series data splitting for training, validation, and testing ML models.

Main Results:

• The ML model achieved a mean absolute error of 2.6% on the test set.
• Demonstrated high sensitivity (93%) and specificity (56%) in predicting GPR.
• Identified an overestimation trend for lower measured GPR values.

Conclusions:

• ML models can be effectively integrated into clinical practice for virtual QA of SRS.
• This approach facilitates more efficient and targeted PSQA strategies.
• Enhances the safety and efficacy of SRS delivery for brain metastases.