AI Tool Enhances Decision-Making in SRS for Brain Metastases
In Partnership With
An AI-driven tool enhances stereotactic radiosurgery for small brain metastases by predicting local failure risks, optimizing radiation dosing, and personalizing MRI follow-ups.
A new machine learning tool is transforming decision-making in stereotactic radiosurgery (SRS) for small brain metastases. The tool aims to evaluate factors such as radiation dose, patient characteristics, and treatment-related factors to determine the likelihood of local failure at 6 months, 1 year, and 2 years post-treatment.1
Brain metastases, especially those that are under 2 cm, present challenges in achieving optimal local control following SRS. Conventional treatment dosing typically consists of 20 Gy, 22 Gy, or 24 Gy, and relies on general guidelines. However, these fail to account for nuanced patient-specific factors.
By integrating AI into the decision-making process, this machine learning model offers clinicians the ability to assess local failure probabilities at 6 months, 1 year, and 2 years post-treatment based on factors such as prescription dose, age, Karnofsky performance score (KPS), and SRS treatment course.
“What we wanted to do is evaluate all of the patient characteristics, other treatment-related factors, and identify what would be the risk of local failure at each dose level, and that is at 6 months, 1 year, and 2 years after treatment,” said Rupesh Kotecha, MD, chief of radiosurgery and director of central nervous system metastasis with Baptist Health Miami Cancer Institute. “This requires evaluation of so much data in addition to our usual methods, from a statistical perspective, to evaluate that risk of local failure.”
According to an abstract presented at the 2024 American Society for Radiation Oncology (ASTRO) meeting, the study analyzed a sizable dataset from 235 patients treated at Miami Cancer Institute between 2017 and 2022, which included 1,503 brain metastasis cases across 358 SRS courses. A rigorous propensity score matching analysis was performed to adjust for confounding variables.
The study cohort had a median age of 65 years (interquartile range [IQR], 55-73), with 61% of the population being female. The median KPS was 90 (IQR, 80-90), and the median number of lesions treated per SRS course was 4 (IQR, 2-7). Further, lung cancer (58.5%)was the most common primary tumor, followed by breast cancer (24.6%), and prescription doses were distributed as 20 Gy for 297 lesions (20%), 22 Gy for 442 lesions (29%), and 24 Gy for 764 lesions (51%).
“We used machine learning algorithms to help us to determine what are those factors that are associated with local failure and how we could potentially predict a patient's risk of local failure after their treatment with radiosurgery,” he added.
With a median follow-up of 12 months (IQR, 4-23), local failure was observed in 138 lesions (9.2%) across 47 patients. Using propensity score matching, 276 lesions from 123 patients were included in the generalized estimating equations model. The machine learning model performed well in terms of accuracy, with an 88% accuracy rate and 91% specificity, particularly in the 1-year model. Additionally, an 0.8 area under the curve was observed.
The AI tool also extends its utility beyond dose optimization, offering practical applications in patient follow-up care. Based on individualized failure risks, clinicians can adjust the frequency of follow-up MRI scans, which reduces imaging for low-risk patients and intensifies monitoring for those at higher risk.
“In this study, we were able to develop an initial machine learning model that can predict local failure as a function of dose. [T]his is useful in 2 ways, directly for clinical implementation,” said Kotecha.
In the future, Kotecha explained that the model’s predictive power may expand with larger, diverse datasets from additional institutions.
“We have a very diverse patient population at Miami Cancer Institute. That helps with generation of models with regards to internal validity, and I think for external validity as well. But as we add additional patient populations or data sets from other institutions, it will help us to identify [if there] are limitations to our particular model when it is applied at different institutions.”
