Using Machine Learning to Assist Radiation Oncology Treatment Ordering for Cancer Patients
Washington University School of Medicine and WWT collaborated to leverage machine learning (ML) to develop methods to predict radiation treatment plans for cancer patients.
Radiation therapy: A crucial element of cancer treatment
Radiation therapy is indicated for more than half of cancer patients. At Washington University School of Medicine (WUSM) in St. Louis, radiation oncologists in the Department of Radiation Oncology assess patients and conduct diagnostic workups in coordination with other oncology specialties (e.g., through a multi-disciplinary tumor board). Where appropriate, they will prescribe treatment with an order including the type of radiation, dose and number of sessions.
While national and international guidelines describe general aspects of the radiation oncology treatment prescription, the details are dictated by individual patient factors and institutional factors, such as physician training and experience, or available equipment. Additionally, as radiation oncology treatments become more complex, there are increased safety risks associated with incorrect ordering, miscommunication and delays in treatment.
Research indicates that artificial intelligence (AI) has the potential to alleviate these pain points by automating aspects of the treatment order process.
ML supporting MDs
Using its rich diagnostic, clinical and treatment data, WUSM worked with WWT to develop machine learning (ML) algorithms to predict treatment orders for head and neck, lung, and prostate cancers.
The first step of development required loading years of prior radiation order data onto hardware built to handle ML and AI workloads; WWT selected Intel Xeon® Scalable Processors for the task. With a robust dataset and sufficient computing power, WWT and WUSM analyzed the variation in treatments and experimented with ML techniques to predict certain components of radiation oncology treatment orders.
Ultimately, WWT took a two-stage development approach. A primary model was developed to specifically predict radiation prescription dose while a second model predicted a larger set of treatment options (e.g., immobilization devices, imaging techniques, patient positioning instructions). The model’s predictions were integrated into a proof of concept application that oncologists could use as a decision support tool.
Traditionally, clinical decision support tools have been developed as rule-based algorithms. This approach requires developers to have an expert-level understanding of every treatment option and spend enormous amounts of time hard-coding each case. Such tools can take years develop and are difficult to update as medical processes and treatments evolve over time.
The next generation of decision support technology will likely be ML-based. ML models can capture patterns from data that may not be obvious to human analysts, allowing for faster development and a broader scope than rule-based programs.
Predicting a radiation treatment
Our clinical support application generates core elements of a radiation oncology treatment order using a limited set of clinical diagnostic information provided by the user.
As seen in the interface below on the left side, a radiation oncologist specifies the clinical factors of the patient. In this example, we have a patient with stage III cancer of the head and neck who was not a candidate for surgery. On the right side, we see the ML prediction: treatment of the target volumes to doses of 7000 and 5600 cGy concurrently, administered in increments of 200 and 160 cGy. A second ML model fills out predictions for additional components of the radiation oncology order, such as the patient position during the therapy sessions and the type and frequency of the imaging requested.
Improving time to treatment
The time from patient admission to treatment is a powerful factor in patient outcomes. Each week that patient treatment is delayed diminishes the likelihood of a positive outcome. WWT's work has the potential to help generate treatment orders faster and reduce the time required by physicians to manually input every aspect of a treatment order.
Further, this tool can also be used as a quality assurance method to see if a patient has an unexpected treatment order given a particular indication. Collaborations like this demonstrate how data experts and healthcare experts can work together to bring about new data-driven technology to serve physicians and their patients.