Utilization of Machine Learning and Proton Collaborative Group Data to Develop a Model for Predictive Prostate Cancer Proton Radiation Therapy Outcomes

Abstract

In the current era of personalized medicine, there remains a need for innovative tools that consider inherent patient radiosensitivity, as well as demographic and clinical level variables such as race, age, and prior treatment for the determination of precise radiotherapeutic treatment of cancer. Demographic and clinical data were obtained from the Proton Collaborative Group Registry (PCG – REG001-09). To achieve the objectives, we measured the resulting PSA levels post proton treatment to determine recurrence vs. positive outcome. We created a new variable “Change in PSA Levels” which allowed us to measure whether or not there was a drastic change in pre- versus post-treatment PSA. To build the Logistic Regression Model, we then created a binary response variable called “Patient Response to Therapy” using the “Change in PSA Level” values. The specified cut-off for “Change in PSA” level was used to categorize “Patients who Responded positively to Therapy” versus “Patients who did not respond positively to Therapy”. Patients were then stratified into quintiles based on the “Change in PSA Levels” and tested in multiple models based on the above stratification. Based on the above stratification, the patient population size that was used for training the model was 569 patients. We used a Machine Learning Pipeline (MLP) that included the below stages to train the model: 1) Missing Value Treatment Stage, 2) Normalizer Stage, 3) Box-Cox Transformer Stage, 4) Feature Assembler Stage, and 5) Ridge Regression Stage. The above MLP was executed using a 2-Fold Cross Validation Method and using Grid Search Technique to tune the Hyper Parameter (λ value) for the Ridge Regression stage of the MLP. Our results demonstrated that prior Dutasteride (-0.142), prior androgen deprivation therapy (-0.835), age at enrollment (0.621), and interruption of treatment (0.273) were significant predictors of patient outcomes (AUC=0.8319). The utilization of machine learning on treatment response data could help us develop tools that could be used by radiation oncologists to accurately predict how a patient may respond to radiation therapy.