Room: Track 2
Purpose: To investigate the impact of external beam radiation therapy (EBRT) treatment planning, treatment delivery, and patient-specific clinical and anatomical factors on biochemical failure-free survival (bFFS) for high-risk prostate cancer patients using conventional and machine learning-based methods.
Methods: Prostate cancer patients with prostate-specific antigen (PSA) at diagnosis =20ng/mL, or Gleason score =8, or T stage =T3a and received curative EBRT between 2010 and 2015 at four institutions were retrospectively included. We analyzed up to five years of post-EBRT PSA follow-up data for each patient. Biochemical failure was defined by the ASTRO-Phoenix definition. Clinically relevant factors for treatment planning, delivery, and patient-specific clinical and anatomical features were identified. Multivariable analysis was performed using the Cox proportional hazard model to identify statistically significant (p=0.05) factors for bFFS. Additionally, random survival forests were performed with 1000 bootstrap replicates to determine variable importance in predicting bFFS.
Results: 777 patients were included in this study (median follow-up 4.35 years). The four-year bFFS for this patient cohort was 85.0% (CI: 82.1%-88.0%). On multivariable analysis, only Gleason score=9 (p<0.001), PTV conformity index defined as V95%/PTV (p=0.005), and PSA at diagnosis (p=0.03) had a statistically significant impact on bFFS. In random survival forests, the three most predictive factors were Gleason score, PSA at diagnosis, and PTV conformity index. T stage, equivalent PTV margin derived from CTV and PTV volumes, PTV D99%, treatment modality, verification imaging modality and frequency, dose calculation algorithm, hormone use, and volumes of the rectum, bladder, CTV, and PTV were not considered predictive by both methods.
Conclusion: This work identified PSA at diagnosis, Gleason score, and PTV conformity index as factors with a significant impact on prostate EBRT outcome using both conventional analysis and random survival forests. The impact of clinical factors will be further explored using machine learning methods with a larger patient cohort.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by the Alberta Innovates - Health. There are no relevant financial disclosures or conflicts of interest to declare.