Introduction:

Accurate risk stratification is critical to guide management decisions in localized prostate cancer (PCa). As previously described, we developed multi-modal artificial intelligence (MMAI) models on digital histopathology and clinical data (age, PSA, clinical T-stage, and Gleason information) and validated in five phase III PCa trials (NRG/RTOG 9202, 9408, 9413, 9910, and 0126) that outperformed standard risk models (NCCN) in prediction of distant metastasis (DM) and prostate cancer-specific mortality (PCSM). Here, we aimed to further validate our MMAI models for their prediction of DM and PCSM in NRG/RTOG 9902, a phase III randomized trial of men with localized or locally advanced high risk prostate cancer.

Methods:

The validation cohort included 318 localized high-risk PCa patients from NRG/RTOG 9902 with histopathology data (84% of all eligible patients). The previously locked models were used to generate MMAI scores and evaluated for their prognostic ability. Individual clinical factors served as comparators. Time-to-event endpoints were summarized using cumulative incidence curves. Univariable Fine-Gray models and area under the time-dependent receiver operating characteristic curves (td-AUCs) were used to assess performance. Hazard ratios (HRs) were reported per standard deviation increase of the score. As there were no significant treatment benefit differences between the two arms of NRG/RTOG 9902 and arms were well-balanced, we considered the entire trial population as one group for outcome analysis.

Results:

The evaluable population included men with median baseline PSA 23.0 ng/mL, 32% of which had cT3-4 disease, and 67% had Grade Group 4 or 5 disease. At the median follow-up of 10.0 years, 42 men had experienced DM and 29 with PCSM. As opposed to clinical and pathological factors, the MMAIs were significantly prognostic across outcome measures (Table 1). The DM MMAI was significantly associated with risk of DM with HR of 2.37, (95% confidence interval [CI] 1.61-3.50, p < 0.001); similarly for PCSM (HR [95%CI] = 2.72 [1.72-4.31], p < 0.001). Using quartile splits on the DM MMAI, the lower 75% of patients had an estimated 5-yr and 10-yr DM rate of 4% and 7%, and the highest quartile had an average 5- and 10-yr DM rate of 19% and 32% (Figure 1). Similar results were observed for the PCSM MMAI.

Conclusion:

MMAI prognostic models outperform clinical and pathological variables for the prediction of DM and PCSM in a population of men at high risk for disease progression. This study provides important evidence for consistent validation of our deep learning MMAI models to improve prognostication and enable more informed decision-making for patient care.

This project was supported by grants U10CA180868, U10CA180822, UG1CA189867, U24CA196067 from the National Cancer Institute.
 

Funding: U10CA180868, U10CA180822, UG1CA189867, U24CA196067 from the National Cancer Institute