External validation of a digital pathology-based multimodal artificial intelligence (MMAI)-derived prognostic biomarker in the randomised phase III CHHiP trial.

Abstract

308 Background: Risk stratification in localised prostate cancer (PCa) based on clinicopathological parameters is inadequate, leading to under- and over-treatment. We used CHHiP trial data to externally validate a previously-developed MMAI prognostic model with potential for cost-effective improved treatment personalisation. Methods: H&E slides from CHHiP translational substudy patients were centrally reviewed by a uro-pathologist between 2013 and 2015, with Gleason grade group (GGG) rescored using contemporary guidelines. GGG was reassigned in 52% of cases. We evaluated the locked ArteraAI prostate MMAI algorithm v1.2 which combines age, T-stage, and PSA with digital prostate biopsy H&E images. Multivariable cox regression analysis of biochemical/clinical recurrence (BCR; trial primary endpoint) and development of distant metastases (DM) were performed with models including age and MMAI as continuous (per 0.1 increase) and categorical (ArteraAI pre-validated 3-tier risk groups) variables. The additional prognostic value of MMAI beyond UK-recommended Cambridge Prognostic Group (CPG) and NCCN risk group models was assessed by change in Concordance Index and Likelihood Ratio Test (LRT). Results: Of 1854 patients with centrally-reviewed pathology, 1797 (97%) had clinical data and H&E with sufficient tumour for MMAI analysis. Median follow up was 14.3 years. Patients were classified as MMAI High (129, 7.2%), Intermediate (885, 49.2%) or Low (783, 43.6%). In univariable analysis, high MMAI score was significantly associated with increased BCR risk, as both a categorical variable (MMAI High risk hazard ratio (HR) = 5.07, 95%CI = 3.77-6.81, p=<0.001, MMAI Intermediate-risk HR = 1.90, 1.52-2.36, p=<0.001), and continuous variable (MMAI raw score HR = 1.55, 1.44-1.67, p=<0.001). MMAI was also significantly associated with DM. Addition of MMAI to CPG and NCCN multivariable models significantly improves discrimination (C-index) and overall fit (change in LRT) for BCR and DM (Table 1). Conclusions: ArteraAI MMAI improves prediction of BCR and DM in CHHiP over standard criteria. This first large-scale external validation in a contemporary UK cohort of localised PCa with rigorously standardized care will inform future prospective trials of MMAI biomarker-guided PCa treatment selection. Endpoint Model C index (MMAI risk group) ΔLRT χ² (p) C index (MMAI raw score) ΔLRT χ² (p) BCR CPG + ageCPG + age + MMAI 0.620.65+0.03 p=<0.001 55.6 p=<0.001 0.620.67+0.05 p=<0.001 67.5 p=<0.001 BCR NCCN + ageNCCN + age + MMAI 0.590.64+0.05 p=<0.001 67.2 p=<0.001 0.590.66+0.07 p=<0.001 80.9 p=<0.001 DM CPG + ageCPG + age + MMAI 0.650.71+0.06 p=0.001 35.3 p=<0.001 0.650.73+0.08p=<0.001 47.0 p=<0.001 DM NCCN + ageNCCN + age + MMAI 0.610.70+0.09 p=<0.001 43.4 p=<0.001 0.610.72+0.11 p=<0.001 55.8 p=<0.001 Total, n=1794. Events BCR, n = 426 (23.7%), DM, n = 121 (6.7%).