Introduction:

Artificial intelligence (AI) digital pathology-based biomarkers are gaining adoption in clinical practice for their ability to practically and efficiently personalize patient care. Here we further demonstrate the value of digital pathology through validation of 1) a multimodal artificial intelligence (MMAI) biomarker utilizing digital pathology images and key clinical variables (age, PSA, T-stage) and 2) a newly developed “image-only” AI biomarker utilizing only digital pathology images. We assess both models for prediction of the primary outcome of 10-year risk of distant metastasis (DM) and secondary outcome of prostate cancer-specific mortality (PCSM) in a contemporary cohort of patients with localized prostate cancer treated at three US institutions (University of California, San Diego; University of California, San Francisco; University of Kansas Medical Center).

Methods:

The MMAI (model version v1.2) and image-only (model version v1.3) models were previously trained and locked prior to this multicenter validation study (doi:10.1089/aipo.2024.0004). Retrospective clinical data and digital pathology from first available routine prostate biopsies were compiled from each of the participating sites between 02/2024-09/2024. Eligible patients were diagnosed between 2005-2020 and had evaluable digital histopathology and complete clinical data. The association between each biomarker and DM or PCSM was evaluated continuously (per standard deviation) and categorically [raw continuous scores, ranging from 0-1, were categorized into 3 risk groups (low, intermediate, high) at the time of development and locked prior to validation] using univariable and multivariable Fine-Gray proportional hazard regression models and cumulative incidence analyses.

Results:

Due to different image pre-processing between the MMAI and image-only models, MMAI scores were generated for 991 patients and image-only scores for 886 patients. The MMAI cohort was composed of 35% NCCN low, 41% intermediate, and 23% high-risk patients with median age 65 years, PSA 6.2 ng/mL, and follow-up of 8.2 years. Primary treatment included active surveillance (36%), radiation therapy +/- ADT (23%), and prostatectomy (41%). Most patients were White (76%) or Black (8%), and Grade Group (GG) distribution was 42% GG1, 25% GG2, 15% GG3, 11% GG4 and 7% GG5 (similar to the image-only cohort). Both MMAI and image-only continuous and categorical scores were significantly associated with 10-year risk of DM and PCSM (Table 1) and remained significant when controlling for NCCN risk group or treatment. The models had similar categorical risk distribution (Figure 1), and there were no notable differences in performance by race.

Conclusion:

This study confirms the generalizability and prognostic utility of the MMAI model in contemporary US clinical settings. While the MMAI model has been validated in numerous other studies, this is the first validation of an image-only model. Further research is needed, but the image-only model appears to offer similar prognostic ability, suggesting promise for scalable, image-based risk stratification that does not require clinical inputs. These findings substantiate the presence of valuable prognostic information within histopathology.

Funding: This study was supported by Artera Inc.