Multimodal AI Biomarker Improves Recurrence/Distant Metastasis Prediction in Prostate Cancer

A digital pathology-based multimodal artificial intelligence (MMAI) biomarker that can facilitate prediction of recurrence and distant metastasis in patients with prostate cancer has been externally validated in a randomized phase III trial.

“We’ve provided external validation that MMAI significantly improves prediction of recurrence and distant metastases over standard tools, and we can use this to better allocate treatment,” said Anna Clare Wilkins, PhD, MBBChir, MBBS, MRCP, Group Leader at the Institute of Cancer Research and Honorary Consultant in Clinical Oncology at the Royal Marsden NHS Foundation Trust, when presenting the findings of the validation at the 2026 ASCO Genitourinary Cancers Symposium (Abstract 308). The study was presented in collaboration with Artera AI.

Background and Model Methods

The global incidence of prostate cancer is projected to double by 2040, highlighting the need for improved treatment strategies for affected patients. To address this growing need, Artera AI developed a prognostic MMAI test as an algorithm to guide treatment selection.

The MMAI test was trained, tested, and validated across several RTOG studies. External validation was done in an international study, the phase III CHHiP trial of conventional or hypofractionated high-dose intensity-modulated radiotherapy for prostate cancer. Recruitment took place between 2002 and 2011, and all study participants received intensity-modulated radiotherapy and consistent androgen-deprivation therapy for 4 to 6 months. Additionally, all biopsies collected for the study underwent central pathology review, including a Gleason rescore by a specialist uropathologist.

Test results from 1,797 of the 1,854 patients included in the study were assessed for the MMAI validation, with a test failure rate of less than 0.5%.

The MMAI test was applied to both National Comprehensive Cancer Network (NCCN) and Cambridge Prognostic Group (CPG) risk stratification systems, which are both commonly used in the United Kingdom to assess risk in patients with localized prostate cancer.

Findings

MMAI risk groups varied with the existing NCCN and CPG risk stratification categories. “Each MMAI risk group is present across all of the different NCCN and CPG systems,” Dr. Wilkins said. “So this is indicating that within each of the risk groups there is biological variation detected by the MMAI test.”

For example, among patients classified as high-risk by the NCCN, only 17.6% were considered high-risk by the MMAI test, while 71.8% and 10.5% were classified as intermediate- and low-risk, respectively. Similarly, among patients in CPG 5, only 29% were classified as high-risk by the MMAI test, whereas 58% and 13% were classified as intermediate- and low-risk, respectively.

Biochemical/clinical recurrence-free survival rates showed a clear separation among the three MMAI risk groups, with the high-risk group experiencing the earliest and highest rate of recurrence. The 10-year biochemical recurrence–free survival rate was 85.4% for the MMAI low-risk group, 73.4% for the intermediate-risk group, and 45.3% for the high-risk group (log-rank P < .0001).

Comparatively, the 10-year biochemical recurrence–free survival rates were 91.6%, 77.7%, and 64.3% for the NCCN low-, intermediate-, and high-risk groups, respectively.

A similar pattern was observed for distant metastasis–free survival across risk groups, with the MMAI high-risk group showing the greatest risk of distant metastasis. The 10-year distant metastasis–free survival rate was 73.1% for the high-risk group, compared with 96.4% and 92.5% for the low- and intermediate-risk groups, respectively (log-rank P < .0001).

As a measure of the model’s discriminatory ability, concordance (C-index) was also assessed. The addition of the MMAI test to the standard risk stratification system significantly improved the C-index for both the CPG and NCCN systems in predicting recurrence (P < .001) and distant metastasis (P < .001). Notably, for distant metastasis prediction using NCCN risk stratification, the C-index increased significantly from 0.61 to 0.72 with the addition of MMAI.

The researchers also looked at the change in likelihood ratio as a second performance measure, noting that the ratio increased with the addition of the MMAI test for both the CPG and NCCN systems (P < .001 for both). “What this is saying is that with MMAI added in, then the model fits the data much better,” Dr. Wilkins explained.

Image-based features accounted for most of the prognostic performance of MMAI (86.3%), capturing established pathologies such as Gleason scores as well as broader aspects of prostate cancer biology, according to Dr. Wilkins.

She added that improvements in prediction were observed across a range of performance metrics, consistent with the European Society for Medical Oncology’s requirements for AI-based biomarkers in oncology.

DISCLOSURE: Dr. Wilkins reported receiving honoraria from Johnson & Johnson and research funding from Artera AI, AstraZeneca, Veracyte, and Genentech/Roche.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.