Validation of a Digital Pathology–Based Multimodal Artificial Intelligence Biomarker in a Prospective, Real-World Prostate Cancer Cohort Treated with Prostatectomy

Prostate cancer is the most common solid malignancy in men and a leading cause of cancer mortality globally (1). A challenge in the management of localized prostate cancer is its heterogeneity both in outcomes and treatments; localized prostate cancer can range in aggressiveness from a low-risk, indolent disease which can be managed with surveillance, to a higher-risk disease which may be managed with more aggressive definitive therapies such as radical prostatectomy (RP) or radiotherapy with androgen deprivation (2). Currently, treatment decisions are based on several factors including age, performance status, previous abdominal surgery, inflammatory bowel disease, and patient preferences, but there is a need to better predict the risk of recurrent disease to improve selection of patients for surgery or radiotherapy. To minimize the risk of over- and undertreatment of prostate cancer, accurate prognostic tools are needed to risk-stratify patients across treatment options.

Advances in artificial intelligence and digital pathology have enabled the development of novel prognostic tools in prostate cancer that can more accurately risk-stratify men with prostate cancer compared with standard clinical variable–based risk models (3, 4). For example, the ArteraAI Prostate Test is a multimodal artificial intelligence (MMAI) deep learning model that incorporates digital histopathology and clinical data and accurately prognosticates across endpoints such as distant metastases and prostate cancer–specific mortality (PCSM). It is recommended in the National Comprehensive Cancer Network (NCCN) Guidelines as an advanced risk stratification tool for localized prostate cancer using prostate biopsy specimens (2) and can be used in addition to NCCN risk, unique patient characteristics, and patient preference to further inform clinical decision-making.

The MMAI model was initially developed and validated using NRG Oncology/Radiation Therapy Oncology Group (RTOG) phase III randomized controlled trial datasets, consisting primarily of North American men treated with definitive radiotherapy regimens. To further characterize the generalizability of this novel technology, this study aimed to evaluate the prognostic validity of the test for the first time on prostate biopsy samples from a large, real-world, prospectively collected dataset of non–North American men with localized prostate cancer who underwent definitive treatment with RP and had RP-specific endpoints such as biochemical recurrence (BCR) and adverse pathology (AP) at RP.

This study involved retrospective analysis of prospectively collected data from men enrolled in Urology Prostate Cancer, a large, community-based biomarker study of men who were referred for biopsy and diagnosed with prostate cancer between 2004 and 2010 at Skåne University Hospital in Malmö, Sweden (5). Patients gave written informed consent and were consecutively enrolled in the cohort. The study was conducted in accordance with recognized ethical guidelines and Declaration of Helsinki and approved by the Lund University Ethical Committee (2016/1030 and 2018/937). Representative hematoxylin and eosin–stained diagnostic biopsy slides from these men were scanned on an Aperio CS2 slide scanner at 40× magnification in 2021. Men who subsequently underwent RP were included in this analysis. Digitized biopsy images were combined with clinical information to generate MMAI scores from the locked MMAI model.

This analysis used the ArteraAI MMAI Prostate Test (version 1.2; ref. 6), which includes a pre-established, locked algorithm developed using data from North American patients with nonmetastatic prostate cancer and has demonstrated prognostic risk stratification for distant metastases and PCSM. The algorithm utilizes digital histopathology images as well as complete clinical data (age, PSA, and tumor stage) to generate a continuous risk score ranging from 0 to 1. Scores were categorized into low-risk, intermediate-risk, and high-risk groups based on pre-established cut points correlating with a 3% and 10% estimated 10-year risk of distant metastases (7).

The primary objective of this study was to validate the previously locked MMAI model developed on biopsy specimens in an external prospectively collected dataset of European men with prostate cancer who subsequently received RP. The primary endpoint was time to BCR, defined as two successive PSA measurements ≥0.2 ng/mL after RP. The secondary objective was to evaluate the biopsy-based model on the surgical endpoint of AP at RP, defined as Gleason grade group (GG) 3 or higher, pT3b or higher, and/or node-positive disease. Men eligible for upstaging (i.e., diagnosed as Gleason GG 2 or lower, cT3 or lower, and node-negative disease) were evaluable for the AP at the RP endpoint. Death without events was considered as a competing risk for the BCR endpoint.

Descriptive statistics of the validation cohort are reported by medians and IQR or frequencies and proportions, as appropriate. Univariable and multivariable Fine–Gray models and logistic regression models were constructed to evaluate the association of the MMAI score with BCR and AP at RP, respectively. Multivariable analyses (MVA) were adjusted for clinical risk nomograms. The MMAI score was evaluated both continuously and categorically; subdistribution HRs (sHR) and ORs and 95% confidence intervals (CI) were reported per SD increase in score or for intermediate to high versus low risk, as appropriate. Estimated event rates at 5 years were estimated by the cumulative incidence function and compared using the Gray test for the BCR endpoint. The distribution of MMAI risk across NCCN (2) and Cancer of the Prostate Risk Assessment (CAPRA; ref. 8) risk groups were compared using the Fisher exact test. Subgroup analyses were performed in which analyses were repeated within NCCN risk groups (NCCN low to favorable intermediate versus NCCN unfavorable intermediate to high).

Statistical analyses were performed using R version 4.2+ (R Foundation for Statistical Computing). All statistical tests were two-sided and considered statistically significant at the 0.05 level. No adjustment for multiple testing was applied. Patients with missing clinical or outcome information were omitted from any relevant analyses, assuming the data were missing completely at random.

Restrictions apply to the availability of these data, which were used under license for this study. Data are available from the authors upon reasonable request with the permission of the regional health care provider, Region Skåne.

Of the patients enrolled in the Urology Prostate Cancer cohort, 293 were diagnosed with prostate cancer on biopsy, of which 235 subsequently underwent RP. Clinical data were available for 230 patients (Supplementary Table S1), but only 154 had digital histopathology images from the diagnostic prostate biopsy available, of which 143 had evaluable images and complete clinical data to generate MMAI scores (Fig. 1).

Table 1 summarizes the demographic and clinical characteristics of the analysis cohort. The median follow-up of censored patients was 8.8 years (IQR, 6.8–10.2). The median age at biopsy was 64 years (IQR, 59–67), and the median PSA was 7.5 ng/mL (IQR, 5.7–12.9) at diagnosis. The Gleason GG at biopsy ranged from 1 to 5 with the majority of patients having 1 (39.9%) or 2 (30.8%). There was cT3 disease in 7.7% of patients and 28.6% were NCCN high risk. On final pathology, 11.2% had Gleason GG 4 to 5, 46.1% had pT3a-b, 35.7% had a positive surgical margin, 40.8% had extraprostatic extension, 11.9% had seminal vesicle invasion, and 9.6% had lymph node invasion. A subset of 88 patients was eligible for upstaging and evaluable for AP at RP.

On univariable analysis (UVA), the continuous MMAI score was significantly associated with BCR (sHR, 2.45; 95% CI, 1.77–3.38; P < 0.001; Table 2). MMAI categorical risk groups were also significantly associated with BCR (sHR, 5.39; 95% CI, 1.91–15.23; P = 0.002; Table 2). On MVA, the continuous MMAI score remained independently prognostic when adjusted for NCCN (sHR, 2.13; 95% CI, 1.44–3.14; P < 0.001) and CAPRA (sHR, 1.99; 95% CI, 1.29–3.07; P = 0.002; Table 3). Consistent results were observed for categorical MMAI risk groups (sHR, 3.45; 95% CI, 1.20–9.86; P = 0.02 for NCCN and sHR, 3.48; 95% CI, 1.24–9.75; P = 0.02 for CAPRA; Table 3). Estimated 5-year BCR rates for MMAI intermediate to high versus low were 25.2% (95% CI, 16.0%–35.5%) versus 3.6% (95% CI, 0.6%–11%), respectively (Fig. 2; P < 0.001).

Among the subset evaluable for AP at RP, the continuous MMAI score was significantly associated with AP (OR, 4.85; 95% CI, 2.54–10.78; P < 0.001; Table 2) on UVA. MMAI categorical risk groups were also significantly associated with AP (OR, 25.86; 95% CI, 6.64–172.95; P < 0.001; Table 2). On MVA, the continuous MMAI score remained independently prognostic when adjusted for NCCN (OR, 5.07; 95% CI, 2.26–13.68; P < 0.001; Table 3) and CAPRA (OR, 4.59; 95% CI, 2.32–10.51; P < 0.001; Table 3). Consistent results were observed for categorical MMAI risk groups (OR, 41.71; 95% CI, 6.47–850.96; P = 0.001 for NCCN and OR, 23.67; 95% CI, 5.87–162.16; P < 0.001 for CAPRA; Table 3).

Similar results were observed within NCCN subgroup analyses (Supplementary Table S2). The continuous MMAI score remained significantly associated with BCR within NCCN low to favorable intermediate patients (sHR, 2.78; 95% CI, 1.11–7.01; P = 0.03) and within NCCN unfavorable intermediate to high patients (sHR, 1.95; 95% CI, 1.39–2.74; P < 0.001) on UVA. With very few events, the categorical MMAI risk group was underpowered for the NCCN low to favorable intermediate patients but remained significantly associated with BCR within the NCCN unfavorable intermediate to high patients (sHR, 3.55; 95% CI, 1.08–11.70; P = 0.04). Among the subset evaluable for AP at RP, the continuous MMAI score remained significantly associated with AP within NCCN low to favorable intermediate patients (OR, 3.92; 95% CI, 1.76–10.87; P = 0.002) and within NCCN unfavorable intermediate to high patients (OR, 7.03; 95% CI, 2.28–36.29; P = 0.004) on UVA. The categorical MMAI risk group remained significantly associated with AP within the NCCN low to favorable intermediate patients (OR, 12.06; 95% CI, 2.43–91.32; P = 0.005) but did not converge within the NCCN unfavorable intermediate to high patients, as there were no AP events in the MMAI low group.

The distribution of the MMAI risk group within NCCN and CAPRA risk groups is shown in Fig. 3 (see Supplementary Fig. S1). Of the 38 (29%) men with NCCN high-risk disease, 4 (10.5%) and 26 (68.4%) were found to be MMAI low and intermediate risk, respectively. Similarly, 2 (6.5%) and 23 (74.2%) of the 31 (21.7%) men with high CAPRA risk were found to be MMAI low and intermediate risk, respectively.

In this study, we found a previously developed digital histopathology–based MMAI biomarker to be prognostic for the endpoint of BCR in European men with localized prostate cancer treated with RP. Cumulative incidence plots showed that the biomarker performed well in stratifying the cohort among the previously defined three-tier MMAI risk groups. In addition, we found the biomarker to be prognostic for AP at RP among the subgroup that had its full RP pathology details available for analysis and was eligible for upstaging.

The clinical significance of these findings is that they represent the first external validation of the MMAI biomarker run on prostate biopsy samples from a cohort of men who received definitive surgical treatment with RP. This is compared with the initial development and validation efforts for the novel MMAI model which were done on the prostate biopsy samples of over 5,600 men enrolled in the NRG Oncology/RTOG phase III randomized controlled trial datasets treated with definitive radiotherapy regimens. Although the outcomes of men with localized prostate cancer treated with definitive radiotherapy regimens versus surgery are thought to be no different (9), this analysis is an important step in establishing generalizability to a population initially managed with surgery alone. In addition, it is reassuring that the model performed well in (i) an independent prospectively collected population, as opposed to the more curated cohorts typical of randomized controlled trials (10) and (ii) a non–North American population, compared with the NRG/RTOG cohorts which are enrolled from sites in the United States and Canada. These observations further support the external generalizability of the MMAI model.

The MMAI model demonstrated prognostic performance for the RP-specific endpoints of BCR and AP at RP It also risk stratified patients within NCCN and CAPRA risk groups, providing an additional layer of personalized information that has the potential to inform treatment escalation or de-escalation strategies in surgically managed men, such as the timing of the use of neoadjuvant or adjuvant therapies in addition to RP. For example, the phase III CALGB 90203 (Alliance) trial evaluated neoadjuvant chemohormonal therapy for high-risk localized prostate cancer before RP (11). Although chemohormonal therapy did not demonstrate an improvement in the primary endpoint of 3-year biochemical progression–free survival over RP alone, there were distinct genomic and transcriptomic features observed that demonstrated useful prognostic features (12). It is plausible that a prognostic biomarker driven by histopathology features, such as this MMAI model, could identify subgroups more likely to benefit from neoadjuvant or adjuvant therapies. Additionally, given its ability to predict BCR, which is valuable in the management of patients after RP, MMAI risk also has the potential to inform PSA monitoring, with more frequent follow-up in patients who have increased MMAI risk. These are important avenues for additional research.

Of note, these findings are consistent with prior preliminary validation efforts of the MMAI biomarker, such as its association with PCSM and overall survival on RP specimens (as opposed to prostate biopsy specimens examined herein; ref. 13) as well as its association with PCSM on prostate biopsy specimens in men from the United Kingdom with high-risk localized or metastatic prostate cancer (14).

This study has some limitations. It is a retrospective study using prospectively collected data from a single institution from a nation with a less racially and ethnically diverse population relative to the United States. Given the secondary, nonrandomized study design, we are unable to control for all potential known and unknown confounders. In addition, not all diagnosed patients from the primary reference cohort had the necessary clinical data or histopathology to generate MMAI scores. Furthermore, some patients had complete clinical or histopathology data but were cT3 or Gleason GG 4 to 5 at the time of biopsy and therefore could not be evaluated for the secondary endpoint of AP at RP. Patients excluded from the primary analysis cohort seemed to have slightly lower risk than those in the analysis cohort (Supplementary Table S1). Any of these limitations could introduce possible bias.

In conclusion, we found the prostate biopsy MMAI biomarker, previously shown to be prognostic for distant metastasis and PCSM in men receiving definitive radiation, to be prognostic for the post-RP endpoints of BCR and AP. This biomarker validation study using a real-world cohort further supports the use of MMAI biomarkers in men with localized prostate cancer outside the United States and Canada and treated with surgery.

A. Bjartell reports personal fees from Accord, Sandoz, European Association of Urology, and IPSEN; grants and personal fees from Astellas, AstraZeneca, and Bayer; nonfinancial support from ESMO; grants from Curasight, Johnson & Johnson, and Spectracure; and other support from LIDDS Pharma, WntResearch, and Glactone Pharma outside the submitted work. V.Y.T. Liu reports other support from Artera outside the submitted work. M. Tierney reports other support from Artera outside the submitted work. T.J. Royce reports other support from Artera outside the submitted work. M. Sjöström reports grants from the Prostate Cancer Foundation and Swedish Cancer Society during the conduct of the study as well as personal fees from Astellas and Veracyte/Adelphi Targis outside the submitted work. E. Chen reports employment with Artera. A. Esteva reports other support from Artera during the conduct of the study as well as a patent for Multiple pending to Andre Esteva/Artera; in addition, A. Esteva is the founder and CEO of Artera and is employed by Artera. F.Y. Feng reports personal fees from Janssen, Bayer, Roivant Sciences, Astellas Pharma, Foundation Medicine, Varian, Novartis, Tempus, and Amgen and other support from PFS Genomics, Bristol Myers Squibb, SerImmune, and Artera outside the submitted work. No disclosures were reported by the other authors.

A. Bjartell: Conceptualization, resources, data curation, supervision, funding acquisition, methodology, writing–original draft, writing–review and editing. A. Krzyzanowska: Resources, data curation, investigation, writing–review and editing. V.Y.T. Liu: Conceptualization, resources, formal analysis, validation, visualization, methodology, writing–original draft, writing–review and editing. M. Tierney: Conceptualization, methodology, writing–original draft, project administration, writing–review and editing. T.J. Royce: Conceptualization, methodology, writing–original draft, writing–review and editing. M. Sjöström: Conceptualization, methodology, writing–original draft, writing–review and editing. M.M. Palominos-Rivera: Resources, data curation, investigation, writing–review and editing. E. Chen: Conceptualization, writing–review and editing. A. Kraft: Resources, data curation, writing–review and editing. A. Esteva: Writing–review and editing. F.Y. Feng: Writing–review and editing.

This study was funded by grants to A. Bjartell from The Swedish Cancer Society (Cancerfonden #2018/522 and #2021/1629), Swedish Scientific Council (Vetenskapsrådet, #2020-02017 and 2023-02671), The Research Funds at Skåne University Hospital, and The Governmental Funding through The Faculty of Medicine, Lund University (contract number 2018/810 and 2022/0105). The authors would like to thank Kristina Ekström-Holka for help with the scanning of the slides and Anna Holst, Anna Stiehm, Emelie Winell, Suzy Lindberg, Annica Löfgren, and Åsa Andersson for help with patient clinical information collection.