ASCO 2023 RECAP Personalizing Androgen Deprivation Therapy in Patients With High-risk Localized Prostate Cancer Using Artificial Intelligence
By: Andrew J. Armstrong, MD, ScM, FACP, Duke University, Durham, North Carolina | Posted on: 25 Oct 2023
At the 2023 ASCO Annual Meeting, we presented data on the first successfully validated predictive biomarker of long-term androgen deprivation therapy (LT-ADT) benefit with radiation therapy (RT) in men with localized high-risk prostate cancer. We developed and validated this biomarker using an artificial intelligence (AI)-derived digital pathology-based platform across multiple NRG cooperative group trials of localized high and intermediate risk prostate cancer, with external validation in the phase 3 NRG/RTOG 9202 trial. The predictive AI biomarker identified 34% of high-risk men that could derive similar benefit with short-term ADT (ST-ADT), thus avoiding the side effects of prolonged ADT.
Currently, men with high-risk, locally advanced prostate cancer who choose to pursue radiotherapy are also treated with LT-ADT.1,2 Despite the proven clinical benefits of LT-ADT on preventing metastasis and improving overall survival in these patients,3 LT-ADT is associated with increased morbidity due to treatment side effects including muscle and bone loss, potential cognitive impacts, fatigue, cardiovascular risks, and hot flashes.4 ST-ADT may have a lower risk of toxic side effects and reduce non-prostate cancer related mortality.2 There is a clear unmet need for predictive biomarkers to identify men with high-risk localized prostate cancer who have an excellent prognosis and do not benefit from LT-ADT, and can thus be spared the risks of LT-ADT. While existing genomic and clinical risk stratification tools are prognostic, they have not shown predictive utility for ADT duration. ArteraAI, a precision medicine company developing AI tests to personalize cancer therapy, utilizes a multimodal artificial intelligence platform that leverages digital pathology and clinical data such as PSA, stage, age, and Gleason sum to provide AI-driven solutions for prognostic5 and predictive6 biomarkers in localized prostate cancer (Figure 1). The ArteraAI Prostate Test is now supported by National Comprehensive Cancer Network guidelines (1.2023) as a risk stratification tool for localized prostate cancer.
Our team leveraged data from 6 prospective phase 3 randomized trials to develop and validate an AI-derived biomarker that can predict which men with higher-risk localized disease are more or less likely to benefit from longer-term ADT with RT (Figure 2). ArteraAI clinical prediction models are intended to support physician decision making by predicting whether a patient will have an improved outcome in response to treatment and are not intended to replace pathologists to diagnose and risk stratify patients. Generalizability is a crucial aspect when developing and evaluating AI models to ensure applicability across populations. The LT-ADT predictive biomarker was developed using data from prostate biopsies across multiple academic and community sites across North America and African American men composed 21% of the cohort.
Digitized whole slide images of H&E-stained biopsies at time of diagnosis, as well as clinical and outcome data (follow-up >8 years) from 2,641 patients were used for model development to predict the benefit of LT-ADT on distant metastasis (DM). We then validated the ArteraAI LT-ADT predictive model using data from NRG/RTOG 9202,7 a phase 3 clinical trial that randomized men with intermediate to high-risk disease to either ST-ADT (4 months) or LT-ADT (28 months). Of note, explainability of AI models is an ongoing area of research and a topic of much debate as AI advances in health care. As a first step towards understanding what components of our model are driving predictive utility, we evaluated the weighted contribution of image and individual clinical components on model performance and found that image features contributed the most to the ArteraAI LT-ADT predictive biomarker (42.6%). Further investigation will be required to assess the underlying biology driving prediction of ADT benefit.
Predictive utility for the ArteraAI model was evaluated for ADT duration with Fine-Gray or Cox PH interaction models. Event rates were estimated by the cumulative incidence method. Results in the overall validation cohort showed estimated 15-year DM risks for the RT+LT-ADT group vs RT+ST-ADT group were 17% vs 26%, respectively (HR 0.64, 95% CI 0.50-0.82, P < .001), similar to the results of the prior long-term report of the clinical study.7 Among patients identified as biomarker positive, the estimated 15-year DM risks for the RT+LT-ADT group vs RT+ST-ADT group was 19% vs 33%, respectively (HR 0.55, 95% CI 0.41-0.73, P < .001, Figure 3). In contrast, patients identified as biomarker negative did not have a significant treatment benefit, where the estimated 15-year DM risk was 11% for both treatment groups (HR 1.06, 95% CI 0.61-1.84, P = .84, Figure 3). A significant interaction between treatment and predictive model for time to DM was observed with a P value of .04 (Figure 3), meaning the test was not only prognostic but also predictive of LTADT benefits.
These results confirm successful validation of this predictive biomarker for LT-ADT benefit with RT in localized high-risk prostate cancer using an AI-derived digital pathology-based platform in the phase 3 NRG/RTOG 9202 trial. The ArteraAI LT-ADT predictive biomarker showed an absolute difference of 14% in 15-year DM estimated risk between RT+LT-ADT and RT+ST-ADT, in the biomarker positive group, with no significant difference observed between treatment groups in biomarker negative patients and identifies 34% of men who could derive similar benefit with ST-ADT, avoiding the side effects of prolonged ADT.
Further clinical impact of this research comes from the observation that approximately 20% of AI biomarker positive men still suffer from distant metastases at 15 years despite receiving LT-ADT (Figure 3). This suggests that this group of men may benefit from further treatment intensification, such as potent AR inhibitors or taxanes or even PET guided radiotherapy and should be the subject of future clinical trial investigation. There is much promise in the use of AI in prostate cancer, and more and more questions are currently being addressed, including the need for potent AR inhibitors or the need for adjuvant radiotherapy.
Future validation studies in prospective clinical trials are needed for these new questions and for the 40% of intermediate-risk men who tested positive for the AI biomarker, suggesting they may benefit from LT-ADT. Despite their resource-intensive nature and rarity, prioritizing such validation work is important to assess the performance and generalizability of AI models in real-time clinical settings.
- Bolla M, de Reijke TM, Van Tienhoven G, et al. Duration of androgen suppression in the treatment of prostate cancer. N Engl J Med. 2009;360(24):2516-2527.
- Roach M, Bae K, Speight J, et al. Short-term neoadjuvant androgen deprivation therapy and external-beam radiotherapy for locally advanced prostate cancer: long-term results of RTOG 8610. JCO. 2008;26(4):585-591.
- Bolla M, Collette L, Blank L, et al. Long-term results with immediate androgen suppression and external irradiation in patients with locally advanced prostate cancer (an EORTC study): a phase III randomised trial. The Lancet. 2002;360(9327):103-108.
- Keating NL, O’Malley AJ, Smith MR. Diabetes and cardiovascular disease during androgen deprivation therapy for prostate cancer. JCO. 2006;24(27):4448-4456.
- Esteva A, Feng J, van der Wal D, et al. Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials. Npj Digit. Med. 2022;5(1):71.
- Spratt DE, Tang S, Sun Y, et al. Artificial Intelligence Predictive Model for Hormone Therapy Use in Prostate Cancer. Preprint. Res Sq. 2023;rs.3.rs-2790858.
- Lawton CA, Lin X, Hanks GE, et al. Duration of androgen deprivation in locally advanced prostate cancer: long-term update of NRG oncology RTOG 9202. Int. J. Radiat. Oncol. 2017;98(2):296-303.