The purpose of our abstract was to address the question of can we identify predictive biomarkers of hormone sensitivity and duration in high-risk patients with localised prostate cancer who have chosen radiotherapy as their treatment. The current standard of care for men with high-risk disease is 2-3 years of androgen deprivation; we call that long-term ADT. Some men need this therapy to help improve their cure rates and some men may be able to get away with less therapy.
As you know, androgen deprivation therapy comes with a lot of long-term side effects – side effects sexually, side effects on their bones, their muscles, their fatigue, their stamina and overall wellbeing. So we sought to develop a pathologic AI biomarker to identify men who may be able to have excellent long-term outcomes with short-term ADT, meaning about six months.
In order to address this question first we had to build the AI biomarker. So, to build this AI biomarker we had to digitise prostate pathology biopsies. We ran these into an artificial intelligence neural network tool that was able to learn through machine learning the important pathologic elements that are driving the differences in outcomes long term, using an endpoint of distant metastasis. Distant metastasis is a key endpoint for men with prostate cancer who are seeking to avoid death or distant metastasis with its treatment burdens. So, we built this AI tool on that endpoint using six phase III trials that were conducted through the NRG National Co-operative Group.
We then externally validated it in the 9202 RTOG phase III study which changed our practice in the United States to favour long-term androgen deprivation therapy over short-term androgen deprivation therapy. We picked this study because it was a positive study and would be a great suitable prospective study to validate our AI biomarker.
We built the biomarker using what we call a multimodal approach. So it incorporates digital pathology, and that’s the most important element of the biomarker, but it also incorporates clinical variables like Gleason score, T stage and PSA and age. So this multimodal AI tool was then locked in as a biomarker and then validated for the predictive accuracy of the need for long-term ADT in preventing distant metastases.
That’s exactly what we did. In the phase III study we externally validated that patients who test positive through this AI biomarker had significant benefits of long-term ADT. These men needed ADT to reduce their risk of death with distant metastases and distant metastases. Conversely, about a third of the patients who were high risk tested negative for the biomarker. These patients did excellent with short-term ADT and had no benefits from long-term ADT. The impact of this is that we could demonstrate a positive interaction, meaning that it was a predictive tool, the p-value was 0.04, suggesting that if a third of men test negative they could be spared the long-term consequences of ADT.
Conversely, we found that intermediate risk patients who would otherwise normally just get short-term ADT, some of these men tested positive with the AI biomarker and a treatment intensification approach may save these men’s lives long-term and reduce their risk of distant metastases. That was about 40% of the intermediate risk patients.
How could this research impact the future treatment of localised high-risk prostate cancer?
In clinical practice, a digital tool could be used to read in prostate biopsies. This could be uploaded for the biomarker assessment, which can be done very rapidly without expending any tissue and the results rapidly returned to the patient who can then make an informed decision about whether they would like to pursue short-term ADT or whether they would need long-term ADT.
There are many ongoing research studies to prospectively validate this but also to identify novel AI biomarkers that may identify men who might need even more intensive therapy, like a potent AR inhibitor in this context.
The final point is that we compared the AI tool with the clinical tools that we currently use. The clinical tools, while they were prognostic, lacked predictive accuracy, meaning that the AI actually improved upon our clinical performance and provided something useful that the current human pathologists cannot detect.