Artificial Intelligence in Medicine: Can AI Predict Prostate Cancer Outcomes?

If there were an accurate crystal ball, would you want to know your future? And if you knew something bad was going to happen, would you try to take steps to avoid it? The ancient Greek notion that you can’t escape your fate is illustrated by the tragic tale of King Oedipus, who was informed by the Oracle that he would murder his father and kill his mother. In spite of everything he did to try to avoid it, he actually (and inadvertently) created the fulfillment of the prophecy.

In today’s medicine, there are no crystal balls. Yet, once a doctor establishes a diagnosis—that is, identifies a disease based on a number of factors—he or she also determines a prognosis, which means the likely outcome of the disease along with the chances of surviving it. As you can see, a diagnosis is based on tangible, physical, objective indications; whereas a prognosis is a matter of probability. To call it an educated guess does not do justice to the years of education, experience, and necessary research a doctor brings to bear in the effort to predict the future as accurately as possible.

This is as true for prostate cancer (PCa) as for any other illness. The more precise the diagnosis, the more accurate the prognosis. Today, we have access than ever before to the clinical factors (age, PSA, MRI PI-RADS, Gleason grade group, genomic biomarkers, etc.) that are the diagnostic formula. We also have access to basic statistical algorithms such as the Partin tables and online calculators to generate a prognosis. For example, a patient diagnosed with early stage, low-risk PCa has a prognosis of nearly 100% survival at 10 years regardless of the treatment he chooses. However, a man with Gleason grade group 4 and PI-RADS 4 has a prognosis that may include high risk for recurrence, regardless of the treatment he chooses. And yet, prognosis influences treatment planning as much as diagnosis does. It’s thus essential that the PCa prognosis be as accurate as possible.

Prognosis using Artificial Intelligence (AI)

There’s another thing to consider: the PCa patient’s viewpoint. Receiving a prognosis can be very emotional. There’s always an element of anxiety before the doctor delivers the news. If the prognosis is good, there’s a wave of relief as worry is lifted. The grateful patient may be willing to face even a painful treatment with slow recovery, buoyed by the knowledge that all will be well when it’s complete. However, if the prognosis is discouraging, the patient faces a psychological process similar to the classic stages of loss: disbelief, anger, bargaining, depression, and acceptance—all the while undergoing one or more treatments in hopes of extending life and quality of life. Therefore, the accuracy of the prognosis, and how early it’s revealed to the patient, make a huge difference in the patient’s PCa journey.

But, as the saying goes, doctors are only human. We invest all we can in predicting the course and outcome of the patient’s cancer, but sometimes there is borderline evidence making a prognosis almost too close to call. For example, molecular biomarkers such as circulating tumor cells or genomics are increasingly being studied, but we don’t yet have a body of Level One research to assure confidence in, say, a patient’s likely survival odds if he carries the breast cancer BRCA gene mutation.

At the 2022 ASCO Genitourinary Cancer Symposium, a UCSF research team presented their new prognostic AI biomarker using multi-modal deep learning with digital histopathology in localized PCa. Their model was trained on pre-treatment biopsy slides from five Phase III randomized PCa radiation clinical trials. In addition to images of the pathology slides, the clinical factors were also available for 5,654 patients enrolled in the trials.

The data they collected was used to train and validate a multi-modal AI architecture. The model was structured to predict the following endpoints:

• 5-year probability of biochemical recurrence (BCR)
• 5-year probability of distant metastasis (DM)
• 10-year prostate cancer-specific survival (PCaSS)
• 10-year overall survival (OS).

The prognostic performance of their AI algorithm was compared with that of the standard National Comprehensive Cancer Network (NCCN) risk-factor model, which includes PSA, T-stage and Gleason score for determining prognosis. The research team found their AI algorithm had “superior performance compared to NCCN risk groups for all clinical endpoints.”^[i] Having demonstrated that AI has better prognostic ability, they noted, “This massively scalable technology is feasible and can help personalize the management of prostate cancer patients.”^[ii]

We congratulate the UCSF team on their contribution to Artificial Intelligence in the landscape of PCa prognosis, and to improving outcomes for PCa patients everywhere.

NOTE: This content is solely for purposes of information and does not substitute for diagnostic or medical advice. Talk to your doctor if you are experiencing pelvic pain, or have any other health concerns or questions of a personal medical nature.

[i] Esteva A, Feng J, Huang S, Van der Wal D et al. Development and validation of a prognostic AI biomarker using
multi-modal deep learning with digital histopathology in localized prostate cancer on NRG Oncology phase III clinical trials. J Clin Oncol 40, 2022 (suppl 6; abstr 222)
[ii] Ibid.