AI predicts prostate cancer severity
There’s an old saying in medicine that, if you live to 80 then 80% of you will get prostate cancer, assuming you’re a man of course. But that doesn’t mean that 80% of men will die of prostate cancer. Some will, but the majority will die of something else completely different. In other words, they don’t need aggressive - potentially quality-of-life-changing prostate cancer treatment. And the problem is that when we find prostate cancer in a person, at the moment we don’t know who does and who doesn’t need that intervention. But that might be about to change. Chris Smith spoke to Graham Pockley and Georgina Cosma, who have used a machine learning algorithm to identify a set of markers present on natural killer cells present in the blood that can much more accurately detect who actually has prostate cancer, and the likely aggressiveness of their disease…
Graham - Prostate cancer affects around 45,000 men each year in the UK. And around 11 or 12,000 men die of the disease. One of the big problems in trying to manage patients with prostate cancer is having a very clear understanding of whether they have the disease in the first place, because the current tests don't work very well. And if they do have the disease, how sort of serious is that disease? Do they need treatment or can they just be left and watched? So the key problem we were trying to solve is to improve the diagnosis of prostate cancer using a simple sort of blood test, rather than having to take tissue out of patients and looking at it under the microscope.
Chris - We do have some blood tests you can use to diagnose prostate cancer or at least monitor it. So what's wrong with that?
Graham - Yes. The most common test is something called prostate specific antigen - PSA test. So this is a blood test that measures the levels of a certain molecule in the blood. The problem with the prostate specific, the PSA test, is the fact that in some cases, the level is normal in men who have prostate cancer. In other cases, it can be raised in men who don't have prostate cancer. So there's a lot of false positives and false negatives, and that causes a problem because it doesn't allow that particular test to be used for screening a large number of men because of the false results it gives.
Chris - And what have you done instead to try to work out a way around that problem?
Graham - The key thing is we still wanted to go with a blood test because that's obviously the easiest thing to do. So we decided to focus on the patient's immune system. As well as protecting people against infection the immune system protects people against cancer and the immune system can recognise the presence of cancer. So we thought that if we took a picture of the immune system in the blood, can that be used to tell whether a man has got prostate cancer? And if he does have prostate cancer, how serious it is.
Chris - Georgina, how did you actually go about doing this?
Georgina - So for this study, we collected and examined the natural killer cells of 72 patients. Amongst these patients, 31 of them had prostate cancer and 41 were healthy patients. So these biological data were then used to produce computer models that can detect the presence of prostate cancer and its severity.
Graham - For the test what we did is we took blood samples and we took immune cells from those patients. And we looked at how they appeared by putting them through an analysis. And so we could identify a whole range of different sorts of characteristics or features of these white blood cells, these natural killer cells. And then we wanted to assess whether there are any differences in the profile of those features or those characteristics between men who did have prostate cancer and men who didn't.
Chris - And that's where you come in Georgina.
Georgina - Yeah. So we initially had a set of 32 biomarkers and the aim was to find the subset of biomarkers that can be used to predict presence and severity of prostate cancer. We ended using a technique, which is a computation optimisation approach, to identify the set of biomarkers that would make good fingerprints for predicting prostate cancer. The fingerprints that we found was then used to build a machine learning model.
Graham - In essence then you know that you've got these features that you can look for that are there, but you don't know which combination are going to be the strongest and most powerful predictors of who's really got disease or not. So having identified the ones you think are going to be the right choices, that's where you then move into building a model so that you can then take this forward and test it against more data effectively to see how good it is, presumably
Georgina - Yes, precisely. So after selecting the best candidate set of biomarkers, we developed machine learning based tools. This resulted in a prediction model that was 12.5% more accurate than the PSA test that's used in clinical practice in detecting prostate cancer. We also developed a second detection tool, which was 99% accurate in predicting the risk of disease in patients with prostate cancer.
Chris - So Graham, basically, you now have a test which is potentially a lot more accurate than just PSA in isolation.
Graham - Yes, that's right. I think generally speaking, you know, the PSA test is only accurate in about 3 out of 10 cases. And so there's lots of men who have a PSA test who are told that they have prostate cancer when they don't. The other issue is that in the vast majority of cases, men die with prostate cancer rather than of it. So it's not in the majority of cases, a life limiting disease, but people are then sort of labelled as having prostate cancer and have to live with that for the rest of their lives. What this test allows us to do is sort of identify the presence of prostate cancer in the first case, but more arguably more importantly is we can categorise that, or we can say whether that disease is something to worry about or something not to worry about. And that allows the clinicians to focus on treating the more aggressive disease.
Chris - And Georgina, presumably you have road tested this by taking people with an unknown diagnosis and then got the predictions from your model and then followed up with a gold standard, perhaps a biopsy or something from those people, to know that you're actually producing accurate predictions.
Georgina - Yes. We also carried out separate tests using a set of records which were not used during the train and test process. So it's like a mini clinical trial basically where we took the finished machine learning model and we input about, say 10 test records, and then looked at the outputs of the model and compared that to their gold standard, which was the biopsy results that we had.
Chris - And critically people are often in a quandary when they're given a diagnosis of prostate cancer as to whether or not they should seek active intervention or a more watch and wait type strategy. Does it help them to make that decision?
Graham - Yeah, absolutely. I think that's the fundamental finding of the project. It allows that decision to be made far more accurately. The approach would reduce the need for about 70% of biopsies, which is quite an unpleasant experience. 1 in 20 men who undergo that procedure get some form of infection. So what it would really allow the clinical team to do is to tell a man, you have a very low grade prostate cancer, but it's absolutely nothing to worry about. There's no need to treat it. Or say you have prostate cancer, it's something we're worried about, and we do need to take it further and then further investigations. But it would remove unnecessary investigations in the large proportion of men. And at the moment that's just not possible.