Last updated January 30, 2018 at 9:51 am
CancerSEEK uses machine learning in what could be a breakthrough in early diagnosis of certain cancers.
Breaking down the facts and figures
With the CancerSEEK blood test:
- Patients with cancer are successfully diagnosed in 70% of cases
- Healthy patients are successfully diagnosed in more than 99% of cases; that is, there is less than one per cent occurrence of false positives
This study examine eight cancers; ovary, liver, stomach, pancreas, oesophagus, colorectum, lung, and breast.
- Breast cancer had the lowest successful diagnosis rate at 33%
- For ovary, liver, stomach, pancreas, and oesophageal cancers there are currently no screening tests available for average risk people. The CancerSEEK test successfully diagnosed these cancers at a range from 69% to 98%
- For oesophageal and stomach cancers, the cancer site can be accurately diagnosed as at one of two anatomical structures in 83% of cases
Machine learning to the rescue
Diagnostic tools like this are crucial, as early cancer detection dramatically improves successful treatment rates.
“The majority of localized cancers can be cured by surgery alone, without any systemic therapy. Once distant metastasis has occurred, however, surgical excision is rarely curative. One major goal in cancer research is therefore the detection of cancers before they metastasize to distant sites,” say the authors in the study.
But it’s also important for generalised tests, like blood tests, to have specific results to avoid false positives.
“New blood tests for cancer must have very high specificity; otherwise, too many healthy individuals will receive positive test results, leading to unnecessary follow-up procedures and anxiety,” they say.
So with this in mind, the researchers had several conflicting challenges. They needed to detect a large number of genetic mutations which could be sequenced thousands of times, without detecting too many genes so they wouldn’t get ‘noisy’ results, all the while being a cost-effective procedure.
And that’s where machine learning could come to the rescue. The researchers could use public available data to craft algorithm that would find the goldilocks solution.
They tested 1005 patients who had been clinically diagnosed with the eight cancers being studied, but whose cancer hadn’t yet metastasized (that is, spread) and who hadn’t yet started treatment. As a control, they studied 850 healthy individuals.
Each subject underwent a non-invasive blood test that asses mutations in 16 cancer genes and 10 circulating protein biomarkers.
As this project works towards the goal of earlier diagnosis, future developments are likely to be less sensitive, as earlier stage cancer is simply less detectable. This is where machine learning will be invaluable to the researchers as a method to refine their process. They predict that CancerSEEK will ultimately cost around $500, which is comparable to current, invasive diagnostic procedures.
This research was published in Science.