A [Better] Machine Learning Approach to Diagnosing Colon Cancer

Going head-to-head with Cologuard

A few weeks ago, I wrote this article regarding the applications of machine learning towards early diagnoses of colorectal cancer (CRC).

Since then, we’ve come a long way.

Once again, I’d like to talk about the problem we’re trying to solve.

The Problem

I’m gonna let the Cleveland Clinic explain this for me:

Colorectal cancer (CRC) is the third most common cancer and the third leading cause of cancer–related deaths in the United States. It is estimated that approximately 135,000 new cases of CRC will be diagnosed in 2016 and that 50,000 deaths from colorectal cancer will occur.1 CRC screening has been associated with a decrease in CRC incidence and mortality.2 Unfortunately, only 65% of eligible Americans are up–to–date with the recommended screening. Colorectal cancers detected on screening are more likely to be early stage and curable compared with cancers detected on an examination done for symptoms related to the tumor. Efforts should be focused on improving the rates of screening, recognizing and mitigating risk factors, adhering to evidence based intervals for colonoscopic surveillance, and enhancing the quality of colonoscopy.

Unfortunately, the current screening tests are either expensive or inaccurate, according to the Colon Cancer Alliance.

Fecal occult blood tests, namely the FIT and the gFOBT, are extremely inexpensive and can be bought for just $5. Naturally, these tests have a high patient compliance due to their ease and inexpensiveness. Results can be seen in just five minutes without any lab work. These tests have a 70% sensitivity, however, meaning that nearly one in three people who take the test will have cancer, but the test will say they don’t.

Stool DNA tests, like Cologuard, are much more accurate, yet they have a price tag upwards of $600. These tests are also conducted in a lab environment and can therefore up to two weeks to get results (source). On top of this, they are available only by prescription, and are not meant for people with a history of cancer or for late stage cancers.

In the end, the need for cheap, accurate, and quick screenings for colon cancer has still not been met.

What we do

Once again, enter CounteractIO.

CounteractIO combines the FIT with powerful AI algorithms to boost the sensitivity and overall accuracy of the test.

A lot of progress has been done since the last post, and that’s what I wanted to share today.

A ROC for the models we used to classify cancer/no cancer.

After 10-fold cross validation scoring, we were able to get the following results:

Specificity (percentage of true positives): 90%
Sensitivity/Recall (percentage of true negatives): 87%
Precision (confidence of each guess): 91%
F1 (harmonic mean of Precision and Recall): 87%
AUC: .97

Comparison with other tests

Cologuard falls under the Stool DNA Test category

I promised you I’d go head-to-head with Cologuard. As you can see, our metrics are very reasonably on par with them, especially given our significant advantages in the game.

CounteractIO is for anyone at any stage of CRC, completely done at home, 60x cheaper, and up to 4,000x faster than Cologuard.

Disclaimer: We’re currently in talks to get funding and clinical trials.