AI is trained to spot danger signs in blood tests

This is the third feature in a six-part series examining how AI will transform medical research and treatment.

Ovarian cancer is “rare, underfunded and deadly,” said Audra Moran, president of the New York-based global philanthropic organization Ocra.

As with all cancers, the earlier it is detected, the better the outcome.

Most ovarian cancers start in the fallopian tubes, so by the time the cancer reaches the ovaries, it may have already spread to other places.

“Five years before symptoms appear is when ovarian cancer may need to be detected to have an impact on mortality,” Moran says.

But new blood tests are emerging that harness the power of artificial intelligence (AI) to detect signs of cancer at a very early stage.

AI can also speed up other blood tests, not only for cancer but also for potentially deadly infections such as pneumonia.

Dr. Daniel Heller is a biomedical engineer at Memorial Sloan Kettering Cancer Center in New York.

His team developed a test technique that uses nanotubes, tiny tubes of carbon that are about 50,000 times smaller than the diameter of a human hair.

About 20 years ago, scientists began discovering nanotubes that emit fluorescence.

Over the past decade, researchers have learned how to modify the properties of these nanotubes so that they react with almost everything in the blood.

It is now possible to put millions of nanotubes into a blood sample and have them emit different wavelengths of light based on what is attached.

But the problem of interpreting the signal remains, which Dr. Heller likens to finding a fingerprint match.

In this case, the fingerprint is a pattern of molecules that bind to the sensor, with different sensitivities and binding strengths.

But the patterns are too subtle for humans to discern.

“You look at the data and you don’t understand it at all,” he says. “AI can only see different patterns.”

Deciphering the nanotube data means loading the data into a machine learning algorithm that tells the algorithm which samples are from ovarian cancer patients and which are from people without ovarian cancer. meant to convey.

These also included blood from people with other forms of cancer and other gynecological conditions that could be confused with ovarian cancer.

A major challenge in using AI to develop blood tests for ovarian cancer research is that the tests are relatively rare, which limits the data on which to train the algorithms.

And much of that data is siled within the hospitals that provide treatment, with minimal data sharing for researchers.

Dr. Heller describes training the algorithm based on available data from just a few hundred patients as a “Hail Mary pass.”

But he said that while AI was able to achieve greater accuracy than the best cancer biomarkers currently available, it was only a first attempt.

Further research is being conducted to see if this system can be improved using a larger set of sensors and samples from more patients. Just as self-driving car algorithms can be improved with more road tests, algorithms can be improved with more data.

Dr. Heller has high hopes for this technology.

“What we want to do is triage all gynecological conditions, so when someone comes in with a complaint, is it more likely to be cancer or more likely not to be cancer? , or can we give doctors a tool that can tell them right away if it’s likely to be cancer?”

Dr. Heller said this could be “three to five years” away.

AI could not only help with early detection, but also speed up other blood tests.

Pneumonia can be fatal for cancer patients, and there are approximately 600 types of microorganisms that can cause pneumonia, so doctors must perform multiple tests to identify an infection.

But new types of blood tests are simplifying and speeding up the process.

California-based Karuis uses artificial intelligence (AI) to accurately identify pneumonia pathogens and select the appropriate antibiotic within 24 hours.

Carius chief executive Alec Ford said: “Prior to our test, patients with pneumonia would undergo 15 to 20 different tests to identify infection in just their first week in hospital. The test will cost about $20,000.”

Carius has a database of microbial DNA containing tens of billions of data points. Test samples from patients can be compared to that database to identify the exact pathogen.

Ford says none of that would be possible without AI.

One challenge is that researchers currently do not necessarily understand all the connections that AI may make between test biomarkers and disease.

Over the past two years, Dr. Slave Petrovsky has developed an AI platform called Milton that uses biomarkers from UK Biobank data to identify 120 diseases with a success rate of over 90%.

Only AI can find patterns in such large amounts of data.

“These are often complex patterns, and there may not be a single biomarker, but the whole pattern needs to be considered,” says Dr. Petrovsky, a researcher at pharmaceutical giant AstraZeneca. Masu.

Dr. Heller uses similar pattern matching techniques in his research on ovarian cancer.

“We know that the sensor binds to and responds to proteins and small molecules in the blood, but we don’t know which proteins and molecules are specific to cancer,” he says.

More extensive data, or lack thereof, remains a drawback.

“People aren’t sharing their data or don’t have the mechanisms to do so,” Moran said.

Ocra is funding a large-scale patient registry with patients’ electronic medical records that will allow researchers to train algorithms on that data.

“It’s still early days. We’re still in the AI ​​wilderness,” Moran says.