Machine learning for cancer screening

Scientists have developed a blood test or “liquid biopsy” that detects lung cancer at an early stage, something that could save more than 11,000 lives in the US alone.

The research, led by radiation oncologist Maximilian Diehn from the Stanford Cancer Institute in California and published in Nature, uses machine learning to drill down on tiny levels of DNA from the tumour in the bloodstream.

The team found the quantity of DNA revealed a host of facts about the most common type of lung cancer, known as non-small cell lung cancer, in people who already had the disease, including its cell type, how advanced it was and how aggressively it was likely to spread.

The researchers then trained a machine learning model to rate the chance that DNA variants in the blood came from lung cancer, a method they call “lung cancer likelihood in plasma” or “Lung-CLiP”.

When they put Lung-CLiP through its paces, it detected lung cancer in 63% of patients who were at Stage 1, the earliest phase when the cancer is confined to one lung and hasn’t spread to lymph nodes or outside the chest.

It’s potentially a critical advance because lung cancer is the leading cause of cancer deaths in the US, killing more than 155,000 people each year. One in five of those deaths is preventable.

If the disease is picked up early, each arm of treatment, including surgical excision, radiotherapy and chemotherapy, becomes more effective. But recent efforts to screen for early-stage lung cancer have been a dismal failure.

In 2013, the US Preventive Services Task Force recommended all high-risk people get screened with a low radiation dose chest CT scan (LDCT). That means if you are between 55 and 80 years old and smoke 30 or more packs of cigarettes a year – including people who quit in the last 15 years – you should get a LDCT every year.

But one study found that in 2015 fewer than four of every 100 eligible candidates were scanned in the preceding year, something the authors put down to doctors and patients not knowing about the test, as well as problems with reimbursement.

LDCT also generates a high number of false positives – in excess of 90% of the scans detect a cancer that’s not actually there.

The Lung-CLiP blood test could refine and jump start that screening process.

“One potential application of Lung-CLiP could be to serve as an initial screen for some of the approximately 95% of high-risk patients in the US who, despite being candidates for LDCT, are not being screened,” the authors write.

People with a positive Lung-CLiP test could then be referred more confidently for LDCT, a “hybrid” approach the authors say could raise the number of lives saved annually in the US from the current 600 to 12,000.

It’s an approach the researchers think could also work for cancers beyond the lung:

“[By] incorporating molecular features appropriate for other cancer types, we expect that it could be feasible to develop CLiP methods for a diverse range of malignancies,” they conclude.