Blood count. Biopsy. Drug cocktails. Snippets like these tell the story of a person鈥檚 experience of cancer. Gather up the stories of hundreds of thousands of people and you could learn about the disease itself.
A team at Memorial Sloan Kettering Cancer Center in New York is training an artificial intelligence to find similarities between cases that human doctors might miss. The software combs through 100 million sentences taken from clinical notes about people with cancer.
鈥淲e鈥檙e looking into the exhaust of all that data to try to find something interesting,鈥� says , who presented the work at the annual meeting of the American Association for the Advancement of Science in Washington DC last week.
Advertisement
His idea is to build computational models that capture how a person is doing, how they compare to others and how their disease is likely to progress in the future. 鈥淥nce we have that, we can think about how to treat the patient best.鈥�
Secret similarities
R盲tsch鈥檚 team built a machine learning algorithm to crunch through anonymised clinical notes from 200,000 people with cancer. Their program sorted millions of sentences 鈥� including patients鈥� symptoms, medical histories and doctors鈥� observations 鈥� into 10,000 related clusters.
Each cluster represented a common observation found across several medical records. For example, a doctor鈥檚 note recommending a particular course of treatment, or picking up on a noteworthy symptom. Connections between clusters were then mapped, showing the relationships between different comments or courses of treatment.
In a second study building on R盲tsch鈥檚 work, the clusters are now being compared against the records of about 2000 people with different types of cancer. The researchers are looking for hidden associations between written notes and patients鈥� gene and blood sequencing. For example, patients with similar genetic results might have the same kind of note in their files. These connections can reveal similarities doctors might not have noticed before.
Digital diagnosis
The hope is that these associations will inspire ideas for research. 鈥淵ou can take the genetic information and make this connection in order to find new hypotheses, which can then be tested,鈥� says R盲tsch.
Machine learning is proving useful in a range of medical applications. For example, computers are also being trained to diagnose problems using biological images like X-rays and MRI scans. Another system at Chicago hospitals learned to predict when people were likely to experience a heart attack in the near future.
鈥淭he human mind is limited, hence you need to use statistics and computer science,鈥� says R盲tsch.
