Canadian researchers have demonstrated for the first time that artificial intelligence (AI) can accurately predict how long a patient will survive any cancer simply by reading the oncologist’s initial consultation report.
The researchers, from the University of British Columbia (UBC), found that natural language processing, a form of machine learning that analyzes the language in documents, was more than 80% accurate at predicting 6-month, 36-month, and 60-month survival.
Typically, AI applications in cancer require data from structured databases and many are specific to one particular cancer. This new approach transcends both.
“Our results suggest it is possible to predict the survival of patients with cancer without having to construct datasets or limiting the predictions to specific types or locations of cancer,” the authors write.
The study was published online February 27 in JAMA Network Open.
First author John-Jose Nunez, MD, a psychiatrist and clinical research fellow at the UBC Mood Disorders Centre and BC Cancer, Vancouver, Canada, was pleasantly surprised by his team’s results: “I thought they would be pretty good [but] I think they were even a bit better than we expected.”
Cancer researchers have been exploring machine learning since the mid-1990s, typically using complex, structured datasets. The first publication of natural language processing in cancer appeared in 2004. However, using unstructured consultation notes is new, and the simplicity of the approach appeals to Nunez.
“The contribution of this paper is that we’re not using specific bloodwork, specific tumor markers, things that can be expensive or hard for people to use,” Nunez said.
“Essentially, everybody receiving cancer care is going to have one of these documents. So this is really easy to use for a wide group of people and for us to keep building the models and treating it with more data. That’s a big advantage.”
Nunez and colleagues pulled together the records of 47,625 patients who started care in one of BC Cancer’s six centers from April 2011 to December 2016. (BC Cancer provides government-funded cancer care to all residents of British Columbia.)
All patients had a medical- or radiation-oncologist consultation report dated within 180 days of diagnosis.
The reports are in English, generated by dictation, and consist of simple paragraphs with headings, BC Cancer’s standard format for initial consultation notes, according to Nunez.
The researchers fed the 47,000-plus reports into four different algorithms developed by AI researchers over the past 15 years.
Based solely on the words used in the oncologists’ reports, all four language-processing models predicted 6-month, 36-month, and 60-month survival with more than 80% accuracy.
It’s similar to how physicians learn, said Nunez. “Throughout med school residency we need to see thousands and thousands of patients, lots and lots of data, to train our own neural networks inside our brain,” he said. “The nice thing is that AI can train much quicker…a couple of days to go through 30,000-plus patients.”
Unlike trainee doctors, however, software does need to sweat over textbooks, said Nunez, because the algorithms are simply looking for correlations. “That’s one of the coolest parts of AI and one of the most interesting,” Nunez said.
Nunez said that BC Cancer has no immediate plans to integrate his simple “survival predictor” into care and that more validation is required.
Although the Canadian team has not yet compared the performance of their model to that of humans, US researchers showed last year that AI can outperform oncologists — at least for 3-month survival in certain cancers.
A team at the City of Hope, Duarte, California, last year published a study in JAMA Network Open that showed their machine-learning model was 60% accurate at predicting 3-month mortality in patients with metastatic solid tumors overall versus 35% in the case of oncologists (P < .001).
However, this did not apply to all cancers: oncologists were just as good as AI at predicting 3-month survival for genitourinary, lung, and rare cancers in this study.
City of Hope has now embedded “mortality modeling” into its electronic medical records (EMR). The intention, said first author Finly Zachariah, MD, an associate clinical professor in the Department of Supportive Care Medicine, Duarte, California, is to ensure that the right conversations are happening at this critical time.
“If there’s a time that we want to know the patient’s preferences, values, and priorities – that would be now,” said Zachariah.
The Californian software doesn’t generate an actual survival prediction, according to Cameron Carlin, the data science manager at City of Hope.
“We are not directly propagating the model’s prediction of likelihood of mortality,” Carlin said. “We are using this to trigger workflows associated with goals of care.”
The City of Hope tool might, for example, remind alert social workers that the patient lacks an advance directive. It also acts as a partner to the oncologist.
Zachariah explained: “If a clinician is in the chart for a patient in the hospital, they get an alert showing them the top six reasons why the model predicted in a given way, and can decide whether they agree and if they should have a discussion with the patient/family.” They would also be prompted, for example, to involve clinical social work or palliative care, Zachariah said.
Nunez agrees that AI is a partner, not a substitute, for the oncologist.
“I don’t think it’s going to replace anybody’s decision, [and] I don’t think it’s going to be used to determine it,” said Nunez. “But one more data point might be helpful for both patients and oncologists…to [have] a more objective entity telling you that time is limited. It will be helpful to move that conversation forward, for people to really understand their current situation.”
JAMA Netw Open. 2023;6:e230813. Full text
Nunez has reported receiving funding for the study from the Pfizer Innovation Fund through the BC Cancer Foundation. Zachariah reported no relevant financial relationships when the study was published. He currently owns a start-up company that is developing machine-learning software.
For more from Medscape Oncology, join us on Twitter and Facebook.
Source: Read Full Article