5 hours Ago
A group of investigators led by Cedars-Sinai have developed and successfully tested a new artificial intelligence (AI) method to make launching cancer clinical trials easier and faster. The method uses patients' pathology reports to automate the classification of patients by the severity of their cancers, potentially shortening the process of selecting candidates for clinical trials. Their achievement, described in Nature Communications , significantly expands AI's health care applications.
The new AI method, also called a model, offers a much-needed alternative to tumor registries, the databases maintained by governments and hospitals. Researchers normally use tumor registries to screen cancer patients for clinical trials . Cancer registries require specially trained employees to manually identify a patient's cancer stage by reviewing laboratory reports, clinicians' notes and other information.
The process can be slow and tedious. "By the time a cancer patient's data is entered into a tumor registry, months may have passed, along with the opportunity for the patient to participate in relevant clinical trials or other treatments," said Nicholas Tatonetti, Ph.D.
, vice chair of Computational Biomedicine at Cedars-Sinai, associate director for Computational Oncology at Cedars-Sinai Cancer and corresponding author of the study. "Our AI model can dramatically reduce that delay, accelerating the pace of research and expanding patients' access to clinical trials." The team's AI model quickly identifies the cancer stage by extracting and interpreting the text of just one element of a patient's electronic health record: the pathology report, which describes the findings of pathologists examining tissue specimens from the patient.
In tests involving thousands of patient records, the study's investigators confirmed that the AI model they created was highly effective in staging patients' cancers. The method is based on a so-called transformer model of AI, which is designed to simulate the complex decision-making power of the human brain. The study team first "trained" the model to stage cancers using publicly available pathology reports from a government database, The Cancer Genome Atlas.
These reports covered nearly 7,000 patients and included 23 types of cancers. To make sure the model worked in various settings, investigators then applied it to nearly 8,000 pathology reports maintained by a single medical center. The results, as measured by a standard statistic for evaluating AI models, rated the method as highly accurate.
"This was an important finding because it means that our AI model is an 'off- the-shelf' tool that can be generalized to other institutions without requiring that it be trained for each location," Tatonetti said. Besides screening patients by their cancer stages for clinical trials, the AI model also can be used to automate classification of patients for observational and retrospective data analysis, and for potential treatments, according to Tatonetti. "Future research could build on our method to integrate the pathology text with other types of clinical data, potentially advancing personalized cancer treatment," he said.
The creation of the AI model was made possible by earlier research , also led by Tatonetti, that removed technical obstacles to computers extracting and analyzing pathologists' notes from electronic health records. In a notable decision, the investigators in the new study have made their AI model, which they named BB-TEN: Big Bird—TNM staging Extracted from Notes, available to other institutions for academic uses and certain other purposes. "By speeding up the selection of candidates for cancer clinical trials, this innovative AI model shows promise for accelerating the development of relevant treatments and making them available to more patients," said Jason Moore, Ph.
D., chair of the Department of Computational Biomedicine at Cedars-Sinai. More information: Jenna Kefeli et al, Generalizable and automated classification of TNM stage from pathology reports with external validation, Nature Communications (2024).
DOI: 10.1038/s41467-024-53190-9.
