By Bob Kronemyer

Originally published by our sister publication, Gastroenterology & Endoscopy News

Advanced machine learning (ML) technology was found beneficial in capturing the time sequence of clinical events for pancreatic cancer and predicting risk over several years, according to a presentation at the 2022 annual meeting of the American Association for Cancer Research, in New Orleans (poster LB55).

The researchers noted that currently, there are no reliable biomarkers or screening tools to detect pancreatic cancer early.

Inspiration to pursue artificial intelligence, according to the investigators, came from the observed similarity between disease trajectories and the sequence of words in natural language. In other words, each diagnosis is considered as a word to form a complete sentence, as opposed to previous AI models that treated diagnoses like a conglomeration of words and did not make use of the sequencing of disease diagnoses in an individual’s medical history.

Through electronic health records, the current AI method identified a subset of patients with a 25-fold risk for developing pancreatic cancer within three to 36 months. The study used, in part, clinical records from the Danish National Patient Registry (NDPR), comprising 41 years (1977-2018) and 6.1 million patients, of whom roughly 24,000 developed pancreatic cancer.

After risk assessment, the investigators tested a range of ML methods to predict cancer occurrence in time intervals of three to 60 months. For cancer occurrence within 36 months, the best ML model showed an odds ratio (OR) of 47.5 for 20% recall and 159.0 for 10% recall.

The researchers validated their findings using electronic medical records from the Mass General Brigham Health Care System, in Boston, for which the data set performed comparably to the NDPR for the best ML model (OR, 112.0 for 20% recall and OR, 162.4 for 10% recall). The area under the receiver operating characteristic curve was also similar: 0.88 for the Boston data set versus 0.87 for the original training set in Denmark. In addition, AI was able to estimate the contribution of individual disease features such as obesity and diabetes to predict the occurrence of cancer.

Study results support the design of future screening trials for high-risk patients, according to the study’s researchers. AI applied to real-world clinical records can also potentially shift focus from treating late-stage to detecting early-stage cancer, thus improving patient life span and quality of life.

The ability of an AI tool to help clinicians identify patients at high risk for pancreatic cancer can lead to increased patient enrollment in prevention or surveillance programs, as well as undergo early treatment.

In conclusion, an AI tool can integrate information pertaining to risk factors in the context of an individual patient’s disease history. The investigators expect to increase prediction accuracy of AI by using data other than disease codes, including prescriptions, laboratory values and images.