New tool outperforms previous models in predicting esophageal, gastric cardia cancer
Click Here to Manage Email Alerts
Key takeaways:
- Researchers developed and validated the Kettles Esophageal and Cardia Adenocarcinoma prediction tool using data from more than 10 million U.S. veterans.
- The new tool was more accurate than previous models.
A new machine-learning tool accurately predicted incident esophageal and gastric cardia adenocarcinoma using electronic health record data, according to a report published in Gastroenterology.
“We designed this study because most providers are unfamiliar with guidelines recommending which patients should be considered for screening for esophageal adenocarcinoma, and previously available prediction tools are not easily accessible at the point of care,” Joel H. Rubenstein, MD, MSc, research scientist at the VA Center for Clinical Management Research and professor of gastroenterology at Michigan Medicine, told Healio. “We envisioned developing a prediction tool that could be seamlessly integrated into electronic health records, using only data from those records to give providers the estimated risk of cancer in their patient in real-time.”
Using data from the Veterans Health Administration Central Cancer Registry, Rubenstein and colleagues identified 8,430 veterans diagnosed with esophageal adenocarcinoma and 2,965 diagnosed with gastric cardia adenocarcinoma, all of whom had at least one medical encounter between 2005 and 2018. More than 10 million individuals were used as controls.
Researchers included demographics, prescriptions, laboratory results and diagnoses 1 to 5 years before index date as predictors and developed the Kettles Esophageal and Cardia Adenocarcinoma prediction tool (K-ECAN), using 50% of the data for training and 25% for preliminary validation and final testing.
According to results, K-ECAN had better discrimination (area under the receiver operating curve [AuROC] = 0.77) compared with previously validated models such as HUNT (AuROC = 0.68) and Kunzmann (AuROC = 0.64), although the model’s accuracy was “slightly diminished” when only using data 3 to 5 years before index (AuROC = 0.75).
“We found that the new tool was more accurate than previously published tools and more accurate than published guidelines,” Rubenstein said. “Although a GERD disease diagnosis was associated with cancer outcome, the tool did not gain all that much information from including a GERD diagnosis; the tool can predict cancer risk even in individuals without a GERD diagnosis.”
Moreover, E-KAN was “still most accurate” compared with HUNT and Kunzmann when using a simulated a non-veteran population (AUROC = 0.85), researchers reported.
“We believe that use of K-ECAN could increase appropriate uptake of screening,” Rubenstein said. “Future research is needed to validate the tool outside of the Veterans Health Administration and to understand how best to implement its use in guiding screening decisions.”
Collapse
Read more about
Click Here to Manage Email Alerts