Chapter 82
Cancer — 9
If we could assess a person’s immune status with a comprehensive test—what I call an “immunome”—assaying our innate and adaptive immunity in depth, it would help us detect individuals with impaired responsiveness, as typically seen with aging. That could trigger use of a vaccine or other strategy to strengthen the immune system in a person at high risk. The only immune marker we have in common practice today is the ratio of white blood cells to lymphocytes, which is rudimentary but predictive of outcomes across all major diseases and in patients with cancer.
Pancreatic cancer is renowned for being diagnosed far too late, with a high fatality rate. The five-year survival rate is less than 10 percent. This hit home recently when my seventy-three-year-old brother-in-law in good health presented to an emergency department for evaluation of acute blurred vision. He was admitted to the hospital and an extensive workup determined that it was due to blood clots from stage 4 pancreatic cancer, for which he had no symptoms or warning. Approximately fifty thousand deaths in the United States are from pancreatic cancer each year, making it the third-leading cause of death from cancer after lung and colon.
In light of that, a major study was conducted in two distinct cohorts: more than eight million people in Denmark and about three million people in the US Veterans Affairs health system. A total of nearly twenty-eight thousand cases of pancreatic cancer were diagnosed. When an AI model combed through all the electronic medical records over several years, it achieved high accuracy for predicting who would get pancreatic cancer. For the highest-risk group, 320 of 1,000 people aged fifty and older developed pancreatic cancer. In a similar study of fifty-five health care organizations in the United States, AI analytics of the electronic records differentiated the 35,000 patients who developed pancreatic cancer from 1.5 million people who did not. In both studies, the model identified tens of features, such as results of lab tests, symptoms, medications, and coexisting conditions, that collectively accounted for its performance. That’s just with the use of time-sequenced electronic records.
What if we add other layers, such as genome sequencing, scans, or blood biomarkers?
If whole genome sequencing was routine, we wouldn’t miss data from high-risk pancreatic cancer predisposition genes such as mutations in BRCA1, BRCA2, PALB2, CDKN2A, ATM, Lynch syndrome (MLH1, MSH2, MSH6, PMS2, and EPCAM), Peutz-Jeghers syndrome (STK11), and Li-Fraumeni syndrome (TP53). Using genomics of the tumor and AI analytics, it is possible to determine a more successful treatment plan and predict recurrence of a tumor more than ten years from initial diagnosis.
Many patients get CT scans of the chest or abdomen without contrast dye for various reasons that have nothing to do with assessment of the pancreas. But an AI model was trained on both types of CT scans to accurately detect pancreatic cancer from these scans, markedly outperforming radiologists, a task considered so difficult that it would defy attempts from a noncontrast scan. This work was independently replicated with ability to pick up preinvasive pancreatic cancer from standard abdominal CTs with high accuracy. The compounds carbohydrate antigen 19.9 (CA19-9) and bilirubin have been shown to correlate with pancreatic cancer and differentiate from benign pancreatic abnormalities. We’ll come back to pancreatic cancer shortly for other layers of data to predict optimal treatment and patient outcomes.