A machine learning program at Penn State Heart and Vascular Institute was able to predict the response of hepatocellular carcinoma (HCC) to transarterial chemoembolization (TACE) in about 80% of the cases. However, the program performed even better when predicting long-term response.
Artificial intelligence (AI) is the science of developing computer systems able to perform tasks previously only performed by humans. Machine learning is a subset of AI that uses algorithms and provides past data to a computer so it can recognize patterns and “learn” to make decisions.
Researchers at the Penn State Heart and Vascular Institute tested how well machine learning could predict the clinical efficacy of TACE as monotherapy to treat HCC, the most common type of primary liver cancer. TACE combines local delivery of high doses of chemotherapy with embolization.
Currently, the Barcelona Clinic Liver Cancer staging system is commonly used for HCC treatment decisions, said presenter Tuan Vu, a third-year medical student at Penn State College of Medicine who plans to become an interventional radiologist. “But with the advancement of machine learning over the years, we can extract more data about the cases in the past … to analyze and find patterns that you and I cannot even see, in order for us to stratify patients.”
The researchers conducted a retrospective review of 124 patients who had undergone TACE as first-line treatment for HCC. The machine was tasked with analyzing data from pre-treatment CT imaging and clinical values to predict immediate treatment response.
The machine accurately predicted TACE effectiveness in 80% of the cases, a percentage similar to previous research using AI prediction for TACE, said Jeff Cruz, MD, the study’s principal investigator and now an assistant professor of clinical radiology at Temple University.
Next, the team used the machine learning model to see if it could predict longitudinal treatment response with follow-up treatment. These secondary endpoints included liver transplantation, surgical resection, time to next treatment and duration of clinical benefit. The machine performed better with secondary endpoints, accurately predicting the type of follow-up treatment 88% of the time.
The team decided to run the program a third time to see if the algorithm would continue its predictions. This is where things got “fascinating,” Dr. Cruz said. The team provided the model with tumor response data at the posttreatment imaging in addition to the pre-treatment imaging and clinical values, approximating information that would be presented at a multidisciplinary conference.
“It actually increased the accuracy of the model to 96%,” Dr. Cruz said. “It was able to better predict subsequent outcomes when it already knew the outcome of one event. … That was not something that was programmed into it.”
That is important, he said, because many conversations at HCC tumor boards focus on “Should we continue on, should we hold off, or should we switch therapy?”
“That’s the interesting part about this—that when you give it the information, the computer was actually able to correspond and be accurate with the decisions made at the tumor board, which is multispecialty,” Dr. Cruz said.
Mr. Vu became interested in AI through one of his best friends from their home country of Vietnam. At the beginning of the COVID-19 pandemic, his friend was part of a team that created a chest X-ray algorithm to determine “regular” pneumonia versus COVID-19 pneumonia. That prompted Mr. Vu to look for AI research opportunities when he started medical school.
“AI is not perfect; it’s not the end-all, be-all,” he said. “However, AI is an excellent assistant in terms of managing oncology patients.”
