Advanced AI algorithms can test for pancreatic cancer and forecast disease development up to three years before a diagnosis.
According to a recent study published in Nature Medicine, an AI tool can identify individuals at risk for pancreatic cancer by analyzing their medical records. The deadly disease is often challenging to detect in its early stages, but the research found that an AI-based population screening could identify potential cases of pancreatic cancer.
Artificial Intelligence Algorithms
The new study trained an AI algorithm on a combination of 9 million patient records from Denmark and the United States, and the model was tasked with identifying patterns that could predict which patients were at risk of developing pancreatic cancer.
The algorithm used combinations of disease codes and their timing to make predictions, with some of the symptoms and codes not being directly related to the pancreas.
Researchers evaluated various AI models to determine their effectiveness in identifying individuals with a higher risk of developing pancreatic cancer over different time frames, ranging from six months to three years.
Compared to population-wide estimates of disease incidence, each version of the AI algorithm showed significantly greater accuracy in predicting which patients would develop pancreatic cancer.
The researchers suggested that the model’s accuracy is on par with current genetic sequencing tests, which are typically only available to a limited number of patients in datasets.
Read more: Harvard Study Reveals Benefits Of Quitting Smoking Early For Lung Cancer Patients
Accuracy of AI Model
The algorithm was trained to recognize patterns that suggested future pancreatic cancer risk based on the sequence of disease conditions over time.
The researchers evaluated the top-performing AI algorithm on a new dataset of over 3 million patient records from the US Veterans Health Administration, including 3,864 pancreatic cancer patients.
The tool’s predictive accuracy was slightly lower on the US dataset, likely due to the data being collected over a shorter time period and including a different patient population than the Danish dataset. However, when the algorithm was retrained on the US dataset, its accuracy improved.
The study highlights the importance of using high-quality and diverse datasets to train AI models and the need for access to large representative datasets globally.
In the absence of globally valid models, local health data should be used to train AI models to reflect the idiosyncrasies of local populations.
Read more: Colorado lawmakers want a property tax return of $661 per homeowner for $2.37 billion