AI Tools Target Early Cancer Detection and Clinical Trial Predictions

A Carnegie Mellon University spinout company is using artificial intelligence to identify patients at high risk for lung, liver and pancreatic cancer by training its prediction tool on millions of patient medical records. Xlue Inc. achieved a prediction accuracy rate of 50% among patients with no history of cancer, while the tool's accuracy rate for patients who've already had cancer reached 70%.

The best shot at curing cancer is catching it early, doctors say, but that doesn't always happen. Between 80% and 85% of pancreatic cancer cases aren't diagnosed until the aggressive and lethal disease has already reached advanced stages, according to a 2023 study in the open access medical journal Cureus. Lung and liver cancer also have vague or mild symptoms that can mask the disease until it's too late.

Xlue's CATCH-FM tool is identifying signals that are associated with developing cancer in the future by mimicking the patient's trajectory, allowing treating doctors to recommend follow-up screening for early diagnosis. The company trained its predictive tool from electronic medical records for millions of patients spanning two decades, which were part of a large-scale Taiwanese national health care claims database. Talks are also underway to scan offline and de-identified medical records stored by hospital giant UPMC, Pennsylvania's biggest health care system.

The Shadyside-based company, which employs six people, was spun out of CMU in 2025. The company, which is not yet profitable, has raised $1.5 million in an early friends and family round. Xlue's technology could be adapted to identify patients with a high risk of suffering a stroke or heart attack.

Xlue's approach could also save money: the average cost to insurers to treat one commercially insured patient with metastatic pancreatic cancer ranges between $95,000 and $116,000, according to a 2021 study in the American Health & Drug Benefit journal. Screening for the three types of cancer could ultimately improve public health. Widespread mammography screening, for example, has helped reduce breast cancer mortality rates between 8% and 40%, according to the Centers for Disease Control and Prevention.

In the U.S., only 18.2% of those eligible underwent lung cancer screening in 2022. Lung cancer is the leading cause of cancer-related death worldwide, according to the World Health Organization's GLOBOCAN 2022 data.

In drug development, predictive analytics are also advancing, particularly in oncology clinical trials. Current predictive models are improving, especially when forecasting success from phase I to phase II or III. Many tools are already demonstrating reasonably strong accuracy in that area, which helps companies prioritize the most promising compounds.

However, predicting success from preclinical stages to clinical outcomes is much more difficult. First-in-class compounds often lack sufficient historical data, making accurate modeling challenging. While there is scientific rationale and mechanism-of-action insights, translating that into reliable clinical predictions remains complex.

The biggest unmet need is predicting which compounds will succeed in the clinic. In the future, predictive analytics may become strong enough to reduce reliance on animal models, especially in oncology, where those models often fail to predict patient outcomes.