REDMOD AI Detects Pancreatic Cancer Signs Nearly 500 Days Before Diagnosis
A new artificial intelligence system, named REDMOD (Radiomics-based Early Detection MODel), has demonstrated the ability to identify subtle tissue changes associated with pancreatic ductal adenocarcinoma (PDAC) on CT scans before a clinical diagnosis is made, according to a study published in the journal Gut.
The model analyzes radiomic patterns in CT images that are not visible to the human eye and does not rely on the presence of a visible tumor. The study found that the AI system detected pre-clinical signs of the disease an average of 475 days before diagnosis.
Study Design and Performance
Researchers trained and tested REDMOD on abdominal CT scans from patients who were later diagnosed with pancreatic cancer, as well as from cancer-free control subjects. The scans were initially read by radiologists as showing no evidence of disease. In total, the study analyzed scans from:
- 219 patients who were later diagnosed with PDAC.
- 1,243 control patients without the disease.
The study involved patients from multiple hospitals, and cancer and control groups were matched by age, sex, and scan date. The average age of the cancer patients was 69.
The performance of REDMOD was compared to human radiologists. Key results include:
- Sensitivity: REDMOD achieved a sensitivity of 73%, compared to 39% for radiologists in detecting early malignant changes.
- Long-term detection: For scans taken more than two years before the clinical diagnosis, REDMOD was nearly three times more accurate than radiologists (68% vs. 23%).
- Scans taken more than 18 months before diagnosis: the AI was reported to be twice as sensitive as radiologists.
- False positive rate: In a test of 63 pre-diagnosis scans and 430 control scans, the false positive rate was 18.8%.
- Consistency: In 90-92% of cases, the model produced consistent results when a patient was scanned months apart.
Limitations and Next Steps
The study authors noted several limitations, stating that the research did not include an ethnically diverse patient group. They emphasized that prospective validation of the model in high-risk patient populations—such as those with new-onset diabetes, unexplained weight loss, or a family history of the disease—is necessary before it can be considered for clinical use.
The AI model is currently being evaluated in a clinical trial that will follow participants for three to five years.