Deep-Learning NLP Tool Helps Catch Health Issues Early
AI Tool Reviews Imaging Findings to Flag High-Risk Conditions
Each year, thousands of imaging reports are generated at hospitals. While primary results address the issue at hand, a full report can contain other significant — yet unexpected — findings.
These incidental findings include lung, adrenal or thyroid nodules, which are common and often benign. Lung nodules, for example, occur in more than 50% of adults with a CT or chest X-ray. Of those, 95% are noncancerous.
However, without a process for identifying and communicating these findings, patients could be at risk for delayed or missed treatment.
In 2020, Northwestern Medicine set out to develop a solution using deep-learning natural language processing (NLP).
The first step was training the algorithm to understand clinical interpretations. This involved manually annotating and classifying thousands of imaging reports and then using them to teach the algorithm to look at word patterns to determine if a radiologist was describing a potential adrenal or lung nodule. Once it’s been trained, the tool then sorts through thousands of reports and identifies imaging findings, both expected and incidental, where follow-up may be needed.
Tool Supports Early Detection
In addition to flagging findings that could otherwise go undetected, the algorithm helps expedite determining next steps in care and supports faster treatment, improving outcomes for many patients.
After the tool’s initial release, the team studied findings to determine its impact. Over 13 months, more than 570,000 imaging studies were screened, of which more than 29,000 were flagged as containing lung-related follow-up recommendations.
The tool demonstrated a positive predictive value of 90.3% for lung findings requiring follow-up.
The Continuing Impact
As of March 2023, the algorithm has reviewed more than 3 million imaging reports, with over 130,000 findings identified and tracked.
3 million
imaging reports reviewed
130,000+
findings identified and tracked
The team determined the system had significant potential and began to focus on expanding its development. The platform can scaled and adapted to other clinical cases with appropriate annotation of the relevant variables.
In 2024, the tool was expanded to identify thyroid findings, with hepatic steatosis added in 2025. AI models for liver tumors, pancreatic cysts, aortic findings and ovarian findings are underway. Early detection for these types of high-risk conditions can improve outcomes.
