“It doubles our efficiency”: built-in AI radiology tools increase productivity by 40% and save lives

In the first real-life deployment of generating artificial intelligence, a new AI system has increased radiological productivity by 40% for medical imaging interpretation.

The tool was developed in Northwest Medicine and tested in 11 hospitals, generating nearly 24,000 radiological reports in five months while maintaining clinical accuracy and identifying life-threatening conditions within seconds of imaging.

https://www.youtube.com/watch?v=wor80psraom

Real world performance exceeded expectations

Unlike previous AI tools that focused on single conditions, Northwest’s system analyzed the entire X-ray and CT scan, resulting in a complete report that radiologists can review and finalize. The average efficiency gain reached 15.5%, and some radiologists improved by up to 40%.

“This is the first use of AI as far as I know, and it significantly improves productivity, especially in healthcare. Even in other areas, I haven’t seen nearly 40% improvements.”

Built from scratch, not large-scale technology

Instead of adapting to existing models like Chatgpt, Northwest engineers built the system from scratch using clinical data from their own hospital networks. This approach creates a lightweight, professional tool with much less computing power than commercial alternatives.

“There is no need for health systems to rely on tech giants,” said Dr. Jonathan Huang, a third-year medical student with a doctorate. In biomedical engineering. “Our research shows that building custom AI models fits the scope of a typical health system without relying on expensive and opaque third-party tools like Chatgpt.”

The speed of life-saving in key cases

The system automatically marks life-threatening conditions, such as radiologists even viewing images before crashing lungs. During the test, it identified 72.7% of clinically significant pneumothorax cases with a specificity of 99.9%, alerting within 24 seconds of image completion compared to a typical delay of 24.5 minutes radiologist notification.

“For me and my colleagues, that’s not to say it doubles our efficiency. It’s a huge advantage and strength multiplier,” said Dr. Samir Abboud, director of Emergency Radiology at Northwestern Medicine.

Exceeded efficiency: Maintain quality

Peer review of 800 studies showed no difference in clinical accuracy or text quality between AI-assisted and traditional reports. The system maintains the same reporting correction rate, with only 0.14% of AI-assisted reporting required appendixes, compared with 0.13% of traditional reporting.

The reason for setting this system is the method of its report generation. Instead of simply marking out abnormalities, it provides 95% full reports for each patient and radiologist’s style, basically like a highly skilled trainee whose job requires final review.

Solve severe shortages

Schedule is not important. By 2033, the United States expects a shortage of 42,000 radiologists as imaging volume increases by 5% and residence locations increase by only 2%. The Northwest Tool’s tool provides a practical solution to help radiologists clear up the backlog and deliver results in hours rather than days.

During the study period, the AI ​​system saved 63 hours of documentation, which was equivalent to reducing the coverage requirement from 79 radiologist shifts to 67. This efficiency improvement becomes particularly valuable in emergencies where every minute is important.

Prevent the diagnosis of the body

The study reveals compelling examples of the system’s potential to save lives. In one case, after preliminary reading, a patient with large pneumothorax was discharged from the emergency department. The AI ​​system was marked immediately, but since the alarm has not been played, the patient was recalled only six hours after the radiologist’s review.

Another patient who was tested for pneumonia had pneumonia until an oxygen desaturation event occurred 11 hours after imaging. The AI ​​determines it within seconds of the scan.

Error rate tells the story

The complexity of the system shows its editing mode. For chest X-rays, the word error rate between the AI ​​draft and the final report is only 0.31, meaning about one-third of the words need to be modified. For non-thoracic studies, the rate was 0.63, reflecting an increase in musculoskeletal and other imaging.

Importantly, the system generates alarms for pneumothorax at just over one rate per day throughout the health system, indicating its ability to minimize alarm fatigue while capturing critical cases.

The future of medical AI

“You still need radiologists as the gold standard,” Abid stressed. “Medical changes – new drugs, new devices, new diagnoses – we must make sure that AI remains the same. Our role is to ensure that every explanation is suitable for the patient.”

The technology has two approved patents, while the others are yet to be audited and are under early commercialization. Etemadi’s team (who he calls it a “Bell Labs style” approach, attracting talent from major technology and financial companies to work in the hospital system.

“We will not only push healthcare AI forward, but also push the fundamentals of AI with a small percentage of the cost of large AI labs. This is the beginning of the DeepSeek moment for healthcare AI,” Etemadi said.

The study, published in JAMA Network Open, represents the first prospective clinical assessment of generative AI for radiological reporting, providing a roadmap for other health systems that want to increase efficiency without damaging patient care.

Related