A new AI-powered cancer screening tool being piloted at Galway University Hospital (GUH) has proved successful in helping to identify lung cancers that might otherwise have been missed.

The Philips Lung Cancer Orchestrator is an integrated lung cancer patient management system that can be applied to CT images generated through lung screening programmes or that reveal incidental pulmonary findings.

In GUH the system forms the basis of a new incidental pulmonary nodule pathway. The orchestrator is being used identify suspicious pulmonary findings in patients that might otherwise be overlooked.

The software uses natural language processing (NLP) to mine radiology reports for relevant keywords to trigger referral for follow-up, management and treatment. The system has the ability to enrol patients in a management protocol with automated notifications, reminders and status updates.

Consultant respiratory physician David Breen has been spearheading the Lung Cancer Orchestrator project at GUH. He says the pilot is already delivering benefits.

“Lung cancer is the leading cause of cancer related death,” Dr Breen said. “It causes more deaths than breast, prostate and colorectal combined. So it’s a major problem within our population. The vast majority of our patients are silent until it’s too late.

“In other words, they present with stage three or four disease. If you diagnose patients early and you get them at stage one, they can have potentially resectable options and their survival goes up to about 92 per cent.”

Data consistently shows that even a four-week delay in treatment is associated with significant increases in mortality, so early detection is crucial.

Unfortunately, there are often lost opportunities to detect lung cancer. Incidental lung nodules are common findings in images requested for other clinical reasons.

Statistics suggest that more than 70 per cent of these incidental lung nodules are not followed and managed appropriately, despite the fact that 25 per cent of actionable incidentally found pulmonary nodules are malignant.

“We know that in our own population there are many cases that come in with symptomatic lung cancer and when you bring them into clinic and review them, they’ll have CT scans done in the past where a nodule has been there and has not been picked up on,” Dr Breen said.

“It was a missed chance for cure and instead they’ve actually presented with late disease.

“Often these patients never come under our care. They are under other teams and based on how it’s reported, [the nodule] may not be something that is a red flag for a GI surgery or a neurologist.

“The Lung Cancer Orchestrator allows these cases to be picked up and brought to our attention and then we can follow up with them.”

The Orchestrator uses NLP to screen every radiology report in the facility’s radiology information system or EMR for certain keywords, such as ‘size 7mm’, ‘nodule’, ‘left lower lobe’ or ‘right upper lobe’, and also the type of nodule.

These keywords are structured and compared to the Fleischner Society pulmonary nodule recommendations on the follow-up and management of indeterminate pulmonary nodules detected incidentally on CT.

Ultimately, the software generates a recommendation for future management, whether that be regular surveillance or further diagnostic tests.

Dr Breen presented the initial findings from the GUH pilot project at last week’s Health Informatics Society of Ireland (HISI) conference on AI in Healthcare.

The latest data show that 5530 patients have been triaged through NLP. Among these there were 185 cases requiring follow up. Six patients were confirmed to have lung cancer.

“That doesn’t sound like a huge amount,” Dr Breen said. “It’s about 0.5 per cent. But if you actually take the most recent data from the UK, 900,000 CT scans have been performed in the UK as part of lung cancer screening and the diagnostic yield is one per cent.

“That’s deemed an appropriate return, so we’re doing well, especially since the system is not yet as efficiently as it can be.

“We’re definitely at the crawl phase. It’s not doing everything that we want and we definitely haven’t used it to the maximum capacity. That’s stage two.

“It should be much more virtual than we’re doing it, because the radiology ordering systems is still at its infancy and there is a fear that patients will be lost.

“There’s a lot of face-to-face interactions with patients which probably isn’t needed in a nodule follow-up program. It’s getting all of that right over the coming years.”