SAN FRANCISCO – Interstitial lung disease is a difficult diagnosis to make, but a combination of artificial intelligence (AI) techniques and automated language processing could help clinicians identify the early signs of ILD and start patients on therapy, investigators say.
For example, applying an AI algorithm to spirometry readings taken from patients whose data were registered in the UK Biobank identified 27% as having ILD, and of this group, 66% had ostensibly normal lung function on spirometry but were later diagnosed with ILD, reported Marko Topalovic, PhD, from the AI company ArtiQ in Leuven, Belgium, at the American Thoracic Society’s international conference.
Neil Osterweil/MDedge News
Dr. Marko Topalovic
“A diagnosis of ILD is very challenging, so you have patients who are going to be misdiagnosed or have a very late diagnosis, so we aimed to apply our AI algorithm on spirometry to see whether we could detect ILD much earlier,” he said in an interview conducted during a poster discussion session.
AI detected ILD up to 6.8 years before a clinician’s diagnosis, Dr. Topalovic said.
Reading between the lines
In a separate study, investigators at the University of California, Davis, used language analysis software to scour electronic health records for words indicative of early ILD, and found that the technique dramatically shortened the median time to a pulmonary referral, compared with historical controls.
“This is a language processing program that can essentially look through the radiology reports and look for the key words that often describe interstitial lung disease, like traction, honeycomb, fibrotic, etc. With those studies being flagged, an actual pulmonologist will then further review the scan, and see whether it meets criteria for one of the interstitial lung diseases,” lead author William Leon, MD, a resident in the department of internal medicine at the University of California, Davis, said in an interview.
Neil Osterweil/MDedge News
Dr. William Leon
“We then sent the primary care doctor a message to say: ‘Hey, this patient has ILD. You need to send them to a pulmonologist,’ ” he added.
Putting it together
Philip L. Molyneaux, MRCP (UK), MBBS, BS (Hons), from Imperial College London, who comoderated the session but was not involved in the studies, speculated that combining these and other, nontechnical interventions also discussed could help to improve diagnosis of ILD and allow clinicians to prescribe therapy earlier in the disease course.
“What’s going to give you the biggest impact for patients? Everyone working individually is coming up with great advances, and if you put them all together it’s going to provide much greater benefit for our patients,” he said in an interview.
AI Spirometry details
In collaboration with colleagues at the Laboratory of Respiratory Disease at University Hospital in Leuven, Dr. Topalovic applied AI to results of spirometry performed prior to diagnosis of ILD among 109 patients registered in the UK Biobank, a repository of information on more than 500,000 volunteers.
The patients selected had ILD listed as their cause of death, had spirometry performed up to 7 years before their deaths, and did not receive a diagnosis of ILD on the day of the index spirometry.
In all 73% of patients were men, 27% women, with an average age of 64.6 years. A large majority of the sample (77.15%) had a history of smoking, and 60 of the patients (55%) died within one year of an ILD diagnosis.
The investigators plugged the spirometry data and each patients demographic information – including gender, age, height, weight, race, and smoking status – into the AI clinical decision support program, which yielded a statistical probability for each subject of having normal lung function, asthma, COPD, ILD, another obstructive disease, or another unidentifiable respiratory disease.
In 29 patients (27%) the software listed ILD as the highest probability, and of this group 19 patients (66%) had normal lung function according to standard interpretation guidelines.
Spirometry parameters among patients identified as having probable ILD were different from those where ILD was not detected. For example, forced vital capacity (FVC) was 76% of predicted among patients with likely ILD versus 87% of predicted in those who had a diagnosis later (P = .003). Similar differences were seen in the forced expiratory volume in 1 second to FVC ratio, at 0.82 vs. 0.75, respectively (P = .007).
There were no differences in mortality or in median time between spirometry and clinician diagnosis between the groups.