LONDON – In a proof-of-principle study, artificial intelligence (AI) led more frequently to the correct diagnosis of underlying lung disease than did physicians’ use of standard algorithms, such as those recommended by the American Thoracic Society and the European Respiratory Society, according to late-breaker data presented at the annual congress of the European Respiratory Society.
“The beauty of this approach is that artificial intelligence can simulate the complex reasoning that clinicians use to reach their diagnosis but in a more standardized and objective fashion, so it removes any bias,” reported Wim Janssens, MD, PhD, of the division of medicine and respiratory rehabilitation at University of Leuven (Belgium).
The AI employed in this study was based on a subfield of computer science that relies on patterns within statistics to build decision trees. Often called machine learning, this type of AI grows smarter as it learns from the patterns it finds in the data provided.
In this case, the AI was designed to provide diagnoses for lung diseases based on patterns drawn from clinical and lung function data. The computer-based choices were compared to diagnoses reached by clinicians. The final diagnoses were validated by a consensus of expert clinicians.
“The computer-based choices were in almost all cases better than the choices made by standard diagnostic algorithms,” reported Marko Topalovic, PhD, a researcher in AI who is affiliated with the University of Leuven. Dr. Topalovic presented the data at the ERS.
The study involved 968 patients presenting with lung symptoms to a pulmonary clinic for the first time. Standard clinical data, such as smoking history, body mass index, and age, were collected. Lung function studies conducted in all patients included spirometry, body plethysmography, and airway diffusion, although participating clinicians were permitted to order additional tests at their own discretion. Clinical diagnoses were separated into 10 predefined disease groups.
The average accuracy of clinicians’ diagnoses was 38%. The clinicians were best at identifying chronic obstructive pulmonary disease (COPD), having accurately diagnosed 74% of the cases of this disease. For other disease groups, the clinician’s accuracy rarely exceeded 50%.
The diagnoses made by AI, on the other hand, on average, were 68% accurate. For diagnosing COPD, the AI achieved a positive predictive value of 83% and a sensitivity of 78%. The positive predictive value and sensitivity of AI for asthma (66% and 82%, respectively) and interstitial lung disease (52% and 59%) were both significantly greater than those achieved by the clinicians.
When findings are ambiguous or there are anomalies in the clinical picture, a final clinical diagnosis can be challenging, according to Dr. Janssens. He suggested that automation eliminates the potential for bias, which often occurs when clinicians inadvertently give more weight to one clinical variable relative to another.
The decision-making system tested in this study was characterized as “a first step to automated interpretation of lung function,” Dr. Topalovic said. He added that he expects the AI to improve as it receives more data.
“Not only do we think this system can help nonexperienced clinicians, but it will help experts reach a diagnosis more quickly,” Dr. Topalovic said. Noting that this same approach is being pursued in other fields of medicine, he said he thinks adding AI to respiratory medicine will reduce the number of redundant tests and, in other ways, introduce opportunities for efficiencies and reduced costs.
Dr. Topalovic and Dr. Janssens reported no relevant financial relationships.