Diagnostic radiological methods play a key role in screening of socially significant diseases. Nevertheless, organization of screening in its current form faces a number of challenges: personnel deficiency and lack of financial resources, insufficient population coverage, lack of examinations’ accessibility and occupational burnout of doctors. Use of artificial intelligence-based services (AI services) with 100.0% sensitivity for autonomous sorting of studies’ results could solve these problems, but it is necessary first of all to make sure that these AI-services provide satisfactory security and screening quality.
OBJECTIVE
To study prospectively the safety and quality of autonomous sorting of preventive imaging studies’ results in real-life clinical practice settings.
MATERIAL AND METHODS
A prospective single-center blind diagnostic study was performed. The study included the results of 209 497 preventive radiographies/fluorographies. All studies were processed with AI-services, configured to 100% sensitivity. The tasks of AI-services included sorting the results into categories «norm» and «not norm». Thereafter, the solutions of AI-services and radiology doctors about categorization were compared.
RESULTS
The proportion of defects of AI-services (false assignment of studies’ results to the «norm» category) was 0.08%, the proportion of clinically significant defects of AI-services — 0.02%. The frequency of false omissions in AI-services omissions (0.04%) was lower than in the average radiologist (8.6%). The consistency in the decisions of AI-based medical products, radiologists, experts is extremely high (Cohen’s kappa >0.99).
CONCLUSIONS
The possibility of autonomous sorting of mass preventive studies’ results has been proved. A new model of organization of medical care on the basis of autonomous sorting of preventive imaging studies’ results by medical product based on artificial intelligence is proposed.
