Auscultation, as one of the main patient examination methods, is used in daily clinical practice. Even though it is a relatively simple diagnostic tool used by all physicians (and in specific situations by nurses as well), the assessment of auscultation itself remains quite subjective, depending on the individual characteristics of the examiner, and requires attentiveness and specific skills that diminish over time. This also significantly complicates the design, development and practical use of electronic stethoscopes. The practical application of artificial intelligence techniques would facilitate the initial assessment of auscultations, complementing the clinical work of doctors and nurses with a tool that is easily accessible, objective, continuously improved and easy to use. The main goal of the project is to identify pathological lung sounds, applying artificial intelligence techniques to create a prototype medical decision support system. Lung auscultation procedure is performed according to a standard examination methodology with a 3M Littmann 3200 type (IEC60601-1-2) e-stethoscope, with the recordings stored in a database developed for this study. These records will be evaluated by a team of medical specialists assembled for this study, and the sounds will be annotated according to the pathologies recorded by the instrumental examinations (chest x-ray, spirograms). Once the initial database is created and the audio recordings are pre-processed, they are transformed into the chosen representation (feature vectors, either feature-engineered manually or automatically extracted using deep learning models) for machine learning-based pathology detection and pathology-type identification tasks. In addition, the experiments will aim at evaluating the feasibility of fine-tuning existing models using known deep learning architectures. This could potentially lead to even higher accuracy with a considerably small data sample, at the same time simplifying and reducing the need for time and resource intensive data collection and manual annotation for further improvements of an already developed model. The summarized results for the automated diagnosis of lung auscultations using artificial intelligence techniques, together with the recommendations and insights for further research, will be published in a scientific paper and a conference report.
Project funding:
KTU Research and Innovation Fund
Project results:
–
Period of project implementation: 2023-04-11 - 2023-12-31
Project partners: Lithuanian University of Health Sciences
