Spiroergometry (SE) is an important diagnostic procedure for objectively assessing physical performance in terms of respiration, cardiovascular function and metabolism, as well as motivation and ‘mental’ performance. It is highly informative in the assessment of patients with diseases of the lungs, heart and musculoskeletal system, as well as in evaluating the performance of athletes. For example, SE is used in the planning of lung surgery or heart and lung transplants. To this end, numerous circulatory and respiratory parameters are measured during physical exertion (usually on a bicycle ergometer). From these values, other parameters such as maximum oxygen uptake or respiratory equivalents for oxygen and carbon dioxide – i.e. ultimately measures of respiratory efficiency – can be derived. The use of spiroergometry is enshrined in numerous guidelines; however, due to the considerable time required and the complexity of the analysis, the test is rarely carried out, meaning that the potential of spiroergometry is far from being fully utilised.
The procedure is relatively straightforward to carry out, but complex to interpret. During interpretation, individual parameters and the correlations between these parameters over time are analysed. Depending on the research question, various deviations from these patterns are identified. In addition, graphical representations of the plotted parameters are interpreted. High-level interpretation is based on extensive specialist knowledge combined with considerable experience. An automated evaluation of SE investigations using an AI model is therefore a very useful aid in practice and is the objective of the SPIRAKI project.
A neuro-symbolic AI – a hybrid of deep learning and logic-based AI – is being developed for the algorithmic implementation. This cutting-edge approach combines the performance of neural networks with the explainability of expert systems. Using deep learning, parameters are derived from the raw data of cardiovascular and respiratory function measurements; these serve as input for the downstream logic-based AI. This system can integrate the complex diagnostic guidelines directly into the AI model as expert knowledge in a user-friendly manner, thereby utilising medical expertise. To this end, SPIRAKI uses a probabilistic logic language that allows both the outputs of the neural networks to be processed in detail and, at the same time, a probability to be assigned to the diagnoses. As logic-based AI is, by its very nature, explainable, its use makes AI-based diagnoses comprehensible and allows doctors to understand them. This fosters medical confidence in and acceptance of the AI application. The final model can be used as a decision support system for ergometry and efficiently assist the doctor conducting the test in making a diagnosis. This makes it easier and safer to carry out examinations, as the software can also detect spiroergometry tests that have been performed incorrectly. SPIRAKI thus enables a significantly broader and simpler use of spiroergometry than before. Thanks to the simplification and time savings, spiroergometry becomes accessible and attractive as a baseline examination for a wide range of patient groups and can therefore reach a broad section of the population. Depending on the examiner’s experience, the analysis typically takes between 20 and 60 minutes. SPIRAKI reduces this time to just a few minutes. This makes it possible to treat large groups of patients with common symptom clusters (shortness of breath, reduced exercise tolerance) and numerous widespread conditions such as heart failure or chronic obstructive pulmonary disease (COPD).
Partners
Universität des Saarlandes, Lehrstuhl für Pneumologie, Allergologie, Beatmungs- und Umweltmedizin
