RP16 - Possible uses of AI to evaluate (AI-based) information at the Point of Care
Increasing digitalization in medicine enables ever better and more individualized treatment of patients supported by technically processed information extraction and knowledge representation. This can involve patient-specific data such as the patient’s personal situation, family history, clinical data, molecular data (omics) and image data, as well as patient-independent knowledge sources such as guidelines, standard operating procedures (SOPs), studies and scientific literature. In addition, the doctors’ individual clinical experience and the treatment of previous patients also play a role. Previously, a dashboard for the Point of Care was developed that integrates the existing processed data. The dashboard was designed and evaluated in collaboration with doctors. The available information is offered to the treating physicians in the dashboard, regardless of whether it contradicts or reinforces each other or whether there is a completely unusual information situation.
The aim of this research project is to offer physicians further support by making the relevance of the information displayed in relation to the diagnosis, and further treatment recognizable, and by identifying other aspects such as contradictions or reinforcements. In addition, the aim is to investigate whether the support of AI-based diagnoses, treatment strategies and prognoses can be improved.
The research project is essentially divided into three sub-projects, each of which is aimed at one research question.
Sub-project: Relevance
Research question: What relevance does the available information have for diagnosis, status determination and treatment? In this sub-project, labels are to be recorded for a regession procedure: The recording is to be tool-based. The data for labeling should be collected both automatically, based on the specific use of the dashboard, and explicitly through a specific assessment by doctors. The labels will be used to implement a regression model, which will be continuously updated through online learning during operation. Finally, as a proof-of-principle a visual implementation of the evaluation of the dashboard components by the model is to be implemented and evaluated within the dashboard.
Sub-project: Diagnosis and status determination
Research question: How well can a machine diagnosis and status determination be made on the basis of very heterogeneous data? Using a classification procedure with soft-max output, a diagnosis and status determination for the current case is to be carried out. The main challenge is to prevent an unbalanced ratio of input data in the model. Particularly with regard to the image data, a feature reduction step must probably be integrated, for example using a U-Net-based encoder-decoder structure, whereby it must be evaluated how many pooling layers are necessary and feasible for the reduction. With regard to explainability, in addition to the actual classification the explanation of which input data has what influence on the result should be part of the output. An additional gain in knowledge can be the comparison of the information used by the doctors and the model in order to understand any deviations with respect to the evaluation. A user study is carried out in order to pass on information to doctors in a targeted manner. The information presented should match the doctor’s experience and potential contradictions and new findings should be highlighted.
Sub-project: Treatment
Research question: What quality can machine suggestions for treatment achieve? In this subproject, a machine model and a software algorithm will be developed to suggest meaningful treatments. The integration of an online learning process makes it possible to reduce the influence on the treatment prognosis based on old data by updating the model and to dynamically incorporate new treatment methods. The results of this sub-project will also be explained and compared with similar cases.