Multi-modal and multi-criteria conflict analysis model based on deep learning and dominance-based rough sets: Application to clinical non-parallel decision problems

The non-parallel disease progression and curative effect are the difficulties of clinical diagnosis and treatment decisions. Experts (doctors) constantly summarize these non-parallel phenomena for more accurate diagnosis and treatment. In order to discover the mechanism of clinical non-parallel decision-making, this paper constructs a multi-modal and multi-criteria conflict analysis method based on deep learning (DL) and dominance-based rough sets (DRSA). First, for multi-modal attribute information, we adopted a deep learning based visual attention distribution to focus on the priority areas of images, a deep residual network is used for a feature extractor. The dominant characteristics of the attributes are considered, and the dominant similarity relationship based on cosine similarity is constructed using DRSA. Second, conditional attributes are used to classify objects and predict clinical progression (outcome). At the same time, the objects are classified according to decision attributes based on DRSA. Third, the Pawlak conflict analysis is introduced to analyze the consistency between the predicted results of conditional attributes and the practical results generated by decision attributes. Finally, four clinically non-parallel decision datasets are used, including colorectal cancer (CRC), membranous nephropathy (MN), rheumatoid arthritis (RA) diagnosis and MN efficacy evaluation, to verify the applicability and validity of the proposed model and discover the non-parallel decision mechanism of different diseases. This paper constructs a data-driven clinical decision research paradigm, and provides a research approach to a wide range of non-parallel decision-making problems.