Probabilistic model of age-related macular degeneration staging

Intelligent systems significantly reduce routine work for doctors and improve diagnostic accuracy in medical diagnostics. However, the algorithms' effectiveness may be hindered by the heterogeneity of medical data and intermediate phases of a patient's clinical presentation. Additionally, not all intelligent systems can provide sufficient interpretability to analyze the clinical correlation between the system's responses and the actual clinical picture, thereby limiting their implementation in real medical practice. This article presents a solution for developing an intelligent algorithm to stage age-related macular degeneration of the fundus, a socially significant disease, including the challenging-to-detect intermediate stage, and an approach to visualize the algorithm's work in predicting disease stages. The algorithm's performance, developed based on a multimodal approach, was compared with deep neural networks and other multimodal approaches with commonly used data fusion algorithms. The comparison results demonstrate the elimination of the class imbalance effect due to the difficulty of detecting the intermediate disease stage, resulting in a reduction in the dynamic range of specificity and sensitivity values for all classes to no more than 5% for the developed approach. Moreover, the accuracy for the intermediate stage increased by 20% compared to existing commonly accepted data fusion algorithms.