Study On Clinical Decision Support Methods Based On Multi-label Learning

With the rapid accumulation of clinical data in recent years,clinical decision support research based on retrospective data has become a significant focus in the field of medical informatics.However,most of the current decision support research focuses on constructing a targeted model for a specific disease;these studies often ignore the widely existing disease relations in clinical practice.A few studies have utilized disease relations by following a multi-label learning scheme;however,the existing generally-purposed algorithms are insufficient to fulfill the specific requirements of clinical decision support.Therefore,it is of great significance to innovate multi-label learning methods targeted at decision support tasks and enhance their ability to utilize disease relations,to improve the rationality,completeness,and clinical application capabilities of clinical decision support research.To achieve such a goal,this paper first proposes a novel multi-label classification algorithm framework.Innovative clinical decision support methods for both clinical static and short sequential data are then constructed referring to the actual clinical needs.The framework and methods enhance their overall ability to utilize the latent information in disease relation in clinical data,providing innovative concepts and theoretical-methodological support for related research in this field.The contents and innovations of this paper mainly include:Propose a multi-label classification algorithm framework based on the concept of label correction.By focusing on the problem of error propagation occurring in stacking-structured multi-label learning methods,two error propagation-related assumptions and delicate output determination methods are proposed,reducing the harmful effects caused by error propagation.An algorithm framework with label correction as the core concept is then introduced.Under the framework,a generalized exemplary algorithm is designed,verifying the universal effectiveness of the framework in multiple fields.Propose a multi-label auxiliary diagnostic method with clinical static data.The method is designed for the demand of risk probability output in the clinical decision support field.It enhances multi-label methods’ability to utilize disease relations by using an innovative possibility correction procedure,with improvements between 0.5%to 3.0%on the index of AUC_PR in the auxiliary diagnostic task for five chronic diseases.The proposed method can be universally utilized in multi-disease auxiliary diagnosis tasks with clinical static data,providing a powerful methodological tool for research under the most common static clinical decision support research paradigm.Propose a multi-label method for the early detection of chronic diseases based on clinical short-sequential data.This method focuses on the specific clinical decision support problem of chronic disease early detection,innovates multiple domain knowledge introducing techniques,and fully utilizes the clinical multi-label short-sequential data.The method significantly improves the ability of chronic disease early detection,with early detection rates of five chronic diseases ranging from 41.8%to82.9%.The study achieves the problem-oriented optimization of the multi-label learning method,providing a case and approach support for specific clinical decision support research based on multi-label learning.This paper studies the clinical decision support methods based on multi-label learning from different perspectives,such as the theoretical algorithm framework,the field requirements-based universal multi-label learning method,and the specific problem-oriented optimization techniques.From different clinical decision support research levels,the outcomes of this paper achieve innovations in methods and enhancements in performances.In general,this study well expands new approaches and theories for the studies in the clinical decision support field.