< Terug naar vorige pagina

Publicatie

An Automated Clinical Decision Support System with Better Interpretability

Boek - Dissertatie

A clinical decision support system (CDSS) is a system which analyzes health data to provide data-driven solutions or suggestions which can help clinicians or other health professionals make decisions. There are mainly two types of CDSS, i.e. knowledge-based and non-knowledge-based. Knowledgebased CDSS stores a set of pre-defined rules to make decisions following expert medical knowledge, while non-knowledge-based CDSS uses artificial intelligence, machine learning models, or statistical methods, to train the computer programs from past experiences and discover patterns or rules from clinical data. Although a rapid growth in the use of non-knowledgebased CDSS can be observed recently, they are still facing the challenge of interpretability and clinical relevance, as many machine learning methods are black-box methods, making it difficult to give clinically meaningful interpretation for the results. This is the barrier that must be overcome before the widespread of non-knowledge-based CDSS.This thesis presents a CDSS with better interpretability and clinical relevance, which covers a full process of clinical decision making, including data cleaning, disease risk factors identification, disease diagnosis and prediction, disease sequence analysis, and the system architecture of a CDSS implementation. We proposed an automated data cleaning method which applied the clinical knowledge to clean the electronic health records data collected from daily consultation. This method largely reduced the manual efforts and improved the time efficiency, without a significant loss in accuracy and completeness. Then we proposed an ensemble feature selection framework to identify risk factors for chronic diseases. The heterogeneous-distributed implementation and the new voting strategy to combine the results led to a set of algorithm-identified risk factors with better interpretability and clinical relevance. We also discussed the prognostic models for chronic disease prediction. The imbalanced data structure and its effects on the predictive performance were not sufficiently discussed in previous studies on prognostic models for disease prediction. We presented a resampling method which could help accurately identify the positive cases and suggested the evaluation measures with a penalty of false positive. Moreover, when studying multimorbidity, we took the sequence of development into consideration for the first time. We proposed Markov chain analysis and Weighted Association Rules Mining, which were never used for the topic of multimorbidity before. Finally, we present the system architecture of a CDSS from MIDAS project, which showed a possibility to integrate data from multiple sources and provide different types of analytic tools to support decision-making.In summary, we proposed a CDSS that could provide clinically relevant results with better interpretability for clinicians or other health professionals who were not proficient in machine learning methods. The CDSS could assist them in many steps of clinical decision making, such as risk factor selection, disease diagnosis and prediction, and disease sequence exploration. The CDSS was an integrated solution, and automation and privacy preserving were well considered during the development and implementation.
Jaar van publicatie:2021
Toegankelijkheid:Open