Improved Causal Models of Alzheimer's Disease

Abstract

A Causal Bayesian Network (CBN) is a popular analytical framework for causal inference. Currently, there are many methods and algorithms available for analyzing and learning causal networks from various datasets. However, the accuracy and fit of the causal network learned from data relies on the accuracy of the algorithm and the completeness of the data. These models do not consider prior expertise from clinicians or other authoritative sources. Biomedical ontologies represent the collective knowledge of the experts who created them. By codifying and standardizing symptom variables, diagnostic information and other clinical information, the biomedical ontologies are sources of prior causal knowledge. This expert knowledge can be used to orient arcs in a CBN, thereby improving their accuracy. This paper proposes a novel method of collecting prior causal knowledge from International Classification of Diseases Version 10 Clinical Modification (ICD-10-CM), which will be used to orient the arcs of a CBN learned from diagnosed patient data from the National Alzheimer's Coordinating Center's (NACC) Uniform Data Set (UDS). To demonstrate the effectiveness of our method, we will first generate a baseline CBN model from the data using the unmodified Max-Min Hill-Climbing (MMHC) algorithm. Then, we will obtain prior knowledge in the form of ordered variable pairs from ICD-10-CM and use these ordered variable pairs to inform the MMHC algorithm. The resulting modified CBN will then be compared to the baseline, and both models will then be analyzed in the context of existing scientific and epidemiological research of Alzheimer's Disease to show the significant impact of prior ontological knowledge in improving CBNs learned from data.