The knowledge representation and algorithm for personalized infectious disease risk prediction

Abstract

Infectious diseases are a major cause of human morbidity. However, in the EU in 2014 more than 40 thousand deaths caused by infectious diseases were considered preventable. Information about infection risk based on personal and environmental attributes, as well as up-to-date infectious disease risk knowledge is expected to make lay people aware of their infection risks. With the emergence of APIs and GPS technology, surrounding location features and weather information can be inferred from a person's position. This offers an opportunity to create a system for personalized infectious disease risk prediction. This thesis presents research towards a system that can predict personalized infectious disease risk (IDR) based on a person's attributes and geo-position by utilizing infectious disease risk knowledge (entitled PROSPECT-IDR: Personalised Prediction of Infectious Disease Risk). A knowledge representation was designed to facilitate epidemiologists to encode infectious disease risk knowledge in a form familiar to them. The generic IDR ontology represents personal and environmental risk factors for all human infectious diseases (n=234). Quantifications for the risk factors (e.g. odds ratios) are encoded using five IDR rule types. This IDR knowledge representation (ontology and rule types) allows encoding of knowledge about risk of infectious diseases prevalent in a region. The IDR ontology can never be complete, as new risk factors for existing diseases, and new diseases, are constantly discovered. The initial generic ontology contains all risk factors found in the Atlas of Human Infectious Diseases, and in factsheets from the CDC and WHO. Each instantiation of knowledge for a specific disease in a region comprises of a subset of risk factors from the generic ontology plus any new risk factors not found there, along with a set of risk quantification rules (instantiations of the five rule types). An algorithm (entitled BN-Builder) converts the knowledge-base into a fully functioning and consistent risk prediction model, a Bayesian Network, which is the core of the PROSPECT-IDR prediction system. The usefulness and completeness of the IDR knowledge representation (initial generic ontology and five rule types) were evaluated using 22 published case-control studies that encode infectious disease risk knowledge. Each case-control study was encoded as one evaluation knowledge-base. With regard to completeness, more than 3/4 of the ontology objects needed to encode the knowledge in the evaluation case-control studies were found in the initial generic ontology. With regard to usefulness, more than 3/5 were used to encode evaluation case-control studies. With regard to completeness and usefulness of the five rule types, all infectious disease risk knowledge in the 22 evaluation case-control studies can be encoded with just those five rule types, and all five rule types were used. To evaluate BN-Builder algorithm, the consistency between the generated BN and the knowledge-base was measured. Chi-square tests for differences were carried out for two evaluation knowledge-bases that covered all functions of the algorithm and all data ranges allowed by the rule types. There was no significant difference between the resulting infectious disease risk prediction and the encoded knowledge (p > .05). Evaluation results suggest that the IDR knowledge representation is useful. Further, statistical findings indicate that the BN-Builder algorithm generates infectious disease risk predictions that are consistent with the encoded risk knowledge. The PROSPECT-IDR system that this IDR-KB and BN-Builder algorithm is designed for is expected to give information about personalized infectious disease risk prediction to lay people. So, the relevant preventive actions can be tailored based on this personalized information, and thus, hopefully will reduce the incidence number of infectious diseases in the world.