Data mining for precision medicine in clinical decision support systems

Data mining for precision medicine in clinical decision support systems

With the addition of new drugs in the market each year, the number of drugs in drug databases is constantly expanding, posing a problem when prescribing medications for patients, especially elderly patients with multiple chronic diseases who often take a large variety of medications. Besides the issue of polypharmacy, the need to handle the rapid increase in the volume and variety of drugs and the associated information exert further pressure on the healthcare professional to make the right decision at point-of-care. Hence, a robust decision support system will enable users of such systems to make decisions on drug prescription quickly and accurately.

Although there are many systems which predict drug interactions, they are not customised to the medical profile of the patient. The work in this study considers the drugs that the patient is taking and the drugs that the patient is allergic to before deciding if a specific drug is safe to be prescribed. To exploit the vast amount of biomedical corpus available, the system uses data mining methods to evaluate the likelihood of a drug interaction of a drug pair based on the textual description that describes the drug pair. These methods lie within the prediction layer of the conceptual three-layer framework proposed in the thesis. This framework enables drug information to be used in a decision support system which associates with the medical profile of the patient. The other two layers are the knowledge layer and the presentation layer. The knowledge layer comprises information on drug properties from drug databases such as DrugBank. The presentation layer presents the results via a user-friendly interface. This layer also obtains information from the user the drug to be prescribed and the medical profile of patients. Models used in these data mining methods include the network approach and the word embedding approach.

Empirical experiments with these models support the hypothesis that drug interactions are associated with similarities derived from their feature vectors, resulting in the deployment of a decision support system for use in dental clinics. A survey conducted on dentists found positive response in the use of such a system in helping them in drug prescription which result in a better treatment outcome. They found the system useful and easy to use. The novel approach of using information on drug interaction through data mining for use in a personalised decision support system has provided a platform for further research on optimising of drug prescription, transforming the clinical workflow at point-of-care within the healthcare domain.