Tesis (Ingeniero Biomédico)

Tuberculosis (TB) is an infectious disease caused by Mycobacterium tuberculosis, which can affect any organ of the body, being pulmonary TB the most frequent form of the disease and the one that causes the most deaths. According to the World Health Organization (WHO), TB is among the 10 leading causes of death worldwide, and in the case of Colombia, TB is a disease of public health interest, due to the high number of cases reported in the territory, compared to other communicable diseases.

One of the main problems for TB management lies in diagnostic methods, which require personnel and infrastructure that are not always available in places with deficient health systems. According to the national protocol for TB detection, the diagnosis of pulmonary TB should be made by microbiological confirmation, for which there are three types of tests: smear microscopy, molecular tests and cultures. All tests have an associated cost and their availability is limited, so the generation of tools that provide support in the diagnosis of TB can help to have a better control of the disease.

Artificial intelligence (AI) is an area of computer science that seeks to endow machines with intelligent behaviors in order to perform a specific task. One of the applications of AI are Decision Support Systems (DSS), these systems applied in healthcare, seek to generate models based on large volumes of data and previous clinical knowledge, to help the physician in making better decisions regarding patients.

In order to generate tools that help in the management of TB, this work uses AI techniques for the development of a DSS that supports TB diagnosis, using the information contained in electronic medical records (EMR). EHRs are sources of information widely used by physicians, in which the health status of patients is recorded, so it is expected that with the information contained in them, a computational tool can be generated to help health professionals in the management of TB.

For the development of the work, a database was built from 151 EHRs of patients suspected of pulmonary TB. The database contains the clinical reports of the patients on dates prior to the diagnostic tests, so that the reports do not contain information on the final diagnosis of TB. For the creation of the diagnostic tool, clinical reports were taken and preprocessed to clean the text, then, features were extracted using 2 methods TF-IDF (term-frequency - inverse document frequency) and Word2Vec; subsequently, machine learning models were used to make TB prediction. Model exploration was performed by cross-validation, finding that the best results are obtained by performing dimensionality reduction of the features obtained with TF-IDF, and using the random tree algorithm for classification. The performance metrics obtained on the test sets with this model are: 0.721, 0.802, 0.462, and 0.723, in accuracy, sensitivity, specificity, and F1-score respectively.

This work was developed within the project Generation of alternative models based on computational intelligence for screening and diagnosis of pulmonary tuberculosis (minciencias, Universidad del Rosario, Universidad Antonio Nariño, Subred Integrada de Servicios de Salud Centro-Oriente-Hospital Santa Clara), which is a project formed by a joint team of physicians and engineers, and aims to generate computational tools that can be used in places with poor infrastructure for the diagnosis of pulmonary TB. Within the project, computational models are being developed using clinical, epidemiological and sociodemographic variables, and it is expected that in the future this work will be integrated with other strategies generated within the project, for the construction of a more robust system that can support the physician in the diagnosis of pulmonary TB.