Artificial Inteligent for Diagnosing Drug-Resistant Tuberculosis

Overview

Title: Artificial Neural Network as Diagnostic Tools For Rifampicin-Resistant Tuberculosis In Indonesia. A Predictive Model Study and Economic Evaluation. Background: Drug-resistant tuberculosis has become a global threat particularly in Indonesia. The need to increase detection, followed by appropriate treatment is a concern in dealing with these cases. The rapid molecular test (specifically for detecting rifampicin-resistant) is now being utilized in health care service, particularly at primary care level with some challenges including the lack of quality control (including how to obtained and treat the specimen properly prior to the examination) which then, affect the reliability of the results. Drug-Susceptibility Test (DST) is still, the gold standard in diagnosing drug-resistant tuberculosis but this procedure is time-consuming and costly. The artificial intelligent including data exploration and modeling is a promising method to classify potential drug-resistant cases based on the association of several factors. Objective : 1. To develop a model using an artificial intelligence approach that is able to classify the possibility of rifampicin-resistant tuberculosis. 2. To assess the diagnostic ability and the accuracy of the model in comparison to existing rapid test and the gold standard 3. To evaluate the cost-effectiveness evaluation of Artificial Neural Network model in Web-Based Application in comparison with the standard diagnostic tools Methodology 1. A cross-sectional study involving all suspected drug-resistant tuberculosis cases that being referred to the study center to undergo rapid molecular test and DST test over the past 5 years. 2. A comprehensive, retrospective medical records assessment and tuberculosis individual report will be performed to obtain a variable of interest. 3. Questionnaire assessment for confirmation of insufficient information. 4. Model Building through machine learning and deep learning procedure 5. Model Validation and testing using training data set and data from the different study center Hypothesis : Artificial Intelligent Model will yield a similar or superior result of diagnostic ability compare the Rapid Molecular Test according to the Drug-Susceptibility Test. (Superiority Trial)

PROCEDURE 1. Under the permission granted by the study centers, the team will obtain the medical records of all eligible cases within the past 5 years 2. The investigators then collect the information of interest variable/parameter which obtained by history taking and further examinations and also medical Billing and Hospital pay per service. For participants with Health Insurance, the direct spending for treatment will be based on INA-CBGs (case-based group) payment. This data then will be recorded in an electronic database. Parameter for model development : Host-based : 1. Presence of Diabetes Mellitus (Including years of being diagnosed, HbA1c Before DST examination and treatment, medication either insulin or oral anti-diabetic) 2. Presence of HIV ((Including years of being diagnosed, CD4 level Before DST examination and treatment, and anti-retroviral medication) 3. Tobacco cessation (Brinkman Index) 4. Alcohol consumption 5. History of Immunosuppressant use (steroid) 6. Presence of other diseases (cancer, stroke, cardiovascular disease) 7. History of drug abuse 8. History of adverse drug reaction during tuberculosis treatment 9. Adherence of previous tuberculosis therapy 10. Presence of COPD 11. Body Mass Index Environment 1. History of Contact with Tuberculosis Patients 2. Healthy Index of Living Environment (Household crowds) Agent 1. Level of Bacterial Smear Before DST 2. Extension of Lesion in Chest X-Ray 3. Presence of Cavitation Sociodemographic Factors 1. Age 2. Gender 3. Education 4. Income Level 5. Health Insurance 6. Marital Status 7. Employment Status 3. For incomplete information, a confirmation to the health center that was referring the cases will be done using the Tuberculosis Registration or questionnaire. 4. The model building will be done using an Artificial Intelligent Model in R. A selected model is an Artificial Neural Network either using Radial Base Function or multi-layer perceptron. Several important procedures including : 1. Determine Significant Parameter 2. Dealing with Insufficient and Imbalanced data class (over or under-sampling) 3. Normalization (Batch, Min-Max) 4. Layer and design 5. Training and test distribution (70:30) 6. Model Selection 5. External Validation will be done to the appointed study center. Precision: (true positive + True Negative)/All cases 6. The Incremental Cost-Effectiveness Ratio Simulation will be done, comparing the best model versus the gold standard and GeneXpert yielding a saving per unit of effectiveness