SYNTHETIC DATA GENERATION AND DATA SCIENCE BASED DEVELOPMENT OF SPECTROSCOPY READINGS TO HBA1C CONVERSION METHODS

• Main Author: Romora Partigor, Nathanael
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/50194
• Institution: Institut Teknologi Bandung

According to the International Diabetes Federation (IDF), total cases of diabetes in Indonesia reached ten million cases in 2017. Total claim from patients suffering diabetes and its complication on BPJS has reached more than 33% of its total budget. Diabetes can be controlled if the patients continually monitor their blood sugar level and consult it to doctors. There are two ways to measure a person’s blood sugar level: blood glucose and HbA1c. Blood glucose test give the instantaneous level of blood sugar from the test subject. Meanwhile, HbA1c test gives a person average blood sugar level from the last two to three months, thus it is recommended to use HbA1c test for longer term monitoring of diabetes patients. The existing problem of this test is that it’s still expensive, invasive, and requires non-mobile equipment. These problems lead to accessibility issues, especially in rural areas, and low percentage of people using the test. This final project focus on making the test less expensive and more portable. The methods we use to measure HbA1c is spectroscopy. This final year project in particular will explain the methods to convert spectroscopy reading to the concentration of HbA1c and its implementation on Android device. To convert the measurements, we could use statistical and machine learning approach. From the authors’s literature study, it’s decided that we will use machine learning as it could accomodate more complex data behaviors. Specifically, the author chose neural network model over linear regression because neural network models could give non-linear models accurately. The study of comparing four most used machine learning framework (CNTK, TensorFlow, PyTorch, and MXNet) leads the author to use TensorFlow for the model building. To answer the problem of portability, the author chose Android over iOS as Android gives much faster prediction time with similar if not better accuracy score. It could be concluded that the following design choices will be used in this projcet: Data Science for the methodology, Machine Learning for the model, TensorFlow to build the model. Due to limitations caused by the COVID-19 pandemic, the input for implementation and testing for this final year project is still using partially simulated synthetic data. Further work need to be done for real patient data realization and testing of the design. The generation of synthetic data was succesful numbering at 95 data pointss. Conversion software was succesfully developed with mean absolute error of 0.055 or equal to 1.099% relative error.