DEVELOPMENT OF M-HEALTH APPLICATION WITH SEVEN-SEGMENT DISPLAY READING FEATURE FROM GLUCOSE METER

• Main Author: Romyz Aufa, Daffa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87584
• Institution: Institut Teknologi Bandung

M-Health application is useful for monitoring blood glucose conditions for diabetics. The entry of blood glucose values from a glucose meter into the application is generally done manually. This process is time-consuming and prone to error. Therefore, developing a prototype m-health application that uses a model to read blood glucose measurement results can be a solution. This paper proposes the development of an m-health (mobile health) application that has 4 features: taking a picture of the glucose meter with a phone camera, reading the glucose value with a model, storing the reading results in a database, and data visualization. The model was trained using 3764 images of blood glucose measurement results from glucose meters. The model ultilized is YOLO11 with small, medium, and large variations. In order to be used in the application, the trained model is converted into TensorFlow Lite format (.tflite). The TensorFlow Lite model is then quantized to 16-bit float precision (FP16) and 8-bit integer precision (INT8) to reduce inference time and model size. Based on the test results, the model chosen that is implemented in the application is the small model with INT8 precision. The model was chosen because it has a small inference time and file size while having an accuracy that are not too far from other model variants. The model has an accuracy of 94.14%, an f1-score of 97.03%, an inference time of 330.7 ms, and a file size of 11.4 MB. Model is tested on the m-health application with 156 images from the test dataset resulted in an accuracy of 97.77%, an f1-score of 98.34%, and an average inference time of 1918.15 ms.