Clinical Research and Promotion of Intelligent Early Identification and Auxiliary Diagnosis System for Pediatric Tuberculosis in Shannan City, Tibet

Period

    2022.1-2024.12

Cooperative unit

    The First Affiliated Hospital of USTC

Introduction

    Tuberculosis is a contagious infection caused by bacteria that primarily affects the lungs, but can also invade various organs throughout the body. Pediatric tuberculosis has a high rate of misdiagnosis due to its low bacillus nature and difficulty in obtaining high-quality respiratory specimens. The mortality rate is nearly 50% worldwide and 62% in China. Tibet is a high-incidence region for tuberculosis, and due to the unique ecological environment of the plateau, certain tuberculosis-susceptible genes may have evolved and may exhibit different clinical characteristics, which makes diagnosis more difficult. In addition, the strength of primary hospitals is weak and primary doctors have a lack of understanding of the disease, so providing an effective diagnostic and screening method for pediatric tuberculosis in Tibet is a major challenge in the prevention and control of the disease.

    Based on the research of the Pediatric tuberculosis Questionnaire, this project aims to use artificial intelligence technology to build an intelligent early identification and auxiliary diagnosis system for pediatric tuberculosis and evaluate its clinical effectiveness, and analyze relevant clinical test data to conduct in-depth molecular diagnostics of the diagnostic value of pediatric tuberculosis biology.

Methodology

    Research on early screening methods for pediatric tuberculosis using deep learning

    A classification model will be trained based on patient data from four primary care hospitals in Shan Nan, and a second model will be trained based on diagnostic data from the University of Science and Technology of China. The two models will be fused using federated learning to generate a global model while ensuring that patient data from the two hospitals is kept private and not shared. To use the trained model, patients will be required to fill out a questionnaire, and the model will output a risk-level assessment based on the extracted features. The diagnosis will be confirmed by a doctor through clinical testing.




    Optimization method of pediatric tuberculosis questionnaire based on deep learning

    Our proposed method for optimizing a questionnaire for pediatric tuberculosis based on deep learning consists of seven steps. First, we determine the target population for screening. Next, we distribute and collect the questionnaire. Then, we extract important feature information from the collected questionnaire and input it into a trained model. The model then outputs feedback on each feature vector. For negative feedback, we remove the corresponding question from the questionnaire to optimize the screening questionnaire. At the same time, the model outputs a diagnosis for the patient, which is verified by a doctor and added to the initial dataset to expand the dataset and further improve the generalization ability of the model.