Mathematical model-based clinical decision support system algorithm design study that can support the diagnosis of celiac disease

Background: Celiac disease (CD) is one of autoimmune disorder. It is caused by consumption of gluten which is a protein found in barley, rye and wheat. Individuals with CD experience an immunity response due to damage to the lining of their small intestines after consuming gluten, causing them to experience a variety of sign and symptoms also potential long-term complications. Currently definitive diagnosis of CD is typically made by small bowel biopsy, which determines the extent of intestinal damage. Diagnosing celiac disease is difficult because some of its signs and symptoms resemble with other autoimmune diseases. Early diagnosis prevents severe progression of the disease by preventing rapid increase in intestinal permeability. In order to overcome the difficulties experienced in the diagnosis of CD and to enable patients to reach a specialist as soon as possible, it is becoming increasingly important to develop systems that can support diagnosis, with the help of developing technology. While research focuses on intestinal epithelium deformation in the field of video capsule image processing, the effect of gluten sensitivity on facial appearance is still an open research topic in the literature. When looking at the development of celiac disease in the body, it is associated with diseases that impact the skin through intestinal inflammation, immune system problems and microbiome dysbiosis. In addition, CD causes changes in skin structure and facial features because it affects the gastro-intestinal system and can cause malabsorption of nutrients. This holds promise for supporting the diagnosis of CD through face image classification by detecting observable facial features or some patterns. Although there are studies in the literature that aim to support early diagnosis using data on disease history and symptoms, there are no hybrid models that use facial images. Based on this, this thesis study was planned under two different research questions. The first research question focused on the changes on the CD face images and asked, "How effective are the transfer learning models VGG16 and MobileNetV2 in classification CD facial images to aid in the diagnosis, and do they support the early diagnosis of CD with significant performance?" It is structured in the format. The second research question asked: "Can CD Marsh (0-4) levels be determined with machine learning models using routine blood test results and demographic data performed in primary health centers?" focused on and compared the performance of deep learning models (PyTorch) and traditional machine learning algorithms. With both research questions, it is aimed to present a new perspective on the literature. Since such a data set does not currently exist, data was collected from patients diagnosed with CD followed in the gastroenterology clinic at Van Training and Research Hospital, Turkey, within the scope of ethical permission 2022/09-02. Purpose: It is obvious that it is important to examine the feasibility of designing a system that aims to ensure early diagnosis of CD, which has similar symptoms to many autoimmune diseases, and to ensure that patients are referred to a specialist as soon as possible, and to examine its effective use in the clinic. Based on this, this thesis first aims to question the ability of transfer learning with two pretrained models to classify the database containing facial images to show meaningful performance in the early diagnosis of CD. Secondly, it is aimed to prevent delays in referring celiac disease to a specialist in primary health care units due to its similar symptomatic features with other diseases. For this purpose, it is aimed to analyze a method that will support the diagnosis of the disease and the determination of Marsh levels by classifying routine blood test and demographic data with Decision Tree, Random Forest, Gradient Boosting, Naive Bayes and deep learning (Pytorch) algorithms. It is also aimed to analyze a method that will support the determination of the celiac level severity. Method: In the thesis study, two different data sets were used and interpreted by applying two methods suitable for the purpose. For the first research question, facial images of celiac patients were classified using the transfer learning method. Two different pre-trained models (VGG16 and MobileNetV2) were used. Then model with the highest classification performance was selected. For the second research question, with the aim of determining intestinal atrophy (marsh level) in celiac patients, clinical signs and symptoms were classified with deep learning (Pytorch) and traditional machine learning algorithms, and the model with the highest accuracy rate was determined. Models with high performance criteria in both stages were selected for the development of models in which hybrid data sets could be used in the future. In the dataset containing face images (DS_Face), there was a CD diagnosis and an equal number of control groups. There are 200 adult facial images. While 70% of the images diagnosed with CD in the DS_Face dataset is female and 30% are male, the distribution of the control group is 60% female and 40% male. Of the patients in the facial images diagnosed with CD, 28 were newly diagnosed and 72 were previously diagnosed. Of the previously diagnosed patients, 25 did not comply with the gluten-free diet. 47 of them are partially adhered to the diet. The second data set (DS_Marsh) was created to determine the marsh level based on routine blood test and demographic information; It consists of patient history, demographic information and blood test results of 182 people. DS_Marsh dataset; It has a distribution of 132 female and 50 male patients. In this data set, the distribution of compliance with the gluten-free diet is as follows: 72 people who do not comply and 106 people who partially comply. Results: In the DS_Face dataset, an accuracy rate of 73% was reached in the classifying of facial images of non-compliant, partially adherent, newly diagnosed patients and the control group using the pre-trained VGG16 model. Precision, F1 and recall score values were found to be 0.54, 0.52 and 0.56, respectively. Moreover, the pre-trained MobileNet.v2 model used by the Medipipe model marker had a test loss of approximately 0.44 and an accuracy rate of 85% on the test dataset with 0.80 precision, 0.81 recall and 0.80 F1 scores. As a result, in the DS_Face dataset, the MobileNet.v2 model showed a stronger performance in CD diagnosis compared to the VGG16 model. On the DS_Marsh dataset, the PyTorch model performed with 80% accuracy in determining Marsh levels. Recall, precision, and F1 scores were measured as 0.80, 0.81 and 0.70, respectively. Recall, precision, and F1 cores were achieved as 1.0, 1.0, 1.0 with the Decision Tree, Random Forest and Gradient Boosting classifiers, each of which showed full performance, 100% accuracy rate and classification capacity. The recall, precision, and F1 scores of the Naive Bayes classifier, which shows a 55% accuracy rate compared to others, are listed as 0.44, 0.67 and 0.53, respectively. Conclusion: The performances of the VGG16 Model and Mobilenet.V2 Model were compared in the DS_Face data set created to examine the first research question. Model VGG16 correctly classified almost one third of the test images. Mobilenet.V2 Model classified 85% of the test images correctly. The VGG16 model's classification performance was lower than the MobilenetV2 model due to its architectural structure and some other features mentioned below. VGG16 is a deeper network with more parameters due to its model architecture and more suitable for big dataset. MobileNetV2, on the other hand, is designed to be more efficient on small data sets with fewer parameters. This caused MobilenetV2 to perform better on the DS_Face dataset. Additionally, since MobileNetV2 was used with the Mediapipe model marker, the mediapipe model marker provided additional preprocessing and feature extraction that improved the performance of the model and contributed to higher accuracy. And finally, VGG16, a more complex model, may have overfitted, especially if the dataset is not large enough. In conclusion, the performance of facial classification models such as Mobilenet.V2 on the changes that occur on the skin and facial features of patients newly diagnosed with CD and previously diagnosed but not complying with the gluten diet provides concrete findings regarding the development of decision support models capable of supporting early diagnosis in these patients. The models used for classification in the DS_Marsh dataset created to examine the second research question, respectively, showed over 80% accuracy in Pytorch, while Decision Tree, Random Forest and Gradient Boosting showed 100% performance. System designs that will allow these routine blood test findings and demographic information to be used even in primary health care units are promising due to their contribution to the quality of life of patients, the patient load in the healthcare system, and the economy. This study, planned with two research questions to improve the patient' life quality and reduce the burden of healthcare services and expenses in the economy, emphasizes the importance of data-based methods in medical diagnoses to improve decision-making and benefit patient outcomes. In future studies, it has been concluded that the use of the Mediapipe model marker with pre-trained MobileNetV2 model and Decision Tree, Random Forest and Gradient Boosting classifiers will provide high performance in hybrid system designs that will accelerate referral to a specialist for early diagnosis by using both facial images, routine blood findings and demographic information. Keywords: Diagnosis of Celiac Disease, Face Image classification, Deep and machine learning, Transfer learning, Mediapipe (MobileNetV2),and VGG16