Cytokines Analysis Combined with Machine Learning for Determining the Etiology of Inaccessible Infection

Date
2025-05-10
Authors
Amal Mazen Aref Dofesh
امل مازن عارف دوفش
Journal Title
Journal ISSN
Volume Title
Publisher
Al-Quds University
Abstract
Accurate diagnosis of viral and bacterial infections in patients with fever is crucial in clinical diagnosis to ensure appropriate treatment. It requires obtaining a high-quality sample from the patient for laboratory testing. In some cases, obtaining a sample from the patient is difficult, especially in cases of inaccessible infections which may require invasive procedures. Routine laboratory tests, such as complete blood cell (CBC) count, C-reactive protein (CRP), etc. lack the high sensitivity to identify the primary cause of the infection associated with fever. This can lead to misdiagnosis in some cases which can impact patients’ health. The goal of this study is to combine cytokine results with machine learning models to develop a unique diagnostic tool that can separate and differentiate between types of infections. We focused on three cytokine indicators (IP-10, IL-6, and TRAIL) and combined them with advanced machine learning techniques such as Principal Component Analysis, Linear Discriminant Analysis, and Support Vector Machine (SVM). Our study included 268 pediatric patients, 245 of them met the inclusion criteria. Blood samples were drawn from the participants after obtaining ethical approval. 97 samples were selected for testing, based on our capabilities. 22 cases were confirmed as having a bacterial infection in the diagnostic microbiological laboratory, 55 were diagnosed with a viral infection (using PCR), 10 were diagnosed with fever of unknown origin, and 10 were diagnosed with an inaccessible infection. 87 controls participated in the study. Cytokine tests were performed using ELISA technique. We initially found the cut-off values for cytokine measurements, but the cut-off values failed to differentiate between bacterial and viral samples due to the high overlap between the results. This finding prompted us to move from using cytokine biomarkers alone to correlating cytokine values with more advanced techniques. PCA and LDA demonstrated the ability of cytokines to separate different infection groups, achieving success rates of up to 90%. However, when WBC and CRP variables were added to the cytokine biomarkers, our results showed lower success rates, dropping to 65%, and the accuracy of the model was reduced. Despite the high success rates of the PCA and LDA models, our data was abnormally distributed. For this reason, we turned to use a more advanced classification technique based on machine learning algorithms. The results of the three cytokines were integrated with the Support Vector Machine (SVM) technique, showing a sensitivity of 87.5% and a specificity of 73.6%, with an AUC of 0.94 for classifying viral infections from bacterial infections. In conclusion, our study demonstrates that using cytokines combined with machine learning algorithms is a faster, newer, and less painful tool for detecting the cause of infection. Thus, directly treating the patient with the correct treatment, as well as improving clinical diagnosis and reducing the excessive consumption of unnecessary antibiotics, which reduces the spread of antibiotic-resistant strains of bacteria and promote effective antibiotic stewardship.
يُعد التمييز الدقيق بين العدوى الفيروسية والبكتيرية لدى المرضى الذين يعانون من الحمى أمراً حاسماً لضمان تقديم العلاج المناسب وتفادي المضاعفات الناتجة عن التشخيص الخاطئ. تعاني الطرق التقليدية، مثل فحص تعداد كريات الدم البيضاء وCRP، من انخفاض الحساسية والدقة، بينما تتطلب طرق مثل PCR الحصول على عينات عالية الجودة، وقد تكون غير ممكنة في بعض الحالات، مثل التهابات الأعضاء الداخلية. استجابة الجسم المناعية للعدوى تختلف حسب نوع المسبب؛ إذ تُفرز السيتوكينات بأنماط مميزة في العدوى الفيروسية مقارنة بالبكتيرية. في هذا السياق، تهدف دراستنا إلى استخدام ثلاثة مؤشرات سيتوكينية (IP-10, IL-6, TRAIL) وربطها بخوارزميات تعلم آلي مثل PCA، LDA، وSVM، بهدف تحسين دقة التشخيص وتقليل الحاجة للإجراءات الجراحية. شملت الدراسة 268 طفلاً، تم تحليل عينات من 97 منهم باستخدام ELISA. أظهرت العدوى البكتيرية ارتفاعًا ملحوظًا في IL-6 وCRP، بينما تميزت العدوى الفيروسية بارتفاع IP-10 وTRAIL. لم تكن قيم الـcut-off كافية للتمييز بين النوعين بسبب التداخل، لكن استخدام خوارزميات PCA وLDA حقق دقة تصنيف تصل إلى 90%. عند استخدام نموذج SVM، ارتفعت حساسية التشخيص إلى 87.5% وخصوصية 73.6%، مع AUC بلغ 0.94. تُظهر نتائجنا أن الجمع بين مؤشرات السيتوكينات وتقنيات الذكاء الاصطناعي يُعد أداة فعّالة، دقيقة، وغير جراحية لتشخيص أنواع العدوى المختلفة، ما يساهم في تحسين رعاية المرضى وتقليل الاستخدام غير الضروري للمضادات الحيوية، وبالتالي الحد من انتشار مقاومة البكتيريا للمضادات.
Description
Keywords
Citation
Dofesh، Amal Mazen. (2025). Cytokines Analysis Combined with Machine Learning for Determining the Etiology of Inaccessible Infection [رسالة ماجستير منشورة، جامعة القدس، فلسطين]. المستودع الرقمي لجامعة القدس.