Appendicitis is a disease that occurs when the internal part of the appendix becomes blocked and requires urgent surgical intervention [1]. It usually presents with symptoms such as abdominal pain, nausea and vomiting, but since these symptoms are similar to many abdominal diseases, diagnosis can be difficult [2], [3]. The diagnosis of appendicitis in children can be more complicated than in adults. Because children have difficulty expressing their complaints and do not cooperate enough during the examination, diagnosis can be delayed [4], [5]. These delays increase the risk of peritonitis, sepsis, perforation or related postoperative complications, especially in pediatric patients, and complicate the treatment process [6], [7]. Therefore, early diagnosis of acute appendicitis is of vital importance for pediatric patients [1]. Early diagnosis can prevent possible complications such as perforation and sepsis and reduce costs by shortening the duration of hospital stay [8], [9].

Various clinical and imaging methods are used in the diagnosis of appendicitis. Physical examination performed by pediatric surgeons contributes significantly to the diagnostic process with high accuracy rates [10]. Ultrasonography (USG) and computerized tomography (CT) are widely used imaging techniques in the diagnostic process. USG is preferred due to its low radiation risk; on the other hand, CT offers higher diagnostic accuracy [11]. The Pediatric Appendicitis Score (PAS), developed based on clinical symptoms, helps in diagnosis and can prevent unnecessary CT use [12]. However, each of these methods has various limitations in terms of sensitivity and specificity. Especially when the entire appendix cannot be visualized, USG may be insufficient in diagnosis due to the risk of missing cases [2], [3].

One of the contemporary approaches used in the diagnosis of diseases is artificial intelligence (AI)-based methods. In recent years, artificial intelligence technologies have achieved remarkable success in different disciplines; especially in the field of medicine, they have become an important tool in diagnosis, disease course (prognosis) and mortality predictions, and in optimizing treatment processes [13], [14]. One of the main reasons for this situation is the capacity of machine learning (ML) algorithms, a sub-branch of artificial intelligence, to reveal hidden patterns in complex data sets [15], [16]. In this way, faster, more reliable and cost-effective solutions can be provided for diseases with similar clinical symptoms, difficult to diagnose and high uncertainty levels [17], [18], [19], [20].

Today, blood, urine and other laboratory samples are analyzed using advanced learning algorithms supported by artificial intelligence; thus, it is possible to achieve predictive health outcomes. In this context, machine learning (ML)-based approaches are increasingly applied and rapidly developed in the field of medicine due to their potential to develop real-time decision support systems in healthcare, reduce drug costs, shorten hospital stays and increase the quality of healthcare [21], [22].

As in this study, it has been reported that routine blood test (RBT) data play a significant role in the cost-effective, reliable, and rapid diagnosis and prognosis of diseases [22]. Therefore, RBT data have been utilized in numerous artificial intelligence studies for the diagnosis and prognosis of various diseases [17], [18], [19], [20]. In this context, various recent clinical studies [21], [22], [23], [24], [25], [26] have emphasized that routine blood tests may offer an effective and cost-efficient alternative for the early detection and prognosis of various diseases, including COVID-19. In a study conducted by Reismann et al. [27], complete blood count, CRP, and ultrasound data obtained from children and adolescents aged 0–17 were used to automatically classify acute appendicitis and complication status through machine learning (ML) and artificial intelligence (AI) algorithms. The model was developed using retrospective data from 590 patients, with only 35% of the data used for training and the remaining 65% for testing. The model achieved 90% accuracy in diagnosing appendicitis (sensitivity 93%, specificity 67%). Although its success in distinguishing complicated appendicitis was limited (51% accuracy), the model is noteworthy for its potential to reduce unnecessary surgical interventions and for demonstrating high performance using routine diagnostic parameters. Gollapalli et al. [28] used a local dataset to distinguish between complicated and uncomplicated appendicitis and to improve the diagnosis of acute appendicitis by applying KNN, decision tree, bagging, and stacking algorithms. The stacking model demonstrated the highest performance, with 97.51% training accuracy, 92.63% test accuracy, 95.29% precision, and a 92.04% F1 score. Explainable Artificial Intelligence (XAI) analyses revealed that the most influential blood features affecting model performance were Neutrophil and WBC count.

Mijwil and Aggarwal [29] compared various machine learning algorithms for the diagnosis of acute appendicitis using a dataset of 625 patients. The Random Forest algorithm achieved the highest performance with an accuracy of 83.75%. The study demonstrated that ML-based models have the potential to improve diagnostic accuracy and reduce unnecessary surgical interventions. An et al. [30] evaluated the effectiveness of AutoML (AutoGluon) and automatic feature engineering (autofeat) methods in the diagnosis of acute appendicitis cases with ambiguous computed tomography (CT) findings. In their study using a dataset of 303 patients, the AutoGluon model built solely with clinical data outperformed traditional methods with an AUROC of 0.785, while the model built with both clinical and CT data outperformed traditional methods with an AUROC of 0.886. The results demonstrate that AutoML-based models have the potential to be used in clinical decision support systems by improving diagnostic accuracy. Kim et al. [31] developed a fully automated 3D convolutional neural network (CNN)-based diagnostic framework for diagnosing and staging appendicitis from contrast-enhanced abdominopelvic CT images. The model, built with the DenseNet169 architecture, achieved 79.5% accuracy and AUC 0.865 for appendicitis diagnosis and 76.1% accuracy and AUC 0.827 for complicated appendicitis discrimination. This study presents a reliable AI tool that can reduce negative appendectomy rates and provide decision support to clinicians.

Shahmoradi et al. [30] developed an artificial neural network-based clinical decision support system (CDSS) for the accurate diagnosis of acute appendicitis. The study demonstrated that the results obtained using the Support Vector Machine (SVM) algorithm were consistent with pathological findings with 91.7% sensitivity, 96.2% specificity, and 95% accuracy, demonstrating that this system may be effective in reducing negative appendectomy rates and improving the clinical decision process. Wei et al. [33] compared nine machine learning models for the early diagnosis of complicated and uncomplicated acute appendicitis in elderly patients and confirmed that the Gradient Boosting Machine (GBM) algorithm performed best (sensitivity: 0.9167, specificity: 0.9739) using SHAP analysis. The study emphasized that the developed Shiny-based prediction tool can assist clinicians in rapid diagnosis and optimal treatment planning. Chan and Yau [34] systematically examined the performance of machine learning models in the diagnosis of acute appendicitis and, in an analysis of 6 existing studies, found that these models provided higher diagnostic accuracy (91.7% sensitivity and 96.2% specificity) compared to traditional scoring systems (such as Alvarado). The study recommends the integration of more comprehensive clinical and biomarkers to increase the diagnostic power of future models. Roshanaei et al. [35] compared four machine learning models to improve the diagnosis of appendicitis in acute abdominal pain and found that Gaussian Naive Bayes performed best (accuracy: 95.03%; sensitivity: 87.18%; specificity: 97.54%).

This study aims to detect pediatric appendicitis by integrating Ensemble ML classifiers with RBT data and develop an applicable decision support system. RBT data, which provides basic biomarkers, plays an important role in the diagnosis, prognosis, and follow-up of various diseases [22], [25], [36] and has been widely used in AI-based medical studies [13], [14], [15], [16], [17], [18].

Selecting an appropriate patient cohort: Determining the patient group and their characteristics to be included in the analysis.

Preliminary screening of RBT variables using logistic regression: Preliminary assessment of significant biomarkers effective in appendicitis.

Training Random Forest, Gradient Boosting, and LightGBM models with and without SMOTE using 10x10-fold stratified cross-validation: Modeling with three different Ensemble Machine Learning algorithms and addressing the imbalanced data problem.

Interpreting features using SHAP analysis: Visualizing the impact of variables to explain model decisions.

Generating all possible arithmetic biomarker combinations of two and three RBT features from the seven RBT features: Deriving new predictive biomarker models

Drawing probability threshold curves for the best biomarker and constructing a three-region decision tree: Developing an intuitive and applicable classification scheme for clinical decision support.