# Brain tumor detection with real-world predictions in Jordan hospitals

## TL;DR

This study uses machine learning to accurately detect brain tumors in real-world clinical data from Jordanian hospitals, showing high performance and potential for practical use in healthcare.

## Contribution

The study benchmarks multiple ML algorithms for brain tumor classification and validates their performance on real-world Jordanian hospital data.

## Key findings

- Neural Network achieved the highest AUC of 0.996 and 95.8% accuracy in brain tumor classification.
- SVM closely followed with an AUC of 0.993 and 94% accuracy.
- Models maintained over 92% accuracy on real-world clinical data from Jordanian hospitals.

## Abstract

The rising incidence of brain tumors and their diverse characteristics make early and accurate diagnosis increasingly challenging. Traditional diagnostic techniques, while effective, often rely on subjective assessment, highlighting the potential of machine learning (ML) to enhance diagnostic accuracy and efficiency. This study evaluates the performance of seven ML algorithms—Decision Tree, AdaBoost, k-Nearest Neighbors (k-NN), Neural Network, Logistic Regression, Random Forest, and Support Vector Machine (SVM)—for brain tumor classification. A comprehensive dataset of 7,023 instances, encompassing glioma, meningioma, pituitary tumors, and healthy samples, was used in a three-way balanced design, with models validated through stratified 10-fold cross-validation. With AUC values near 1.00, Specifically, the Neural Network achieved the highest performance with AUC = 0.996, accuracy = 0.958, F1 = 0.958, precision = 0.958, and recall = 0.958, followed closely by SVM (AUC = 0.993, accuracy = 0.940). the results show that sophisticated models like SVM and neural networks perform better in terms of prediction than more straightforward models like AdaBoost and Decision Trees. The work investigates data augmentation strategies like SMOTE to alleviate class imbalances and further improve model resilience. It also talks about how interpretable AI techniques like SHAP and LIME can be included to increase clinical acceptance and trust. In order to solve ethical issues with algorithmic bias and data protection, federated learning is also taken into consideration for safe multi-institutional collaboration. Notably, our models showed excellent dependability in correctly categorizing tumors when evaluated on actual clinical cases from Jordanian hospitals, highlighting their potential for practical implementation in rural healthcare settings. This research establishes benchmarks for ML-based tumor classification, paving the way for improved diagnostic capabilities in diverse and resource-constrained clinical environments, Validation on retrospective, anonymized cases from Jordanian hospitals confirmed clinical applicability, with models maintaining > 92% accuracy on real-world data.