به كارگيري الگوريتم‌هاي يادگيري ماشين در بهبود تشخيص و كاهش خطاهاي انساني در خدمات مراقبت‌هاي بهداشتي

Cardiovascular diseases an‎d diabetes are major public health challenges, an‎d their early detection can help reduce complications an‎d treatment costs. The complexity of clinical data an‎d the potential for errors in their interpretation highlight the need to investigate clinical decision-support tools. This study aimed to comparatively eva‎luate machine learning algorithms for identifying heart disease an‎d diabetes an‎d to examine their potential to support diagnostic decision-making. Five algorithms, Ran‎dom Forest (RF), Artificial Neural Network (ANN), One-Dimensional Convolutional Neural Network (1D-CNN), Support Vector Machine with a Radial Basis Function kernel (SVM-RBF), an‎d Extreme Gradient Boosting (XGBoost), were implemented an‎d eva‎luated in Python. After excluding six incomplete records, the UCI Heart Disease dataset comprised 297 samples with 13 clinical features. The balanced version of the BRFSS 2015 diabetes dataset contained 70,692 samples with 21 health-related, lifestyle, an‎d demographic features. Both tasks were formulated as binary classification problems. Each dataset was divided into training an‎d test sets using an 80:20 stratified split an‎d a fixed ran‎dom seed of 42. Stan‎dardization parameters were estimated exclusively from the training data an‎d subsequently applied to the test data to prevent data leakage. Model performance was eva‎luated using accuracy, precision, recall, F1-score, an‎d the area under the receiver operating characteristic curve (ROC-AUC). On the heart disease dataset, SVM-RBF an‎d XGBoost jointly achieved the highest accuracy of 85%, while SVM-RBF attained the highest reported AUC of 0.954. SVM-RBF demonstrated relative accuracy improvements of 2% over both RF an‎d ANN an‎d approximately 4.08% over 1D-CNN, while matching the accuracy of XGBoost. On the diabetes dataset, calculations based on the confusion matrices showed that XGBoost achieved the highest numerical accuracy, 75.20%, with an AUC of 0.830. Its relative accuracy improvements over 1D-CNN, ANN, SVM-RBF, an‎d RF were approximately 0.01%, 0.24%, 0.29%, an‎d 0.35%, respectively, indicating closely comparable performance across the models on this dataset. These findings support the potential of the eva‎luated algorithms for developing diagnostic support tools. However, the statistical significance of the observed performance differences an‎d any actual reduction in human diagnostic errors require further statistical an‎d clinical eva‎luation.

محمدرضا احمدزاده , محمد داورپناه جزي

الهام محمودزاده , مينا اميري