Algoritmo do vizinho mais próximo para predição de doenças

Abstract

The general objective of this study was to evaluate the efficiency of the K-Nearest Neighbor (KNN) algorithm for disease prediction, conducting tests for early diabetes detection using an approach that combines a literature review and experimental research. During the study, it was observed that the KNN model, applied to a dataset composed of categorical and numerical variables obtained from patients at the Sylhet Diabetes Hospital in Bangladesh, achieved a performance accuracy of up to 72.4%. This result was based on scenarios where preprocessing included variable selection aimed at improving accuracy, such as the patient's age. The preprocessing stages involved data normalization, handling missing values, and class balancing, which were crucial for achieving better outcomes. Additionally, the appropriate choice of the "K" hyperparameter was critical for balancing model accuracy and stability, with different values tested to identify the most suitable one for the dataset used. In this study, a K value of 3 yielded the best results. Metrics such as accuracy, confusion matrix, and precision were used to evaluate the model's effectiveness. The study concluded that the KNN model is an efficient approach for disease prediction, provided the data is properly preprocessed and the K parameter is carefully selected. Despite its simplicity and versatility, KNN faces limitations, such as sensitivity to data dimensionality and class imbalance. Techniques like normalization and feature selection help mitigate these issues, but distance calculation in large datasets remains a challenge in complex contexts, highlighting the need for careful evaluation of its application in clinical problems.