IMPACT OF DISTANCE METRICS ON KNN ALGORITHM:CASE STUDY OF A HEART DISEASE PREDICTION

  ABSTRACT The k-Nearest Neighbours (kNN) algorithm is a fundamental machine learning method whose effectiveness relies heavily on the choice of distance metric within the underlying metric space. This study used a heart-disease dataset gotten from Kaggle containing 500 patients with seven features. The dataset was split into two parts: 70% for training and 30% for testing. It was also screened for basic data issues (e.g., missing values), and the training set was balanced using the Synthetic Minority Over-sampling Technique (SMOTE). The study compared five distance metrics: Euclidean, Manhattan, Minkowski, Mahalanobis, and Cosine, and evaluated accuracy, precision, recall, F1-score, and confusion matrices. Overall, Manhattan performed best, providing the strongest balance between identifying true cases and avoiding false alarms. Euclidean and Cosine were strong; Euclidean achieved the highest Macro-F1 and accuracy. However, Manhattan offered the most favourable clinical trade-off, higher recall for no-disease and highest precision for disease reducing missed cases and false alarms. The study conclude that metric choice is a key design decision; recommending Manhattan when minimizing false negatives is paramount, with Euclidean or Cosine as strong alternatives.