Determination of Usual Metrics From the Confusion Matrix of Multiclass Classifiers in Intelligent Algorithms in Human Movement Sciences

Intelligent algorithms that use classifiers are typically evaluated by determining specific metrics derived from the confusion matrix. The association of two or more of these metrics is crucial because, especially in multiclass classifiers with an n × n confusion matrix for n ≥ 3, the number of false positives (FP) and false negatives (FN) can be high. Consequently, results that are a priori considered satisfactory for specificity and sensitivity may lead to misinterpretations. This article aims to present different multiclass matrices, highlighting the impact of the number of true positives (TP), true negatives (TN), as well as FP and FN. To achieve this, eight different scenarios of 4x4 confusion matrices (walking, running, jumping, squatting) and 5x5 confusion matrices (walking, running, jumping, squatting, throwing) were simulated. The data were organized in Excel®, where the metrics of interest were calculated: Accuracy, Precision, Sensitivity, Specificity, Fowlkes-Mallows Index (FMI), and Balanced Accuracy. The results of the different simulations evidenced the variability among the metrics as the values of TPs, TNs, FPs, and FNs were altered. Another point to highlight is the possibility of obtaining excellent metrics while having a poor classifier.