The study presents a novel scheme that recognizes and classifies different sub-phases within the involuntary breathing movement (IBM) phase during breath-holding (BH). We collected force data from eight recreational divers until the conventional breakpoint (CB). They were in supine positions on force plates. We segmented their data into no-movement (NM) phases, i.e., the easy phase (EP) and IBM phase (comprising several events or sub-phases of IBM). Acceleration and jerk were estimated from the data to quantify the IBMs, and phase portraits were developed to select and extract specific features. K means clustering was performed on these features to recognize different sub-phases within the IBM phase. We found five–six optimal clusters separating different sub-phases within the IBM phase. These clusters separating different sub-phases have physiological relevance to internal struggles and were labeled as classes for classification using support vector machine (SVM), naive Bayes (NB), decision tree (DT), and K-nearest neighbor (K-NN). In comparison with no feature selection and extraction, we found that our phase portrait method of feature selection and extraction had low computational costs and high robustness of 96–99% accuracy.