Freezing of gait (FoG) leads to imbalance and falls in Parkinson’s disease. Cues potentially prevent FoG occurrence and unfreeze the patient from FoG. Although FoG detection and its prediction have been intensively studied, its termination is mostly neglected. Continuous cueing after the termination of freezing is annoying and has potential side effects in the long term. In this article, for the first time, we attempt to develop a machine learning approach for the prediction of start and termination of freezing, which can potentially provide automated controlled cueing in Parkinson’s individuals. We hypothesize certain attributes that correspond to the transition from walking to freezing and vice versa. To this end, we propose unique labeling of classes to predict freezing termination as follows: no FoG, pre-FoG (immediate state preceding FoG), FoG, and pre of post FoG (state just before unfreezing). Daphnet dataset, freely available online, was utilized to develop customized attributes measured from accelerometers with the data recorded from ten participants. The high-dimensional attributes were reduced using a principal component (PC) analysis before being fed to the k-nearest neighbor (kNN) classifier for prediction. With 45 PCs, we achieved average (SD) precision, sensitivity, specificity, f1 score, and accuracy of 95.55% (4.60%), 94.97% (4.86%), 99.19% (0.85%), 95.25% (4.72%), and 98.92% (1.56%), respectively, across the four classes while utilizing the attributes from 60 data points ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\approx 0.93$ </tex-math></inline-formula> s) prior to a given instant. Specifically, for the newly introduced labeling—pre of post FoG, we observed average (SD) precision, sensitivity, specificity, and f1 score of 92.73% (10.15%), 91.50% (10.34%), 99.83% (0.32%), and 92.10% (10.25%), respectively, with 60 previous data points and 45 PCs. These results show the potential of the present approach in the timely deactivation of cues in real time to avoid any side effects of cueing. Enhanced clinical benefit for Parkinson’s patients is the major contribution of this study.