This paper presents a novel technique to predict freezing of gait in advance-stage Parkinsonian patients using movement data from wearable sensors. A two-class approach is presented which consists of autoregressive predictive models to project the feature time series, followed by machine learning based classifiers to discriminate freezing from nonfreezing based on the predicted features. To implement and validate our technique a set of time domain and frequency domain features were extracted from the 3D acceleration data, which was then analyzed using information theoretic and feature selection approaches to determine the most discriminative features. Predictive models were trained to predict the features from their past values, then fed into binary classifiers based on support vector machines and probabilistic neural networks which were rigorously cross validated. We compared the results of this approach with a three-class classification approach proposed in previous literature, in which a pre-freezing class was introduced and the problem of prediction of the gait freezing incident was reduced to solving a three-class classification problem. The twoclass approach resulted in a sensitivity of 93±4%, specificity of 91±6%, with an expected prediction horizon of 1.72 seconds. Our subject-specific gait freezing prediction algorithm outperformed existing algorithms, yields consistent results across different subjects and is robust against the choice of classifier, with slight variations in the selected features. In addition, we analyzed the merits and limitations of different families of features to predict gait freezing.