Wearable watches provide very useful linear acceleration information that can be use to detect falls. However falls not from a standing position are difficult to spot among other normal activities. This paper describes methods, based on pattern recognition using machine learning, to improve the detection of "soft falls". The values of the linear accelerometers are combined in a robust vector that will be presented as input to the algorithms. The performance of these different machine learning algorithms is discussed and then, based on the best scoring method, the size of the time window fed to the system is studied. The best experiments lead to results showing more than 0.9 AUC on a real dataset. In a second part, a prototype implementation on an Android platform using the best results obtained during the experiments is described.