Wearable watches provide very useful linear acceleration information that can be use to detect falls. Howeverfalls not from a standing position are difficult to spot amongother normal activities. This paper describes methods, basedon pattern recognition using machine learning, to improve thedetection of "soft falls". The values of the linear accelerometersare combined in a robust vector that will be presented as inputto the algorithms. The performance of these different machinelearning algorithms is discussed and then, based on the bestscoring method, the size of the time window fed to the systemis studied. The best experiments lead to results showing morethan 0.9 AUC on a real dataset. In a second part, a prototypeimplementation on an Android platform using the best resultsobtained during the experiments is described.