Disclosed are systems and methods for detecting a fall of a user of a wearable device. The described system samples signal data from an accelerometer of the wearable device and performs a preliminary determination of fall detection using pattern recognition. This preliminary detection may be based on whether a first portion of the signal data exceeds a first threshold indicative of a fall of a wearer of the electronic device and whether a second portion of the signal data is within a range indicative of motion stillness following the fall. The described system further applies a machine learning classifier to the signal data associated with the preliminary indication to generate a fall classification indicating the signal data represents a fall of the wearer of the electronic device.
