Human　joint　moment　plays　an　important　role　in　quantitative　rehabilitation　assessment　and　exoskeleton　robot　control.　However,　the　existing　moment　prediction　methods　require　kinematic　and　kinetic　data　of　human　body　as　input,　and　the　measurement　of　them　needs　special　equipment,　which　makes　them　unable　to　be　used　in　an　unconstrained　environment.　According　to　the　situation,　this　paper　develops　a　novel　method　where　a　small　number　of　input　variables　selected　by　Elastic　Net　are　used　as　the　input　of　artificial　neural　network　(ANN)　to　predict　joint　moments,　which　makes　the　prediction　in　daily　life　possible.　The　method　is　tested　on　the　experimental　data　collected　from　eight　healthy　subjects　that　are　running　on　a　treadmill　at　a　speed　of　2,　3,　4,　and　5　m/s,　respectively.　Taking　the　right　lower　limb＇s　10　electromyography　(EMG)　and　5　joints　angle　data　as　candidate　variable　sets,　Elastic　Net　is　used　to　obtain　the　variable　coefficients　of　the　right　lower　limb＇s　four　joint　moments.　The　inputs　of　the　ANN　determined　by　the　variable　coefficients　are　used　to　train　and　predict　the　joint　moments.　Prediction　accuracy　is　evaluated　by　using　the　normalized　root-mean-square　error　(NRMSE　%)　and　cross　correlation　coefficient　(rho)　between　the　predicted　joint　moment　and　multi-body　dynamics　moment.　Results　of　our　study　suggest　that　the　method　can　accurately　predict　joint　moment　(NRMSE　＜　7.89%,　rho　＞　0.9633)　with　only　5-6　EMG　signals.　In　conclusion,　this　method　can　effectively　reduce　the　input　variables　while　keeping　a　certain　precision,　which　makes　the　joint　moment　prediction　simple　and　out　of　equipment　limitation.　This　method　may　facilitate　the　researches　on　real-time　gait　analysis　and　exoskeleton　robot　control　in　motor　rehabilitation.