According to the fall phenomenon of the elderly, this paper makes suggestion for improvement. The multilayer perceptron neural network is used to conduct weight analysis of the physical factors affecting the balance of the elderly body, and the probability for the fall of the elderly is predicted based on the forward propagation principle. By taking all experimental indicators of standard old people as the basis, balance comparison template is set up to calculate displacementtime function relation of body feature points of standard old people and sample old people in metric space, and MATLAB software is used to draw a graph of it. By comparison, the positions affecting the balance of the elderly are found, and the suggestions for their improvement according to different positions are put forward. Results show that the accurate rate of the prediction by neural network reaches 80 percent, that more than half of the elderly in 76 samples has weak balance ability and has high risk of the fall, and that the observation point for abnormal positions of the elderly involves in their shoulder joint, back brain, spine and so on. The results obtained are of certain reference value to preventing the fall of the elderly.