One of the consequences of aging is the increased risk of falls, especially when someone walks in unknown or uncontrolled environments. Usually, gait is evaluated through observation and clinical assessment scales to identify the state and deterioration of the patient's postural control. Lately, technological systems for bio-mechanical analysis have been used to determine abnormal gait states being expensive, difficult to use, and impossible to apply in real conditions. In this article, we explore the ability of a system based on a single inertial measurement unit located in the lower back to estimate spatio-temporal gait parameters by analyzing the signals available in the Physionet repository "Long Term Movement Monitoring Database" which, together with automatic classification algorithms, allow predicting the risk of falls in the elderly population. Different classification algorithms were trained and evaluated, being the Support Vector Machine classifier with a third-degree polynomial kernel, cost function C = 2 with the best performance, with an Accuracy = 59%, Recall = 91%, and F1- score = 71%, providing promising results regarding a proposal for the quantitative, online and realistic evaluation of gait during activities of daily living, which is where falls actually occur in the target population. Clinical Relevance - This work proposes an early risk of falls detection tool, essential to start preventive treatment strategies to maintain the independence of the elderly through a non-invasive, simple, and low-cost alternative.