A large number of human diseases result from disruptions to protein structure and function caused by missense mutations. Computational methods are frequently employed to assist in the prediction of protein stability upon mutation. These methods utilize a combination of protein sequence data, protein structure data, empirical energy functions, and physicochemical properties of amino acids. In this work, we present the first use of dynamic protein structural features in order to improve stability predictions upon mutation. This is achieved through the use of a set of timeseries extracted from microsecond timescale atomistic molecular dynamics simulations of proteins. Standard machine learning algorithms using mean, variance, and histograms of these timeseries were found to be 60-70% accurate in stability classification based on experimental ΔΔG or protein-chaperone interaction measurements. A recurrent neural network with full treatment of timeseries data was found to be 80% accurate according the F1 score. The performance of our models was found to be equal or better than two recently developed machine learning methods for binary classification as well as two industry-standard stability prediction algorithms. In addition to classification, understanding the molecular basis of protein stability disruption due to disease-causing mutations is a significant challenge that impedes the development of drugs and therapies that may be used treat genetic diseases. The use of dynamic structural features allows for novel insight into the molecular basis of protein disruption by mutation in a diverse set of soluble proteins. To assist in the interpretation of machine learning results, we present a technique for determining the importance of features to a recurrent neural network using Garson's method. We propose a novel extension of neural interpretation diagrams by implementing Garson's method to scale each node in the neural interpretation diagram according to its relative importance to the network.