An important area in clinical applications of artificial intelligence is the development of models to assist with the automatic identification and classification of disease. It is hoped that these models could improve diagnostic processes, and ultimately the quality of life for patients. Automatic classification of Alzheimer's Dementia is one example of this task, and has been recently showcased by the Interspeech 2020 ADReSS and 2021 ADReSSo Challenges, which attempted to improve the ability of machine learning models to predict the presence of the disorder by using speech and text samples. Using a dataset from the 2020 Challenge, with 25-dimension eGeMAPS acoustic features extracted, this study implements two neural models as binary classifiers (a BiLSTM with Attention and a CNN) as the basis for an explainable investigation into the acoustic features that are most important for this prediction. These models are used in a series of experiments using the SHAP model explanation framework, as well as feature ablation methods to investigate feature importance in terms of ranking and how feature values affect importance. Comparisons are then made across explanation methods and model types, for both feature importance and related feature value effects, and the data provided are analyzed to determine patterns in how the acoustic features behave during prediction.