Automatic screening of Alzheimer's Dementia (AD) can have significant impact on society and the well-being of the patients. Early detection of AD from spontaneous speech offers great potential for inexpensive and convenient casual testing. We propose our deep neural network architecture that leverages acoustic, linguistic, and demographic features to build a model for dementia screening for the biographic interview speech corpus of Interdisciplinary Longitudinal Study on Adult Development and Aging (ILSE). We oversample non-sequential and sequential features using well-known oversampling techniques and adapted data augmentation techniques to overcome the challenge of the imbalanced dataset, since the distribution of the diagnostic groups in ILSE corresponds to the prevalence of dementia. Our system achieves 70.6% of unweighted average recall on a 3-class classification problem. Moreover, we also investigate the feature importances to the model prediction to identify the most relevant indicators for AD detection, which may contribute to interpreting signs of cognitive decline and thus supporting clinicians in the diagnosis of dementia.