Lightweight automatic diagnostic tools for Amy-otrophic Lateral Sclerosis (ALS) and the associated dysarthria are essential for deployment in resource-limited platforms like mobile phones or general purpose computers. This study performs speech-based low-complexity classification of ALS and healthy subjects by cutting down (1) model complexity and (2) input feature dimensionality. Low complexity Dense Neural Network (DNN) models with 2 or less hidden layers are explored in comparison with the highly complex state-of-the-art Convolutional Neural Network (CNN) with Bidirectional Long Short-Term Memory (BiLSTM) architecture. On the other hand, various temporal statistics (standard deviation, autocorrelation at varying lags) obtained from the commonly used Mel-Frequency Cepstral Coefficients (MFCC) or its individual coefficients are investigated as the low dimensional features. Experiments with 72 ALS and 55 healthy subjects using Spontaneous Speech (SPON) and Diadochokinetic Rate (DIDK) tasks indicate the following. Model complexity reduction with DNN architectures gives comparable, or in some cases better performance, w.r.t. the CNN-BiLSTM model. DNN architectures, with lag 1 au-tocorrelation of MFCC (along with its delta and double delta coefficients) as the input feature vector for SPON task and standard deviation of the same for DIDK task, can respectively achieve 5.67% and 6.59% higher mean classification accuracies than the CNN-BiLSTM model with entire MFCC sequence as input while causing 99.99% reduction in the model parameter count. Moreover, using single dimensional standard deviation feature of the first delta coefficient for SPON and that of the second delta coefficient for DIDK, together with the DNN models, achieve 94.59% further reduction in the model parameter count while incurring only 1.76% and 5.17% further decrease, respectively, in the classification performance.