We present an efficient utilization of a machine learning (ML) method concentrating on the 'AI for social good' application. We develop a digital dementia biomarker for early-onset dementia forecast. The paper demonstrates encouraging preliminary results of EEG-wearable-based signal analysis and a subsequent classification adopting a signal complexity test of a multifractal detrended fluctuation analysis (MFDFA) in emotional faces working memory training and evaluation tasks. For the digital biomarker of dementia onset detection, we examine shallow- and deep-learning machine learning models. We report the best median accuracies in a range of 90% for random forest and fully connected neural network classifier models in both emotional faces learning and evaluation experimental tasks. In addition, the classifiers are trained in a ten-fold cross-validation regime to discriminate normal versus mild cognitive impairment (MCI) cognition stages using MFDFA patterns from four-channel EEG recordings. Thirty-five volunteer elderly subjects participate in the current study concentrating on simple wearable EEG-based objective dementia biomarker progression. The reported outcomes showcase an essential social benefit of artificial intelligence (AI) employment for early dementia prediction. Furthermore, we improve ML employment for the succeeding application in an uncomplicated and applied EEG-wearable examination.