An efficient machine learning (ML) implementation in the so-called 'AI for social good' domain shall contribute to dementia digital neuro-biomarker development for early-onset prognosis of a possible cognitive decline. We report encouraging initial developments of wearable EEG-derived theta-band fluctuations examination and a successive classification embracing a time-series complexity examination with a multifractal detrended fluctuation analysis (MFDFA) in the face or emotion video-clip identification short-term oddball memory tasks. We also report findings from a thirty-five elderly volunteer pilot study that EEG responses to instructed to ignore (inhibited) oddball paradigm stimulation results in more informative MFDFA features, leading to better machine learning classification results. The reported pilot project showcases vital social assistance of artificial intelligence (AI) application for an early-onset dementia prognosis. Clinical Relevance- This introduces a candidate for an objective digital neuro-biomarker from theta-band EEG recorded by a wearable for a plausible replacement of biased 'paper & pencil' tests for a mild cognitive impairment (MCI) evaluation.