Using artificial intelligence (AI) to predict the conversion of cognitively unimpaired individuals to mild cognitive impairment

Abstract Background Mild cognitive impairment (MCI) is the objective decline in neurocognitive functioning, but without significant impairment of the individual’s ability to perform the usual instrumental activities of daily living (1). Diagnosing MCI can be done using a combination of different methods such as cognitive testing, structural neuroimaging, nuclear imaging, and cerebrospinal fluids, and plasma biomarkers. Some of these methods are invasive and expensive. Machine learning (ML) algorithms can be trained to predict the onset of MCI, using data from non‐invasive methods. Method Four databases (WoS, MEDLINE, EMBASE, and CINAHL) were searched using search string of relevant terms (cognitive decline, artificial intelligence, prediction, and cognitively unimpaired). Articles that reportedtrained ML algorithms using non‐invasive predictors of MCI in cognitively healthy adults (≥ 18 years old) were included. The review was registered with PROSPERO (CRD42022379027) and PRISMA guidelines were followed. The Newcastle‐Ottawa Quality Assessment scale was used to assess the quality of studies. Result Of the 1,098 articles identified and screened, 22 studies were included. 373 non‐invasive predictors of MCI were identified. The most prevalent predictors included: demographic information (age, sex, years in education), voice and speech parameters (sentence repeating, semantic fluency), gait parameters (velocity, cadence, step time), performance on cognitive tests (MMSE, K‐MoCA, CDR, AD8), eye‐movements (pupil diameter, saccade orientation) and Instrumental activity of daily livings (medication, finance management). The maximum and minimum number of predictors used to develop a ML algorithm was 121 and 2, respectively. The ML algorithms developed in the included studies had (average sensitivity and specificity of 77.90%±22 and 80.84%±18, respectively) exceeded the performance of algorithms that used invasive and expensive predictors such as nuclear imaging and neuroimaging which achieved accuracy scores (60%‐77%) (2). Conclusion Non‐invasive predictors can be used to train ML algorithms to predict the onset of MCI in cognitively healthy individuals with good accuracy scores exceeding 70%. These algorithms may can aid clinician decision making, thus setting early treatment plans, and allowing individuals to be involved in care planning before progression of MCI to dementia AD.