There is a growing interest in the introduction of artificial intelligence in clinical medicine (1–3). Efforts are underway to predict and diagnose prodromal or early-stage dementia (4–7) at home and in clinical settings (8, 9).
Dementia is a state of cognitive impairment (CI) including loss of memory, language, problem solving, and executive functions that is severe enough to impair daily life. Many studies have demonstrated the presence of mild cognitive impairment (MCI), which is not severe enough to impair social and occupational activities, in the prodromal stage of dementia (10, 11). Neuropsychological testing is required in people with complaints of cognitive problems for an establishing objective diagnosis (12). Most of the screening tools have been constructed based on neuropsychological tests. The Rey–Osterrieth Complex Figure Test (RCFT) is widely used by neuropsychologists to assess cognitive function. The test was first developed by Rey in 1941 (13) and has proved to be a useful tool for analyzing visuospatial construction, perceptual organization, and visual memory in clinical evaluations and research studies (14). Patients with parieto-occipital lesions, especially on the right side, have difficulties in spatial organization of the drawing, probably because of visual disorientation (15), while patients with frontal lobe damage show impairment in programming abilities in figure reproduction (16, 17). Early-stage Alzheimer’s disease (AD) patients perform poorly on this test (18). Seo et al. showed that that the copy condition of the test was associated with spatial organization and planning and significantly predicted the conversion to pre-MCI or MCI (19). The salience of visuospatial and organizational skills as evaluated by the copy condition of the RCFT differed according to the level of intelligence (20). To obtain a more quantitative value for the accuracy of a participant’s drawing, many researchers use the RCFT based on the Osterrieth scoring criteria to diagnose CI (21).
The Clock Drawing Test (CDT) is also widely used as a screening test for dementia patients because of the following advantages: simple to use and reflection of a variety of cognitive functions, including visuospatial function, frontal lobe execution, and memory of clock concepts. The CDT requires the subject to draw the hour and minute hands of the clock to show the time "11:10." This can lead to a "self-stimulating error response," one of the errors in drawing the clock. It is necessary to suppress the tendency to draw a needle in a number called "10" driven by the perceptual level of information instead of "2" which is a meaningful number to indicate "10" (22). Studies related to dementia have reported that CDT is useful in screening of cognitive impairment (23, 24), and that is can be used for screening MCI . The CDT has a variety of scoring systems (25). Among them, the Consortium to Establish a Registry for Alzheimer's Disease Clock Drawing Test (CERAD-CDT) (26) is known as the simplest method with high diagnostic efficiency (27).
Detection of the severity of dementia is important for clinical and research purposes, and the Clinical Dementia Rating Scale (CDR) is one of the most commonly used tools for assessment. The CDR comprises the global and sum of boxes (SOB) scores. The CDR-SOB is considered a more detailed quantitative index than the global score and provides more information regarding patients with mild dementia. Studies have shown that CDR-SOB scores may have the potential for discriminating between patients with MCI and patients with a very early stage of AD dementia who are assigned a global CDR score of 0.5. Patients with MCI were assigned a CDR-SOB score of 1.8 ± 0.8, and very mild AD patients were assigned a CDR-SOB score of 3.0 ± 0.8 (28). O'Bryant et al. suggested the following staging system based on SOB scores: CDR-SOB 0, normal; 0.5–2.5, questionable impairment; and 3.0–4.0, very mild dementia (29). We also assessed the severity of dementia based on CDR-SOB score.
Several researchers have demonstrated that a digital CDT of limited number of subjects was able to differentiate patients with AD and other dementia syndromes from normal controls using machine learning (30, 31). However, digital CDT needs special equipment, and in deep learning, more data of good quality can expect better results. Therefore, we predicted cognitive impairment with deep learning based on more drawing test data than previous study. Conventional CDT and RCFT are drawing tests that evaluate cognitive changes and are constituents of neuropsychological test batteries. We questioned whether these two simple drawing tests can be used as screening tests to predict CI using convolutional neural network (CNN) algorithms. We also questioned whether the CDT, which measures various cognitive functions, could be better in predicting CI than the RCFT. Our objective was to evaluate the prediction accuracies of these two tests for CI and to compare them.