The clinical significance of an AI-based assumption model for neurocognitive diseases using a novel dual-task system

Background Dual-task composed of gait or stepping tasks combined with cognitive tasks has been well-established as valuable tools for detecting neurocognitive disorders such as mild cognitive impairment and early-stage Alzheimer’s disease. We previously developed a novel dual-task system with high accuracy for differentiating patients with neurocognitive disorders from healthy controls. In this study, we aimed to elucidate whether the output value obtained through artificial intelligence assumptions has clinical meaning other than diagnosis labelling. Methods This is a retrospective cross-sectional study. Patients with Alzheimer’s disease dementia, dementia with Lewy bodies, or mild cognitive impairment who participated in our previous dual-task experiment and completed all routine neuropsychological assessments at our hospital within one year of the experimental date were eligible for inclusion in the neurocognitive disorders group. Participants in the healthy control group were recruited from community-dwelling older adults. The correlation between the output value, “y-value”, and each neuropsychological test: Mini-Mental State Examination (MMSE), Addenbrook’s Cognitive Examination, Logical Memory tests, Frontal Assessment Battery, and digit span were assessed by Pearson’s correlation coefficient. We also evaluated the correlation between the MMSE and those neurocognitive tests. To elucidate the diagnostic availability of the dual-task system and the MMSE on this dataset, we conducted a receiver operating characteristic analysis. Results We enrolled 97 participants in the neurocognitive disorders group: 42 with Alzheimer’s disease dementia, 11 with dementia with Lewy bodies, and 44 with mild cognitive impairment. Additionally, 249 participants were included in the healthy control group. Although the y-value showed significant correlations with several tests, the MMSE demonstrated much stronger significant correlations with a broader range of cognitive tests. Meanwhile, its sensitivity and specificity were 0.969 and 0.912, respectively, and the area under the curve was 0.981, which was higher than the 0.934 of the MMSE. Conclusion Our new AI-driven dual-task system has a high ability to predict neurocognitive disorders. However, the clinical significance of its output values is limited to screening for neurocognitive disorders and does not extend to estimating cognitive function. When using this system in clinical practice, it is essential to understand its limitations and select the appropriate usage scenarios.