Using deep learning model on drawing tasks as a screening tool for mild cognitive impairment in older adults with sensory impairments

Abstract

AbstractBackgroundMild cognitive impairment (MCI) and sensory impairments commonly co‐exist in an aging population (e.g., Rong et al. 2020). A previous study has shown that patients with hearing impairment tend to do worse on verbal tests (Utoomprurkporn 2020). The reason is that the patients fail to hear the task instructions and have to put in more effort which is cognitively taxing. Therefore, a non‐verbal cognitive task for MCI screening is needed.MethodWe developed a deep learning model for MCI screening which requires only drawing tasks (trail‐making task, cube‐copying task, clock drawing task) on a digital tablet as its inputs. We cross‐validated the model on 981 subjects (age 55‐89, 77% female, 29% MCI) from our healthy aging clinic. The model used VGG16 as a backbone with self‐attention layers which expect to both improve the model performance and provide visual explanations.ResultThe model can discriminate between MCI and healthy controls with an AUC of 0.84. Our model provided better visual explanations that are compatible with how experienced clinical psychologists interpret the drawings. To increase the generalizability, we further developed a second model that aims to automate the clock drawing scoring. The models were tested in patients with and without sensory impairments.ConclusionWe investigated the feasibility of our deep learning models on drawings as a potential tool for MCI screening in an aging population with sensory impairments. The tool requires little or no human clinical experience and potentially can be used as a part of the population‐based MCI screening program.