Inspecting Linguistic Features in Interaction Transcripts of Patients with Alzheimer’s Disease through Interpretable Machine Learning

Alzheimer’s disease (AD) is the most common type of dementia, affecting over 5 million Americans every year, a figure that is set to more than double by 2050. Although there currently is no approved treatment either to halt AD or to slow down its progressive course, early diagnosis can help improve the lives of the patients as well as those of their caregivers and families as coping strategies take time to be implemented. In recent years, computer science has supported the diagnosis of dementia through machine learning and deep learning techniques. Algorithms have been developed with the aim of automatically classifying and early identifying patients affected by dementia. In these experiments, the primary concern has been the improvement of classification performance, while less attention has been devoted to examining the decisions made by the machine in order to advance knowledge by unravelling new facts regarding the way in which AD patients use language. Thus, the present study sets out to identify the most predictive features that characterise the speech of patients suffering from AD by inspect- ing the choices made by an algorithm trained to discriminate between transcripts of AD patients and of healthy subjects with the tools provided by linguistics and within an interpretable machine learning paradigm.