Latent class cluster analysis to capture language profiles in the acute phase post-stroke

Background Although timely and accurate detection of language deficits is an essential part of quality stroke care, little is known about acute – in contrast to chronic – post-stroke language profiles. Traditional aphasia subtypes are solely based on behavioral data, thereby neglecting the inherent multifactorial nature of aphasia. Such multivariate approaches are however crucial to study post-stroke language deficits, as these capture the complex interplay between the neural and behavioral level, as well as stroke-related comorbidities that are especially relevant in the acute stage post-stroke. Aims In this study, we investigate whether a data-driven approach including clinical data beyond language can provide additional insight in acute post-stroke language profiles. Methods & procedures Behavioral data from the Language Screening Test (LAST), medical information (stroke symptoms, lesion location) and demographic data were collected in 452 stroke patients. Patients were grouped in acute language profiles using Latent Class Cluster Analysis, which accounts for interdependence among the observed variables in a person-centered way. To validate the obtained acute language profiles, the scores on two more extensive language assessments (ScreeLing and ANELT) that were administered in a subset of the sample (NScreeLing = 87, NANELT = 60) were compared between the acute language profiles using a non-parametric ANOVA. Outcomes & results Data-driven clustering divided the stroke patients into five acute language profiles which we labeled as follows: No language deficit (37%), productive language deficit (24%), mild language deficit (10%), severe language deficit (12%) and comorbid language deficit (17%). Solely relying on the language screening (LAST), stroke patients with mild and comorbid language deficits did not show significant differences in naming, automatic speech and verbal instruction scores. However, our data-driven approach suggested that in acute stroke, apparent language difficulties in the comorbid language deficit profile were rather due to advanced age and neglect. Post-hoc validation showed that the no, productive, mild and severe language deficit groups scored significantly different on an impairment-based language assessment (ScreeLing) which was not included in the initial cluster analysis. Conclusion An automatic integration of post-stroke symptoms and comorbidities thus seems a promising tool to improve detection and follow-up of post-stroke language deficits.