Predicting individual long-term prognosis of spatial neglect based on acute stroke patient data.

One of the most pressing questions after a stroke is whether an individual patient will recover in the long term. Previous studies demonstrated that spatial neglect-a common cognitive deficit after right hemispheric stroke-is a strong predictor for poor performance on a wide range of everyday tasks and for resistance to rehabilitation. The possibility of predicting long-term prognosis of spatial neglect is therefore of great relevance. The aim of the present study was to test the prognostic value of different imaging and non-imaging features from right hemispheric stroke patients: individual demographics (age, sex), initial neglect severity and acute lesion information (size, location). Patients' behaviour was tested twice in the acute and the chronic phases of stroke and prediction models were built using machine learning-based algorithms with repeated nested cross-validation and feature selection. Model performances indicate that demographic information seemed less beneficial. The best variable combination comprised individual neglect severity in the acute phase of stroke, together with lesion location and size. The latter were based on individual lesion overlaps with a previously proposed chronic neglect region of interest that covers anterior parts of the superior and middle temporal gyri and the basal ganglia. These variables achieved a remarkably high level of accuracy by explaining 66% of the total variance of neglect patients, making them promising features in the prediction of individual outcome prognosis. An online tool is provided with which our algorithm can be used for individual outcome predictions (https://niivue.github.io/niivue-neglect/).