Digitalisation of the Brief Visuospatial Memory Test-Revised and Evaluation with a Machine Learning Algorithm.

The disease multiple sclerosis (MS) is characterized by various neurological symptoms. This paper deals with a novel tool to assess cognitive dysfunction. The Brief Visuospatial Memory Test-Revised (BVMT-R) is a recognized method to measure optical recognition deficits and their progression. Typically, the test is carried out on paper. We present a way to make this process more efficient, without losing quality by having the patients using a tablet App and having the drawings rated with the use of a machine learning (ML) algorithm. A dataset of 1'525 drawings were digitalized and then randomly split in a training dataset and in a test dataset. In addition to the training dataset the already trained drawings from a preliminary paper were added to the training dataset. The ratings done by two neuropsychologists matched for 81% of the test dataset. The ratings done automatically with the ML algorithm matched 72% with the ones of the first neuropsychologist and 79% of the ones of the second neuropsychologist. For a semi-automated rating we defined a threshold value for the reliability of the rating of 78.8%, under which the drawing is routed for manual rating. With this threshold value the ML algorithm matched 80.3% and 86.6% of the ratings of the first and second neuropsychologists. The neuropsychologists have in that case to manually check 17.4% of the drawings. With our results is it possible to execute the BVMT-R Test in a digital way. We found out, that our ML algorithms have with the semi-automated method the similar matching as the two professional raters.

Automated Rating of Multiple Sclerosis Test Results Using a Convolutional Neural Network.

This work concerns methods for automated rating of the progression of Multiple Sclerosis (MS). Often, MS patients develop cognitive deficits. The Brief Visuospatial Memory Test-Revised (BVMT-R) is a recognized method to measure optical recognition deficits and their progression. Typically, the test is carried out on paper using geometric figures which the patient should recognize and trace. The results are rated manually by a physician. The goal of this work was to digitize the BVMT-R and to support the interpretation of the test results using a machine learning (ML) algorithm. A convolutional neural network (CNN) was used to rate the drawings of a patient. As a result, the correct point value of the BVMT-R could be determined with an accuracy between 57% and 76% based on a training set of 624 patient drawings obtained from 135 patients. These drawings had been previously physician rated to serve as a gold standard. In our experiment, we obtained reasonable accuracy above 80% when more than 40 drawings were available, but our training sample was too small for more detailed analysis. Conclusion: At the currently achieved classification accuracy, results analysis will remain a physician task, potentially supported with ML based preclassification, but there is hope that ML accuracy can be further improved to enable automated follow-ups.