Mild Cognitive Impairment Diagnosis and Detecting Possible Labeling Errors in Alzheimer’s Disease with an Unsupervised Learning-based Approach

The diagnosis of cognitive disabilities like dementia and mild cognitive impairment is challenging because several factors are non-linearly related to these pathologies. Thereby, classification errors committed by a specialist can become more frequent. In this work, we propose a methodology for detecting possible labeling errors. For this, we use a classification technique capable of learning patient profiles in an unsupervised manner, then add semantic value to each profile by applying the majority voting technique. Our goal is to make a tool robust against labeling errors present in the data. We achieved a mean accuracy of 89.33%, which is not an improvement considering this as a standalone tool. Then, we compare the labels with the labels provided by an artificial neural network trained in a supervised manner. We could experimentally find pieces of evidence of possible labeling errors in 9.14% of this dataset samples. It shows that our contribution is valuable since it can indicate possible labeling errors.