abstract
- Discovering and characterizing reproducible disease subtypes results is one of the most demanding and fundamental tasks in many fields, such as bioinformatics and health informatics. It could facilitate diagnosis and is a vital step toward more individualized therapy. This paper aims to analyze the ability of unsupervised learning methods to identify a small collection of reliable and stable subtypes of subjects with mild cognitive impairment (MCI) and to discover the primary prodromal Alzheimer's disease (AD) stages in subjects with MCI to AD conversion risk. We present a novel unsupervised learning methodology to identify the notable stable and reproducible subtypes. The proposed method takes advantage of the consensus clustering of unsupervised clustering methods. For this mean, we obtained the data from the Alzheimer's disease Neuroimaging Initiative study. 346 features, including demographic information and MRI-derived features, described 839 subjects with early MCI. We randomly split the data into discovery (70%) and validation (30%) sets. The discovery set was analyzed using five unsupervised clustering methods, and robust consensus clustering was used to determine the most stable and reliable subtypes. The results show that the proposed method identified four different MCI patient subtypes. After discovery, subtypes were predicted in the testing set and associated with MCI conversion. One subtype had a high-risk (OR = 2.99, 95%CI = 1.65 to 5.41) of converting to Alzheimer's disease.