Descriptive and predictive models for Guillain-Barre syndrome based on clinical data using machine learning algorithms

Abstract

Guillain-Barré Syndrome (GBS) is an autoimmune neuropathy of fast evolution and potentially fatal. The exact cause of GBS is unknown; however, it is frequently preceded either by a respiratory or a gastrointestinal infection. The diagnosis of GBS includes clinical, serological and electrophysiological criteria (nerve conduction or electrodiagnostic test) [67]. The severity of GBS varies among subtypes, which can be mainly Acute In ammatory Demyelinating Polyneuropathy (AIDP), Acute Motor Ax onal Neuropathy (AMAN), Acute Motor Sensory Axonal Neuropathy (AMSAN) and Miller-Fisher Syndrome [59]. A better understanding of the di erences in the GBS subtypes is critical for the implementation of appropriate treatments for total recovery and, in certain cases for the survival of patients. Hos pitalization time and the cost of treatments vary according to the severity of the speci c subtype. Finding a minimum feature subset to accurately identify GBS subtypes could lead to a simpli ed and cheaper process of diagnosis and treatment of the GBS case. The ultimate goal of a physician is to get patients to a full recovery. This can be more e ectively achieved when an early diagnosis of the case is performed using a minimum number of medical features. There are many successful applications of machine learning in medicine, biology and genetics Also, the machine learning techniques have been applied successfully in the early diagnosis of diseases, including neurological disorders [5, 93]. Besides, the identi cation of a reduced number of relevant features to predict GBS subtypes could guide physicians to design a faster, simpler and cheaper diagnosis of the case. This dissertation applies machine learning techniques to create two models for GBS, a descriptive model and a predictive model. The former aimed at identifying a minimum set of relevant fea tures that builds four clusters, each corresponding to a speci c GBS subtype. This model uses two approaches. In the rst approach, the problem is tackled by using lter methods along with a clus tering algorithm called Partitions Around Medoids (PAM). In the second approach, a novel method termed QSA-PAM (Quenching Simulated Annealing-Partitions Around Medoids) is introduced in order to solve the same problem. Several experiments are conducted with a real dataset. Results from descriptive model allowed the identi cation of 16 relevant features selected out of an original 356-feature dataset. The goal of the predictive model was to apply the 16 relevant featuresto classify GBS subtypes. Again, two approaches were used. In the rst approach, classi cation experiments were conducted using single classi ers. The second approach applied ensemble methods. Three types of experiments were performed in both approaches: four classes classi cation, One versus All classi cation and One vs One classi cation. A comparison of performance between both approaches was made. Successful results in experiments in both approaches resulted in creating the rst predictive model for GBS using machine learning techniques. Besides, the analysis performed in this work provides insight about the best methods for each classi cation case.