Machine learning classifier to identify clinical and radiological features relevant to disability progression in multiple sclerosis

Silvia Tommasin,Carlo Pozzilli,Alessandro Taloni,Giuseppe Pontillo,Serena Ruggieri,Patrizia Pantano,Sirio Cocozza,Maria Petracca,Laura De Giglio,Arturo Brunetti,Costanza Giannì,Nikolaos Petsas

doi:10.1007/s00415-021-10605-7

Abstract

ObjectivesTo evaluate the accuracy of a data-driven approach, such as machine learning classification, in predicting disability progression in MS.MethodsWe analyzed structural brain images of 163 subjects diagnosed with MS acquired at two different sites. Participants were followed up for 2–6 years, with disability progression defined according to the expanded disability status scale (EDSS) increment at follow-up. T2-weighted lesion load (T2LL), thalamic and cerebellar gray matter (GM) volumes, fractional anisotropy of the normal appearing white matter were calculated at baseline and included in supervised machine learning classifiers. Age, sex, phenotype, EDSS at baseline, therapy and time to follow-up period were also included. Classes were labeled as stable or progressed disability. Participants were randomly chosen from both sites to build a sample including 50% patients showing disability progression and 50% patients being stable. One-thousand machine learning classifiers were applied to the resulting sample, and after testing for overfitting, classifier confusion matrix, relative metrics and feature importance were evaluated.ResultsAt follow-up, 36% of participants showed disability progression. The classifier with the highest resulting metrics had accuracy of 0.79, area under the true positive versus false positive rates curve of 0.81, sensitivity of 0.90 and specificity of 0.71. T2LL, thalamic volume, disability at baseline and administered therapy were identified as important features in predicting disability progression. Classifiers built on radiological features had higher accuracy than those built on clinical features.ConclusionsDisability progression in MS may be predicted via machine learning classifiers, mostly evaluating neuroradiological features.

Highlights

Materials and methodsClinical progression and disability accumulation are highly heterogeneous in all phenotypes of multiple sclerosis (MS) [1]
Since patients who remained stable at follow-up were about 2/3 of both samples, and we aimed at balancing size of classes, at each of the 1000 performed models we picked all the patients who experienced disability progression from both sites’ sample and randomly picked an equal number of patients who remained stable, reaching a final number of 72 participants from Site 1 and 44 participants from Site 2
We found that classifiers built on neuroradiological features were more accurate and sensitive than those built on clinical features or on mixed clinical-neuroradiological features, showing better accuracy, AUC, sensitivity and specificity values (Table 4; Fig. 3)

Summary

Introduction

Materials and methodsClinical progression and disability accumulation are highly heterogeneous in all phenotypes of multiple sclerosis (MS) [1]. Several disease modifying treatments are available to improve long-term prognosis of patients with MS [3]. Several MRI studies have highlighted various aspects of tissue damage in MS [8] demonstrating a prognostic role of T2-hyperintense lesions, global and cortical atrophy [9], as well as that of damage to some key structures, such as thalamus [10] and cerebellum [11, 12]. Lesion burden appeared to be a relevant predictor of long-term cognitive outcome [13] and disease progression [14]. Further factors anticipating disability progression in MS have been shown to be structural and microstructural damage in the cerebellum [15], thalamus [16] and normal appearing white matter (NAWM) [17]

Objectives

Results

Discussion

Conclusion