Comparison of Machine Learning Methods Using Spectralis OCT for Diagnosis and Disability Progression Prognosis in Multiple Sclerosis

Abstract

Machine learning approaches in diagnosis and prognosis of multiple sclerosis (MS) were analysed using retinal nerve fiber layer (RNFL) thickness, measured by optical coherence tomography (OCT). A cross-sectional study (72 MS patients and 30 healthy controls) was used for diagnosis. These 72 MS patients were involved in a 10-year longitudinal follow-up study for prognostic purposes. Structural measurements of RNFL thickness were performed using different Spectralis OCT protocols: fast macular thickness protocol to measure macular RNFL, and fast RNFL thickness protocol and fast RNFL-N thickness protocol to measure peripapillary RNFL. Binary classifiers such as multiple linear regression (MLR), support vector machines (SVM), decision tree (DT), k-nearest neighbours (k-NN), Naïve Bayes (NB), ensemble classifier (EC) and long short-term memory (LSTM) recurrent neural network were tested. For MS diagnosis, the best acquisition protocol was fast macular thickness protocol using k-NN (accuracy: 95.8%; sensitivity: 94.4%; specificity: 97.2%; precision: 97.1%; AUC: 0.958). For MS prognosis, our model with a 3-year follow up to predict disability progression 8 years later was the best predictive model. DT performed best for fast macular thickness protocol (accuracy: 91.3%; sensitivity: 90.0%; specificity: 92.5%; precision: 92.3%; AUC: 0.913) and SVM for fast RNFL-N thickness protocol (accuracy: 91.3%; sensitivity: 87.5%; specificity: 95.0%; precision: 94.6%; AUC: 0.913). This work concludes that measurements of RNFL thickness obtained with Spectralis OCT have a good ability to diagnose MS and to predict disability progression in MS patients. This machine learning approach would help clinicians to have valuable information.

Keywords: Machine learning; Multiple sclerosis; Optical coherence tomography; Retinal nerve fiber layer.

Figure 1

Schematic representation of Spectralis OCT acquisition protocols on a right eye retina. Fast macular thickness protocol measures total volume and macular RNFL (mRNFL) thickness in nine sectors (CF central fovea, IN inner nasal, ON outer nasal, IS inner superior, OS outer superior, IT inner temporal, OT outer temporal, II inner inferior, OI outer inferior). Fast RNFL thickness protocol measures peripapillary RNFL (pRNFL) thickness by providing mean pRNFL thickness (G) and pRNFL thickness in six sectors NS superonasal, N nasal, NI inferonasal, TI inferotemporal, T temporal, TS superotemporal. This protocol also generates 768 pRNFL thickness measurements with a circular sweep from temporal to temporal (counterclockwise). Fast RNFL-N thickness protocol differs from the previous one in two aspects: it adds papillomacular bundle (PMB) thickness and nasal/temporal (N/T) ratio, and the circular sweep is performed from nasal to nasal (clockwise). OD right eye, OCT optical coherence tomography, RNFL retinal nerve fiber layer.

Figure 2

Machine learning pipeline of the proposed method consists of five steps: data preprocessing, variable selection, 10-fold cross-validation, model building and model assessment.

Figure 3

Variable selection for multiple sclerosis (MS) diagnosis model after applying least absolute shrinkage and selection operator (LASSO) to balanced data with 72 MS patients and 72 healthy controls. Raw dataset 1 included general data and fast macular thickness protocol (13 features), raw dataset 2 included general data and fast retinal nerve fiber layer (RNFL) thickness protocol (778 features), and raw dataset 3 included general data and fast RNFL-N thickness protocol (780 features). CF central fovea, OS outer superior, II inner inferior, OI outer inferior.

Figure 4

Variable selection for multiple sclerosis (MS) prognosis models after applying least absolute shrinkage and selection operator (LASSO) to balanced data with 40 MS patients with ΔEDSS ≥ criteria and 40 MS patients with ΔEDSS <  criteria. Raw dataset 4 included general data, MS data and fast macular thickness protocol (19 features with 2-year follow-up); raw dataset 5 included general data, MS data and fast retinal nerve fiber layer (RNFL) thickness protocol (784 features with 2-year follow-up): and raw dataset 6 included general data, MS data and fast RNFL-N thickness protocol (786 features with 2-year follow-up). Raw dataset 7 included general data, MS data and fast macular thickness protocol (19 features with 2-year follow-up); raw dataset 8 included general data, MS data and fast retinal nerve fiber layer (RNFL) thickness protocol (784 features with 2-year follow-up): and raw dataset 9 included general data, MS data and fast RNFL-N thickness protocol (786 features with 2-year follow-up). Values 0, 1 and 2 on the x-axis represent the years of the 10-year follow-up. EDSS expanded disability status scale, ON outer nasal, IS inner superior, CF central fovea, IT inner temporal.

Figure 5

Accuracy of different classifiers for multiple sclerosis (MS) diagnosis and MS prognosis models. Datasets 1, 4 and 7 (brown colour) correspond to clinical data and fast macular thickness protocol. Datasets 2, 5 and 8 (grey colour) correspond to clinical data and fast retinal nerve fiber layer (RNFL) thickness protocol. Datasets 3, 6 and 9 (blue colour) correspond to clinical data and fast RNFL-N thickness protocol. The tested algorithms were: MLR multiple linear regression, SVM support vector machine, k-NN k-nearest neighbours, DT decision tree, NB Naïve Bayes, EC ensemble classifier, LSTM long short-term memory, neural network.

Figure 6

Confusion matrix of the best classifier for each predictive model using different datasets. Top: results for multiple sclerosis (MS) diagnosis. Middle: results for MS prognosis with 2-year follow-up. Bottom: results for MS prognosis with 3-year follow-up. The best classifier and several parameters to analyse the model performance for each dataset were shown in Table 4. (ΔEDSS expanded disability status scale variation).

Figure 7

Receiver operating characteristic (ROC) curve with area under curve (AUC) of the best classification algorithm for multiple sclerosis (MS) diagnosis and MS prognosis using different datasets. The best classifier and several parameters to analyse the model performance for each dataset were shown in Table 4.