Table 1.
Summary of the studies dealing with prognostication of multiple sclerosis (MS) course.
| Reference | Subjects (Records) | Endpoint. Data Used | Model (Best Performing in Bold) | Most Informative Features for the Best Performing Model | Metrics for the Best Model |
|---|---|---|---|---|---|
| Bejarano, 2011 [37] | 71 + 96 (1/patient) |
ΔEDSS > 1 + EDSS range 2 years later + relapse occurrence. Clinical, MRI, MEPs |
Naïve Bayes, DT, LogR, NN |
EDSS, MEPs |
EDSS range: Acc = 80%, Sens = 92%, Spec = 61% AUC = 76 % ΔEDSS > 1: Acc = 75%, Sens = 87%, Spec = 52%, AUC = 74 Relapses: Acc = 67%, Sens = 53%, Spec = 77%, AUC = 65% |
| Wottschel, 2015 [38] | 74 (1/patient) |
CIS converts to MS in 1 or 3 years. Clinical, MRI |
L-SVM |
CIS to MS at 1 year: lesion load, type of presentation, gender CIS to MS at 3 years: age, EDSS at onset; lesion characteristics: count, average proton density, average distance from brain center, shortest horizontal distance from the vertical axis |
CIS to MS in 1 y: Sens = 77%, Spec = 66% CIS to MS in 3 y: Sens = 60%, Spec = 66% |
| Yoo, 2017 [39] | 140 (1/patient) |
CIS converts to MS in 2 years. Clinical, MRI |
LogR, RF, CNN |
Not assessed | Spec = 70.4%, Sens = 78.7%, Acc = 75.0%, AUC = 74.6% |
| Zhao, 2017 [40] | up to 1693 (1/patient) |
ΔEDSS ≥ 1.5 at 5 years. Clinical, ±MRI |
LogR, L-SVM |
Non progressive cases: EDSS at 0, 6, 12 months; disease activity at 0, 6, 12 months; race, ethnicity, family history, brain parenchymal fraction Progressive cases: ΔEDSS; disease activity; pyramidal function and change at 1 y; disease active at baseline, T2 lesion volume |
Spec = 59%, Sens = 81%, Acc = 67% |
| Law, 2019 [41] | 485 (1/patient) | ΔEDSS ≥ 1 at 2 years in SP MS. Clinical, MRI |
Individual and ensemble LogR, L-SVM, DT, RF, ADB |
EDSS, 9-Hole Peg Test, Timed 25-Foot Walk | Spec = 61% (RF), Sens = 59%, PPV = 32.1%, NPV = 82.8 |
| Seccia, 2020 [42] | up to 1515 (up to 14,923) | RR converts to SP at 0.5 to 2 years. Clinical, ±MRI |
NL-SVM, RF, ADB, KNN, CNN |
Not assessed |
RR to SP at 2 y (RF): Spec = 86.2%, Sens = 84.1%, Acc = 86.2%, PPV = 8.9% RR to SP at 2 y (NN): Spec = 98.5%, Sens = 67.3%, Acc = 98%, PPV =42.7% |
| Brichetto, 2020 [43] | 810 (up to 3398) | RR converts to SP within 4 months. Clinical, patient reported outcomes |
LogR, L-SVM, KNN and other linear classifiers |
Not reported | Acc = 82.6% |
| Zhao, 2020 [44] | 724 (CLIMB dataset) + 400 (EPIC dataset) (1/patient) | ΔEDSS ≥ at 5 years. Clinical, MRI |
LogR, L-SVM, ensemble models (RF, boosting methods) |
Value at a given time or change in 2 years of: EDSS, pyramidal function, disease category (RR, SP etc.), MRI lesions, ambulatory index, cerebellar function |
CLIMB dataset, XGBoost Spec = 69%, Sens = 79%, Acc = 71%, AUC = 78% |
| Pinto, 2020 [45] | up to 187 | RR to SP @ 5 years and EDSS > 3 at 6 or 10 years. Clinical, MRI | KNN, DT, LogR, L-SVM |
SP development: EDSS, FS scores (sensory, brainstem, cerebellar and mental), CNS involvement in relapses (pyramidal tract, neuropsychological and brainstem), age at onset. Disease severity:EDSS, FS scores and CNS affected functions during relapses |
RR to SP: Spec = 77%, Sens = 76%, AUC = 86%, EDSS > 3 @ 6 y: Spec = 81%, Sens = 84%, AUC = 89% EDSS > 3 at 10 y: Spec = 79%, Sens = 77%, AUC = 85% |
Abbreviations used: ADB: Adaboost; Acc: accuracy; AUC: area under the receiver-operated curve; CNN: convolutional neural network; CNS: central nervous system; DT: decision tree; EDSS: Expanded Disability Status Scale; FS: functional system; KNN: k-nearest neighbors; LogR: logistic regression; MEP: motor-evoked potential; NN: neural network; RF: random forest; Sens: sensitivity; Spec: specificity; (L-, NL-)SVM: (linear, nonlinear) support vector machine; y: year(s); ±MRI: MRI data not always used; ΔEDSS: change of EDSS.