Harvard Medical School, Boston, MA, USA.
Brigham Multiple Sclerosis Center & Ann Romney Center for Neurologic Diseases, Department of Neurology, Brigham and Women's Hospital, Boston, MA, USA.
Ann Neurol. 2022 Jul;92(1):87-96. doi: 10.1002/ana.26370. Epub 2022 May 4.
The objective of this study was to identify predictors in common between different clinical and magnetic resonance imaging (MRI) outcomes in multiple sclerosis (MS) by comparing predictive models.
We analyzed 704 patients from our center seen at MS onset, measuring 37 baseline demographic, clinical, treatment, and MRI predictors, and 10-year outcomes. Our primary aim was identifying predictors in common among clinical outcomes: aggressive MS, benign MS, and secondary-progressive (SP)MS. We also investigated MRI outcomes: T2 lesion volume (T2LV) and brain parenchymal fraction (BPF). The performance of the full 37-predictor model was compared with a least absolute shrinkage and selection operator (LASSO)-selected model of predictors in common between each outcome by the area under the receiver operating characteristic curves (AUCs).
The full 37-predictor model was highly predictive of clinical outcomes: in-sample AUC was 0.91 for aggressive MS, 0.81 for benign MS, and 0.81 for SPMS. After variable selection, 10 LASSO-selected predictors were in common between each clinical outcome: age, Expanded Disability Status Scale, pyramidal, cerebellar, sensory and bowel/bladder signs, timed 25-foot walk ≥6 seconds, poor attack recovery, no sensory attacks, and time-to-treatment. This reduced model had comparable cross-validation AUC as the full 37-predictor model: 0.84 versus 0.81 for aggressive MS, 0.75 versus 0.73 for benign MS, and 0.76 versus 0.75 for SPMS, respectively. In contrast, 10-year MRI outcomes were more strongly influenced by initial T2LV and BPF than clinical outcomes.
Early prognostication of MS is possible using LASSO modeling to identify a limited set of accessible clinical features. These predictive models can be clinically usable in treatment decision making once implemented into web-based calculators. ANN NEUROL 2022;92:87-96.
本研究旨在通过比较预测模型，确定多发性硬化症（MS）不同临床和磁共振成像（MRI）结局之间的共同预测因子。
我们分析了 704 名在 MS 发病时来我院就诊的患者，共测量了 37 项基线人口统计学、临床、治疗和 MRI 预测因子，以及 10 年的结局。我们的主要目的是确定常见的临床结局（侵袭性 MS、良性 MS 和继发性进展性 MS）的共同预测因子。我们还研究了 MRI 结局：T2 病变体积（T2LV）和脑实质分数（BPF）。通过受试者工作特征曲线（AUC）下面积，比较完整的 37 个预测因子模型与针对每个结局共同预测因子的最小绝对值收缩和选择算子（LASSO）选择模型的性能。
完整的 37 个预测因子模型对临床结局具有高度预测性：侵袭性 MS、良性 MS 和继发性进展性 MS 的样本内 AUC 分别为 0.91、0.81 和 0.81。经过变量选择，每个临床结局之间有 10 个 LASSO 选择的共同预测因子：年龄、扩展残疾状态量表、锥体束、小脑、感觉和肠/膀胱症状、25 英尺步行时间≥6 秒、攻击恢复差、无感觉攻击和治疗时间。这个简化模型的交叉验证 AUC 与完整的 37 个预测因子模型相当：侵袭性 MS 为 0.84 比 0.81，良性 MS 为 0.75 比 0.73，继发性进展性 MS 为 0.76 比 0.75。相比之下，10 年的 MRI 结局比临床结局更受初始 T2LV 和 BPF 的影响。
使用 LASSO 建模可以识别有限的一组可获得的临床特征，从而对 MS 进行早期预后。这些预测模型可以在纳入基于网络的计算器后在治疗决策中得到临床应用。
