The Prognostic Value of Paramagnetic Rim Lesions in People With Multiple Sclerosis

Abstract

Background and Objectives

Paramagnetic rim lesions (PRLs) have been established as a radiologic surrogate for chronic active inflammation in multiple sclerosis (MS). Although associations between PRLs and disability have been observed, data on their prognostic value for long-term risk of clinical worsening, especially in the form of progression independent of relapse activity (PIRA), remain limited. In this study, we assessed the association of PRLs with future disability accumulation over a median follow-up of 5 years and their relation to other (para) clinical parameters.

Methods

From a large, ongoing single-center cohort study, people with MS were included according to the availability of susceptibility-sensitive MRI. PRLs were segmented on reconstructed quantitative susceptibility mapping scans using a deep learning–based segmentation model, visually inspected for plausibility, and corrected if necessary. Multivariable Cox proportional hazards models were used to investigate the prognostic value of PRLs for disability outcomes, including confirmed disability accumulation (CDA) and PIRA. Sensitivity analyses based on PRL count, MS diagnosis, prior disability accumulation, and treatment during follow-up were performed to evaluate the robustness of our findings. In addition, associations between PRL status and CSF parameters as well as future brain atrophy, measured by annual percentage brain volume change, were assessed in an exploratory manner.

Results

Among 862 participants (mean age 40.3 years, 64.7% female), 395 (45.8%) had at least 1 PRL. Over a median follow-up of 5.1 years, 142 of 801 individuals with clinical follow-up for longitudinal analysis experienced CDA (17.7%), and 104 of these CDA events (73.2%) occurred as PIRA. The presence of PRLs was associated with a higher risk of PIRA (adjusted hazard ratio 1.8, 95% CI 1.1–2.8; p = 0.02). Exploratory subgroup analyses showed increased markers of chronic intrathecal inflammation in individuals with PRLs as well as a correlation between the number of PRLs and future brain atrophy (PBVC: β = −0.012%, 95% CI −0.018% to −0.005%; p < 0.001).

Discussion

PRLs are associated with an increased risk of future PIRA and accelerated rates of brain atrophy. These findings emphasize their potential as a prognostic biomarker for risk stratification in MS.

Introduction

White matter lesions are the pathophysiologic hallmark of multiple sclerosis (MS), an inflammatory and neurodegenerative disorder of the CNS. Although the formation of new lesions is typically linked to clinical relapses, recent pathologic and radiologic studies have uncovered a number of potential long-term lesion trajectories, comprising remyelinating, chronic inactive, and chronic active lesions.¹ Within the latter group, paramagnetic rim lesions (PRLs) have emerged as a highly MS-specific substrate of smoldering CNS inflammation.²

As their name suggests, PRLs are characterized by a ring of iron-laden (micro)glia and macrophages surrounding a largely demyelinated lesion core, rendering them visible on susceptibility-sensitive MRI.^3-5 Although earlier studies relied on high-field MRI, detection rates are comparable in clinical 3-T settings.⁶ Quantitative susceptibility mapping (QSM), a technique derived from raw phase images through a series of advanced postprocessing steps, has proven both sensitive and specific for this purpose.^7,8

The diagnostic utility of PRLs as a specific marker of MS pathology led to their inclusion in the recently revised McDonald criteria and associated Magnetic Resonance Imaging Network in Multiple Sclerosis (MAGNIMS)–Consortium of Multiple Sclerosis Centers (CMSC)–North American Imaging in Multiple Sclerosis Cooperative (NAIMS) consensus recommendations.^9,10 Aside from this, the presence and quantity of PRLs have been linked to worse clinical outcomes.^11-15 A handful of studies specifically examined the association with progression independent of relapse activity (PIRA), the presumed clinical correlate of neurodegenerative processes possibly driven by ongoing CNS inflammation. However, the existing evidence is characterized by several limitations. Small cohort sizes with few progression events lead to considerable statistical uncertainty in some studies.^16-18 One group relied on standard susceptibility-weighted imaging (SWI) for PRL identification,¹⁹ which is known to lack sensitivity compared with QSM.²⁰ In the largest study on PRLs and PIRA to date, disability progression events were recorded before QSM scans, so no prognostic information can be extrapolated.²¹ Finally, one retrospective study with a mean follow-up of 10 years did not find an association between the presence of PRLs and subsequent PIRA.²²

The aim of this study was to investigate the prognostic value of PRLs with respect to future disability worsening in MS. To this end, we identified PRLs on baseline QSM imaging and assessed their association with confirmed disability accumulation (CDA) and more specifically PIRA over a median follow-up of 5 years. Exploratory analyses examined the association of PRLs with CSF markers of intrathecal inflammation and with rates of future brain atrophy in the form of annual percentage brain volume change (PBVC).

Methods

Participants

Inclusion criteria for this retrospective, single-center study based on a prospectively recruited and maintained cohort were: (1) diagnosis of relapsing-remitting MS, primary/secondary progressive MS, or clinically isolated syndrome according to the 2017 revisions of the McDonald criteria²³; (2) age ≥18 years; (3) availability of multi-echo gradient echo (GRE) MRI for QSM reconstruction (earliest eligible scan selected); and (4) availability of data to confirm PRL chronicity according to NAIMS and MAGNIMS consensus recommendations.^1,10 Exclusion criteria were: (1) significant imaging artifacts; (2) corticosteroid treatment within 1 month before baseline MRI; and (3) different scanner/scan parameters. Additional subanalyses required: (1) availability of at least 2 clinical follow-up visits (≥ 6 months apart) with Expanded Disability Status Scale (EDSS) data for the CDA and PIRA analysis; (2) availability of CSF data including white cell count, oligoclonal bands (OCBs), immunoglobulin (Ig), and albumin measures; and (3) availability of a follow-up T1-weighted (T1w) scan ≥ 1 year after the QSM scan to determine annual PBVC (latest eligible scan selected). QSM imaging was performed from 2019 to 2020; clinical data were collected between 2009 and 2025.

Clinical Scores and Definition of Disability Accumulation

Baseline EDSS scores were determined at the time of the QSM scan. In case of active relapses, scores after clinical resolution/stabilization were used. CDA was defined as an EDSS increase of 1.5 for baseline EDSS scores of 0, an increase of 1 for baseline EDSS scores between 1.0 and 5.0, and an increase of 0.5 for baseline EDSS scores of 5.5 or higher; the increase had to be confirmed on clinical follow-up over at least 6 months. PIRA events required both the initial and confirmatory visits to occur outside a window of 30 days before and 90 days after the onset of a relapse event. Otherwise, CDA was classified as relapse-associated worsening (RAW). In addition, a roving baseline approach was applied that sets a new reference score every time the EDSS decreases below the previous score (confirmed at the subsequent visit) or a CDA event is registered.²⁴ Manual verification of CDA events was performed by 1 experienced investigator (M.L.) while blinded to imaging and CSF data.

MRI Acquisition and Processing

Scans were acquired on a single 3-T scanner (Ingenia, Philips Healthcare, Best, the Netherlands) with a 32-channel head coil. No hardware or major software upgrades were implemented during the entire follow-up period. The protocol consisted of 3 main sequences: (1) 3D T1w magnetization-prepared rapid GRE (MPRAGE): repetition time (TR) = 9 ms, echo time (TE) = 4 ms, flip angle (FA) = 8°, voxel size = 0.75 × 0.75 × 0.75 mm³, plane = sagittal; (2) 3D fluid-attenuated inversion recovery (FLAIR): TR = 4,800 ms, TE = 317 ms, inversion time (TI) = 1,650 ms, FA = 90°, voxel size = 0.75 × 0.75 × 0.75 mm³, plane = sagittal; and (3) 3D multi-echo GRE sequence (for QSM reconstruction): TR = 41.75 ms, first TE = 6 ms, echo spacing = 6.4 ms, 6 monopolar echoes, FA = 20°, field of view = 230 × 187 × 170, acquired matrix = 328 × 267, acquisition voxel size = 0.7 × 0.7 × 1.5 mm³, reconstructed voxel size = 0.5 × 0.5 × 0.75 mm³, pixel bandwidth = 217 Hz/pixel, flow compensation = first echo, CS-SENSE = 8, elliptical k-space shutter = yes, scan duration = 3 minutes, plane = axial. Additional gadolinium-enhanced T1w scans were acquired with the same MPRAGE sequence.

SWI was processed using the scanner's standard reconstruction pipeline, whereas QSM was generated using a multi-echo total field inversion algorithm. This method directly models the complex multi-echo signal through non-convex optimization and total variation regularization, thereby bypassing the systematic errors typically associated with conventional, sequential QSM processing steps. This approach yields more robust QSM images that are less prone to artifacts than alternative methods. Further details on QSM reconstruction and pipeline comparisons were previously described.²⁵

Lesion Segmentation and Image Analysis

After skull-stripping using High-Definition Brain Extraction (HD-BET),²⁶ T1w, FLAIR, and SWI sequences were rigidly coregistered to QSM space using Greedy,²⁷ with the first-echo magnitude image as reference. PRLs were initially manually delineated on QSM parameter maps, supported by SWI for improved vascular contrast. NAIMS and MAGNIMS consensus recommendations were adhered to for the classification of PRLs.^1,10 These segmentations served as ground truth for an nnU-Net–based segmentation model trained on 127 multimodal data sets, comprising QSM, SWI, T1w, and FLAIR sequences.²⁸ Model training was performed in 3 iterations, using progressively larger subsets (43, 86, and finally 127). Following each iteration, the model-generated PRL masks were visually inspected and corrected to align with the consensus criteria. Uncertain cases (n = 33 lesions) were jointly reviewed and discussed until a consensus was reached. The refined masks were then reintroduced into the model for subsequent training. A total of 178 (20.6%) of the final PRL masks required manual correction, primarily for the removal of false-positive PRLs and edge refinement of existing PRLs. A conservative approach was aded when delineating and identifying PRLs, guided by lesion boundaries defined using T1w MRI. Only in cases where the rim clearly extended beyond the T1w hypointense region, the corresponding hyperintense region on FLAIR was used to guide rim delineation. If multiple distinct rims were evident within a single confluent lesion complex, they were counted as independent PRLs. To confirm PRL chronicity, reference was made to either (1) a concurrent gadolinium-enhanced T1w sequence to verify nonenhancement, (2) a FLAIR scan obtained at least 6 months earlier to evaluate prior lesion existence, or (3) a follow-up QSM scan demonstrating PRL persistence over at least 6 months (Figure 1 for examples of true PRLs and false positives). PRL delineation and classification was performed by C.V. while blinded to clinical data and under the supervision of an experienced neuroradiologist (D.S.) and neurologist (M.M).

Figure 1. Identification Examples of PRLs and False Positives.

(A) Example of a PRL displaying hyperintense rims on QSM across 3 orthogonal planes (axial, sagittal, and coronal), consistent with NAIMS criteria. (B and C) Examples of false-positive PRLs (marked in red), identified by the automatic segmentation nnU-Net model. The enlarged QSM and SWI views show that (B) represents a vein-driven hyperintensity and (C) corresponds to hemosiderin-related susceptibility without distinct rim or hypointense core, both excluded from PRL classification. (D) Example of confirmed chronic PRL showing a persistent lesion on prior FLAIR imaging, a hyperintense rim on QSM, a subtle hypointense signal on SWI, and colocalization with a T1w hypointense lesion core. (E) Example of a “possible PRL” exhibiting a hyperintense rim on QSM, a corresponding hypointense rim on SWI, and lesion colocalization on T1w. However, the lesion was absent on the prior FLAIR scan, leading to exclusion from the confirmed PRL set. Red regions indicate the overlaid lesion mask used for reference. FLAIR = fluid-attenuated inversion recovery; NAIMS = North American Imaging in Multiple Sclerosis; PRL = paramagnetic rim lesion; QSM = quantitative susceptibility mapping; SWI = susceptibility-weighted imaging; T1w = T1-weighted.

Brain lesions were segmented from FLAIR and T1w images using Lesion Segmentation Toolbox (LST)-AI (version 1.1.0).²⁹ PBVC was calculated from 2 T1w scans (baseline and latest available scan for each individual with an inter-scan interval ≥ 1 year) using the Structural Image Evaluation, using Normalization, of Atrophy (SIENA) software (version 6.04).³⁰ Normalized brain volumes were extracted using the Sequence Adaptive Multimodal SEGmentation (SAMSEG) tool integrated into Freesurfer (version 7.3.2).³¹

CSF Parameters

Lumbar puncture was performed at the time of initial diagnosis in the majority of cases (84.8%). Where multiple CSF samples were available, the one collected closest to the baseline imaging date was selected to minimize the temporal interval. CSF and serum concentrations for albumin as well as IgG, IgM, and IgA were measured in parallel by standard nephelometric assays, and the respective CSF-to-serum quotients (Q_IgG, Q_IgM, Q_IgA, and Q_alb) were calculated. Ig indices were defined as Q_IgX/Q_alb. Intrathecal Ig synthesis was determined following the formulae proposed by Reiber.³² Detection of OCBs was performed using isoelectric focusing followed by immunoblotting or immunofixation. OCBs were considered positive if patterns 2 or 3, according to the 2005 consensus statement, and 3 or more CSF-specific bands were present.³³

Statistical Analyses

Between-group comparisons were performed using Fisher exact tests for categorical variables. For continuous variables, two-group comparisons were conducted using two-sided independent t tests for normally distributed data and Wilcoxon-Mann-Whitney tests for non-normally distributed data. For comparisons involving more than 2 groups, 1-way analysis of variance was used for normally distributed variables, while Kruskal-Wallis H tests were applied in cases of non-normality. Where significant group differences were identified, Dunn post hoc tests with Holm-Bonferroni correction were performed to identify specific pairwise differences. Continuous variables are given as mean with SD if normally distributed, otherwise as median with interquartile range (IQR). Aside from the presence or absence of PRLs, 2-group comparisons were also performed using a threshold of 4 PRLs, a value previously found to yield a good trade-off between sensitivity and specificity.^{11,13,16,19,22}

The prognostic value of PRLs for disability outcomes was estimated using time-to-event Cox proportional hazards models. Six baseline covariates were chosen based on data from previous large-scale disability outcome studies^34–38: age, disease duration, sex, EDSS, T2 lesion volume (T2LV), and high-efficacy therapy (HET) during the majority of the 6 months preceding imaging (as a dichotomous variable; Table 1 for substances included in this category). T2LV was log-transformed to account for its right-skewed distribution. Variance inflation factors (VIFs) were calculated to assess potential collinearity. Model estimates are given as adjusted hazard ratios (aHRs) with 95% CIs. Participants lost to follow-up or who did not experience an event by the end of the study period were censored at the date of their last clinical assessment. Sensitivity analyses were conducted in subgroups defined by PRL count, MS diagnosis, and prior disability accumulation. An additional sensitivity analysis was performed with respect to treatment category at baseline and during the entire follow-up including potential switches (no therapy/platform therapy to HET or vice versa). For this purpose, Cox proportional hazards models with HET status as a time-varying covariate were used.

Table 1.

Baseline Characteristics of Study Participants

	Total (N = 862)	PRL− (n = 467)	PRL+ (n = 395)	p Value
Age (y), mean (SD)	40.3 (11.2)	40.3 (11.3)	40.3 (11.0)	0.95
Sex (female), n (%)	558 (64.7)	315 (67.5)	243 (61.5)	0.07
EDSS, median (IQR)	1.5 (0–2.0)	1.0 (0–2.0)	1.5 (1.0–2.5)	<0.001a
Disease duration (y), median (IQR)	5.1 (1.7–9.2)	4.5 (1.5–8.2)	5.4 (1.9–9.7)	0.03a
MS type
CIS, n (%)	40 (4.6)	33 (7.1)	7 (1.8)	<0.001a
RRMS, n (%)	754 (87.5)	406 (86.9)	348 (88.1)	0.68
PMS, n (%)	68 (7.9)	28 (6.0)	40 (10.1)	0.03a
Imaging parameters
Number of T2 lesions, median (IQR)	20 (9–37)	13 (6–25)	30 (17–49)	<0.001a
T2 lesion volume (mm3), median (IQR)	2,268 (686–7,418)	998 (390–2,704)	5,767 (2,257–14,201)	<0.001a
Number of PRLs, median (IQR)	0 (0–2)	0	2 (1–4)	—
Disease-modifying therapy, n (%)
Time on latest therapy (y), median (IQR)	1.9 (0.8–3.9)	1.9 (0.9–4.0)	1.8 (0.7–3.9)	0.31
High-efficacy therapy, n (%)	264 (30.6)	107 (22.9)	157 (39.7)	<0.001a
S1P receptor modulators, n (% of group)	114 (43.2)	53 (49.5)	61 (38.9)
Anti-CD20, n (% of group)	112 (42.4)	42 (39.3)	70 (44.6)
Natalizumab, n (% of group)	31 (11.7)	10 (9.3)	21 (13.4)
Others, n (% of group)	7 (2.7)	2 (1.9)	5 (3.2)
Platform therapy, n (%)	375 (43.5)	227 (48.6)	153 (38.7)	0.007a
Dimethyl fumarate, n (% of group)	125 (33.3)	70 (30.8)	55 (35.9)
Interferon, n (% of group)	112 (29.9)	70 (30.8)	42 (27.5)
Glatiramer acetate, n (% of group)	111 (29.6)	73 (32.2)	38 (24.8)
Teriflunomide, n (% of group)	27 (7.2)	10 (4.4)	17 (11.1)
None, n (%)	223 (25.9)	137 (29.3)	86 (21.8)	0.01a

Abbreviations: CIS = clinically isolated syndrome; EDSS = Expanded Disability Status Scale; IQR = interquartile range; n = number; MS = multiple sclerosis; PMS = progressive multiple sclerosis; PRL = paramagnetic rim lesion; RRMS = relapsing-remitting multiple sclerosis; S1P receptor modulators = sphingosine 1-phosphate receptor modulators.

Disease-modifying therapies add up to 100% within each group. Two-sided Fisher exact tests were performed for categorical variables; two-sided, independent t tests were performed for continuous variables if normally distributed, otherwise Wilcoxon-Mann-Whitney tests were chosen.

^aSignificant p values.

Associations between PRL measures and subsequent brain atrophy were assessed through linear regression models. In addition to the 6 covariates from the Cox proportional hazards analyses, baseline normalized brain volume was included in the models because it is known to influence PBVC measures.³⁹ Model estimates are given as β coefficients with 95% CIs.

The study follows the “STrengthening the Reporting of OBservational studies in Epidemiology” (STROBE) reporting guidelines.⁴⁰

Two-sided p values <0.05 were considered statistically significant. All analyses were performed in R (version 4.4.0), using the survival (version 3.8-6), survminer (version 0.5.2), and tidyverse (version 2.0.0) packages.

Standard Protocol Approvals, Registrations, and Patient Consents

This study involves human participants and was approved by the internal review board of the Technical University of Munich (TUM) (reference number: 5848/13). Participants gave written informed consent to use their data for research purposes.

Data Availability

The anonymized data that supported the findings of this study are available upon reasonable request.

Results

Sample Characteristics

A flowchart of patient selection and analysis groups is shown in Figure 2. Of the 862 patients included in the study (mean age 40.3 years), 558 (64.7%) were women. There were 395 (45.8%) patients who had at least 1 PRL on imaging and 116 (13.5%) had 4 PRLs or more. As summarized in Table 1, individuals with PRLs had longer disease courses (median 5.4 vs 4.5 years), higher EDSS scores (median 1.5 vs 1.0), and higher total T2 lesion counts (median 30 vs 13) and volumes (median 5,767 vs 998 mm³). In addition, the presence of PRLs was associated with an increased likelihood of receiving HET, particularly anti-CD20 agents, at the time of MRI (39.7% vs 22.9%). There was no age difference between groups based on PRL status (p = 0.95).

Figure 2. Participant Flowchart.

The flow of the primary disability progression analysis is indicated by solid blue boxes, while exploratory subcohorts are denoted by white boxes. CIS = clinically isolated syndrome; FLAIR = fluid-attenuated inversion recovery; MS = multiple sclerosis; PBVC = percentage brain volume change; PRL = paramagnetic rim lesion; QSM = quantitative susceptibility mapping; T1w = T1-weighted.

The subgroup with available CSF data (n = 541) was representative of the full cohort in terms of age (mean 39.5 years), sex (64.7% female), and PRL prevalence (44.5% with at least 1 PRL). CSF collection preceded imaging by a median of 2.4 (IQR 0.6–6.1) years, without significant differences between individuals with and without PRLs (p = 0.19). Patients with PRLs had higher albumin quotients (median 5.6 vs 5.0) and IgG indices (median 0.74 vs 0.63) than those without. Moreover, intrathecal IgG synthesis (52.3% vs 41.7%) and positive OCBs (93.8% vs 81.0%) were more frequently found in individuals with PRLs (eTable 1).

Association of PRLs With Disability Worsening During Follow-Up

Clinical follow-up was available for 801 participants (92.9%), with a median follow-up of 5.1 years for both the CDA (IQR 3.6–5.9) and PIRA (IQR 4.0–5.9) analyses and a median visit density of 2.0 (IQR 1.4–2.6) per year (including baseline). A total of 142 patients (17.7%) experienced CDA, mainly in the form of PIRA (104 events, 73.2% of all CDA). The presence of PRLs on baseline MRI was a significant predictor of CDA (HR 2.3, 95% CI 1.7–3.3; p < 0.001), even after adjusting for age, disease duration, sex, baseline EDSS, T2LV, and HET status at baseline (aHR 1.6, 95% CI 1.1–2.3; p = 0.03; Figure 3A and eTable 2). This association was predominantly driven by PIRA (aHR 1.8, 95% CI 1.1–2.8; p = 0.02; Figure 3B and eTable 3; RAW: aHR 1.2, 95% CI 0.5–2.5; p = 0.71; eFigure 1). Correlations among model parameters did not raise concerns regarding collinearity (each VIF <2). Using a threshold of 4 or more PRLs yielded similar results for CDA (aHR 1.7, 95% CI 1.1–2.6; p = 0.01; Figure 3A) and PIRA outcomes (aHR 1.9, 95% CI 1.1–3.1; p = 0.01; Figure 3B). For each additional PRL, the adjusted hazard increased by 4.5% for CDA (aHR 1.045, 95% CI 1.004–1.087; p = 0.03) and by 6.8% for PIRA (aHR 1.068, 95% CI 1.021–1.117; p = 0.004).

Figure 3. Survival Curves for Disability Accumulation Outcomes by PRL Status (2 Subgroups).

(A) Risk of CDA by PRL status. (B) Risk of PIRA by PRL status. Cox proportional hazards models are adjusted for age, disease duration, sex, Expanded Disability Status Scale score, T2 lesion volume, and high-efficacy therapy. The y-axis has been cropped for better discrimination of the curves. aHR = adjusted hazard ratio; CDA = confirmed disability accumulation; PIRA = progression independent of relapse activity; PRL = paramagnetic rim lesion; QSM = quantitative susceptibility mapping.

Sensitivity Analyses

To evaluate the consistency of our findings, several sensitivity analyses were conducted. We first divided the cohort into 3 subgroups to assess the effect of PRL burden granularity on disability outcomes: 0 PRLs, 1–3 PRLs, and ≥4 PRLs (eTable 4). This revealed a stepwise increase in hazard across subgroups, with a significant linear trend for both the CDA (1–3 PRLs: aHR 1.4, 95% CI 0.9–2.2; ≥4 PRLs: aHR 2.2, 95% CI 1.3–3.6; p_trend = 0.004; Figure 4A) and PIRA end points (1–3 PRLs: aHR 1.7, 95% CI 1.0–2.8; ≥4 PRLs: aHR 2.7, 95% CI 1.4–5.0; p_trend = 0.002; Figure 4B).

Figure 4. Survival Curves for Disability Accumulation Outcomes by PRL Status (3 Subgroups).

(A) Risk of CDA by PRL status. (B) Risk of PIRA by PRL status. Cox proportional hazards models are adjusted for age, disease duration, sex, Expanded Disability Status Scale score, T2 lesion volume, and high-efficacy therapy. The y-axis has been cropped for better discrimination of the curves. CDA = confirmed disability accumulation; PIRA = progression independent of relapse activity; PRL = paramagnetic rim lesion; QSM = quantitative susceptibility mapping.

In a subanalysis of 704 patients with a diagnosis of relapsing-remitting MS at baseline, the associations between PRL presence and future CDA (aHR 2.2, 95% CI 1.3–3.5; p = 0.001) or PIRA (aHR 3.2, 95% CI 1.7–6.1; p < 0.001) remained significant. Similar results were obtained using a ≥4 PRLs threshold (CDA: aHR 2.1, 95% CI 1.3–3.5; p = 0.004; PIRA: aHR 2.4, 95% CI 1.3–4.6; p = 0.006). In contrast, among the 61 patients with progressive MS, no significant association was observed between PRLs and disability accumulation (which occurred exclusively as PIRA) for either threshold (≥1 PRLs: p = 0.15; ≥4 PRLs: p = 0.98). Finally, in a subgroup of 159 patients imaged within 1 year of MS diagnosis, only the ≥4 PRLs threshold significantly predicted CDA (aHR 3.5, 95% CI 1.0–12.8; p = 0.05) and PIRA (aHR 7.2, 95% CI 1.4–37.7; p = 0.02).

After excluding any individuals with documented CDA (697 remaining participants for the CDA end point) or PIRA events (749 remaining participants for the PIRA end point) before baseline MRI, results did not change appreciably for the PIRA end point (PIRA: aHR 2.3, 95% CI 1.3–4.0; p = 0.003), whereas significance was lost for CDA (aHR 1.5, 95% CI 0.9–2.4; p = 0.09). Both end points showed significant associations for a threshold of ≥4 PRLs (CDA: aHR 1.8, 95% CI 1.1–2.9; p = 0.02; PIRA: aHR 2.2, 95% CI 1.2–3.8; p = 0.007).

Treatment escalation to HET during follow-up occurred in 115 of 547 patients (21.0%) for the CDA end point and 124 patients (22.7%) for the PIRA end point. Treatment de-escalation from HET during follow-up occurred in 43 of 254 patients (16.9%) for both the CDA and PIRA outcomes. Taking into account these treatment switches did not change the associations between PRLs and disability outcomes in any relevant way for either the ≥1 PRLs threshold (CDA: aHR 1.6, 95% CI 1.1–2.3; p = 0.02; PIRA: aHR 1.8, 95% CI 1.1–2.9; p = 0.02) or the ≥4 PRLs threshold (CDA: aHR 1.7, 95% CI 1.1–2.6; p = 0.01; PIRA: aHR 1.9, 95% CI 1.1–3.1; p = 0.01).

PRLs and Subsequent Brain Atrophy

Follow-up MRI taken a median of 4.1 (IQR 3.6–4.6) years after baseline was available for 747 participants (86.7%), with no significant differences in interval between participants with and without PRLs (p = 0.57). Annual PBVC was more pronounced in people with PRLs (median −0.29%, IQR −0.44% to −0.14% vs −0.23%, IQR −0.37% to −0.11%; p = 0.001). Of note, PRL count explained nearly 4 times more variance in annual PBVC than the binary presence or absence of PRLs (R² 0.013 vs 0.048; p < 0.001). Consequently, the difference in atrophy rates was emphasized when comparing individuals with 4 or more PRLs to those with fewer lesions (median −0.33%, IQR −0.50% to −0.14% vs −0.24%, IQR −0.40% to −0.12%; p = 0.005). After adjusting for age, disease duration, sex, EDSS, T2LV, baseline HET status, and baseline normalized brain volume, only the number of PRLs remained as a significant predictor of PBVC (β = −0.012%, 95% CI −0.018% to −0.005%; p < 0.001).

Discussion

In this large cohort study of more than 800 people with MS, chronic active brain lesions in the form of PRLs were associated with both cross-sectional and longitudinal measures of disease severity. Over a median 5-year follow-up, the presence and number of PRLs significantly predicted CDA, primarily driven by PIRA. In addition, individuals with PRLs were characterized by increased markers of chronic intrathecal inflammation, and higher PRL counts were related to accelerated brain atrophy.

Our observed PRL prevalence (per-patient: 0.46; per-lesion: 0.06) is at the lower end of estimates given by a recent meta-analysis (per-patient: 0.52, 95% CI 0.47–0.58; per-lesion: 0.12, 95% CI 0.09–0.16).⁴¹ This likely reflects our conservative approach to PRL classification. Moreover, 4 studies evaluated for the meta-analysis reported disproportionately high per-lesion prevalences (between 0.49 and 0.53), skewing the results toward higher numbers. Finally, the proportion of patients with progressive MS in our cohort (7.9%) was relatively low compared with similar studies.^11,21,22 This might have been another contributing factor because progressive MS is characterized by relatively higher PRL prevalences.⁴¹

Individuals with PRLs experienced PIRA at nearly double the rate of those without after adjusting for potential confounders—a finding that remained largely consistent across different lesion thresholds and sensitivity analyses. The percentages of patients with CDA (17.7%) and PIRA (13.0%) in our cohort were comparable with another larger study with similar follow-up times (n = 445, CDA: 20.4%, PIRA: 16.6%).²¹ There was a trend toward higher HRs when choosing a previously suggested threshold of ≥4 PRLs compared with a present/absent classification. This became especially apparent in sensitivity analyses with low sample sizes, where significance was retained for a ≥4 PRLs threshold in each case, while this was not always true for a present/absent classification. Supporting this observation, further subdivision into groups with 0, 1–3, and 4 PRLs yielded a stepwise increase in hazard across the 3 groups, with a significant linear trend. Although this trend was investigated and partially confirmed in other studies,^11,22 the size of our cohort allowed for sufficient statistical power to perform reliable subgroup analyses. Notably, the only study that failed to find an association between the presence of PRLs and subsequent PIRA reported a significant association with the number of PRLs.²² Taken together, the observation that PRL count provides incremental prognostic value over a binary (present/absent) classification suggests that the cumulative burden of chronic active inflammation is a key driver of disability in MS. This has practical implications for clinical risk stratification: rather than a one-size-fits-all approach, different PRL thresholds may be required depending on the clinical objective. For instance, a lower threshold might be preferred for high-sensitivity screening to identify at-risk patients, whereas a higher threshold might be necessary to achieve the specificity required for treatment escalation or enrichment in clinical trials.

We investigated associations of PRLs with established markers of disease severity (CSF measures and brain atrophy) in an exploratory fashion. Previous studies on the relation between PRLs and CSF measures have reported conflicting results: One study found an association of PRLs with increased albumin quotients and OCBs but not IgG synthesis,⁴² while a second study reported the opposite pattern.⁴³ We observed both higher rates of CSF-specific OCBs and IgG synthesis as well as higher albumin quotients in individuals with PRLs, which suggests that PRLs may be associated with a more inflammatory phenotype. Since existing data on PRLs and CSF measures (including ours) are merely associative in nature, however, these findings must be interpreted with caution and warrant additional systematic investigation.

Data on the relationship between PRLs and brain atrophy rates are similarly conflicting. One study found a significantly higher annual PBVC among individuals with PRLs,⁴⁴ while another did not.¹² A third study observed a trend of more severe atrophy in patients that developed new PRLs over time.¹⁸ In our cohort, PRLs were associated with higher rates of brain atrophy, although only the number of PRLs survived adjusting for potential confounders, once again indicating that total chronic active lesion burden is of prognostic relevance. However, there are some caveats when interpreting these findings: Absolute rates of annual PBVC were low in our cohort (median −0.25%, IQR −0.41% to −0.12%), bordering those seen in healthy aging.⁴⁵ And while relative between-group differences were notable (26% for a ≥1 PRLs threshold, 38% for a ≥4 PRLs threshold), the long-term clinical impact of these observations remains unclear. Moreover, factors not controlled for in our analysis could have influenced the rate of atrophy: There might have been a higher rate of silent T2 lesion accumulation in individuals with PRLs during follow-up, causing secondary atrophy. In addition, a recent meta-analysis demonstrated that different disease-modifying therapies exert varying effects on brain atrophy—a factor that might have influenced outcomes in our cohort.⁴⁶

Our study has some limitations. The availability of only a single QSM scan per person precluded us from assessing the dynamics of PRL development and disappearance, which have previously been shown to have significant effects on disability worsening.¹⁸ In addition, the associations between PRLs and markers of chronic intrathecal inflammation have to be interpreted with caution. Although CSF immunoglobulin measures are assumed to be relatively stable, the heterogeneous intervals between CSF acquisition and MRI—and the potential effects of disease-modifying therapy on parameters of intrathecal inflammation—might have influenced our observations. Finally, despite our large cohort size and robust sensitivity analyses, the single-center nature of our study potentially weakens the generalizability of our findings.

In conclusion, the presence and number of PRLs was found to have a negative prognostic effect on subsequent disability accumulation, mainly in the form of PIRA. This relationship is supported by associations of PRLs with established MRI (brain atrophy) and CSF markers of chronic CNS inflammation. Taken together, these observations add to the existing evidence for PRLs as a radiologic surrogate for chronic active inflammation in MS and emphasize their potential for prognostic risk stratification.

Glossary

aHR

adjusted hazard ratio

CDA

confirmed disability accumulation

CMSC

Consortium of Multiple Sclerosis Center

EDSS

Expanded Disability Status Scale

FA

flip angle

FLAIR

fluid-attenuated inversion recovery

GRE

gradient echo

HD-BET

High-Definition Brain Extraction

HET

high-efficacy therapy

Ig

immunoglobulin

IQR

interquartile range

MAGNIMS

Magnetic Resonance Imaging Network in Multiple Sclerosis

MPRAGE

magnetization-prepared rapid GRE

MS

multiple sclerosis

NAIMS

North American Imaging in Multiple Sclerosis Cooperative

OCB

oligoclonal band

PBVC

percentage brain volume change

PIRA

progression independent of relapse activity

PRL

paramagnetic rim lesion

QSM

quantitative susceptibility mapping

RAW

relapse-associated worsening

SAMSEG

Sequence Adaptive Multimodal SEGmentation

SWI

susceptibility-weighted imaging

T1w

T1-weighted

T2LV

T2 lesion volume

TE

echo time

TI

inversion time

TR

repetition time

TUM

Technical University of Munich

VIF

Variance inflation factor

Author Contributions

M. Lauerer: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design; analysis or interpretation of data. C.C. Voon: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design; analysis or interpretation of data. J. McGinnis: major role in the acquisition of data; study concept or design. D. Sepp: major role in the acquisition of data; study concept or design. J. Meineke: major role in the acquisition of data; study concept or design. T. Wiltgen: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. R.C. Berg: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. C. Preibisch: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. B. Wiestler: drafting/revision of the manuscript for content, including medical writing for content; study concept or design. C. Engl: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. P. Uibel: major role in the acquisition of data. A. Berthele: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. J.S. Kirschke: major role in the acquisition of data. B. Hemmer: drafting/revision of the manuscript for content, including medical writing for content. M. Mühlau: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design; analysis or interpretation of data.

Study Funding

M. Lauerer has received funding through the “Kommission für klinische Forschung” (KKF) of the School of Medicine and Health, Technical University of Munich (TUM). M. Mühlau was supported by research grant 428223038 of the German Research Foundation, DFG Priority Programme 2177, Radiomics: Next Generation of Biomedical Imaging. B. Hemmer was supported by the Munich Cluster for Systems Neurology (SyNergy).

Disclosure

M. Lauerer, C.C. Voon, J. McGinnis, D. Sepp, J. Meineke, T. Wiltgen, R.C. Berg, C. Preibisch, B. Wiestler, C. Engl, and P. Uibel report no disclosures relevant to the manuscript; A. Berthele has received consulting and/or speaker fees from Alexion, Biogen, Celgene, Horizon, Novartis, Roche and Sandoz/Hexal, his institution has received compensation for clinical trials from Alexion, Biogen, Merck, Novartis, Roche and Sanofi Genzyme; J.S. Kirschke has received speaker fees for Novartis, he is a shareholder of Bonescreen. B. Hemmer has served on scientific advisory boards for Novartis, he has served as DMSC member for AllergyCare, Sandoz, Polpharma, Biocon and TG therapeutics, his institution received research grants from Roche for MS research, he has received honoraria for counselling (Gerson Lehrmann Group), he holds part of 2 patents, 1 for the detection of antibodies against KIR4.1 in a subpopulation of patients with MS and 1 for genetic determinants of neutralising antibodies to interferon; M. Mühlau has received research support from the German Research Foundation (DFG Priority Programme 2177, Radiomics: Next Generation of Biomedical Imaging, grant 428223038), the Bavarian State Ministry for Science and Art (Collaborative Bilateral Research Program Bavaria–QC: AI in medicine, grant F.4-V0134. K5.1/86/34), the German Federal Ministry of Education and Research (BMBF, Medical Informatics Initiative, DIFUTURE consortium, grants 01ZZ1603[A-D] and 01ZZ1804[A-I]), and the NIH (grant 1R01NS112161-01), he has received honoraria for lecturing from Merck Serono. Go to Neurology.org/N for full disclosures.