Детерминанты, движущие эволюцией новых множественных склерозовых повреждений в хронических активных или ремиелинизированных состояниях.

Introduction

Once formed, focal lesions that develop in patients with multiple sclerosis (MS) can follow different trajectories. Some of these lesions undergo complete resolution of inflammation and exhibit extensive remyelination, while some lesions remain chronically inflamed and progressively accumulate tissue damage ( Luchetti et al., 2018 , Kuhlmann et al., 2017 , Bagnato et al., 2011 , Frischer et al., 2015 ). The heterogeneity of lesions can be studied in vivo using Quantitative Susceptibility Mapping (QSM) MRI, a technique that allows for the identification of Paramagnetic Rim Lesions (PRLs), which represent chronic active lesions ( Bagnato et al., 2024 , Dal-Bianco et al., 2021 , Absinta et al., 2021 ), as well as QSM isointense lesions, which have recently been proposed as biomarkers for remyelinated lesions ( Rahmanzadeh et al., 2022 , Chen et al., 2014 , Gillen et al., 2021 ). Determining the processes of chronic inflammation and remyelination of lesions in vivo is of great importance, as the different composition of lesions in an individual plays a crucial role in the development of neurological disability ( Absinta et al., 2019 , Maggi et al., 2021 , Cagol et al., 2024 , Reeves et al., 2024 ). However, identifying PRLs and isointense remyelinated lesions with QSM is often impractical in clinical settings, as it requires advanced MRI sequences, specialized processing, and experienced professionals, which limits the applicability of these biomarkers. Preliminary studies have shown that PRLs can be identified early, at the time of their formation, by analyzing contrast enhancement patterns, diffusion and susceptibility values, normalized magnetization transfer ratio, or T1 relaxation times ( Elliott et al., 2023 , Clark et al., 2023 , Wenzel et al., 2022 , Elliott et al., 2021 , Elkady et al., 2024 , Al Gburi et al., 2025 , Zhang et al., 2019 , Absinta et al., 2016 ). However, many other factors, both radiological and clinical, such as lesion location, the presence of perilesional oedema, the patient's age, the use of corticosteroids and the class of therapy which is initiated after lesion formation, may also play a role.

Moreover, factors that may predict the evolution of a lesion into a QSM isointense, remyelinated state are still unknown.

Being able to determine early on the likelihood of a new lesion developing into a PRL or becoming QSM isointense would have significant practical implications, as it could guide therapeutic approaches.

In this study of patients with newly formed contrast-enhancing (CE) lesions, we aimed to identify the clinical and radiological variables present at the time of lesion formation associated with the development of PRLs or QSM isointense lesions after 12 months. Subsequently, we developed a scoring system, using clinical available features, to calculate the probability of a new lesion to become chronic active or remyelinated.

Methods

Standard protocol Approvals, Registrations, and patient Consents

This study was approved by the institutional review board (ID 10346). Participants gave written consent to participate in the study.

Study design, MRI protocol and lesion assessment

Data were collected from people with MS (PwMS) enrolled in a prospective longitudinal research MRI study developing at least one CE MS lesion and for whom a 12-month follow-up MRI scan was available. Subjects were imaged with a 3 T brain MRI (Siemens Prisma) at both time points with the following MRI protocol: • 3-dimensional (3D) sagittal T2-fluid-attenuated inversion recovery (FLAIR) (repetition time [TR]/inversion time [TI]/echo time [TE] 5000 ms/1800 ms/393 ms; original resolution: 0.8 × 0.8 × 1 mm 3 ; resolution of reconstructed images: 0.4 × 0.4 × 1 mm 3 ); • 3D sagittal T1-magnetization prepared rapid gradient-echo (TR/TI/TE 2300 ms/919 ms/2.96 ms; resolution 1 × 1 × 1 mm 3 ); • 3D sagittal-segmented echo-planar-imaging (EPI) (TR/TE 64 ms/35 ms; Flip Angle = 10°; resolution 0.65 × 0.65 × 0.65 mm 3 ) providing signal magnitude and phase. • 3D turbo spin-echo T1-weighted after injection of a single dose of Gadoteridol (0.2 mL/kg) (TR/TE 700 ms/12 ms; resolution: 1 × 1 × 1 mm 3 );

3-dimensional (3D) sagittal T2-fluid-attenuated inversion recovery (FLAIR) (repetition time [TR]/inversion time [TI]/echo time [TE] 5000 ms/1800 ms/393 ms; original resolution: 0.8 × 0.8 × 1 mm 3 ; resolution of reconstructed images: 0.4 × 0.4 × 1 mm 3 );

3D sagittal T1-magnetization prepared rapid gradient-echo (TR/TI/TE 2300 ms/919 ms/2.96 ms; resolution 1 × 1 × 1 mm 3 );

3D sagittal-segmented echo-planar-imaging (EPI) (TR/TE 64 ms/35 ms; Flip Angle = 10°; resolution 0.65 × 0.65 × 0.65 mm 3 ) providing signal magnitude and phase.

3D turbo spin-echo T1-weighted after injection of a single dose of Gadoteridol (0.2 mL/kg) (TR/TE 700 ms/12 ms; resolution: 1 × 1 × 1 mm 3 );

CE lesions were identified on baseline post contrast T1 images and manually segmented. The 12-months follow-up MRI scan was used for the assessment of PRL and QSM isointense lesions (ISO). QSM and Susceptibility Mapping Weighted Images (SMWI) were obtained from the 3D EPI, using STI Suite (v3.0, UC Berkley,CA). The processing pipeline included image phase-unwrapping, background phase removal and dipole deconvolution, according to international recommendations ( Bilgic et al., 2024 ) Quantitative susceptibility values obtained via this pipeline are implicitly referenced to the whole-brain mask. T1 images were rigidly registered to the EPI magnitude images using ANTs ( https://stnava.github.io/ANTs/ ) and the derived transformations were used to register the CE lesion masks to the QSM/SMWI space with nearest-neighbor interpolation. PRL assessment was performed according to international guidelines ( Bagnato et al., 2024 ) by GB and DB and SMWI was used to exclude the presence of veins running alongside the lesion’s border which may resemble a rim. ISO lesion were defined as lesions with equal QSM signal with respect to surrounding WM. The remaining lesions, characterized by diffuse QSM hyperintensity within the whole lesion, were classified as HYPER ( Boccia et al., 2024 ).

Statistical analysis

Mixed-effect modelling and multi-class classification

All statistical analyses and visualizations were performed using Python. Univariate generalized linear mixed-effects modeling was conducted to evaluate the associations between various predictors of interest and the QSM classification at 12 months. Each model included a subject-specific random intercept to account for multiple lesions per subject and a fixed-effect term for each predictor of interest. A corresponding fixed-effects-only model was also fitted. The Likelihood Ratio Test (LRT) was then computed to compare the two models, assessing whether the inclusion of subject-specific random effects significantly improved the model fit. List of predictors (both lesion-specific and subject-specific) included: • Lesion volume • Lesion topography (periventricular/juxta-subcortical/deep white matter, labelled as 1/2/3) • Pattern of contrast enhancement (ring/mixed/nodular, labelled as 1/2/3) • Presence of oedema (no/yes, labelled as 0/1) • Patient age at lesion onset • Sex (F/M, labelled as 0/1) • Use of corticosteroids within the previous 30 days (no/yes, labelled as 0/1) • DMT class at lesion onset (no therapy/low-medium efficacy DMT/high-efficacy DMT, defined according to ( Filippi et al., 2022 ), labelled as 0/1/2) • DMT switch during follow-up (low-medium efficacy DMT/high-efficacy DMT, labelled as 1/2).

Lesion volume

Lesion topography (periventricular/juxta-subcortical/deep white matter, labelled as 1/2/3)

Pattern of contrast enhancement (ring/mixed/nodular, labelled as 1/2/3)

Presence of oedema (no/yes, labelled as 0/1)

Patient age at lesion onset

Sex (F/M, labelled as 0/1)

Use of corticosteroids within the previous 30 days (no/yes, labelled as 0/1)

DMT class at lesion onset (no therapy/low-medium efficacy DMT/high-efficacy DMT, defined according to ( Filippi et al., 2022 ), labelled as 0/1/2)

DMT switch during follow-up (low-medium efficacy DMT/high-efficacy DMT, labelled as 1/2).

The predictors were used for a subsequent multiclass classification task on QSM classification at 12 months with two of the most common machine-learning models: Random Forest (RF) and XGBoost (XGB) algorithms ( Hastie T, TR, FJ, 2013 ). After random splitting of the dataset into an 80–20 stratified proportion for training and testing, the models were fine-tuned through GridSearchCV for hyperparameter optimization (with 5-fold cross validation). The testing dataset contained exclusively lesions from subjects who were excluded from the training dataset. A class weighting strategy was implemented during training using the inverse class frequency technique to weight the classes based on their frequency. For each random split of the database (iterated 500 times), the best model was selected according to the area under the Receiver Operating Characteristic Curve (ROC-AUC), selected after rigorous evaluation for various performance metrics (accuracy, balanced accuracy, f1, precision, average precision, recall).

Feature importance analysis with SHAP and risk score computation

For the best model with the highest ROC-AUC score among the splits, we implemented SHAP (SHapley Additive exPlanations). SHAP is a feature importance tool based on a game-theoretic approach used in machine-learning for explaining the output of a model by quantifying the importance of each feature ( Lundberg, 2014 ). SHAP was used to identify the most relevant features contributing to the model’s predictions by calculating SHAP values for each feature in the dataset. Features with higher SHAP values were considered more influential in the model’s predictions, while features with lower SHAP values had less impact. Finally, a risk scoring approach was established to assess the likelihood of lesions being classified in the ISO or PRL class. This approach leveraged the “four most influential features”, identified through the SHAP analysis by selecting the features with higher influence on the prediction of ISO or PRL classes. To determine class-specific and feature-specific risk thresholds, the value of each feature associated with a SHAP value equal to 0 in the training set was computed. After identifying the threshold and determining whether continuous variables should exceed or fall below it, as well as defining the significant categories for categorical variables (see Fig. 1 , Fig. 2 ), we calculated the probability that the established condition for each variable was met by computing the ratio of favorable cases to total cases. Subsequently, the probability of a new lesion being classified in a particular class, given that the values of its associated four features were above or below the specific thresholds, was calculated. Finally, a risk score was derived by mapping the probability onto a 0–3 scale as: • If none of the required conditions are met, score = 0 • 0–50 % probability, score = 1 • 50–75 % probability, score = 2 • >75 % probability: score = 3 Fig. 1 Shap analysis for predicting PRL class. Features with higher SHAP values are considered more influential in the model’s predictions, while those with lower SHAP values contribute less significantly. The color-coded bar on the right represents the value of the features of interest. Categorial variables have been coded as follows: pattern of contrast enhancement (ring/mixed/nodular, labelled as 1/2/3), sex (F/M, labelled as 0/1), lesion topography (periventricular/juxta-subcortical/deep white matter, labelled as 1/2/3), disease modifying treatments −DMT- class at lesion onset (no therapy/low-medium efficacy DMT/high-efficacy DMT, labelled as 0/1/2), presence of oedema (no/yes, labelled as 0/1), use of corticosteroids (no/yes, labelled as 0/1), DMT switch during follow-up (low-medium efficacy DMT/high-efficacy DMT, labelled as 1/2). Fig. 2 Shap analysis for predicting ISO class. Features with higher SHAP values are considered more influential in the model’s predictions, while those with lower SHAP values contribute less significantly. The color-coded bar on the right represents the value of the features of interest. Categorial variables have been coded as follows: sex (F/M, labelled as 0/1), lesion topography (periventricular/juxta-subcortical/deep white matter, labelled as 1/2/3), disease modifying treatments −DMT- class at lesion onset (no therapy/low-medium efficacy DMT/high-efficacy DMT, labelled as 0/1/2), presence of oedema (no/yes, labelled as 0/1), DMT switch during follow-up (low-medium efficacy DMT/high-efficacy DMT, labelled as 1/2), pattern of contrast enhancement (ring/mixed/nodular, labelled as 1/2/3), use of corticosteroids (no/yes, labelled as 0/1).

If none of the required conditions are met, score = 0

0–50 % probability, score = 1

50–75 % probability, score = 2

>75 % probability: score = 3

Shap analysis for predicting PRL class. Features with higher SHAP values are considered more influential in the model’s predictions, while those with lower SHAP values contribute less significantly. The color-coded bar on the right represents the value of the features of interest. Categorial variables have been coded as follows: pattern of contrast enhancement (ring/mixed/nodular, labelled as 1/2/3), sex (F/M, labelled as 0/1), lesion topography (periventricular/juxta-subcortical/deep white matter, labelled as 1/2/3), disease modifying treatments −DMT- class at lesion onset (no therapy/low-medium efficacy DMT/high-efficacy DMT, labelled as 0/1/2), presence of oedema (no/yes, labelled as 0/1), use of corticosteroids (no/yes, labelled as 0/1), DMT switch during follow-up (low-medium efficacy DMT/high-efficacy DMT, labelled as 1/2).

Shap analysis for predicting ISO class. Features with higher SHAP values are considered more influential in the model’s predictions, while those with lower SHAP values contribute less significantly. The color-coded bar on the right represents the value of the features of interest. Categorial variables have been coded as follows: sex (F/M, labelled as 0/1), lesion topography (periventricular/juxta-subcortical/deep white matter, labelled as 1/2/3), disease modifying treatments −DMT- class at lesion onset (no therapy/low-medium efficacy DMT/high-efficacy DMT, labelled as 0/1/2), presence of oedema (no/yes, labelled as 0/1), DMT switch during follow-up (low-medium efficacy DMT/high-efficacy DMT, labelled as 1/2), pattern of contrast enhancement (ring/mixed/nodular, labelled as 1/2/3), use of corticosteroids (no/yes, labelled as 0/1).

The accuracy of this scoring system was assessed using the test dataset.

Data availability

Anonymized data not published within this article will be made available by request from any qualified investigator.

Results

Lesion’s phenotypes

We included 44 PwMS (mean age = 36 years, SD = 11; median EDSS = 2.5, range = 0–6.5; Table 1 ) with evidence of at least one CE lesion, for a total of 173 CE lesions. 43 lesions had to be excluded because they were not analyzable at QSM at 12 months because of lesion localization/presence of artifacts (n = 20) or dimension (n = 23), according to international recommendations ( Bagnato et al., 2024 ). Nineteen CE lesions were excluded because they were a reactivation of a pre-existing MS lesion. A total of 111 CE lesions were included in the final analysis (mean number of lesions per subject = 3, SD = 3). At 12 months, 15 (13 %) lesions were classified as PRLs, 50 (45 %) as ISO and 46 (42 %) as HYPER. None of these lesions had CE at 12 months follow-up. Table 1 Demographic and clinical characteristics of the study population. Participants, n 44 Age, years (mean, SD) 35.7 (10.7) Sex, female (n, %) 30 (68.2) EDSS (median, range) 2.5 (0–6.5) Disease duration, years (mean, SD) 7.6 (7.0) Corticosteroids use (n, %) 25 (56.8) DMT at lesion onset Treatment naive Low/medium efficacy DMT High efficacy DMT 13 (29.5) 21 (47.7) 10 (22.7) DMT switch after lesion onset Low/medium efficacy DMT High efficacy DMT 12 (27.3) 32 (72.7) Number of new CE lesion per subject (mean, SD) 2.9 (2.8) SD = standard deviation; EDSS = Expanded-Disability-Status-Scale; DMT = disease modifying treatments; CE = contrast enhancing.

Table 1 Demographic and clinical characteristics of the study population. Participants, n 44 Age, years (mean, SD) 35.7 (10.7) Sex, female (n, %) 30 (68.2) EDSS (median, range) 2.5 (0–6.5) Disease duration, years (mean, SD) 7.6 (7.0) Corticosteroids use (n, %) 25 (56.8) DMT at lesion onset Treatment naive Low/medium efficacy DMT High efficacy DMT 13 (29.5) 21 (47.7) 10 (22.7) DMT switch after lesion onset Low/medium efficacy DMT High efficacy DMT 12 (27.3) 32 (72.7) Number of new CE lesion per subject (mean, SD) 2.9 (2.8) SD = standard deviation; EDSS = Expanded-Disability-Status-Scale; DMT = disease modifying treatments; CE = contrast enhancing.

Demographic and clinical characteristics of the study population.

SD = standard deviation; EDSS = Expanded-Disability-Status-Scale; DMT = disease modifying treatments; CE = contrast enhancing.

Mixed-effect modeling and multi-class classification

The fit of both the mixed-effect and the fixed-effect models returned an LRT statistic with p-value = 1, which implies that the mixed-effect model did not significantly improve over the null model. This ensures that treating each lesion independently for classification does not result in substantial loss of information. ROC-AUC scores for the 100 splits for both RF and XGB are displayed in Supplementary Fig. S1 , while the best 10 results obtained for both algorithms are reported in Supplementary Table S1 . The average ROC-AUC score is 0.72 (std = 0.07) for XGBoost Classifier 0.7 (std = 0.07) for Random Forest Classifier.

Feature importance analysis with SHAP

SHAP on the overall best model (XGBoost classifier with ROC-AUC = 0,90) shows that the most significant features in predicting lesion classification are similar for the two classes (lesion volume, patient age and sex) except for the pattern of CE (which was selected for PRLs classification) and the lesion topography (which was selected for ISO classification). Fig. 1 and Fig. 2 represent for each feature which values impact most on the prediction for PRLs and ISO respectively. Similar results were obtained including in the analyses the CE lesions (n = 19) which represented a reactivation of a pre-existing MS lesion and excluded from the main analyses.

Risk score computation

Class-specific and feature-specific threshold for computing the risk scores were calculated on the training set. The class-specific thresholds for the four most influential features for PRLs and ISO classes are listed in Table 2 , together with the associated probability and weight of being classified in the PRLs vs ISO class. Table 3 reports the normalized sum of probabilities for all the possible combinations and the associated risk scores for PRL and ISO classes Table 4 . Table 2 Thresholds and probabilities/weights associated with the risk scoring system. PRL ISO Feature name Threshold Probability Feature name Threshold Probability Pattern of Contrast Enhancement Ring-pattern 0.89 Lesion Volume <0.17 mL 0.65 Lesion Volume >5 mL 0.78 Patient Age <36 yo 0.74 Patient Age >48 yo 0.78 Patient Sex Female 0.46 Patient Sex Male 0.33 Lesion Topography Juxta-subcortical/Deep white matter 0.43 The threshold value corresponds to the value that has shap value equal to 0, and it is above or below the value according to which values are more predictive with respect to the classification task (for the categorical features it is uniquely the most influential category for prediction). The probability was calculated as the ratio between the number of favorable events meeting the threshold and the total number of events. Table 3 Normalized sum of probabilities with the associated PRL risk scores in different scenarios. Feature name Normalised sum of probabilities Associated risk score None 0 0 Pattern of CE 0.23 1 Age 0.18 1 Volume 0.18 1 Sex 0 1 Age + Volume 0.50 1 Volume + Sex 0.32 1 Age + Sex 0.20 1 Pattern of CE + Sex 0.36 1 Pattern of CE + Age 0.54 2 Pattern of CE + Volume 0.54 2 Pattern of CE + Volume + Sex 0.68 2 Age + Sex + Volume 0.64 2 Pattern of CE + Age + Sex 0.68 2 Pattern of CE + Age + Volume 0.86 3 Pattern of CE + Age + Volume + Sex 1 3 The probabilities were calculated in accordance with Table 2 as a combination of the probability values associated with the four most relevant features. These values were rescaled to obtain a probability range between 0 and 1. The table reports all possible combinations of the four features, providing a comprehensive view of how different feature scenarios contribute to PRL risk classification. Table 4 Normalized sum of probabilities with the associated ISO risk scores in different scenarios. Feature name Normalised sum of probabilities Associated risk score None 0 0 Volume 0.11 1 Age 0.17 1 Sex 0.02 1 Topography 0 1 Age + Sex 0.42 1 Volume + Sex 0.36 1 Volume + Topography 0.35 1 Age + Topography 0.40 1 Sex + Topography 0.25 1 Volume + Age 0.52 2 Age + Sex + Topography 0.65 2 Volume + Sex + Topography 0.60 2 Volume + Age + Topography 0.75 2 Volume + Age + Sex 0.77 3 Volume + Age + Sex + Topography 1 3 The probabilities were calculated in accordance with Table 2 as a combination of the probability values associated with the four most relevant features. These values were rescaled to obtain a probability range between 0 and 1. The table reports all possible combinations of the four features, providing a comprehensive view of how different feature scenarios contribute to ISO risk classification.

Table 2 Thresholds and probabilities/weights associated with the risk scoring system. PRL ISO Feature name Threshold Probability Feature name Threshold Probability Pattern of Contrast Enhancement Ring-pattern 0.89 Lesion Volume <0.17 mL 0.65 Lesion Volume >5 mL 0.78 Patient Age <36 yo 0.74 Patient Age >48 yo 0.78 Patient Sex Female 0.46 Patient Sex Male 0.33 Lesion Topography Juxta-subcortical/Deep white matter 0.43 The threshold value corresponds to the value that has shap value equal to 0, and it is above or below the value according to which values are more predictive with respect to the classification task (for the categorical features it is uniquely the most influential category for prediction). The probability was calculated as the ratio between the number of favorable events meeting the threshold and the total number of events.

Thresholds and probabilities/weights associated with the risk scoring system.

The threshold value corresponds to the value that has shap value equal to 0, and it is above or below the value according to which values are more predictive with respect to the classification task (for the categorical features it is uniquely the most influential category for prediction). The probability was calculated as the ratio between the number of favorable events meeting the threshold and the total number of events.

Table 3 Normalized sum of probabilities with the associated PRL risk scores in different scenarios. Feature name Normalised sum of probabilities Associated risk score None 0 0 Pattern of CE 0.23 1 Age 0.18 1 Volume 0.18 1 Sex 0 1 Age + Volume 0.50 1 Volume + Sex 0.32 1 Age + Sex 0.20 1 Pattern of CE + Sex 0.36 1 Pattern of CE + Age 0.54 2 Pattern of CE + Volume 0.54 2 Pattern of CE + Volume + Sex 0.68 2 Age + Sex + Volume 0.64 2 Pattern of CE + Age + Sex 0.68 2 Pattern of CE + Age + Volume 0.86 3 Pattern of CE + Age + Volume + Sex 1 3 The probabilities were calculated in accordance with Table 2 as a combination of the probability values associated with the four most relevant features. These values were rescaled to obtain a probability range between 0 and 1. The table reports all possible combinations of the four features, providing a comprehensive view of how different feature scenarios contribute to PRL risk classification.

Normalized sum of probabilities with the associated PRL risk scores in different scenarios.

The probabilities were calculated in accordance with Table 2 as a combination of the probability values associated with the four most relevant features. These values were rescaled to obtain a probability range between 0 and 1. The table reports all possible combinations of the four features, providing a comprehensive view of how different feature scenarios contribute to PRL risk classification.

Table 4 Normalized sum of probabilities with the associated ISO risk scores in different scenarios. Feature name Normalised sum of probabilities Associated risk score None 0 0 Volume 0.11 1 Age 0.17 1 Sex 0.02 1 Topography 0 1 Age + Sex 0.42 1 Volume + Sex 0.36 1 Volume + Topography 0.35 1 Age + Topography 0.40 1 Sex + Topography 0.25 1 Volume + Age 0.52 2 Age + Sex + Topography 0.65 2 Volume + Sex + Topography 0.60 2 Volume + Age + Topography 0.75 2 Volume + Age + Sex 0.77 3 Volume + Age + Sex + Topography 1 3 The probabilities were calculated in accordance with Table 2 as a combination of the probability values associated with the four most relevant features. These values were rescaled to obtain a probability range between 0 and 1. The table reports all possible combinations of the four features, providing a comprehensive view of how different feature scenarios contribute to ISO risk classification.

Normalized sum of probabilities with the associated ISO risk scores in different scenarios.

The probabilities were calculated in accordance with Table 2 as a combination of the probability values associated with the four most relevant features. These values were rescaled to obtain a probability range between 0 and 1. The table reports all possible combinations of the four features, providing a comprehensive view of how different feature scenarios contribute to ISO risk classification.

The accuracy of the score system was tested using the testing dataset. Supplementary Table S3 shows the predicted risk scores for each lesion, and the true classification. The scoring system correctly identifies as at higher risk the true class in 17 out of 18 cases and equal risk score to both classes in 1 case.

Discussion

Investigating the in vivo heterogeneity of MS lesions remains a significant challenge. In this study, we aimed to assess whether standard clinical and radiological features could assist clinicians in predicting whether a new acute MS lesion would evolve into a chronically active or inactive, remyelinated state. Notably, identifying paramagnetic rim lesions (PRLs) and isointense remyelinated lesions (ISOs) with quantitative susceptibility mapping (QSM) is often impractical in clinical settings, as it requires advanced MRI sequences, specialized processing, and experienced professionals, which limits the applicability of these biomarkers. Therefore, our goal was to develop a simple, clinically accessible score to predict the likelihood of an acute MS lesion progressing to either a PRL or ISO. Using a machine learning approach, we identified five key features associated with these lesion classes: patient age, lesion volume, pattern of contrast enhacement, patient sex, and lesion topography.

Age

We found that new CE MS lesions occurring in older people have a higher chance of becoming PRLs, indicating that age is one of the most robust predictors of chronic neuroinflammation. Vice versa, a younger age is associated with a higher chance of inflammation resolution and remyelination. These results are in line with a body of pathological ( Luchetti et al., 2018 , Frischer et al., 2015 ), MRI ( Absinta et al., 2019 ) and PET ( Nylund et al., 2022 ) ( Hamzaoui et al., 2023 ) evidence, indicating that the risk of experiencing chronic inflammation increases over time, while the remyelination potential decreases with ageing ( Müller et al., 2024 , Kolb et al., 2021 , Bodini et al., 2016 ). We identified an age > 48 years as the optimal cut-off to assign a higher risk of PRL evolution, and an age < 36 years as the cut-off for a higher chance of ISO evolution. Interestingly, analyzing 120 MS autopsies, Frischer et al. ( Frischer et al., 2015 ) found that smoldering plaques begin to peak around age 47. However, it is important to note that younger patients tend to develop a greater number of new CE lesions compared to older people and therefore still face a considerable (sometimes higher ( Dal-Bianco et al., 2021 , Hemond et al., 2022 ) cumulative risk of PRL formation ( Martire et al., 2022 ). Indeed, recent MRI studies have shown that PRLs can also be identified in earlier stages of MS ( Dal-Bianco et al., 2021 ) ( Hemond et al., 2022 ), including radiologically isolated syndrome ( Suthiphosuwan et al., 2020 ) and pediatric cases ( Boccia et al., 2024 ) ( Margoni et al., 2023 ). In conclusion, although new lesions in older individuals are more likely to evolve into PRLs, younger patients may have a higher overall risk of new lesion formation, and consequently, greater PRL accumulation over time.

Lesion volume

Our results confirm previous findings ( Elliott et al., 2023 ) ( Clark et al., 2023 , Katz et al., 1993 , Zhang et al., 2019 , Absinta et al., 2016 ), demonstrating that lesion volume is a significant factor in determining lesion outcome. Larger lesions are more likely to sustain inflammatory activity, while in smaller lesions the control of acute inflammation seems to occur more effectively. Similarly, smaller lesions are more likely to become QSM isointense lesions, suggesting that the physical scale of acute inflammation may directly impact the regenerative processes at lesion’s borders ( Kolb et al., 2021 ) ( Gaitán et al., 2011 ), ( Patrikios et al., 2006 ) limiting local repair mechanisms in larger lesions. Absinta et al. ( Absinta et al., 2016 ) found that, regardless of the initial contrast enhancement pattern or the patient’s age, a PRL is likely to be present when the lesion volume is ≥ 0.16–0.22 mL (depending on the presence or absence of a baseline T2* rim, respectively). These volumetric thresholds are slightly lower than the 5 mL observed in our study, likely due to differences in image resolution (7 T vs. 3 T) and the potential underdetection of very small CE and PRL lesions in our dataset. Nonetheless, findings from Absinta et al. ( Absinta et al., 2016 ) are consistent with the threshold we identified for ISO classification, where only lesions with a volume < 0.17 mL are assigned a higher probability of being classified as ISO.

Pattern of contrast enhancement

As previously reported ( Clark et al., 2023 ) ( Wenzel et al., 2022 , Al Gburi et al., 2025 ), we found that the ring-enhancing pattern is associated with a higher risk of developing PRLs. This observation is in line with findings from Absinta et al. ( Absinta et al., 2016 ) who demonstrated that lesions showing a centripetal pattern at dynamic contrast-enhanced MRI (which is the case for most ring-enhancing lesions ( Davis et al., 2010 ), are likely to develop a persistent rim at the lesion border. In line with this, ring-enhancing lesions are characterized by more extensive tissue damage compared to nodular enhancing lesions ( Tazza et al., 2024 ), as is the case for PRLs compared to non-PRLs lesions ( Rahmanzadeh et al., 2022 ) ( Tolaymat et al., 2020 ).

Sex

Our analyses showed that sex plays a significant, although minor, role in lesion fate. We found that males are at higher risk of developing PRLs while new MS lesions in females have higher chances of becoming ISO. These findings align with pathological evidence showing that male sex is associated with a higher prevalence of chronic active lesions ( Luchetti et al., 2018 ) ( Frischer et al., 2015 ). Accordingly, it has been previously demonstrated by some ( Choi et al., 2024 , Clarke et al., 2024 , Dal-Bianco et al., 2017 ) but no all ( Absinta et al., 2019 ) ( Manouchehrinia et al., 2019 ) in vivo MRI studies, that men are at higher risk of PRL formation. It is noteworthy to remember that male sex has been reported in several studies as an independent predictor of poorer outcomes ( Iacobaeus et al., 2020 ) ( Kalaitzidis and Calabresi, 2021 ), which may be driven, at least in part, by a higher degree of chronic neuroinflammation. These studies indicate potential gender-related differences in MS progression and pathology, possibly influenced by sex hormones. However, the evidence is not conclusive, and further research is needed to fully understand the role of gender in MS inflammation and neurodegeneration.

Lesion topography

In addition to lesion volume, topography also appears to play a role, albeit modest, in determining a lesion fate. Our analysis shows that periventricular lesions are less likely to become ISO than paraventricular or juxta-subcortical lesions. This result is in line with the evidence that certain anatomical locations might inherently predispose lesions to inefficient repair and/or chronic inflammation ( Kolb et al., 2021 ) ( Gaitán et al., 2011 , Tonietto et al., 2022 , Vaneckova et al., 2022 ), impacting the overall trajectory of MS. For example, a detrimental effect of cerebrospinal fluid on remyelination potential has been postulated ( Toubasi et al., 2024 ) ( Kappos et al., 2020 ) and a recent work has suggested that PRLs are more frequent in watershed areas of the brain, hinting at a potential role of lesion perfusion in determining chronic lesion fate ( Lublin et al., 2022 ).

Considering that most neurological disability in PwMS on modern DMTs appears to accumulate independently of acute inflammation ( Maggi et al., 2023 ) ( Alsema et al., 2024 ), it is essential to assess the impact of treatments on chronic neuroinflammation. Preliminary data suggest that therapies, including high-efficacy DMTs, are not able to abolish already formed PRLs[56]. However, it has not been systematically investigated whether DMTs might reduce the risk of a newly formed lesion evolving into a PRL. Here, we sought to examine the impact of high efficacy DMT, as well as the use of high dose corticosteroids, on PRL vs ISO evolution risk. No significant impact of DMTs was observed in modulating this probability. Similarly, high-dose corticosteroid treatment did not appear to influence lesion course. Our findings seem to confirm that the biological features determining lesion evolution may be less responsive to available DMTs.

Interestingly, in our cohort, the inclusion of subject-specific random effects did not significantly improve the model fit, meaning that each new CE lesion can be analyzed independently. This result suggests that lesion-specific features, together with demographic characteristics, play a more relevant role in determining a new MS lesion’s fate than individual features (e.g., genetic predisposing factors). It is still unknown whether specific genetic signatures are associated with a higher risk of brain chronic inflammation. Recently, it has been demonstrated that specific expression changes in astrocyte-specific, oligodendrocyte-specific and microglia-specific gene sets characterize the active lesion rims[57]. In the future, identifying genetic signatures linked to the risk of PRL formation could enhance our ability to identify patients at higher risk for chronic neuroinflammation.

The main limitation of this study is the relatively small sample size. However, in the era of high-efficacy therapies, obtaining large samples of CE lesions for research MRI protocols presents a challenge. Despite this, the lesions were well-distributed across our cohort of PwMS, which is representative of a typical MS population. The second limitation is that, although our scoring system was validated within the testing dataset, additional validation in independent cohorts is essential to further assess its accuracy and generalizability. Finally, we recognize that additional quantitative MRI data (e.g. diffusion, susceptibility, perfusion and myelin MRI, T1 relaxometry) altogether with genetic and /or immunological signatures could undoubtedly improve lesion outcome prediction ( Clark et al., 2023 ) ( Wenzel et al., 2022 ). In particular, the presence of a rim on SWI or QSM at the time of contrast enhancement has been associated with a higher risk of subsequent PRL formation ( Zhang et al., 2019 , Absinta et al., 2016 ). As SWI is likely to be incorporated into most MRI protocols for MS in the near future, detecting an early rim sign at the time of lesion formation could further enhance predictive analyses. However, the primary aim of this study was to develop a scoring system based on clinically accessible features, enabling practical application by neurologists in everyday clinical settings. Future studies that incorporate more advanced imaging modalities could provide additional insights into the pathophysiological mechanisms underlying lesion evolution in MS, which could ultimately aid in the development of targeted interventions to modulate MS lesion fate.

In conclusion, we propose a simple 5-item score to predict the likelihood of a new MS lesion evolving into a chronic active or inactive remyelinated state, which, if externally validated, could offer clinicians a valuable tool for patient prognostication and monitoring.

CRediT authorship contribution statement

G. Boffa: Writing – review & editing, Writing – original draft, Methodology, Formal analysis, Data curation, Conceptualization. C. Razzetta: Writing – review & editing, Writing – original draft, Methodology, Formal analysis. D. Boccia: Writing – review & editing, Formal analysis, Data curation. S. Garbarino: Writing – review & editing, Writing – original draft, Methodology, Formal analysis. E. Cipriano: Writing – review & editing, Formal analysis, Data curation. A. Uccelli: Writing – review & editing, Resources. C. Lapucci: Writing – review & editing, Data curation. M. Piana: Methodology, Formal analysis. M. Inglese: Writing – review & editing, Supervision, Conceptualization.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

GB reports consultant fees from Roche, Novartis, Merck and Janssen. CL has received honoraria for speaking, travel grants, and for participating in the advisory board from Merck, Sanofi, Novartis, Roche, Alexion. AU has received grants (to his institution) from FISM, Biogen, Roche, Alexion, and Merck Serono and has participated on a data safety monitoring board or advisory board (to his institution) for BD, Biogen, Iqvia, Sanofi, Roche, Alexion, and Bristol Myers Squibb. MI received grants from NIH, NMSS, and FISM; received fees for consultation from BMS, Janssen, Roche, Genzyme, Merck, Biogen and Novartis. DB, EC, SG, CR and MP report no conflicts of interest.