Identifying multiple sclerosis subtypes using unsupervised machine learning and MRI data

Multiple sclerosis (MS) is a neurological disorder affecting millions globally. Current classification relies on clinical symptoms (clinically isolated syndrome (CIS), relapsing-remitting MS (RRMS), primary-progressive MS (PPMS), and secondary progressive MS (SPMS)), which don't fully capture the underlying pathological heterogeneity. This study aimed to redefine MS subtypes based on MRI data using unsupervised machine learning to identify groups with similar pathobiological mechanisms, potentially leading to better treatment stratification. The researchers hypothesized that an MRI-based model would improve understanding of MS disease progression compared to clinical data alone. Existing clinical phenotypes and their descriptors (disease activity and disability progression) are insufficient for accurate patient stratification in clinical trials, as they lack alignment with underlying pathobiological processes. The transition between phenotypes is also difficult to pinpoint clinically. MRI offers a superior reflection of MS pathogenic mechanisms, making it ideal for data-driven disease classification. However, identifying subtypes in neurodegenerative disorders is challenging due to their long prodromal periods and the need to integrate observations from cross-sectional and longitudinal studies. The researchers employed the Subtype and Staging Inference (SuStaIn) algorithm, capable of disentangling temporal and phenotypic heterogeneity to identify MRI-based subtypes with distinct temporal progression patterns.

The researchers reviewed existing literature on MS classification, highlighting the limitations of current clinical phenotypes in capturing the underlying pathological heterogeneity. They emphasized the potential of MRI as a more accurate reflection of MS pathogenic mechanisms and discussed previous research using machine learning for identifying disease subtypes in neurodegenerative disorders. The literature review underscored the need for a data-driven approach that accounts for both phenotypic and temporal heterogeneity in MS.

The study used a large dataset comprising MRI scans and clinical data from 19 datasets (14 for training, 5 for validation). These datasets included data from 16 multiple sclerosis (MS) randomized controlled trials (RCTs) and three observational cohorts. The training dataset consisted of 6322 MS patients, while the validation dataset included 3068 patients. Brain MRI scans (T1-weighted, T2-weighted, and T2-FLAIR) were processed using a standardized pipeline to extract 18 features: grey matter volumes in various brain regions, total T2 lesion volume, and regional T1/T2 ratios in normal-appearing white matter. A Bayesian linear regression model was used to adjust for the total intracranial volume, age and age squared. The Subtype and Staging Inference (SuStaIn) algorithm was applied to the training dataset to identify MRI-based subtypes, with leave-one-dataset-out cross-validation used for model selection and internal validation. The optimal model was then applied to the independent validation dataset to test its generalizability. The researchers compared patient characteristics (age, sex, EDSS, and disease duration) between the training and validation datasets. Thirteen MRI features, which significantly differed between the MS training dataset and a control group, were selected and used in the SuStaIn model. Analyses involved comparing clinical characteristics (EDSS, disease duration, lesion load, etc.) across subtypes; assessing disability progression using Cox regression models; analyzing disease activity (relapse rate and enhancing lesions); and evaluating treatment response in selected RCTs using linear mixed-effects models. The concordance index was used to assess the predictive performance of models combining MRI-based subtypes with clinical data. Statistical analyses included Chi-square tests, general linear models, mixed-effects models, and log-rank tests.

SuStaIn identified three MRI-based MS subtypes: cortex-led, NAWM-led, and lesion-led. These subtypes exhibited distinct patterns of MRI abnormality evolution over time. The lesion-led subtype was characterized by early and extensive lesion load, rapid lesion accumulation, high relapse rates, high contrast-enhancing lesion counts, and the highest risk of 24-week confirmed disability progression (CDP). This subtype also showed the highest SuStaIn stage at baseline and the fastest annual increase in stage. In contrast, the cortex-led subtype showed early cortical atrophy, followed by atrophy in other grey matter regions, T2 lesion accrual and late NAWM abnormalities. The NAWM-led subtype showed early reduction in T1/T2 ratio in specific white matter tracts, followed by broader white matter changes and later grey matter atrophy. The lesion-led subtype demonstrated a significant treatment response in both RRMS and progressive MS trials, suggesting that it might be particularly responsive to treatments targeting inflammatory lesion activity. The cortex-led subtype was the most frequent in both datasets, suggesting a more insidious, neurodegenerative component of MS, which is challenging to treat. MRI-based subtypes and stages were more strongly associated with EDSS worsening than baseline EDSS or clinical phenotypes, even when EDSS was similar across subtypes in the validation dataset. Combining MRI-based subtypes with clinical data improved the prediction accuracy of 24-week CDP (concordance index increased from 0.55 to 0.63).

The study's findings strongly support the use of MRI-based subtyping to improve the understanding of MS pathogenesis and to better predict disease course and treatment response. The identification of three distinct subtypes with unique temporal trajectories of MRI abnormalities provides insights into the diverse mechanisms underlying MS. The superior predictive power of MRI-based subtypes compared to clinical phenotypes underscores the importance of incorporating imaging biomarkers for accurate patient stratification in clinical trials and personalized medicine. The significant treatment response observed in the lesion-led subtype, which showed substantial focal inflammatory demyelination and secondary deep grey matter neurodegeneration, supports the idea that targeted therapies can effectively treat inflammatory activity in this subgroup. The prevalence of the cortex-led subtype, marked by early cortical atrophy and insidious inflammation, highlights the need for treatments addressing neurodegenerative processes. The study's results necessitate a shift towards a more biologically driven classification system that complements clinical phenotypes to improve clinical trial design and outcomes. The inclusion of clinical data along with MRI-based subtypes improved predictive accuracy. This study focused on biologically driven stratification by not incorporating clinical data in the model training, in order to enhance biological understanding of the disease progression.

This study successfully identified three distinct MRI-based MS subtypes using unsupervised machine learning, providing valuable insights into MS pathophysiology. These subtypes, characterized by distinct patterns of MRI abnormality evolution, effectively predict disease activity, disability progression, and treatment response, outperforming conventional clinical phenotypes. The lesion-led subtype demonstrated a particularly strong treatment response. The findings advocate for incorporating MRI-based subtyping into clinical trials to improve patient stratification, leading to more effective and targeted treatment strategies. Future research should focus on validating these subtypes in larger, more diverse cohorts and exploring potential therapeutic implications for each subtype. Additionally, investigating the role of spinal cord MRI in enhancing subtype classification could yield further insights.

The study's reliance on data from clinical trials and observational cohorts may introduce biases due to specific selection criteria and treatment effects. The availability of routine MRI measures, rather than more advanced techniques, might have limited the sensitivity of some assessments. Although the researchers performed rigorous analyses to reduce the impacts of differences in scanners and acquisition procedures, some effects might persist. Further validation in real-world clinical settings is necessary to ensure the generalizability of the findings. The study did not include longitudinal clinical and MRI data from a single cohort with the long follow-up required to track patients throughout the natural progression of their MS.