Identifying multiple sclerosis subtypes using unsupervised machine learning and MRI data

Fig. 1. MRI-based subtypes. the evolution of MRI abnormalities in each of the three MRI-based subtypes. The colour shade ranges from blue to pink which represents the probability of abnormality (it can be interpreted as the degree of abnormality) (mild, moderate or severe which approximates 1, 2 and 3 sigma). The cortex-led subtype (left) showed cortical atrophy in the occipital, parietal and frontal cortex in the early stages of the sequences, and a reduction in T1/T2 ratio in the NAWM in the later stages. The normal-appearing white matter (NAWM)-led subtype (middle) showed a reduction in T1/T2 ratio of the cingulate bundle and corpus callosum in the earlier stages of the sequence, and deep grey matter and temporal grey matter atrophy in the later stages. The lesion-led subtype (right) shows early and extensive accumulation of lesions in the earlier stages of the sequence, and a reduction in the T1/T2 ratio in the NAWM in the later stages. The numbers on the left side represent SuStaIn stages. The minimum stage is 1 and the maximum stage is 39 (based on 13 variables that show mild (sigma = 1), moderate (sigma = 2) and severe abnormality (sigma = 3)). Acronyms: NAWM, normal-appearing…

Fig. 2. Subtype membership in the training, and validation datasets. a MRI-based subtypes in the training and internal validation dataset and b the validation dataset. Assignability of the disease subtype, or membership probability, is shown as the distance from each vertex of the triangle. Each of vertices represent the point at which membership of a given subtype is at its maximum (100%). We assigned each subject to one subtype (shown in red, green and blue) based on their maximum probability. c, d The 80th and 90th percentiles for the probability of assignment to the dominant subtype was 99.98% and 99.99% (indistinguishable in the figure) in the training dataset. In the validation dataset these percentiles were 99.45% and 99.97% respectively.

Fig. 3. MRI-based subtypes and disability progression in the placebo arms. The lesion-led subtype had a faster EDSS progression than the other two MS subtypes in both the training (a) and validation (b) sets. Only placebo arms (or comparator arms) of the clinical trials are included. In both a and b we used the log-rank test, with two-sided p-value, to compare survival curves (no correction for multiple comparisons). EDSS expanded disability status scale.

Fig. 4. MRI-based subtypes predicted disease activity in the placebo arms of the validation dataset. The average annual relapse rate for each MRI-based subtype. The centre shows the estimated average of annual relapse rate and error bars represent the standard error. The lesion-led subtype had significantly higher annual relapse rate than the cortex-led subtype (n = 1663 patients).

Fig. 5. Stratification predicts disability progression. Higher SuStaIn stage at baseline predicted time to disability progression: the higher the stage at baseline, the shorter the time to reach 24-week confirmed EDSS progression. p-Value is two-sided (p < 0.0001). When we repeated this analysis inside each MRI-based subtype we found similar results (results not shown). EDSS expanded disability status scale.

Fig. 6. Predicting treatment response with MRI-based subtyping in selected RCTs. Shows the change in EDSS worsening in MRI-based subtypes in the pooled treatment arms of the ORATORIO, ASCEND and OLYMPUS trials (n = 2099 patients) compared to the corresponding subtypes in the pooled placebo arms (e.g., lesion-led subtype on treatment vs. lesion led subtype on placebo and so forth). Patients in the lesion-led subtype had the largest reduction in the rate of EDSS worsening and were the only group who had a significant treatment response. The circle at the centre of each line represents the model-estimated average of percentage EDSS change. Error bars represent the standard error. Abbreviations: 9HPT, 9-Hole Peg test; NAMW, normal-appearing white matter; EDSS, Expanded Disability Status Scale; RRMS, relapsing-remitting multiple sclerosis; PPMS, primary progressive multiple sclerosis; RCT, randomised controlled trial.

Fig. 7. Model development and validation. We processed MRI with a uniform processing pipeline. We trained the model using 14 data sets, and applied the trained model on five independent, unseen data sets for validation. We demonstrated if MRI-based subtyping at baseline, could predict EDSS progression and disease activity. In selected RCTs, from RRMS and progressive (SP and PP MS), we looked at if MRI-based subtyping could predict treatment response. EDSS expanded disability status scale, RCT randomised controlled trial, Tx treatment, PASAT paced auditory serial addition test.