AI-based characterization of multiple sclerosis clinical symptomatology using lesion parenchymal fraction: empirical support for the topographical model of MS
Stephen Krieger1, Thibo Billiet2, Celine Maes2, Nuno De Barros2, Wim Van Hecke2, Annemie Ribbens2, Tim Wang3, Kain Kyle4, Linda Ly4, Justin Garber3, Michael Barnett3
1Mount Sinai Dept of Neurology, 2Icometrix, 3Brain and Mind Centre, 4Sydney Neuroimaging Analysis Centre
Objective:
To evaluate the relationship between lesion location, functional systems, and disability using an artificial intelligence (AI) model.
Background:
The topographical model of MS depicts localized lesion burden in a pool of variable neurological reserve, with the cervical cord at the shallow end.
Design/Methods:
63 people with MS (30 RRMS, 33 SPMS, 51 female, baseline age 45.8±10.0 yrs) were longitudinally evaluated with MRI and clinical metrics (EDSS and functional system scores [FSS]) with median 2 timepoints, max. 8, (total: 154), follow-up 2.1±2.5 years. Imaging: icobrain was used to measure T2-FLAIR lesions on brain MRI, and adapted to measure T2 lesion volumes in cervical cord. To create a lesion/reserve ratio, localized lesion volume percentiles were divided by regional parenchymal volume percentile (cerebral, infratentorial, cervical cord), yielding regionally-specific lesion parenchymal fraction (LPF). Modeling: in 500 iterations, Random Forest (RF) regression models were trained to estimate cross-sectional EDSS and FSS. Estimations were Spearman correlated with actual scores; median correlation was assessed. To evaluate role of LPF, two estimator sets were used: 1) simple model: age and sex, 2) full model: age, sex, and topographical LPF.
Results:
Spearman correlations between fitted and true clinical scores were highest for overall EDSS (R=0.29), pyramidal (R=0.37), bowel/bladder (R=0.38), and cerebellar (R=0.27). Models including LPF outperformed the simple model in 100%, 100%, 98%, and 91% of iterations, respectively. The EDSS and pyramidal FSS were most strongly associated with infratentorial LPF, which also most strongly predicted conversion to SPMS.
Conclusions:
Lesion localization can be mapped to specific functional systems to explicate MS heterogeneity. Lesion parenchymal fraction, an assessment of lesion load/parenchymal volume ratio, is a quantified expression of the topographical model of MS and may be a useful MRI correlate of MS clinical symptoms. Using this metric, AI techniques can be refined to identify specific thresholds for conversion to SPMS.