Changes in the ability to walk and talk lie at the core of neurological conditions such as MS. To measure these changes, the centuries-old neurological examination is plagued by issues of sensitivity, objectivity, meaningfulness and reproducibility. Digital tools that quantify precise aspects of neurological function (gait, speech) offer advantages in capturing subtle changes in function. However, these tools must be validated and anchored against gold standard functional assessments.

A convenience cohort of 176 adults with MS, EDSS 0-6.5, were enrolled in a longitudinal digital phenotyping study. Clinical testing included Expanded Disability Status Scale (EDSS) and MS Functional Composite [MSFC: walking speed (T25FW), processing speed (SDMT), dexterity (9HPT)]. Then, participants completed digital gait and speech  tasks (<15 min). Gait (Protokinetics Zeno) and speech (k:elements) metrics were extracted utilising proprietary software packages. For analysis, the cohort was split (66%/34%) into Discovery/Validation. Key features selected for each clinical outcome utilizing ML tools.

In the validation cohort, for continuous outcomes, model performance of the selected features was: R2=0.79, p<0.0001 for EDSS, R2=0.95, p<0.001 for T25FW, R2=0.78, p<0.001 for SDMT, and R2=0.77, p<0.001 for 9HPT. For categorical outcomes, models yielded ROC = 1.00/0.96/0.99 for EDSS (0-2, 2.5-4, 4.5-6.5), ROC=0.97 for T25FW>5 seconds, ROC = 0.96 for SDMT<50, and ROC=0.99 for 9HPT>25 seconds. Features were selected from both speech and gait tasks for each outcome.

Gait and speech features extracted using ML can precisely quantify their respective domains, but can further be integrated to recapitulate the major established clinical measures in MS. Longitudinal validation data (acquired, N=100) will be presented to inform their utility in monitoring of subtle progression independent of relapse activity (PIRA).