Abstract
Introduction
Spiral drawing is a clinical diagnostic tool to aid classification of action tremor types, which are typical of essential tremor and other neurodegenerative disorders of complex etiology. However, objective tremor classification in the population is scant. In the Cooperative Health Research in South Tyrol (CHRIS) study, we derived novel metrics through digital spiral analysis (DSA) and assessed their performance for tremor classification.
Methods
Archimedes’ digital spiral drawing was performed with both hands by 10,983 adult CHRIS participants, resulting into multiple DSA measures of tremor amplitude, frequency, acceleration, speed, pressure, direction, and power. Both random and opportunistic sampling of 5,810 paired spirals from 2,905 participants (median age 50 years; age range:18-93 years; 53% females) were visually rated on screen in random order by an expert neurologist on a scale between 0 (no tremor) and 9 (maximum tremor) with 1 to 3 assignments per spiral.
Aggregating small relative frequencies in extreme scores, we performed ordinal random forest (ORF) analysis on four ordinal outcomes: scoreH,5={<2,2,3,4,5+}, and scoreH,6={<2,2,3,4,5,6+}, taking the highest value; scoreR,5={<2,2,3,4,5+}, and scoreR,6={<2,2,3,4,5,6+}, taking one value at random, over repeated assignments per spiral.
We investigated the classification performance of a selection of 37 DSA heterogeneous metrics, and both task-related and individual tremor determinants, totalling to 66 features. Five-fold cross-validation scheme was used to assess ORF model performances based on the rank probability score function. Linearly weighted kappa statistics between the observed and the predicted classes were obtained in the testing folds for average classification performance and based on individual class probability distributions over 1000 simulations for individual class forecasting performance. For each outcome, features were ranked by importance (I) for classification accuracy. We then run a final optimized ORF for out-of-sample classification on the whole dataset.
Results
Spiral scores ranged between 0 and 9 (median=2; IQR=2-3). Kappa statistics were fair and internally consistent for both scoreH (prediction range=0.397-0.483, forecasting range=0.281-0.342) and scoreR (prediction range=0.373-0.443, forecasting range=0.256-0.326). Out of 20 top-ranked variables for classification accuracy, 13 to 14 were in the amplitude domain for either scoreH or scoreR (I=0.0026-0.0168). Importance values of age, sex, and drawing-hand dominance had generally lower rankings (I<0.0045). As expected by opportunistic sampling design, final out-of-sample best class predictions were lower than the top-tremor classes for all score classifications.
Conclusion
A mixture of DSA semi-automated derived metrics can build a stable classification model for action tremor in large scale studies of tremor related disorders.