Table 2 Proposed methods in the literature for estimating the severity of PD represented by UPDRS III
From: Ensemble deep model for continuous estimation of Unified Parkinson’s Disease Rating Scale III
Reference | PwP | Sensors location | Method | Unobtrusive | Estimated metric | Gold-standard label | Validation Method | r | MAE |
|---|---|---|---|---|---|---|---|---|---|
Griffiths et al. [8] | 25 | Wrist | Statistical approach | Yes | Bradykinesia score | UPDRS III | Held-out testing set | 0.64 | 18 |
Parisi et al. [17] | 34 | Chest, left and right thighs | Multiple k-Nearest Neighbors models to estimate LA, S2S and G. | No (task-dependent) | Sum of leg agility, sit-to- stand and gait items of UPDRS III | Sum of leg agility, sit-to- stand and gait items of UPDRS III | LOOCV | 0.79 | - |
Rodriguez-Molinero et al. [15] | 75 | Waist | Linear regression | No (task-dependent) | Gait item of UPDRS III | UPDRS III | Held-out testing set | -0.56 | - |
Pulliam et al. [13] | 13 | Wrist and ankle | Multiple linear regression models to estimate tremor, bradykinesia and dyskinesia | Yes | Radar chart of PD tremor, bradykinesia and dyskinesia | UPDRS III | - | 0.81 | - |
Zhan et al. [18] | 152 | Smartphone | Rank-based framework for disease severity score [30] | No (task-dependent) | Mobile PD score | UPDRS III | Held-out testing set | 0.88 | - |
Abrami et al. [28] | 60 | Both wrists | Clustering and Markov-Chain | Yes | Multi-dimensional scale | Sum of tremor, bradykinesia and gait items of UPDRS III | Held-out testing set | \(r^2\) = 0.64 in clinic \(r^2\) = 0.43 at home | |
Butt et al. [25] | 64 | Wrist, fingers, and foot | Adaptive neuro- fuzzy inference system | No (task-dependent) | UPDRS III | UPDRS III | Tenfold cross validation | 0.81 | - |
The developed approach in this study | 24 | Wrist and ankle | Ensemble of dual- Channel LSTM, CNN-LSTM using raw signals and CNN-LSTM using spectrogram | Yes | UPDRS III | UPDRS III | LOOCV | 0.79 | 5.95 |