Parkinsons disease (PD) is a long-term degenerative disorder of the central nervous system. Approximately 60,000 Americans are diagnosed with PD each year. However, the testing for PD is a challenge and cause trouble to explore the association with covariates. In this study, we use wearable sensors to obtain the data testing the vertical ground reaction force (VGRF), in order to investigate the most influential features, including demographics and motor related biomedical signals, for PD prediction. The result selects 7 features which worth being used to do prediction and the accuracy is quite high.
Watch Prediction of Parkinson's Disease With a Regression Model Using Vertical Ground Reaction Force Data as presented by the authors on the SAS Users YouTube Channel.
Figure 1. Data collection steps and pipeline of analysis.
For each of the 18 time series variables, we considered 16 features, including location measures, dispersion measures, shape measures and autocorrelation, resulting in 288 features in total. By using a regression-based method with AIC as the model selection criteria, with age, gender and study group indicator being fixed in the model, 7 among 288 features were selected, including standard deviation of L1 sensor, skewness of L6 sensor, range of R1 sensor, median of R4 sensor, autocorrelation (ACF) with 30 lags of R4 sensor, range of R6 sensor, and coefficient of variation of R7 sensor. The ACF was rescaled by multiplying 100 due to its relatively small scale.
Table 1 showed the odds ratio estimates and 95% confidence intervals obtained from the logistic regression model which was performed on the entire dataset with 165 subjects. After adjusting for other covariates, age was not significantly associated with Parkinson disease status, but gender had a significant effect. The probability of developing Parkinson disease for a male is 5.793 times than that for a female, with 95% CI (1.704, 19.700). After adjusting for covariates, except the range of R1 sensor, other selected features were all significantly associated with Parkinson disease status. Standard deviation of L1 sensor was the most significant feature. Adjusting for other covariates, if a patient’s standard deviation of L1 sensor was getting one unit larger, the probability that he/she develops Parkinson disease would be 0.963 times as before, with 95% CI (0.943, 0.984).
Table 1. Odds ratio estimates and 95% confidence intervals obtained from logistic regression model.
|Effect||Odds ratio||Lower 95% CI||Upper 95% CI||Pr > ChiSq|
|Study group 1 vs 3||1.107||0.301||4.068||0.410|
|Study group 2 vs 3||3.105||0.626||15.406||0.122|
|Gender male vs female||5.793||1.704||19.700||0.005|
|L1 std deviation||0.963||0.943||0.984||0.001|
|100x (R4 ACF 30)||1.086||1.019||1.158||0.011|
To evaluate the prediction performance of the 7 selected features, we compared the final model with a baseline model. Figure 2 (A) showed the ROC curve of the final model including the 7 selected features as well as age, gender and group indicator. Figure 2 (B) showed the ROC curve of the baseline model which only included age, gender and group indicator. Both models were performed on the entire dataset. Evidently, our final model (AUC=0.9329) was considerably better than the baseline model (AUC=0.6134).
Figure 2. (A, left) ROC curve and AUC of final model with 7 selected features as well as age, gender and group indicator, on the entire dataset. (B, right) ROC curve and AUC of the baseline model with age, gender and group indicator, on the entire dataset.
We also compared the two models on the test sets using a resampling method that randomly split the dataset 100 times. Table 2 showed the averaged AUCs and the standard deviations of the two models on test sets. The averaged AUC of our final model was 0.9039, while the averaged AUC of the baseline model was only 0.5303, which demonstrated that the selected features were highly important in terms of Parkinson disease prediction. It was also suggested that age, gender and group indicator only had limited predictability.
Table 2. Averaged AUCs and 95% confidence intervals of the final model and the baseline model on test sets, with 100 times repeated cross validation.
|Model||Mean of AUCs||Standard deviance of AUCs|
To further explore the selected features, we performed Principal component analysis (PCA) on all selected features, and visualized the data by using the first two principal components in Figure 3. It was clear that patients with Parkinson disease and healthy controls could be separated by the first two principal components, which further demonstrated the importance of our selected features.
Figure 3. Visualization on the data by using the first two principal components obtained from the selected features.
Vertical ground reaction force (VGRF) records in this study reflect the jitter and strength of patients’ legs. The jitter frequency is associated with the status of PD. The result follows the common sense of PD in most cases. First, the less a patient’s jitter dispersion and dispersion frequency are, the higher probability this patient may have PD. This is because PD patients invest much less force in striding than healthy controls, so that patients with high jitter frequency are not inclined to be PD. In addition, VGRF of PD patients is more likely to have a narrower range and be symmetric distributed due to the smaller skewness, when compared to healthy controls. The median of VGRF in PD patients tends to be higher than that in healthy controls. The autocorrelation across the time is higher in PD patients than that in the control group. This may because PD patients have a reduced stride length and a short average swing time, resulting in that the walking force of PD patients is relatively stable and predictable from the past records [Frenkel-Toledo et al., 2005].
Our study demonstrates a simple way to predict PD status from the walking condition of patients, which is usually obtained by accessible devices with force sensors. Specifically, we found that logistic regression, a conventional approach shows satisfactory results with a high predicting accuracy with AUC 0.9039 when 10 features are used. With the increasing trend of portable devices and digital health, we believe that our method with simple features can be widely used to predict clinical outcomes such as Parkinson disease that is potentially related to time series predictors.
 Mirelman, A., Bonato, P., Camicioli, R., Ellis, T. D., Giladi, N., Hamilton, J. L., ... & Almeida, Q. J. (2019). Gait impairments in Parkinson's disease. The Lancet Neurology,
 Parkinson’s Foundation. https://www.parkinson.org/Understanding-
 Di Biase, L., Di Santo, A., Caminiti, M. L., De Liso, A., Shah, S. A., Ricci, L., & Di Lazzaro, V. (2020). Gait analysis in Parkinson’s disease: an overview of the most accurate markers for diagnosis and symptoms monitoring. Sensors, 20(12), 3529.
 Shah, J., Pillai, L., Williams, D. K., Doerhoff, S. M., Larson-Prior, L., Garcia-Rill, E., & Virmani, T. (2018). Increased foot strike variability in Parkinson's disease patients with freezing of gait. Parkinsonism & related disorders, 53, 58-63.
 Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff, J. M., Ivanov, P. C., Mark, R. G., ... & Stanley, H. E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. circulation, 101(23), e215-e220.
 Goldsack, J., Chasse, R. A., & Wood, W. A. (2019). Digital endpoints library can aid clinical trials for new medicines-STAT. STAT https://www. statnews.
 Psaltos, D., Chappie, K., Karahanoglu, F. I., Chasse, R., Demanuele, C., Kelekar, A., ... & Cai, X. (2019). Multimodal wearable sensors to measure gait and voice. Digital biomarkers, 3(3), 133-144.
 Snyder, P., Tummalacherla, M., Perumal, T., & Omberg, L. (2020). mhealthtools: A Modular R Package for Extracting Features from Mobile and Wearable Sensor Data. Journal of Open Source Software, 5(47), 2106.
 Frenkel-Toledo, S., Giladi, N., Peretz, C., Herman, T., Gruendlinger, L., & Hausdorff, J.M. (2005). Effect of gait speed on gait rhythmicity in Parkinson's disease: variability of stride time and swing time respond differently. Journal of neuroengineering and rehabilitation, 2(1), 1-7.
 Colyer S.L., Evans M., Cosker D.P., Salo A.I.T. A Review of the Evolution of Vision-Based Motion Analysis and the Integration of Advanced Computer Vision Methods Towards Developing a Markerless System. Sports Med. -Open. 2018;4:24.
 Derawi, M. O., Ali, H., & Cheikh, F. A. (2011). Gait recognition using time-of-flight sensor. BIOSIG 2011–Proceedings of the Biometrics Special Interest Group.
 Khoury, N., Attal, F., Amirat, Y., Oukhellou, L., & Mohammed, S. (2019). Data-driven based approach to aid Parkinson’s disease diagnosis. Sensors, 19(2), 242.
 Perumal, S. V., & Sankar, R. (2016, November). Gait monitoring system for patients with Parkinson's disease using wearable sensors. In 2016 IEEE Healthcare Innovation Point-Of-Care Technologies Conference (HI-POCT) (pp. 21-24). IEEE.
We appreciated the help from our friends Yun He, Zhiji Zhang, Kun Qian, Siqi Huang and advisor Prof. Yuanjia Wang.
1. Base SAS® Procedures Guide
2. SAS® For Dummies®
Registration is open! SAS is returning to Vegas for an AI and analytics experience like no other! Whether you're an executive, manager, end user or SAS partner, SAS Innovate is designed for everyone on your team. Register for just $495 by 12/31/2023.
If you are interested in speaking, there is still time to submit a session idea. More details are posted on the website.