Posters
Presentation Type
Poster
Discipline Track
Clinical Science
Abstract Type
Research/Clinical
Abstract
Background: Parkinson's Disease (PD) is characterized by both motor and non-motor symptoms, and its diagnosis primarily relies on clinical presentation. There is a growing need for diagnostic tools to identify the early signs of PD, particularly the initial motor impairments often manifested as gait abnormalities. Here we seek to present preliminary findings to address this need. Our study focuses on using Machine Learning techniques (ML) to predict the PD clinical stage most efficiently and accurately. Specifically, we have sought to evaluate how spatiotemporal characteristics and other locomotor performance variables obtained on a walkway system can be utilized to identify the Hoehn and Yahr (HY) score in PD.
Methods: Six individuals with PD and 6 Healthy individuals participated in the study. PD patients were classified on the HY scale by a physician (score range 0-5). Participants completed eight passes on the Zeno Walkway while Protokinetics Movement Analysis Software recorded and calculated the temporal, spatial, and pressure measurements of within-step recordings. Data preprocessing and predictive modeling were analyzed using R and the caret package. Multiple regression, utilizing predictors such as gait speed, left and right steps, and walking methods (socks, shoes, and barefoot), were employed to normalize the data. The data was underfitted using 5% for training and 95% for testing and used three repeated 10-fold cross-validations. Models included Random Forest, Neural Networks, Naive Bayes, Support Vector Machines with Linear Kernel (SVM), and Penalized Multinomial Logistic Regression. Models were compared based on a weighted rank system, prioritizing successful prediction of 6th PD patient, followed by full data partition computational efficiency, model accuracy, interclass balanced accuracy, kappa, and weighted averages of area under the curve (AUC) for HY ratings.
Results: The SVM algorithm demonstrated the best ability to predict HY scores, achieving an overall underfitted model accuracy of 76%, interclass balanced accuracy of 82%, weighted AUC of 55%, Kappa of 62%, and full data partition learning of 2.8 seconds. Multimodal Logistic Regression (MLR) demonstrated the next best performance and achieved an overall underfitted model accuracy of 75% with interclass balanced accuracy of 84%, weighted AUC of 59%, Kappa of 68%, and full data partition learning of 87 seconds. The top predictor outcomes in SVM were Stride Velocity, Stride Length, Step Length, and Single Support Center of Pressure Distance. Moreover, MLR, Neural Networks, and SVM algorithms were successful in predicting the correct HY score (3) of the final recruited PD participant using only a quarter of the requested gait protocol data.
Conclusions: Our data highlights the need to test the accuracy and efficiency of multiple models to provide real-time learning in clinical populations. Furthermore, the success of the deployed ML algorithms in this study motivates further exploration to identify the economic feasibility of early detection of PD. The Gait mat and programmed software may assist patients in accessing affordable, validated, and reliable clinical assessment for early-stage Parkinson’s Disease.
Recommended Citation
Salinas, Daniel; Medellin, Gerardo; Bolado, Katherine; Gomez, Tomas; Potter-Baker, Kelsey; Hack, Nawaz Khan Abdul; and Vadukapuram, Ramu, "Assessing Gait Metrics for Early Parkinson's Disease Prediction: A Preliminary Analysis of Underfit Models" (2024). Research Symposium. 11.
https://scholarworks.utrgv.edu/somrs/2024/posters/11
Included in
Assessing Gait Metrics for Early Parkinson's Disease Prediction: A Preliminary Analysis of Underfit Models
Background: Parkinson's Disease (PD) is characterized by both motor and non-motor symptoms, and its diagnosis primarily relies on clinical presentation. There is a growing need for diagnostic tools to identify the early signs of PD, particularly the initial motor impairments often manifested as gait abnormalities. Here we seek to present preliminary findings to address this need. Our study focuses on using Machine Learning techniques (ML) to predict the PD clinical stage most efficiently and accurately. Specifically, we have sought to evaluate how spatiotemporal characteristics and other locomotor performance variables obtained on a walkway system can be utilized to identify the Hoehn and Yahr (HY) score in PD.
Methods: Six individuals with PD and 6 Healthy individuals participated in the study. PD patients were classified on the HY scale by a physician (score range 0-5). Participants completed eight passes on the Zeno Walkway while Protokinetics Movement Analysis Software recorded and calculated the temporal, spatial, and pressure measurements of within-step recordings. Data preprocessing and predictive modeling were analyzed using R and the caret package. Multiple regression, utilizing predictors such as gait speed, left and right steps, and walking methods (socks, shoes, and barefoot), were employed to normalize the data. The data was underfitted using 5% for training and 95% for testing and used three repeated 10-fold cross-validations. Models included Random Forest, Neural Networks, Naive Bayes, Support Vector Machines with Linear Kernel (SVM), and Penalized Multinomial Logistic Regression. Models were compared based on a weighted rank system, prioritizing successful prediction of 6th PD patient, followed by full data partition computational efficiency, model accuracy, interclass balanced accuracy, kappa, and weighted averages of area under the curve (AUC) for HY ratings.
Results: The SVM algorithm demonstrated the best ability to predict HY scores, achieving an overall underfitted model accuracy of 76%, interclass balanced accuracy of 82%, weighted AUC of 55%, Kappa of 62%, and full data partition learning of 2.8 seconds. Multimodal Logistic Regression (MLR) demonstrated the next best performance and achieved an overall underfitted model accuracy of 75% with interclass balanced accuracy of 84%, weighted AUC of 59%, Kappa of 68%, and full data partition learning of 87 seconds. The top predictor outcomes in SVM were Stride Velocity, Stride Length, Step Length, and Single Support Center of Pressure Distance. Moreover, MLR, Neural Networks, and SVM algorithms were successful in predicting the correct HY score (3) of the final recruited PD participant using only a quarter of the requested gait protocol data.
Conclusions: Our data highlights the need to test the accuracy and efficiency of multiple models to provide real-time learning in clinical populations. Furthermore, the success of the deployed ML algorithms in this study motivates further exploration to identify the economic feasibility of early detection of PD. The Gait mat and programmed software may assist patients in accessing affordable, validated, and reliable clinical assessment for early-stage Parkinson’s Disease.