Stage-Wise Data Balancing Promoting Toe-tapping-based Classification of Parkinson's Disease Progression using Smart Insoles

doi:10.21203/rs.3.rs-3645803/v1

Download PDF

Research Article

Stage-Wise Data Balancing Promoting Toe-tapping-based Classification of Parkinson's Disease Progression using Smart Insoles

https://doi.org/10.21203/rs.3.rs-3645803/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Given the slow progression nature of Parkinson's Disease (PD), accurate stage classification is predominant for effective disease management. Traditional clinical evaluations, often based on brief physician-patient interactions, can miss nuanced disease progressions. With their continuous monitoring capabilities and bolstered by recent machine learning (ML) advancements, wearable devices such as monitoring insoles (MONI) present a promising solution. However, capturing a comprehensive dataset spanning all PD stages is challenging, leading to data imbalances. These imbalances can cause ML models to favor the majority class, achieving high accuracy but compromising clinical relevance. To address this inherent challenge in PD data, we identified two most effective sampling methods: Synthetic Minority Over-sampling Technique (SMOTE) and Tomek-Links. We compared their data balance performance using toe-tapping datasets collected from PD patients and their age-matched healthy controls (HC) wearing MONI. Additionally, we utilized SHapley Additive exPlanations (SHAP) for ranking to ascertain each technique's efficacy and verify the top factors contributing to PD stage classification, relating these factors to the underlying mechanisms or symptoms of PD. Data balancing enhanced stage classification by 5-7% compared to the original dataset. Our findings show that with SMOTE and Tomek-Links balanced data, we were able not only to improve the accuracy and robustness of stage classification but also observed the model emphasizing features that have a direct relation to each stage rather than favoring the HC majority class, thus ensuring a model is tailored to the PD domain.

Given the slow progression nature of Parkinson's Disease (PD), accurate stage classification is essential as it provides insights into the severity of the disease and the associated inflammatory response in addition to the potential PD progression effects [1]. PD is expected to impact over 1.2 million individuals in the U.S. by 2030 [2]. This growing prevalence underscores the urgency for enhanced diagnostic tools. Traditional diagnostic methods, anchored on the MDS-Unified Parkinson's Disease Rating Scale (UPDRS) clinical evaluations, can overlook subtle indicators of PD stages due to their dependence on transient physician-patient interactions.

Considering the limitations of traditional methods, wearable devices have become a promising solution by providing continuous and detailed monitoring for PD patients in daily life. MONI, an innovative smart insole, [3] recently developed by our team, can track and monitor their foot motion behaviors using machine learning (ML) techniques. [4, 5]. By eliminating the need for regular clinical visits, MONI streamlines the monitoring process in daily life and provides objective assessments, such as medication effectiveness [6] and risks of falling [7, 8].

However, data imbalance is a persistent challenge because it is difficult, if not impossible, to collect a balanced dataset in daily life. Such imbalances can skew ML models, leading them to favor the majority potentially [9]. Consequently, even if a model trained on skewed data achieves high accuracy, it might frequently predict the majority class simply because it's statistically more likely to be correct. However, this approach can overlook the importance of the minority class, especially when it holds significant clinical relevance, leading to a biased representation against the minority [10, 11]. Refining data balancing strategies tailored for PD-specific data is paramount, considering the above challenge. By achieving a harmonious data representation, we can unlock the full potential of ML models, ensuring they strike a balance between accuracy and clinical significance. Such a balance augments model efficacy, guaranteeing that derived insights resonate with the multifarious symptomatology of PD in real-life contexts. [12–14].

In addressing the challenges posed by data imbalances, several techniques have emerged. The Synthetic Minority Over-sampling Technique (SMOTE) and Tomek-Links are foundational strategies. Building on these core methods, researchers have developed domain-specific adaptations, such as minimizing gender bias in the gene expression data [15] and increasing PD detection accuracy by 7.51% using speech signal data. Mamidi et al. [16] explored using a hybrid SMOTE to detect PD through speech data, achieving an improvement of about 4% in accuracy. Képešiová et al. also introduced different sampling techniques to balance a speech dataset for PD prediction models, improving accuracy by about 3–5% compared to similar methods.

This paper delves deeper into an imbalanced dataset of toe-tapping motion performed by 6 early-stage and 5 mid-late-stage PD patients and 10 age-matched healthy controls (HC), each wearing a pair of MONI. The PD stage is determined by clinician evaluation and patient self-assessment using MDS-UPDRS scores. According to the Hoehn & Yahr staging system, early-stage PD encompasses stages 1 and 2, where symptoms are milder and primarily unilateral (stage 1) or beginning to affect both sides of the body without impairment of balance (stage 2). Mid-late-stage PD, as considered in this paper, covers stages 2 to 4, where symptoms progressively worsen, affecting balance (stage 3) and rendering patients significantly disabled but still able to stand and walk without assistance (stage 4). Therefore, for the purpose of this study, early-stage PD is categorized as Hoehn & Yahr stage 1, while mid-late-stage PD encompasses Hoehn & Yahr stages 2 to 4 [17, 18].

We propose to under-sample the majority class (HC) and over-sample the minority classes (early and Mid-to-late stages) to classify PD stages accurately. Different sampling approaches allow us to holistically evaluate the impact of these data-balancing techniques on the quality and reliability of the ML model-derived insights to PD stage progression classification. The remainder of this paper is organized as follows: Section II details the methods and framework employed, including data collection, preprocessing, sampling, and evaluation. In Section III, we present the results, highlighting model performance and feature importance across different datasets. Section IV discusses the implications of our findings, the significance of motion patterns, and the representation of the SMOTE dataset. Finally, we conclude in Section V, summarizing the key takeaways and potential future directions.

Our ML framework, outlined in Fig. 1, is organized into distinct phases: data collection, data preprocessing focusing on segmentation, sampling techniques, and model evaluation. In the data collection phase, we employed the MONI insoles, depicted in Fig. 2(a), to capture toe-tapping data. This dataset comprises recordings from 11 PD patients (6 early-stage and 5 mid-late-stage) and 10 HC. Detailed demographics and characteristics of these participants are provided in Table I. Each MONI insole is equipped with two accelerometers. These are strategically positioned at the first metatarsal ('T') and the heel area ('H'). These accelerometers operate at a sampling frequency of 150 Hz, as shown in Fig. 2 (b). Participants were guided to execute two specific 10-second toe-tapping exercises for the study. The first, termed "In-phase", required synchronized toe-tapping between the left and right leg. The second, "Out-phase", involved an alternating toe-tapping pattern where the two legs were out of sync.

Table I. Subject Characteristics

Healthy control

Early stage

Mid-late stage

Age

4 (61–70)

5 (71–80)

1 (over 80)

1 (51–60)

3 (61–70)

1 (71–80)

1 (over 80)

1(51–60)

3(71–80)

1(over 80)

Sex*

2 (M), 8 (F)

4 (M), 1 (F)

5 (M), 0 (F)

*Male (M), Female (F)

A. preprocessing

As illustrated in Fig. 1, our ML framework initiates with data preprocessing, wherein it is meticulously filtered using a 4th -order low-pass filter with a cutoff frequency of 12.5 Hz [19]. This cutoff frequency is specifically chosen to capture the essential frequencies associated with PD to eliminate any high-frequency disturbances while retaining the frequencies most indicative of PD symptoms. After this, the data is segmented into distinct motion patterns by pinpointing specific lift-toe and drop-toe events, notably the commencement and culmination of toe-tapping activities. Each of these segments is characterized by the notation\(\left({x}_{T}, {y}_{T}, {z}_{T},{std}_{T},{x}_{H}, {y}_{H},{z}_{H},{std}_{H}\right)\)where \(x, y \text{a}\text{n}\text{d} z\) symbolize the 3-axis acceleration data (orientation as shown in Fig. 2. (c)), and \(std\) denotes the standard deviation across these three axes. In this representation, \(T\) indicates the first metatarsal position, while \(H\) signifies the heel position.

From this preprocessed data, we extrapolate a comprehensive set of 120 features in total from 8 axes of \(\left({x}_{T}, {y}_{T}, {z}_{T},{std}_{T},{x}_{H}, {y}_{H},{z}_{H},{std}_{H}\right)\). From [7, 20], these features had the potential to capture the nuanced movement characteristics affected by PD. These encompass (i) 10 highly-ranked statistical features per axis in the time domain: Mean, Standard deviation (σ), Minimum (Min), Maximum (Max), Range, Root Mean Square (RMS), Skewness, Kurtosis, and Zero Crossing Rate (ZCR) and the Mean Crossing Rate (MCR) per axis; and (ii) 5 highly-ranked features in frequency domain: Peak frequency, Cepstral coefficients (comprising of mean, max, and min coefficients), and Spectral entropy. We utilize the Fast Fourier Transform to transform the signal from time to frequency domain. We denoted the feature with the position and the axis. For example, the mean of the standard deviation axis at the first metatarsal position is \({Mean}_{{T}_{std}}\).

We first established the features and then created four distinct datasets. This includes the original unbalanced dataset and three others designed to address this imbalance. As illustrated in Fig. 1, we applied sampling methods: Tomek-Links, SMOTE, and a combination of SMOTE-Tomek. These methods either reduce the majority classes or synthesize data resembling the minority classes. After preparing the datasets, we trained and evaluated them using a random forest (RF) model. We applied a 5-fold cross-validation (CV) process for validation, but only on the original data. Finally, we assessed the validation performance using precision, recall, F1-score, and AUC/ROC metrics.

B. Sampling Methods

To address the imbalance problem, we employ three data sampling methods. In addition to the original dataset, we created three sampling datasets, as depicted in Fig. 3. The original dataset consists of 714 samples from HC, 424 samples from early-stage PD patients, and 288 samples from mid-late-stage PD patients. Each of these samples is derived from features extracted from raw data files. Specifically, the 714 samples from HC come from 714 individual files. Every file represents a specific combination of subjects, tasks, segments, and sides. The segmentation was based on distinct toe-tapping cycles, identifying segments from the commencement (lift-toe) to the culmination (drop-toe) of toe-tapping activities. This approach yielded 714 segments for the HC group. Segments with general noise or without clear toe lift-toe drop events were excluded from the analysis. To illustrate further, if each file represents one segment, then 714 files would equate to 714 segments, starting from the left toe and ending at the toe drop for each segment.

Given the inherent imbalance in the dataset, the 424 samples from early-stage PD and the 288 samples from mid-late-stage PD will be oversampled to match the majority count of 714 from HC. Conversely, under-sampling will be applied to the 714 HC samples and the 424 early-stage PD samples to match the lowest count of 288 from mid-late-stage PD. A significant observation from this dataset is its inherent imbalance; the data from HC is disproportionately represented compared to the PD stages.

The Tomek-Links Dataset focuses on under-sampling as a strategy to balance the dataset. The primary objective of under-sampling is to diminish the number of samples from classes that are over-represented. Using the Tomek-Links method, unique pairs of instances from different classes are identified through Condensed Nearest Neighbor, ensuring no closer instances exist between them. Once this method is applied, each class has a uniform representation of 288 samples. This count mirrors the data points from the minority class in the original dataset, which indicates the mid to late stage. Under-sampling was explicitly applied to the other two predominant classes to achieve this balance.

The SMOTE Dataset employs over-sampling as its primary strategy. Over-sampling seeks to balance the dataset by augmenting the number of samples in classes that are under-represented. Rather than merely duplicating the samples of the minority classes (early and mid-late stages), the SMOTE method synthesizes new samples that are proximate to the existing ones within the feature space, which bolsters the robustness of the classifier. After the application of SMOTE, every class is populated with 714 samples.

The SMOTE-Tomek Dataset combines the strengths of both SMOTE and Tomek-Links. SMOTE creates synthetic samples for the minority class, while Tomek-Links remove samples from the majority class close to the minority class samples. This helps create a clearer distinct boundary between the two classes. With this combined method, the dataset is balanced, with each class having 714 samples.

A. Model Performance

Table II shows the results of a 5-fold CV evaluation on RF. The original dataset, inherently imbalanced, registers a precision of 88.12% and a recall of 87.36%. Incorporating Tomek-Links, an under-sampling technique, results in a notable increase in precision (95.00%) and recall (93.69%). The SMOTE method, which oversamples the minority class (Early and Mid-late stages) by creating synthetic instances, achieves a precision of 94.84% and a recall of 93.46%. When combining both Tomek-Links and SMOTE, the results from the SMOTE-Tomek dataset are closely aligned with those from using SMOTE alone, showcasing a precision of 94.67%, a recall of 93.18%, and an AUC-ROC of 99.86%. The slight variations in performance metrics between the two methods might not be statistically significant, given their proximity. Furthermore, the combined approach might offer improvements, but the subtle differences might be challenging to discern with the current sample size.

Table II. 5 folds Cross-validation (Validation on Real Data)

Dataset	Precision (%)	Recall (%)	F1-Score (%)	AUC-ROC (%)
Original	88.12	87.36	87.21	97.36
Tomek-Links	95.00	93.69	93.51	99.62
SMOTE	94.84	93.46	92.91	99.89
SMOTE-Tomek	94.67	93.18	92.48	99.86

The F1-score, a harmonized measure of precision and recall, and the AUC-ROC, indicative of the model's capability to distinguish between classes, further corroborate the observations. The original dataset's AUC-ROC of 97.36% experienced a significant rise with both the Tomek-Links (99.62%) and SMOTE (99.89%) techniques. Such elevated AUC-ROC values, especially in the SMOTE dataset, suggest that the model adeptly differentiates between the positive and negative classes, even when synthetic instances are incorporated.

B. Feature Importance Across Different Datasets

SHAP (SHapley Additive exPlanations) analysis [21] calculates the contribution of each feature to every possible prediction in our Random Forest model with 1000 estimators. The importance ranking of features can vary between classes due to distinct characteristics captured by different features. While SHAP doesn't have tunable parameters, the underlying RF model does. Adjusting parameters like the number of estimators or max depth might influence the SHAP values, but with 1000 estimators, our model showed no shifts in feature importance. Our results in Fig. 4 display the top 7 features' importance across datasets, highlighting their role in motion pattern differentiation and the impact of dataset balancing.

1) Original Dataset

The top features predominantly revolved around statistical measures. The 5 top features are more important for HC than early- and mid-late-stage PD, while the \({\text{M}\text{i}\text{n}}_{{\text{H}}_{\text{s}\text{t}\text{d}}}\) and \({\text{M}\text{i}\text{n}}_{{\text{H}}_{\text{z}}}\) favors early-stage PD over the other two classes. This reflects the issue of the imbalanced dataset where HC has the most samples and mid-late-stage PD has the least (Fig. 4).

2) Tomek-Links Dataset

With under-sampling, the top features change from the results of the original dataset. It shifts the feature importance towards the minority class, where the top features share importance to all three classes. The total importance of all 7 features is closely balanced for each class.

3) SMOTE Dataset

With oversampling, the top features also change from the original dataset and differ from the Tomek-Links dataset. Its emphasis shifts towards the mid-late-stage PD more.

4) SMOTE-Tomek Dataset

The feature importance results from the combined approach are similar to the results of SMOTE. Therefore, it achieves a similar performance to the SMOTE dataset.

A. Impact of Imbalanced Datasets

The challenges posed by imbalanced datasets in ML are multifaceted, particularly when it comes to classification tasks. When a particular class dominates, it can inadvertently introduce a bias, overshadowing the subtleties of the minority class. In the original dataset Fig. 4(a), the HC class heavily influenced the model, emphasizing features like the \({\text{M}\text{e}\text{a}\text{n}}_{{\text{T}}_{\text{S}\text{t}\text{d}}}\). This overemphasis on certain features, predominantly from the majority class, risked the model's ability to generalize, potentially sidelining critical patterns from the minority classes.

However, balancing techniques of Tomek-Links and SMOTE bring about a noticeable shift in feature prioritization. The Tomek-Links dataset Fig. 4(b), despite its smaller size of 288 samples, shows some changes in feature emphasis. In contrast, the SMOTE dataset and the combined SMOTE-Tomek approach, both with 714 samples, exhibit more consistent and stable feature importance. For instance, in the SMOTE dataset, features like the \({\text{R}\text{M}\text{S}}_{{\text{T}}_{\text{X}}}\)and the \({\text{S}\text{p}\text{e}\text{c}\text{t}\text{r}\text{a}\text{l} \text{E}\text{n}\text{t}\text{r}\text{o}\text{p}\text{y}}_{{\text{T}}_{\text{X}}}\)became more prominent.

The shift in feature importance post-balancing offers a more consistent and unbiased view of PD motion feature importance. The features that emerged as significant post-balancing emphasize the importance of capturing the multifaceted PD in real-life contexts with no biases toward the majority class. These shifts indicate the model's evolving understanding, highlighting patterns previously overshadowed by the majority class's dominant features. This recalibration towards a balanced perspective ensures the model's predictions are not only statistically accurate but also aligned with the data's actual underlying patterns, enhancing its relevance in all classes of PD.

B. Feature Significance in Motion Patterns

Post-balancing, certain features emerge as significant, suggesting their importance in capturing PD's motor symptoms. The 4 features listed below are prominent in the balanced datasets.

1) RMS_TX

This is one of the most prominent features of post-balancing across datasets Fig. 4 (c & d), which captures the consistency and magnitude of toe movements, potentially highlighting irregular motion patterns typical of PD patients.

Its consistent prominence across balanced datasets underscores its importance in differentiating HC and PD motion patterns. Interestingly, the significance of RMS is also emphasized in evaluating the regional spiking activity in the Subthalamic Nucleus (STN) of PD patients, where it is used to gauge the consistency of neuronal activity [22].

2) Min_H_std

This indicates the range or depth of heel movements, with lower values potentially pointing towards restricted or limited motion, a characteristic symptom of PD.

3) Spectral Entropy_TX

This feature is prominent in the SMOTE dataset, measuring the toe-tapping motion's rhythmic consistency. A reduced spectral entropy could indicate motor coordination issues in PD patients, emphasizing the importance of rhythmic irregularities in distinguishing between healthy individuals and PD patients. This rhythmic irregularity might manifest as disruptions in the STN's neuronal activity, especially when influenced by sedative drugs like remifentanil, which interfere with identifying the dorsolateral oscillatory region in the STN [22].

4) Kurtosis_HX

It identifies sharp deviations from typical motion patterns, becoming more evident in the balanced datasets. This suggests that these deviations, which might have been diluted in the original dataset, are now recognized as significant PD biomarkers.

In all datasets, post-balancing, there's a clear trend: the top features are more stable and consistently contributed by the same classes. This stability, especially in the oversampled datasets, underscores the importance of balanced datasets in capturing PD's multifaceted symptomatology. The under sampling might have achieved slightly higher accuracy due to its smaller sample size. Still, with their relatively larger sample size, the oversampling and combined techniques provided a more comprehensive view, free from the dominant influence of the HC class.

C. Representation of SMOTE Dataset

From the model performance and feature importance analysis, SMOTE and SMOTE-Tomek oversampling methods emerge as the most suitable balancing methods for our original dataset. To better understand what oversampling entails, Fig. 5 delineates the distribution of real versus synthetic data in the SMOTE dataset.

Figure 5(a) showcases a t-SNE visualization of the "Early stage" class. This visualization distinctly marks the locations of actual data points with the synthetic ones. The proximity and overlap between the actual and synthetic data points in this visualization underscore the accuracy with which SMOTE replicates the data distribution for the early stages of the condition. Figure 5(b) presents a similar t-SNE visualization for the "Mid-late stage" class. Here again, the synthetic data points generated by SMOTE align closely with the real data patterns, suggesting that the intrinsic characteristics of the late-stage class are well-represented by the synthetic samples.

These visualizations reveal that synthetic samples produced by SMOTE closely mirror real data patterns. Such alignment indicates that the synthetic data aptly represents the intrinsic data distribution, reinforcing SMOTE's potential in addressing class imbalances. When trained with SMOTE, the visual congruence of synthetic and real instances suggests that the model capitalizes on authentic data patterns, leading to its enhanced performance metrics.

D. Limitation and Future Work

The capabilities of the MONI smart insole as a daily assessment tool have shown promising results. It is designed for continuous monitoring and assessment in real-world scenarios. Its primary objective is to facilitate the tracking of PD progression, providing a nuanced understanding of the patient's daily motor experiences. While our current research has primarily focused on toe-tapping motions, PD presents a range of motor symptoms, including freezing of gait, tremors, and bradykinesia. Expanding the MONI’s monitoring capabilities to include these varied motor impairments alongside toe-tapping motion could significantly enhance its utility.

However, our study has its limitations. The dataset we used is somewhat constrained, which might introduce biases and reduce the depth of our insights. To mitigate this, future work focusses on including a more diverse range of patient demographics and symptom severities. This would not only improve the model's predictive accuracy but also contribute to a more holistic understanding of PD and its impact on patients' lives. Additionally, exploring the integration of other relevant data, such as environmental factors, could provide valuable insights into the progression and variability of PD symptoms

In conclusion, with SMOTE and SMOTE-Tomek, the model performance is improved by 5-7% compared to the original dataset. More importantly, our findings emphasize the importance of addressing dataset imbalances in PD monitoring. By achieving a harmonious data representation, we not only enhance the statistical robustness of our model but also ensure its relevance. The integration of wearable devices like MONI, combined with sophisticated data analysis techniques, paves the way for a more accurate and timely assessment of PD, promising improved patient life quality and health outcomes.

A. Ethics Approval and Consent to Participate

Our research involved human subjects, and we adhered to all necessary ethical considerations. The Human Research Protection Program at the Division of Research, Texas A&M University, provided full approval for our ethical and experimental procedures (Approval No. IRB2019-0368F). This study was conducted in strict accordance with the principles outlined in the Socio-Behavioral Protocol. We obtained informed consent from all participants or their legal guardians, ensuring compliance with both ethical standards and journal guidelines.

B. Consent for Publication

Not applicable.

C. Availability of Data and Materials

The datasets generated and/or analyzed during this study are currently not publicly available. However, they can be made accessible upon reasonable request. For any inquiries regarding the data, please contact the corresponding author.

D. Competing Interests

The authors declare no competing interests.

E. Funding

Not applicable.

F. Authors' Contributions

Ibrahim K. Almuteb: Assisted in data collection and experimental setups, conducted data analysis, wrote the manuscript, and supervised the data interpretation process.

Dr. Rui Hua: Conducted data collection and experimental setups, contributed to manuscript editing and review, and assisted in data analysis.

Dr. Ya Wang: Oversaw the review process, contributed to experimental designs, provided strategic direction for the research, and ensured compliance with standards.

Z. Fan et al., "Systemic activation of NLRP3 inflammasome and plasma α-synuclein levels are correlated with motor severity and progression in Parkinson’s disease," Journal of Neuroinflammation, vol. 17, no. 1, p. 11, 2020/01/08 2020,
C. Marras et al., "Prevalence of Parkinson’s disease across North America," npj Parkinson's Disease, vol. 4, no. 1, p. 21, 2018/07/10 2018.
R. Hua and Y. Wang, "Monitoring Insole (MONI): A Low Power Solution Toward Daily Gait Monitoring and Analysis," IEEE Sensors Journal, vol. 19, no. 15, pp. 6410-6420, 2019.
R. Hua and Y. Wang, "A Customized Convolutional Neural Network Model Integrated With Acceleration-Based Smart Insole Toward Personalized Foot Gesture Recognition," IEEE Sensors Letters, vol. 4, no. 4, pp. 1-4, 2020.
R. Hua and Y. Wang, "Robust Foot Motion Recognition Using Stride Detection and Weak Supervision-Based Fast Labeling," IEEE Sensors Journal, vol. 21, no. 14, pp. 16245-16255, 2021.
R. Hua and Y. Wang, "Distinguishing Medication ON and OFF for Parkinson’s Disease Through Repetitive Foot Motion Recognition and Analysis," IEEE Sensors Journal, vol. 22, no. 12, pp. 12219-12227, 2022..
R. Hua, Y. Wang, D. M. Kennedy, J. E. Hubbard, and Y. Wang, "Toe Tapping Based Falling Risk Evaluation for Patients With Parkinson's Disease Using Monitoring Insoles," IEEE Sensors Letters, vol. 6, no. 6, pp. 1-4, 2022.
R. Hua, N. Nguyen, A. Jawad, K. Kazmierski, B. White, and Y. Wang, "Insole-Based Falling Risk Evaluation for Patients With Parkinson's Disease Through Walking While Holding a Cup of Water," IEEE Sensors Letters, vol. 6, no. 6, pp. 1-4, 2022.
S. Moon et al., "Classification of Parkinson’s disease and essential tremor based on balance and gait characteristics from wearable motion sensors via machine learning techniques: a data-driven approach," Journal of NeuroEngineering and Rehabilitation, vol. 17, no. 1, p. 125, 2020/09/11 2020,.
C. L. Curchoe, A. Flores-Saiffe Farias, G. Mendizabal-Ruiz, and A. Chavez-Badiola, "Evaluating predictive models in reproductive medicine," Fertility and Sterility, vol. 114, no. 5, pp. 921-926, 2020/11/01/ 2020.
H. Gunduz, "Deep Learning-Based Parkinson’s Disease Classification Using Vocal Feature Sets," IEEE Access, vol. 7, pp. 115540-115551, 2019.
H. Fröhlich et al., "Leveraging the Potential of Digital Technology for Better Individualized Treatment of Parkinson's Disease," (in eng), Front Neurol, vol. 13, p. 788427, 2022.
F. Alharbi, L. Ouarbya, and J. A. Ward, "Synthetic Sensor Data for Human Activity Recognition," in 2020 International Joint Conference on Neural Networks (IJCNN), 19-24 July 2020 2020, pp. 1-9.
X. Yang, Q. Ye, G. Cai, Y. Wang, and G. Cai, "PD-ResNet for Classification of Parkinson’s Disease From Gait," IEEE Journal of Translational Engineering in Health and Medicine, vol. 10, pp. 1-11, 2022, doi: 10.1109/JTEHM.2022.3180933.
E. Flynn, A. Chang, and R. B. Altman, "Large-scale labeling and assessment of sex bias in publicly available expression data," BMC bioinformatics, vol. 22, no. 1, pp. 1-23, 2021.
N. T. Mamidi, G. F. S. Medikonda, S. SanghaviValmeti, and V. K. Vadladi, "A Hybrid Approach to Parkinson's Disease Detection using Speech Attributes: The Combination of SMOTE and Active Learning," in 2023 2nd International Conference on Applied Artificial Intelligence and Computing (ICAAIC), 4-6 May 2023 2023, pp. 350-356.
J. M. Templeton, C. Poellabauer, and S. Schneider, "Classification of Parkinson’s disease and its stages using machine learning," Scientific Reports, vol. 12, no. 1, p. 14036, 2022/08/18 2022.
P. Martínez-Martín et al., "Parkinson's disease severity levels and MDS-Unified Parkinson's Disease Rating Scale," Parkinsonism & Related Disorders, vol. 21, no. 1, pp. 50-54, 2015/01/01/ 2015.
T. Heida, E. C. Wentink, and E. Marani, "Power spectral density analysis of physiological, rest and action tremor in Parkinson’s disease patients treated with deep brain stimulation," Journal of NeuroEngineering and Rehabilitation, vol. 10, no. 1, p. 70, 2013/07/08 2013.
K. Ren, Z. Chen, Y. Ling, and J. Zhao, "Recognition of freezing of gait in Parkinson’s disease based on combined wearable sensors," BMC Neurology, vol. 22, no. 1, p. 229, 2022/06/21 2022,.
C. M. Scavuzzo et al., "Feature importance: Opening a soil-transmitted helminth machine learning model via SHAP," Infectious Disease Modelling, vol. 7, no. 1, pp. 262-276, 2022/03/01/ 2022,.
A. Benady et al., "Sedative drugs modulate the neuronal activity in the subthalamic nucleus of parkinsonian patients," Scientific Reports, vol. 10, no. 1, p. 14536, 2020/09/03 2020, .

No competing interests reported.

Download PDF

Editorial decision: Revision requested
06 May, 2024
Reviews received at journal
24 Apr, 2024
Reviewers agreed at journal
12 Apr, 2024
Reviews received at journal
12 Dec, 2023
Reviewers agreed at journal
03 Dec, 2023
Reviewers invited by journal
29 Nov, 2023
Editor assigned by journal
22 Nov, 2023
Submission checks completed at journal
22 Nov, 2023
First submitted to journal
21 Nov, 2023

You are reading this latest preprint version

Stage-Wise Data Balancing Promoting Toe-tapping-based Classification of Parkinson's Disease Progression using Smart Insoles

Status:

Version 1

Abstract

Figures

I. INTRODUCTION

II. METHODS AND FRAMEWORK

A. preprocessing

B. Sampling Methods

III. RESULTS

A. Model Performance

B. Feature Importance Across Different Datasets

1) Original Dataset

2) Tomek-Links Dataset

3) SMOTE Dataset

4) SMOTE-Tomek Dataset

IV. DISCUSSION

A. Impact of Imbalanced Datasets

C. Representation of SMOTE Dataset

D. Limitation and Future Work

V. CONCLUSION

Declarations

References

Additional Declarations

Status:

Version 1