Aravalli Sainath Chaithanya; Nadipudi Kiran Kumar; Gugulothu Venkatesh Prasad; Bejawada Keerthana

doi:10.4258/hir.2025.31.3.274

Journal List > Healthc Inform Res > v.31(3) > 1516092146

Go to TopGo to Top Go to BottomGo to Bottom

TOOLS

Chaithanya, Kumar, Prasad, and Keerthana: Advancements in Parkinson’s Disease Prediction Using Machine Learning: A Neurological Perspective

Original Article

Healthcare Informatics Research 2025; 31(3): 274-283.

Published online: 31 July 2025

DOI: https://doi.org/10.4258/hir.2025.31.3.274

Advancements in Parkinson’s Disease Prediction Using Machine Learning: A Neurological Perspective

Aravalli Sainath Chaithanya

, Nadipudi Kiran Kumar

, Gugulothu Venkatesh Prasad

, Bejawada Keerthana

Department of Electronics and Communication Engineering, Rajiv Gandhi University of Knowledge Technologies, Basar, Telangana, India

Corresponding Author: Aravalli Sainath Chaithanya, Department of Electronics and Communication Engineering, Rajiv Gandhi University of Knowledge Technologies, Basar-504107, Telangana, India. Tel: +91-8520948605, Email: chaitanya.aravalli@gmail.com (https://orcid.org/0000-0002-8317-0827)

Received 17 July 2024 Revised 11 February 2025 Accepted 28 February 2025

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

Objectives

This study aims to predict the severity of Parkinson’s disease (PD) by leveraging a comprehensive dataset integrating cerebrospinal fluid protein and peptide data sourced from UniProt, normalized protein expression metrics, clinical assessments, and gait data. The dataset comprised 248 PD patients monitored longitudinally, with periodic evaluations including 227 proteins, 971 peptides, gait parameters, and Movement Disorder Society-sponsored revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) scores at baseline 0, 6, 12, and 24 months.

Methods

A multifaceted machine learning framework was employed, consisting of random forest, TensorFlow decision forests, and a custom-developed phase-shift ensembling model. Additionally, regression techniques such as linear regression, random forest regressor, decision tree regressor, and K-nearest neighbors were utilized to support the predictions. These models aimed to forecast PD severity as reflected by UPDRS scores.

Results

The custom phase-shift ensembling model demonstrated superior predictive performance, achieving an average symmetric mean absolute percentage error (sMAPE) of 55 across all UPDRS sections. Notably, the random forest regressor excelled in predicting motor function severity (UPDRS-III), attaining an sMAPE of 77.32, indicating its ability to model complex disease progression dynamics effectively.

Conclusions

Integrating biological markers, clinical scores, and gait dynamics facilitates accurate modeling of PD progression. The ensemble-based approach, particularly phase-shift ensembling, improves prediction robustness and interpretability, offering a powerful strategy for the early prediction of PD severity. This study highlights the value of multi-source data fusion and advanced machine learning techniques in supporting early diagnosis and informed treatment planning for neurodegenerative diseases.

Keywords: Biomarkers, Gait, Machine Learning, Neurodegenerative Diseases, Parkinson Disease, Proteomics, Severity of Illness Index

I. Introduction

Parkinson’s disease (PD) [1] is a chronic neurodegenerative disorder characterized primarily by the degeneration of dopamine-producing neurons. Clinically, PD manifests motor symptoms, including tremors, bradykinesia, rigidity, and postural instability, alongside non-motor symptoms such as depression, anxiety, and cognitive impairment [2] (Figure 1). PD is the second most prevalent neurodegenerative condition after Alzheimer’s disease, significantly impacting patients’ quality of life and posing considerable socioeconomic challenges. Despite extensive research efforts, the exact etiology of PD remains unclear.

Previous studies have utilized various modalities, such as voice recordings, spiral drawings, gait analysis, and clinical assessments, for the diagnosis and prediction of PD severity. Machine learning (ML) models have shown substantial potential in these tasks. For instance, Saeed et al. [3] employed voice data to develop predictive models, achieving high accuracy but facing challenges related to small datasets and overfitting. Similarly, Sadek et al. [4] applied artificial neural networks (ANNs) to voice recordings, attaining near-perfect accuracy but necessitating stronger validation to minimize bias. Nawi et al. [5] demonstrated that neural networks outperformed decision trees on dysphonia-based features but suggested the need for improved feature selection. Likewise, Arthi and Sylviaa [6] utilized magnetic resonance imaging data combined with ANN models for PD staging, demonstrating potential yet lacking statistical rigor and comprehensive clinical benchmarks.

Building upon prior research, Nilashi et al. [7] employed clustering and predictive learning for PD severity estimation, underscoring challenges related to feature extraction and limited datasets. Likewise, Alshammri et al. [8] combined ML and deep learning for voice-based PD prediction, emphasizing the importance of refined feature selection. Ahmadi Rastegar et al. [9] and Alzubaidi et al. [10] explored inflammatory cytokines with deep neural networks for early diagnosis, indicating promise but requiring validation in larger cohorts. Notably, Johri and Tripathi [11] analyzed gait and speech impairments using neural networks, achieving high accuracy and proposing the inclusion of additional features to enhance prediction further.

Emerging research increasingly advocates for integrating multimodal data—incorporating both biological and motor function metrics—to improve predictive accuracy. For example, Sarankumar et al. [12] demonstrated that the DBIN model outperformed traditional voice-based prediction methods, while Varghese et al. [13] identified support vector regression as an effective tool for severity estimation. In telemedicine, Govindu and Palwe [14] reported the strong performance of a random forest classifier for PD detection, suggesting the potential for further improvement by combining audio and REM sleep data.

Given the progressive nature of PD and the limitations inherent to single-modality models, this study integrates cerebrospinal fluid (CSF) mass spectrometry data with gait analysis for comprehensive severity prediction. This multimodal approach captures critical motor features complementary to biomarker-based metrics. Unlike classification tasks that rely predominantly on F1-score metrics [15], our regression-based model employed mean squared error (MSE) and symmetric mean absolute percentage error (sMAPE), alongside a novel acceptable prediction range (APR), improving clinical relevance and interpretability. By leveraging protein and peptide datasets, clinical assessments, and gait analysis, we aim to develop an accurate and interpretable model to support clinical decision-making and enhance patient care.

II. Methods

The process flow for this study is shown in Figure 2.

1. Data Collection and Integration

PD is a progressive neurodegenerative disorder characterized by the loss of dopaminergic neurons, leading to motor impairments such as tremors, bradykinesia, postural instability, and gait abnormalities. The severity of these symptoms is quantified using the Movement Disorder Society-sponsored Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) [16], which supports diagnosis and informs treatment decisions.

This study aimed to predict PD severity at specific time points (0, 6, 12, and 24 months) using ML techniques. We utilized a multimodal dataset comprising biomarker, clinical, and motor function data [17]. The dataset integrates CSF mass spectrometry (CSF-MS) readings, clinical assessments, and gait analysis to comprehensively evaluate PD progression.

The dataset consisted of four components:

(1) Protein dataset: The protein dataset included data from UniProt [18] and normalized protein expression (NPX) metrics [19].
- UniProt is a comprehensive database of protein sequences, functions, and annotations, comprising UniProt Knowledgebase (UniProtKB), UniProt Reference Clusters (Uni-Ref), and UniProt Archive (UniParc).
- NPX values quantified alpha-synuclein levels in Olink’s Log2 scale, providing standardized protein concentrations (Figure 3).
- Differentially expressed proteins reflected neurodegeneration, providing insights into PD modeling.
(2) Peptides dataset
- Peptides are short amino acid sequences essential in biological functions such as signaling, enzymatic reactions, and hormone regulation.
- The dataset captured peptide concentrations via mass spectrometry, providing insights into protein expression levels crucial for understanding PD mechanisms (Figure 4).
(3) Clinical dataset: The clinical dataset (Figure 5) included semi-annual and annual CSF-MS results and MDS-UPDRS scores [20, 21] from 248 patients, tracked at 0, 6, 12, and 24 months to support disease severity prediction.
- It included measurements of 227 proteins and 971 peptides from patient CSF samples.
- The MDS-UPDRS includes four sections:
  - ○ UPDRS I: Mentation, behavior, and mood.
  - ○ UPDRS II: Daily activities (speech, swallowing, dressing, hygiene, walking).
  - ○ UPDRS III: Clinician-scored motor function evaluation.
  - ○ UPDRS IV: Therapy-related complications.
(4) Gait dataset & integration with clinical data (updated clinical data): The clinical dataset was expanded with gait analysis data [22], incorporating gait speed, step length, freezing episodes, stride variability, and balance metrics. These parameters were aligned with UPDRS III scores, serving as the primary merging metric due to their comprehensive evaluation of motor symptoms like bradykinesia, rigidity, tremor, and postural instability (Figure 6). Unlike variables such as patient ID or visit month, which might have missing data, UPDRS III provides a consistent reference point for integrating clinical and gait data. As a quantitative measure of motor impairment, UPDRS III facilitates direct comparison with gait-derived features, improving alignment and prediction accuracy. Using UPDRS III as a reference ensures that gait metrics are meaningfully associated with motor severity and disease progression.
(5) Integration of UniProt protein dataset with clinical data: To predict MDS-UPDRS scores (updrs_1, updrs_2, updrs_3, updrs_4), protein and peptide data were integrated with clinical assessments through the following steps:
- 1. Aggregation of protein & peptide data
  - Protein Data (UniProt & NPX): Grouped by visit_id and protein_id, replacing NPX values with the mean per protein for each visit.
  - Peptide Data: Grouped by visit_id and peptide_id, replacing peptide abundance with the mean per peptide.
- 2. Pivoting Data for structured representation
  - Protein and peptide data were converted into a wide-format table, where each row represented a unique patient visit, and each column corresponded to a specific protein or peptide.
- 3. Merging protein-peptide data
  - Protein and peptide datasets were combined based on visit_id.
- 4. Merging with clinical data
  - The protein-peptide dataset was integrated with clinical records using visit_id and patient_id.

By integrating molecular, clinical, and gait data, the final dataset [23] supported a multimodal approach to predicting PD severity, enhancing early detection and enabling personalized treatment strategies. Gait metrics, such as step irregularity and freezing episodes, reflected real-world neurodegenerative symptoms and correlated with protein expression patterns. This structured integration facilitated:

Direct Patient Association: All datasets shared the patient_id, linking protein and peptide expressions to individual clinical profiles.
Temporal Tracking: visit_id aligned molecular and clinical data with specific visits for longitudinal modeling
Improved Predictive Power:
- ○ Molecular Features: Protein & peptide expression levels (UniProt & NPX).
- ○ Clinical Labels: MDS-UPDRS scores for PD severity.
- ○ Gait Features: Motor impairments such as step count and freezing episodes.

This multimodal integration reduced bias from single-source data and improved model generalization. A comparative analysis showed that including gait data significantly improved performance, especially in UPDRS-III predictions.

To achieve highly accurate predictions, the model requires completed data across all three domains—molecular, clinical, and gait. When data were missing, imputation techniques were applied, although this may reduce predictive precision. Feature importance analysis showed that each dataset contributed uniquely:

Protein data revealed underlying biological changes,
Gait data captured motor function impairments,
Clinical data provided a holistic view of disease severity.

Initially, models used only molecular and clinical datasets, effectively identifying early-stage biomarkers. However, adding gait data substantially improved performance by capturing key motor symptoms, underscoring the value of a multifaceted dataset for reliable PD progression modeling.

2. Data Preprocessing

To enhance data quality and model reliability, preprocessing steps included data restructuring, handling missing values, outlier detection, and feature engineering.

1) Data restructuring

1. Grouping & aggregation
- Rows were grouped by visit_id and corresponding protein_id (UniProt) or peptide_id.
- NPX and peptide abundance values were averaged per group to ensure consistency.
2. Pivoting & merging
- Protein and peptide datasets were pivoted using visit_id as the index, with each UniProt or peptide ID as a column [24].
- The pivoted datasets were merged on visit_id, creating a unified molecular profile for each patient visit (Figure 7).

2) Data cleaning & quality control

1. Handling missing values:
- UPDRS scores & gait metrics:
  - ○ If these values were missing at random, they were imputed using mean/mode values.
  - ○ If there were insufficient data, the entry was removed.
- Protein & peptide expression:
  - ○ Imputed using K-nearest neighbors (KNN) to maintain biological variability.
- Gait parameters (e.g., step count, freezing episodes):
  - ○ Imputed using the last observation carried forward and multivariate imputation by chained equations methods
- Protein detection threshold:
  - ○ Proteins rarely detected across samples were removed to reduce noise.
2. Handling outliers:
- Gait metrics:
  - ○ Outliers were identified via the Z-score and the inter-quartile range.
- Protein & peptide expression:
  - ○ Detected using log transformation and boxplot analysis to filter abnormal concentrations.

These steps ensured that the dataset was clean, well-structured, and ready for robust machine learning applications.

3. Feature Selection

To improve robustness and prevent overfitting, dimensionality reduction was performed while preserving predictive value. After preprocessing, the dataset contained 1,195 features from CSF-MS, exceeding the 1,068 observations, posing an overfitting risk and limiting generalizability.

1) Dimensionality reduction: from PCA to MI

Principal component analysis (PCA) was initially explored but replaced with mutual information (MI), as PCA transforms features into components, reducing biological interpretability. MI was preferred due to its ability to:

Capture nonlinear dependencies between features and target variables.
Preserve biological interpretability.
Model complex relationships in biological data, such as protein expression linked to clinical outcomes.

MI ranking identified 129 features (36 proteins and 93 peptides) with the strongest predictive correlations to MDS-UPDRS scores.

2) Encoding clinical treatment information

Levodopa treatment information [25] from the UPDRS 23b column was categorized into:

On – Ingested, with a good prognosis.
Off – Ingested, but poor prognosis.
No – No ingestion or missing data.

These categories were one-hot encoded [26], generating a binary matrix to avoid implying an ordinal relationship and ensure model compatibility. This added three features, bringing the total to 132.

The refined feature set provided a balanced and interpretable input for ML models, improving predictive accuracy while mitigating issues of high dimensionality.

4. Model Training and Validation

This section outlines the ML, training-validation processes, and accuracy enhancement strategies for predicting PD severity.

1) ML based regression models

The final dataset, integrating clinical, protein, peptide, and gait data via visit_id, was divided into training and validation sets at an 80:20 ratio. Given the goal of predicting 16 score-timestamp pairs (4 UPDRS scores × 4 time points), multiple regression models were utilized.

A TensorFlow decision forests (TFDF) model [27], incorporating random forests, was selected due to its suitability for nonlinear data and because it does not require feature scaling. Additional regression models were also tested to enable performance comparison.

2) Training workflow

Data Structuring: Merged features using visit identifiers.
Cleaning: Removed rows with missing targets.
Formatting: Converted training data to TensorFlow dataset format for GPU/TPU efficiency.
Model Setup: Initialized random forest, linear regression, decision tree, and KNN with MSE as the loss metric.

3) Internal validation

Models were evaluated using:

MSE for numerical accuracy.
sMAPE for scale-aware prediction, aiding clinical interpretation.

4) Custom ensemble model

An ensemble approach was adopted to leverage different model strengths:

Linear regression: Served as a baseline model.
XGBRegressor: Captured nonlinear relationships and boosted weak learners.
Random forest & gradient boosting: Robust to noise and biologically interpretable.

GridSearchCV was used to tune hyperparameters for XGBRegressor based on a custom scoring function.

5) Phase-shift ensembling for temporal modeling

Phase-shift ensembling aligned model predictions across timestamps to effectively capture the progression dynamics of PD. Ensemble predictions were adjusted based on protein and peptide temporal trends [28]. This approach assumes patients with similar UPDRS-III scores exhibit analogous gait profiles at specific times. While effective for static data integration, it lacks the capacity to capture sequential dependencies.

6) Comparison with time-series ML techniques

To address this limitation, we compared phase-shift ensembling with time-series models:

Recurrent neural networks: Captures sequential dependencies but suffers from vanishing gradient problems for long-term predictions.
Long short-term memory (LSTM) networks: Retain long-term dependencies, suitable for modeling progression, but are data and computationally intensive.
Transformer models (e.g., temporal fusion transformer): Use self-attention to model complex, nonlinear disease trends; robust to missing data but require substantial hyperparameter tuning.

Findings from the comparison

Phase shift suits static clinical/gait data.
LSTM/transformers outperform in long-term disease modeling.
A hybrid strategy—static feature modeling via phase shift and long-term modeling via LSTMs/transformers—offers an optimal trade-off between interpretability and accuracy.

7) Model optimization & custom loss function

A refined ensemble, Model_Optim_Median, was developed to minimize a custom loss function that incorporated both prediction accuracy and biological variation:

Custom Loss = \sum_{i = 1}^{n} α {(Y i - \hat{Y} i)}^{2} + β (\hat{Y} i - Y i)

where n is the number of observations; Yi and Ŷi refer to actual and predicted values; and α and β are weights balancing MSE and sMAPE, respectively.

8) Final model validation

To ensure reliable results, stratified K-fold cross-validation [29] was applied, preserving UPDRS score distribution across folds and improving generalization across disease severity levels.

5. Testing

During the testing phase, the model predicted UPDRS scores at 6, 12, and 24 months using only baseline biomarker data. A randomly selected test subset, withheld from training, assessed the model’s generalization capability.

Model predictions were evaluated against actual clinical scores using the MSE, sMAPE, and APR, offering insights into both numerical accuracy and clinical relevance.

The results demonstrated a strong correlation between baseline protein/peptide profiles and subsequent UPDRS scores, reinforcing their potential as prognostic biomarkers for PD progression.

6. Performance Metrics

PD severity prediction was treated as a multi-output regression task, aiming to estimate MDS-UPDRS scores (UPDRS I–IV) at 0, 6, 12, and 24 months. Each dataset D(t), with t ∈ {0,6,12,24} contained input-output pairs (x, y) with x ∈ R¹¹⁹⁵ representing features and y ∈ R⁴ denoting UPDRS scores. The model f(θ, x) parameterized by θ was trained to minimize prediction error.

To evaluate performance, we used three metrics: MSE, sMAPE, and APR [30]. These measure numerical accuracy and clinical reliability.

MSE quantifies average squared error, penalizing large deviations:

M S E = \frac{1}{n} \sum_{i = 1}^{n} {(Y i - \hat{Y} i)}^{2}

where Yi and Ŷi are actual and predicted values, respectively.

sMAPE captures percentage error with equal weight to over- and under-predictions.

s M A P E = \frac{100}{n} \sum_{i = 1}^{n} (\frac{| \hat{Y} i - Y i |}{\frac{(| \hat{Y} i | + | Y i |)}{2}})

APR evaluates clinical validity, marking predictions as accurate if they fall within a specified range (e.g., ±10%) of actual values.

APR = \frac{1}{N} \sum_{i = 1}^{N} I (\frac{| (Y i - \hat{Y} i) |}{| Y i |} \leq R) T h r e s h o l d

where R is the threshold (e.g., 0.10), and I is an indicator function.

These metrics jointly assess predictive precision and clinical applicability, supporting interpretability in PD progression modeling.

III. Results

Table 1 summarizes the performance and training durations of various models employed for predicting PD severity. The TFDF model achieved the lowest average sMAPE of 55.009 and MSE of 20.87, with a training time of 353.4 seconds. The custom model using phase-shift ensembling performed competitively, demonstrating a lower sMAPE of 40.197 and MSE of 20.012, though with a longer training duration of 766.7 seconds.

Linear regression exhibited the shortest training time (1 second) but yielded higher error metrics. The random forest regressor balanced computational efficiency and prediction accuracy, achieving an average sMAPE of 57.76, MSE of 27.618, and training time of 99.2 seconds. Notably, this model achieved an sMAPE of 77.32 specifically for predicting UPDRS-III, highlighting its effectiveness in modeling motor function severity.

IV. Discussion

The results affirm the effectiveness of TFDF and the custom model employing phase-shift ensembling for accurately predicting UPDRS scores. TFDF achieved the lowest average sMAPE and MSE, indicating superior predictive accuracy. Similarly, phase-shift ensembling performed robustly, attaining an average sMAPE of approximately 55 across all UPDRS sections, underscoring its ability to model the complex progression dynamics of PD.

Conversely, linear regression and ridge regression resulted in higher sMAPE values, suggesting limitations in capturing the nuanced progression of PD symptoms. Random forest regressor and KNN, while outperforming linear models, did not match the predictive accuracy of TFDF or the custom ensemble. Notably, the random forest method achieved a particularly strong performance, with an sMAPE of 77.32 for UPDRS-III, demonstrating its suitability for modeling motor impairment.

Our findings highlight the utility of integrating clinical biomarkers and gait data with advanced machine learning methods, resulting in improved predictive outcomes. This multimodal integration produced lower MSE and sMAPE values, enhancing clinical interpretability. Additionally, employing the APR metric, which is intuitive and clinician-friendly, complements MSE and sMAPE, providing comprehensive model evaluation that supports clinical decision-making and personalized treatment strategies.

Future improvements could involve refining ensemble strategies, adding features, and shortening the training time for real-time clinical deployment. These models may also be extended to other neurological disorders, such as Alzheimer’s disease, potentially transforming diagnosis and prognosis.

1. Predicting Early-Stage PD

The developed predictive tool shows promising capabilities in early-stage PD prediction, although predictive accuracy typically diminishes in earlier stages due to subtler symptom manifestations. Phase-shift ensembling effectively captures UPDRS trends by adjusting predictions according to temporal shifts in protein and peptide concentrations. While gait data predominantly informs motor-function-related predictions, biomarker and clinical pattern analyses play a greater role in early-stage prediction, potentially guiding earlier clinical interventions.

2. Handling Zero UPDRS Scores: sMAPE Limitations

Linear regression yielded lower MSE but this requires caution due to sMAPE’s limitations with zero actual values, which can distort performance assessment. To mitigate this, we applied a modified sMAPE (msMAPE) by adding a small constant ɛ to the denominator:

m s M A P E = \frac{100}{n} \sum (\frac{| \hat{Y} i - Y i |}{\frac{(| \hat{Y} i | + | Y i |) + e ɛ}{2}})

This adjustment improves the stability of the metric without sacrificing interpretability. While linear regression still excelled in MSE, the custom ensemble outperformed under msMAPE and APR, especially for scores near zero.

Overall, employing this combination of metrics ensures accurate model evaluation across all score ranges, thereby improving reliability and facilitating the development of more effective predictive tools for monitoring PD progression.

Notes

Conflict of Interest

No potential conflict of interest relevant to this article was reported.

Acknowledgments

The authors sincerely thank Teresa Maycas for providing the dataset on “Parkinson’s Disease and Facial Bradykinesia”, available via Mendeley Data. We also extend our gratitude to the “AMP Parkinson’s Disease Progression Prediction” datasets, hosted on Kaggle.

References

1. Mandal A. What is Parkinson’s disease? [Internet]. Manchester, UK: News-Medical;2023. [cited at 2024 Mar 2]. Available from: https://www.news-medical.net/health/What-is-Parkinsons-Disease.aspx.

2. Saravanan S, Ramkumar K, Adalarasu K, Sivanandam V, Kumar SR, Stalin S, et al. A systematic review of artificial intelligence (AI) based approaches for the diagnosis of Parkinson’s disease. Arch Comput Methods Eng. 2022; 29(6):3639–53. https://doi.org/10.1007/s11831-022-09710-1.

3. Saeed F, Al-Sarem M, Al-Mohaimeed M, Emara A, Boulila W, Alasli M, et al. Enhancing Parkinson’s disease prediction using machine learning and feature selection methods. Comput Mater Contin. 2022; 71(3):5639–58. https://doi.org/10.32604/cmc.2022.023124.

4. Sadek RM, Mohammed SA, Abunbehan AR, Ghattas AK, Badawi MR, Mortaja MN, et al. Parkinson’s disease prediction using artificial neural network. Int J Acad Health Med Res. 2019; 3(1):1–8.

5. Nawi NM, Makhtar M, Afip ZA, Salikon MZ. Parkinson disease classification: a comparative analysis on classification techniques. Indones J Electr Eng Comput Sci. 2020; 18(3):1351–8. https://doi.org/10.11591/ijeecs.v18.i3.pp1351-1358.

6. Arthi I, Sylviaa SM. Parkinson’s disease prediction using AI techniques. Int J Creat Res Thoughts. 2023; 11(5):d586–93.

7. Nilashi M, Abumalloh RA, Minaei-Bidgoli B, Samad S, Yousoof Ismail M, Alhargan A, et al. Predicting Parkinson’s disease progression: evaluation of ensemble methods in machine learning. J Healthc Eng. 2022; 2022:2793361. https://doi.org/10.1155/2022/2793361.

8. Alshammri R, Alharbi G, Alharbi E, Almubark I. Machine learning approaches to identify Parkinson’s disease using voice signal features. Front Artif Intell. 2023; 6:1084001. https://doi.org/10.3389/frai.2023.1084001.

9. Ahmadi Rastegar D, Ho N, Halliday GM, Dzamko N. Parkinson’s progression prediction using machine learning and serum cytokines. NPJ Parkinsons Dis. 2019; 5:14. https://doi.org/10.1038/s41531-019-0086-4.

10. Alzubaidi MS, Shah U, Dhia Zubaydi H, Dolaat K, Abd-Alrazaq AA, Ahmed A, et al. The role of neural network for the detection of Parkinson’s disease: a scoping review. Healthcare (Basel). 2021; 9:740. https://doi.org/10.3390/healthcare9060740.

11. Johri A, Tripathi A. Parkinson disease detection using deep neural networks. In : Proceedings of the 12th International Conference on Contemporary Computing (IC3); 2019 Aug 8–10; Noida, India. p. 1–4. https://doi.org/10.1109/IC3.2019.8844941.

12. Sarankumar R, Vinod D, Anitha K, Manohar G, Vijayanand KS, Pant B, et al. Severity prediction over Parkinson’s disease prediction by using the deep Brooke inception net classifier. Comput Intell Neurosci. 2022; 2022:7223197. https://doi.org/10.1155/2022/7223197.

13. Varghese BK, Amali D, Devi KS. Prediction of Parkinson’s disease using machine learning techniques on speech dataset. Res J Pharm Technol. 2019; 12(2):644–8. https://doi.org/10.5958/0974-360X.2019.00114.8.

14. Govindu A, Palwe S. Early detection of Parkinson’s disease using machine learning. Procedia Comput Sci. 2023; 218:249–61. https://doi.org/10.1016/j.procs.2023.01.007.

15. Sainath Chaithanya A, Rachana M. Identification of diseased papaya leaf through transfer learning. Indian J Sci Technol. 2023; 16(48):4676–87. https://doi.org/10.17485/IJST/v16i48.2690.

16. MDS-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) [Internet]. Milwaukee (WI): International Parkinson and Movement Disorder Society;2024. [cited at 2024 May 5]. Available from: https://www.movement-disorders.org/MDS/MDS-Rating-Scales/MDS-Unified-Parkinsons-Disease-Rating-Scale-MDS-UPDRS.htm.

17. Kaggle. AMP Parkinson’s Disease Progression Prediction data [Internet]. Mountain View (CA): Kaggle;2023. [cited at 2024 May 11]. Available from: https://www.kaggle.com/competitions/amp-parkinsons-disease-progression-prediction/data.

18. UniProt Consortium. The universal protein resource (UniProt). Nucleic Acids Res. 2008; 36(Database issue):D190–5. https://doi.org/10.1093/nar/gkm895.

19. Olink What is NPX? [Internet]. Uppsala Sweden: Olink Proteomics AB;2024. [cited at 2024 May 21]. Available from: https://olink.com/knowledge/faq?query=npx.

20. Hrishi AP, Sethuraman M. Cerebrospinal fluid (CSF) analysis and interpretation in neurocritical care for acute neurological conditions. Indian J Crit Care Med. 2019; 23(Suppl 2):S115–9. https://doi.org/10.5005/jp-journals-10071-23187.

21. American Physical Therapy Association. Unified Parkinson’s Disease Rating Scale (UPDRS), Movement Disorders Society (MDS) Modified Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) [Internet]. Alexandria (VA): American Physical Therapy Association;2024. [cited at 2024 Jun 19]. Available from: https://www.apta.org/patient-care/evidence-based-practice-resources/test-measures/unified-parkinsons-disease-rating-scale-updrs-movement-disorders-society-mds-modified-unified-parkinsons-disease-rating-scale-mds-updrs.

22. Maycas T. Parkinson’s disease and facial bradykinesia [Internet]. London, UK: Mendeley Data;2020. [cited at 2024 Jun 19]. Available from: https://doi.org/10.17632/4tjd8jrhp8.1.

23. Chaithanya AS. Comprehensive multi-modal dataset for Parkinson’s disease prediction [Internet]. Geneva, Switzerland: Zenodo;2025. [cited at 2025 Jul 1]. Available from: https://doi.org/10.5281/zenodo.14848598.

24. Microsoft. Group or ungroup data in a PivotTable [Internet]. Redmond (WA): Microsoft;2024. [cited at 2024 Jun 3]. Available from: https://support.microsoft.com/en-us/office/group-or-ungroup-data-in-a-pivottable-c9d1ddd0-6580-47d1-82bc-c84a5a340725.

25. Hauser RA, LeWitt PA, Comella CL. On demand therapy for Parkinson’s disease patients: Opportunities and choices. Postgrad Med. 2021; 133(7):721–7. https://doi.org/10.1080/00325481.2021.1936087.

26. Hancock JT, Khoshgoftaar TM. Survey on categorical data for neural networks. J Big Data. 2020; 7:28. https://doi.org/10.1186/s40537-020-00305-w.

27. Guillame-Bert M, Bruch S, Stotz R, Pfeifer J. Yggdrasil decision forests: a fast and extensible decision forests library. In : Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining; 2023 Aug 6–10; Long Beach, CA, USA. p. 4068–77. https://doi.org/10.1145/3580305.3599933.

28. Naderalvojoud B, Hernandez-Boussard T. Improving machine learning with ensemble learning on observational healthcare data. AMIA Annu Symp Proc. 2024; 2023:521–9.

29. Szeghalmy S, Fazekas A. A comparative study of the use of stratified cross-validation and distribution-balanced stratified cross-validation in imbalanced learning. Sensors (Basel). 2023; 23(4):2333. https://doi.org/10.3390/s23042333.

30. Plevris V, Solorzano G, Bakas NP, Ben Seghier ME. Investigation of performance metrics in regression analysis and machine learning-based prediction models. In : Proceedings of the 8th European Congress on Computational Methods in Applied Sciences and Engineering (ECCOMAS Congress); 2022 Jun 5–9; Oslo, Norway. p. 1–25. https://doi.org/10.23967/eccomas.2022.155.

Figure 1

Parkinson’s disease symptoms.

Figure 2

Overall flow of the study. ML: machine learning, sMAPE: symmetric mean absolute percentage error, MSE: mean squared error, APR: acceptable prediction range.

Figure 3

Protein dataset sample.

Figure 4

Peptide dataset sample.

Figure 5

Clinical dataset sample.

Figure 6

Updated clinical dataset with gait sample.

Figure 7

Sample of the resultant dataset.

Table 1

Performance metrics of algorithms

Algorithm	Train time (s)	UPDRS-I		UPDRS-II		UPDRS-III		UPDRS-IV		Average

		sMAPE (%)	MSE	sMAPE (%)	MSE	sMAPE (%)	MSE	sMAPE (%)	MSE	sMAPE (%)	MSE
TensorFlow decision forests (TFDF)	353.4	65.521	20.870	90.218	27.060	82.396	167.540	97.939	4.170	84.019	55.009

Ridge model	2.1	66.268	22.940	94.638	28.023	90.096	185.732	175.130	4.554	106.528	60.314

Custom model phase shift ensembling	766.7	54.918	20.012	64.718	20.104	57.898	114.620	43.137	6.052	55.168	40.197

Linear regression	1	71.830	19.230	105.958	19.747	95.247	119.624	189.580	1.181	115.656	39.945

RandomForestRegressor	99.2	69.871	27.618	99.373	29.385	86.858	165.395	17.955	5.612	68.420	57.760

DecisionTreeRegressor	4.3	86.149	44.115	92.652	47.168	92.044	164.694	15.096	8.837	64.730	92.190

KneighboursRegressor	3	69.731	26.673	97.250	28.482	86.956	159.396	17.254	5.536	67.790	55.010

TOOLS

Similar articles