STEP-PD: Stage-Aware and Explainable Parkinson's Disease Severity Classification Using Multimodal Clinical Assessments

Abstract: Parkinson's disease (PD) is a progressive disorder in which symptom burden and functional impairment evolve over time, making severity staging essential for clinical monitoring and treatment planning. However, many computational studies emphasize binary PD detection and do not fully use repeated follow-up clinical assessments for stage-aware prediction. This study proposes STEP-PD, a severity-aware machine learning framework to classify PD severity using clinically interpretable boundaries. It leverages all available visits from the Parkinson's Progression Markers Initiative (PPMI) and integrates routinely collected subjective questionnaires and objective clinician-assessed measures. Disease severity is defined using Hoehn and Yahr staging and grouped into three clinically meaningful categories: Healthy, Mild PD (stages 1-2), and Moderate-to-Severe PD (stages 3-5). Three binary classification problems and a three-class severity task were evaluated using stratified cross-validation with imbalance-aware training. To enhance interpretability, SHAP was used to provide global explanations and local patient-level waterfall explanations. Across all tasks, XGBoost achieved the strongest and most stable performance, with accuracies of 95.48% (Healthy vs. Mild), 99.44% (Healthy vs. Moderate-to-Severe), and 96.78% (Mild vs. Moderate-to-Severe), and 94.14% accuracy with 0.8775 Macro-F1 for three-class severity classification. Explainability results highlight a shift from early motor features to progression-related axial and balance impairments. These findings show that multimodal clinical assessments within the PPMI cohort can support accurate and interpretable visit-level PD severity stratification.

• The paper introduces STEP-PD, a multimodal framework that classifies Parkinson's disease severity using 15,606 clinical samples, achieving up to 99.44% accuracy across defined stages.
• It benchmarks five machine learning models with XGBoost leading, and employs SHAP-based explainability to reveal stage-dependent feature importance.
• The study aligns computational insights with clinical practice, highlighting improved diagnostic precision and suggesting future directions for longitudinal and generalizable modeling.

STEP-PD: Stage-Aware and Explainable Parkinson’s Disease Severity Classification Using Multimodal Clinical Assessments

Introduction

The classification and staging of Parkinson’s disease (PD) remain core challenges in neurology due to the progressive and heterogeneous nature of neurodegeneration. Clinical management and therapeutic planning critically depend on accurate, stage-aware assessments of disease severity. The STEP-PD framework directly addresses several deficits in prior computational approaches by integrating multimodal, longitudinal clinical assessments from the Parkinson’s Progression Markers Initiative (PPMI) for visit-level, explainable PD severity classification. This strategy contrasts with models focusing solely on binary detection and single-modality signals, facilitating alignment with clinical realities and decision-making workflows.

Methodology

Data and Labels

STEP-PD utilizes PPMI data spanning 16,162 visit-level samples, ensuring inclusion across all Hoehn and Yahr (H&Y) stages (1–5). After excluding records with missing or indeterminate data, 15,606 distinct samples remain, categorized into Healthy (7,689), Mild PD (stages 1–2; 7,364), and Moderate-to-Severe PD (stages 3–5; 553). This design leverages repeated assessments and ensures robust representation across the PD severity spectrum.

The multimodal feature set comprises 208 variables derived from standardized subjective questionnaires and objective, expert-administered motor and neuropsychological rating instruments such as MDS-UPDRS, QUIP, and MoCA. All numeric features are harmonized through z-score normalization.

Class Formulation and Benchmarking

Four essential classification tasks are explored:

• Healthy vs. Mild
• Healthy vs. Moderate-to-Severe
• Mild vs. Moderate-to-Severe
• Three-class (Healthy, Mild, Moderate-to-Severe)

Five supervised learning algorithms are benchmarked: Logistic Regression, SVM (RBF kernel), KNN, Random Forest, and XGBoost. Imbalanced class proportions, especially evident in advanced-stage PD, are managed with class-aware training (i.e., weighted loss functions and scale_pos_weight in XGBoost).

Optimization and model selection employ 5-fold stratified cross-validation on an 80/20 train–test split, with GridSearchCV for hyperparameter tuning using F1 or macro-F1 as the primary metric.

Explainability

Model transparency is a core design criterion. STEP-PD implements SHapley Additive exPlanations (SHAP) for:

• Global explanations: Mean absolute SHAP values per cohort elucidate feature importance across levels of severity.
• Cross-task heatmaps: Visualize the salience of features at clinically relevant transitions.
• Local (patient-level) explanations: SHAP waterfall plots reveal individualized symptom attribution profiles, essential for actionable clinical interpretation.

Experimental Results

Model Performance

XGBoost delivers the strongest and most stable performance, reflected in high accuracy and F1 metrics for all severity boundaries. Key results include:

• Healthy vs. Mild: 95.48% accuracy, ROC-AUC 0.9888
• Healthy vs. Moderate-to-Severe: 99.44% accuracy, ROC-AUC 0.9983
• Mild vs. Moderate-to-Severe: 96.78% accuracy, F1 0.7661, MCC 0.7516
• Three-class severity: 94.14% accuracy, macro-F1 0.8775

Confusion matrices support the stability of these results, with misclassifications primarily occurring near the transition boundaries (e.g., Mild vs. Moderate-to-Severe), a reflection of underlying clinical ambiguity.

Interpretability and Clinical Alignment

Global SHAP analysis reveals a stage-dependent shift in predictive features:

• Early separation (Healthy vs. Mild): Dominated by motor features—bradykinesia (NP3BRADY), rigidity, tremor (NP3RTCON), and fine motor tasks.
• Advanced separation (Healthy vs. Moderate-to-Severe; Mild vs. Moderate-to-Severe): Postural instability (NP3PSTBL), gait dysfunction (NP2WALK, NP3GAIT), and balance impairment become salient.

Non-motor symptoms (e.g., medication overuse, anxiety markers from QUIP and STAIAD scales) and cognitive tests (Benton Judgment, semantic fluency) gain relevance in advanced-stage discrimination. This aligns with the transition from predominantly dopaminergic to more complex, multifactorial symptomatology as PD progresses.

SHAP heatmaps succinctly visualize the functional domains driving decision boundaries, clarifying which neurocognitive functions are operational at each stage.

Local SHAP analysis demonstrates the plausibility of individual-level predictions. For “prototype” subjects in the Healthy and Mild groups, model attribution corresponds to clinically expected symptom patterns.

Comparison to Prior Work

STEP-PD outperforms or matches the accuracy of prior PD classification frameworks, especially those trained on constrained modalities (e.g., speech, wearables, imaging) or focused solely on binary discrimination (Islam et al., 30 Jan 2026, 2604.17611). Unlike most published studies, STEP-PD provides granular, clinically aligned stage prediction and detailed interpretability, thus advancing the translational relevance of ML in PD.

Notably, STEP-PD avoids synthetic oversampling during training, relying instead on class-aware weighting to preserve the representational integrity of severity boundaries—an approach aligned with best practice recommendations for imbalanced biomedical data [islam2026scope_pd].

Limitations and Future Directions

Despite its strengths, STEP-PD is subject to several constraints:

• Target–feature overlap: Some influential predictors are derived from MDS-UPDRS items closely related to H&Y staging, which can induce target leakage.
• Visit-level granularity: Modeling is performed at the visit, not subject, level. Future work should incorporate temporal models (e.g., RNNs, survival analysis) to capture patient-specific disease trajectories.
• Mod-Severe class size: Limited Moderate-to-Severe samples (especially H&Y stage 4/5) pose ongoing challenges for robust minority-class performance.
• Cohort generalizability: Current results are restricted to PPMI data; rigorous external and subject-agnostic validation is required prior to deployment.
• Independence assumptions: Repeated measures from the same subject may appear in both training and validation folds, encouraging over-optimistic performance estimates.

Implications and Future Applications

STEP-PD demonstrates that multimodal, repeated clinical assessments—interpreted through transparent, severity-aware models—can enable both accurate and explainable PD staging. This approach facilitates:

• Precision clinic triage (early recognition versus advanced disease),
• Progression monitoring using routine clinical data,
• Cohort stratification for clinical trials or observational research.

The framework establishes a foundation for developing personalized monitoring platforms capable of integrating additional data modalities (e.g., genetics, imaging, digital phenotyping) and temporally explicit modeling. Future developments may enable real-time, interpretable AI-assisted feedback during clinic visits, supporting nuanced and data-driven treatment adaptation.

Conclusion

STEP-PD advances the paradigm for computational Parkinson’s disease severity classification by uniting robust multimodal data use, severity-stage awareness, and state-of-the-art model explainability. The system exhibits strong performance at all diagnostic boundaries with transparent rationales for its predictions. While additional validation on independent cohorts and further attention to longitudinal modeling are warranted, STEP-PD substantiates the potential of ML-driven, stage-aware, interpretable decision support within PD clinical care (2604.17611).