Current Trends in AI Gait Analysis for the Detection and Assessment of Parkinson’s Disease Severity: Systematic Review and Meta-Analysis of Performance Using Logit Transformation

Background/Objectives: Artificial intelligence (AI) offers a promising approach for detecting and classifying symptom severity in patients with Parkinson’s disease (PD). The objective was to provide an overview of AI methods performance used for this classification through a systematic review and meta-analysis conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. Methods: The Google Scholar, IEEE Xplore, PubMed/MedLine, and ScienceDirect databases were searched for the period 2015–2025. The studies included were original, peer-reviewed studies written in English that addressed an AI method based on machine learning (ML) or deep learning (DL) for the classification of PD patients. The dataset used had to be “Gait in Parkinson’s Disease,” in which the severity of disease symptoms was assessed using the Hoehn and Yahr (H&Y) scale. Studies had to report at least one of the five performance metrics: accuracy, sensitivity, specificity, precision, and F1 score. Two reviewers independently selected articles, assessed the risk of bias using PROBAST (Prediction Model Study Risk of Bias Assessment Tool), and extracted data. The logit-transformed values were pooled separately by performance metrics and by severity level using a random-effects model. Cochran’s Q test, the I² statistic, and inter-study variability (τ²), computed using the generalized inverse variance method with the restricted maximum likelihood model, were used to assess heterogeneity. Forest plots with 95% confidence intervals were used to present the results. Possible causes of heterogeneity were explored using a subgroup analysis (ML vs. DL) and a sensitivity analysis. Finally, publication bias (Egger’s test) and the certainty of the evidence (using GRADE—Grading of Recommendations Assessment, Development, and Evaluation) were assessed to verify the generalizability of the results. Results: Among the 257 unique records, 12 studies were included. The methods demonstrated very high overall performance (>92%): accuracy (96.4%, 95% CI: 95.9–96.9%), specificity (97.7%, 95% CI: 97.3–98.1%), sensitivity (94.0%, 95% CI: 92.7–95.2%), precision (93.4%, 95% CI: 92.0–94.6%), F1 score (92.1%, 95% CI: 90.6–93.4%). Accuracy, specificity, and precision were high for all H&Y levels. However, the more advanced the symptoms, the lower the sensitivity (97.3% for H&Y0 vs. 92.1% for H&Y3). ML models achieved the best results for classifying healthy patients (H&Y0: 95.7% to 98.2%), while DL approaches performed better for classifying higher severity levels (>92%). Heterogeneity and inter-study variability were moderate (I²: 40–50% and τ²: 0.3–0.4) for precision and F1 score, and high (I² > 90% and τ² > 0.6) for accuracy, specificity, and sensitivity. The GRADE analysis revealed low-quality evidence for precision and F1 score and very-low quality for accuracy, specificity, and sensitivity. Conclusions: Thus, AI-based wearable gait assessment devices show great promise in terms of aiding clinical decision-making and treatment personalization. However, further research using a rigorous methodology (PROBAST) is needed to ensure the generalizability of the results and the clinical viability of the proposed solutions.