SGC-RML: A reliable and interpretable longitudinal assessment for PD in real-world DNS

SGC-RML is a new AI framework that improves Parkinson's disease assessment by combining speech, gait, and wearable sensor data while providing reliability estimates and confidence measures. The model achieves strong predictive performance across multiple datasets and can reject uncertain assessments or recommend retesting, addressing critical gaps in real-world digital health monitoring.

SGC-RML represents a significant advancement in digital health assessment by tackling a fundamental problem in clinical AI: determining not just what to predict, but when predictions are reliable enough for clinical use. Traditional machine learning models optimize for average accuracy while ignoring the heterogeneity inherent in real-world medical data—different devices, incomplete records, and patient variability. This framework introduces three critical mechanisms: uncertainty quantification, conformal calibration, and selective rejection, creating an auditable decision-making process essential for clinical deployment.

The research addresses a growing industry need as healthcare systems increasingly adopt digital biomarkers for chronic disease management. Parkinson's disease represents an ideal testbed because it presents both motor and non-motor symptoms across multiple measurable dimensions. By mapping diverse sensor modalities into a unified 8-dimensional symptom space, SGC-RML creates interpretability alongside accuracy—clinicians can understand which symptom domains drive predictions rather than treating the model as a black box.

The experimental validation demonstrates practical utility: with just five subject-specific calibration points, longitudinal prediction error drops from nearly non-functional (MAE 8.38) to clinically relevant levels (MAE 3.24). This calibration efficiency matters significantly for adoption, as collecting extensive baseline data burdens both patients and healthcare providers.

The framework's broader implications extend beyond Parkinson's assessment to any chronic disease requiring multimodal monitoring. Digital therapeutics companies, remote patient monitoring platforms, and clinical research organizations could leverage this paradigm to deploy more trustworthy AI systems. The emphasis on rejection mechanisms and calibration coverage reflects regulatory expectations emerging around AI in healthcare, positioning SGC-RML as aligned with forthcoming FDA and EMA guidance on clinical decision support systems.

• →SGC-RML unifies heterogeneous medical data (speech, gait, wearables, clinical variables) into an interpretable 8-dimensional symptom representation for Parkinson's disease assessment.
• →The framework achieves strong predictive performance (MAE 4.579, R² 0.772 on PPMI; AUC 0.953 on mPower) while maintaining calibrated confidence estimates and rejection capabilities.
• →With minimal subject-specific calibration data, longitudinal prediction error improves dramatically from non-functional to clinically viable levels across multi-site datasets.
• →Built-in uncertainty estimation and conformal calibration enable clinicians to reject unreliable assessments and recommend retesting based on evidence quality.
• →The architecture addresses real-world deployment barriers—cross-device bias, incomplete labeling, and heterogeneous modalities—making it production-ready for digital health platforms.