A clinic-updated digital twin for Parkinson’s disease progression: Governed Bayesian forecasting with uncertainty-gated reporting

Abstract Background Clinical digital twins hold considerable promise for forecasting disease progression, yet the question of when a model’s outputs should be withheld remains largely unaddressed. A predictive model qualifies as a governed reporting system only when it specifies the operational boundaries under which its outputs are reliable and enforces criteria for suppressing results that fall outside those bounds. Methods We present a governed Bayesian digital twin for multi-domain Parkinson’s disease (PD) progression, tracking motor function (MDS-UPDRS Part III), cognition (Montreal Cognitive Assessment, MoCA), and autonomic function (SCOPA-AUT). A monotone latent state-space model captures disease progression under four architectural constraints: non-decreasing latent severity, visit-triggered updating, full posterior uncertainty propagation, and non-causal scope. A six-rule confidence gate evaluates each forecast before release; when evidence is insufficient, the gate suppresses the output and returns a structured reason code. We evaluated the framework on the Parkinson’s Progression Markers Initiative (PPMI), a multicentre longitudinal observational study ( N =4,628 participants; 28,185 visits), using five-fold cross-validation with independent model refits, equity analysis, and coupling-topology sensitivity assessment. The framework is available at https://gitlab.com/ahmed.hemedan/symphony-dt , with a research prototype at https://symphony-dt.com/ . Results Predictive interval coverage at the 95% level ranged from 94% to 96% across all three endpoints, compared with 64–69% for linear mixed-effects baselines. The confidence gate released governed forecasts at 32.7% of visits under strict three-domain requirements, increasing to 48.1% under a validated partial-observation extension. Suppression was predominantly driven by incomplete clinical assessment (51.5%) rather than model uncertainty (0.2%), and operated equitably across sexes (Cramér’s V =0.049). Five of six cross-domain coupling parameters were identified from the data (sign probability ≥ 0.99; contraction ratios 0.19–0.35), with all cross-domain forecast correlations matching the directions predicted by the coupling topology. The framework’s own diagnostics localised two observation-model limitations, Prodromal motor heteroscedasticity and medication-burden sensitivity, to a single model layer and specified their resolution. Conclusions Governed silence, defined as the rule-based suppression of predictions when reliability conditions are not met, can be embedded in clinical prediction architecture, quantified as a pipeline output, and audited for equity. This work demonstrates the technical executability of governed digital twin architecture at cohort scale and provides a foundation for prospective deployment under routine clinical conditions.