Predictability-Aware Multimodel Forecasting of Parkinson’s Disease Progression: Quantifying Model Complementarity and Oracle Recoverability
Year: 2026
Type: article
Abstract: Parkinson’s disease (PD) progression is heterogeneous, which makes individual prognosis difficult. We developed a multimodel framework to carry out approximately one-year forecasting of MDS-UPDRS Parts I–III in participants from the Parkinson’s Progression Markers Initiative (PPMI; 49% enrolled with PD), using patient-disjoint evaluation with leakage controls. The calibration-selected deployable strategy achieved MAEs of 1.188 (95% CI 1.113–1.269), 2.223 (2.070–2.379), and 5.110 (4.759–5.478) for Parts I–III (R2 of 0.541, 0.825, and 0.802). Retrospective Oracle selection among four candidate models lowered the MAE by about 31–35% relative to the best single candidate chosen on calibration data. However, the deployable strategy recovered only 4.4% (95% CI −1.2 to 8.9), 7.9% (3.4 to 11.6), and 5.2% (0.4 to 9.3) of this headroom, indicating large model complementarity but limited predictability in terms of which model is best for a given patient. In a complementary Hoehn–Yahr (HY) analysis, 83.11% of 9–15-month transitions kept the same stage. Among HY-stable transitions, 70.17% showed some degree of worsening in at least one MDS-UPDRS domain, but only 33.41% exceeded the published minimal clinically important difference. A calibrated model for one-year HY progression achieved ROC-AUC of 0.700±0.016 and PR-AUC of 0.235±0.017 at 11.23% prevalence, with a Brier score of 0.094±0.004. These findings separate retrospective complementarity from prospective recoverability.
Source: Bioengineering
Language: English
Subfield: Clinical Neurology
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $3,258
OpenAlex ID: W7235634685
