Work

Dynamic prediction of bidirectional CDR-stage transitions: locked external validation and calibration in ADNI and OASIS-3

Year: 2026
Type: article
Abstract: Alzheimer’s disease prediction research seldom addresses bidirectional clinical-stage transitions, dynamic updating, and external calibration together. We evaluated the transportability of longitudinal predictions of improvement, persistence, and progression from Clinical Dementia Rating (CDR) 0.5. We conducted a retrospective prediction-model development and external-validation study using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) for development, Open Access Series of Imaging Studies-3 (OASIS-3) for locked external validation, and the Australian Imaging, Biomarkers and Lifestyle study for descriptive feasibility assessment. Baseline and 90-day landmark models estimated first observed transitions to CDR 0 or CDR ≥ 1 and persistence at CDR 0.5 over 12 and 36 months. The dynamic model D1 incorporated updated Mini-Mental State Examination (MMSE) information into ridge-penalized cause-specific Cox models. Frozen external predictions were evaluated for discrimination, calibration, and prediction error; recalibration was assessed separately. ADNI included 2,447 index participants and 4,517 landmarks (1,412 participants); OASIS-3 included 491 index participants and 845 D1 landmarks (437 participants). Untouched D1 improvement/progression areas under the curve (AUCs) were 0.680/0.721 at 12 months and 0.747/0.729 at 36 months. Versus the static-information comparator (D0) on the same 845 OASIS-3 landmarks, D1 improved 12-month progression AUC by 0.071 (95% CI 0.024–0.129). At 36 months, improvement and progression AUC differences were 0.024 (0.001–0.051) and 0.029 (0.003–0.051), with a multistate Brier difference of − 0.014 (− 0.027 to − 0.002). D1 persistence discrimination was poor (AUC 0.502 at 12 months and 0.541 at 36 months). D1 improvement-risk observed-to-expected ratios were 2.154 at 12 months and 2.283 at 36 months, indicating substantial underprediction. Cross-fitted local recalibration reduced the 36-month multistate Brier score from 0.707 to 0.629. In ADNI, agreement between CDR improvement and sustained diagnosis reversion was limited (Jaccard 0.254; 0.456 under parallel sustained definitions). Limitations include clinical rather than biomarker-confirmed staging, research-cohort sampling, and incomplete death and dropout information. Updated cognition improved external prediction of selected CDR-stage transitions, but persistence discrimination was poor and absolute improvement risk did not transport adequately without updating. The results do not establish clinical deployment readiness. Observed CDR-stage improvement may be transient and must not be interpreted as biological recovery from Alzheimer’s disease.
Institution: Chosun University
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $4,290
OpenAlex ID: W7234512551
FWCI:
Cites:
Cited by:
Related to: 10