GRAD: A Two-Stage Algorithm for Resolving Diagnostic Uncertainty in the Plasma p-tau217 Gray Zone

Introduction: Plasma phosphorylated tau-217 is widely used as a plasma-based biomarker for Alzheimer's Disease detection, demonstrating superior accuracy for detecting brain amyloid pathology. However, 30-50% of patients fall within an intermediate diagnostic "gray zone" where biomarker results are indeterminate, often decreasing physician confidence and requiring subsequent diagnostic workup. To address this, we developed a two-stage machine learning algorithm 'GRAD: Gatekeeper & Reflex for Alzheimer's Disease' to increase clinical confidence and reduce the AD health economic burden. Methods: We initially analyzed 320 participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) with plasma biomarkers and amyloid PET. We then built a two-stage machine learning classifier mimicking real clinical workflow where the stage 1 'Gatekeeper' used the gold-standard marker: p-tau217 with respective 25%/75% probability thresholds. The stage 2 'Reflex' step applied Random Forest multi-marker classification (p-tau 217, AB42/40, NFL, GFAP) for difficult-to-diagnose gray zone cases. To ensure statistical robustness, leave-one-out cross-validation with bootstrap confidence intervals was used. We externally validated the GRAD architecture on 1,644 A4 Study participants, with MRI enhancement analysis in 1,044 gray zone cases. To measure cost-effectiveness we compared our GRAD-staged testing to universal PET. Results: The model's 'Gatekeeper' resolved 55.6% of ADNI cases with 88.8% accuracy (NPV 91.8%, PPV 85.0%). The complete pipeline achieved AUC 0.867 (95% CI: 0.825-0.904), with 80.6% sensitivity, 80.0% specificity, LR+ 4.03, LR- 0.24. For the difficult-to-diagnose gray zone cases, the 'Reflex' machine learning model achieved AUC 0.755. In our A4 validation, the predictions correlated strongly with centiloid (r= 0.693). Expanding beyond plasma biomarkers, MRI integration improved gray zone classification from AUC 0.829 to 0.853 (p=0.014). The cost modeling analysis projected a 67% reduction in spending versus the current standard of universal PET. Discussion: Our clinically-staged diagnostic algorithm, 'GRAD', provides actionable classifications for the majority of patients while routing uncertain cases for additional workup. The GRAD framework offers a practical, cost-effective approach for implementing plasma biomarkers in clinical practice. Future iterations of this framework, with integration of novel biomarkers like MTBR-tau243 present a significant opportunity to alleviate the AD health-economic burden and eliminate expensive but unnecessary diagnostic measures.