We work toward a world without dementia by connecting quantitative neuroscience to real-world dementia care. Current work includes OASIS-3 secondary analyses integrating PET and psychometrics, multicenter stroke lesion segmentation supporting voxel-based lesion–symptom mapping, physician-scored evaluation of clinical AI (BooksMed/ExpertMedQA), and the Express Dementia Care (EDCC) implementation pilot focused on workflow, readiness, and equity.
Selected evidence
Each highlight links to a manuscript, preprint, or abstract so collaborators can evaluate methods and results.
99.4% → 4.7%
Simulated false-positive risk under the null (reuse vs confirmation)
Methods preprint on two common failure modes in biomarker trials and trial-ready cohorts: estimand drift from enrichment/range restriction and selection–inference reuse (“double dipping”). Includes an audit of 72 studies (10/72 with possible/definite non-independence) plus mechanistic simulations, with practical safeguards for design and reporting.
1,098
OASIS-3 longitudinal records analyzed
Secondary analysis of 1,098 OASIS-3 longitudinal records integrating amyloid PET and neuropsychological testing to model continuous dementia severity and cognitive decline—supporting more precise phenotyping and earlier detection (AUC > 0.90 in secondary analysis).
21 / 7
Physicians / countries in validation cohort
Twenty-one physicians across seven countries evaluated expert neurology scenarios to test whether clinical AI can be trusted in high-stakes reasoning; BooksMed/ExpertMedQA achieved state-of-the-art accuracy compared with Med-PaLM 2, Almanac, and GPT-4.
Diagnostics 2022
Multicenter validated model
Peer-reviewed multicenter lesion segmentation enabling reproducible quantitative imaging and scalable voxel-based lesion–symptom mapping workflows.
EDCC
Implementation pilot
EDCC implementation pilot integrates triage workflows, caregiver navigation, and anti-amyloid readiness/equity measurement—designed to extend evidence-based dementia care beyond major centers.
Methods preprint
Figures from our latest methods preprint on enrichment, estimand mismatch, and selection–inference reuse (“double dipping”).
Fig. 1 — AUC and null p-values
Enrichment shifts measured performance and distorts null p-values under weak association.
Open figure →
Fig. 2 — Double dipping vs split-sample
Selection–inference reuse inflates false positives relative to split-sample confirmation.
Open figure →
Fig. S1 — Simulation grid (A)
Grid summary of deployment vs in-enriched evaluation differences (supplement).
Open figure →
Fig. S2 — Sensitivity analysis (E)
False-positive risk sensitivity to the number of candidate biomarkers and sample size (supplement).
Open figure →
Preprints are shared for transparency and feedback and have not been certified by peer review; they should not be used to guide clinical practice.
Integrated Biomarkers
Longitudinal, multimodal modeling of secondary datasets (e.g., OASIS-3) integrating amyloid PET, structural MRI, and psychometrics to quantify continuous dementia severity and trajectories across distributed cortical and subcortical systems.
Related publications
Care Pathways
Closed-loop Express Dementia Care (EDCC) pathway focused on workflow design, caregiver navigation, and system-level measurement of access and readiness for disease-modifying therapies in rural East Texas.
Systems & AI
Combining primate systems neuroscience with clinically grounded AI, spanning lesion–symptom mapping, physician-scored benchmarks (ExpertMedQA), and BooksMed decision support.
Related publications
