Runs entirely in your browser · CSV files are never uploaded
FAIRNESS-AWARE NEUROIMAGING
Fair SPARE-AD.
Explore an AD-like imaging signature.
Place new scans in the context of the reference patterns.
RegAlign · XGBoost
REFERENCE EXPLORER
Score across age
Choose diagnoses, ethnicity & age range
Filters apply to reference and uploaded rows. Uploaded diagnoses come from DX; model predictions do not replace missing diagnoses.
Click legend entries to hide a group; double-click to isolate. Hover for scan details; drag to zoom. Trends are fixed reference fits, using straight lines or quadratic curves. Uploaded scans never enter the fits. Filters change visibility only.
BATCH OUTPUT
Your scored scans
Upload a CSV to return all three scores, the model output and the three-way trend prediction and original binary prediction.
| Scan | Age | DX | Index | Logit | Age-Z | Trend prediction | Binary model | Trend status |
|---|
Methods, interpretation & provenance
The reference contains precomputed scores for CN, MCI and DA scans. Subject identifiers, cohort information and source-partition labels have been removed, and the row order has been shuffled separately for each model. Original identity mappings are never delivered to the browser.
Best model uses the frozen selected model and its saved binary cutoff. Average model takes the arithmetic mean of probabilities from all 100 saved models before deriving scores, with a binary cutoff of 0.5. The browser selects the 145 features in their saved order, applies half-to-even rounding and float32 ICV normalization, and evaluates the original decision trees. CSV files and calculations stay on your device.
Some reference scans had training exposure or contributed to model adaptation or threshold selection. These displays are descriptive, not independent validation. Your features must already use the same upstream harmonization and volume units; the app does not fit a new harmonizer.
The reference contains 6,223 CN, 5,590 MCI and 2,784 DA scans. Records can include repeated visits; reference counts are not independent participant counts. Higher scores indicate a more DA-like imaging pattern. DA denotes Dementia of Alzheimer; AD labels are displayed as DA.
Trend prediction compares absolute score differences from the fixed CN, MCI and DA trends at the same age. Both straight and quadratic trends are available. A missing age or a nearest-trend tie leaves the trend prediction unavailable and retains the binary model label as a fallback. Extrapolation is flagged in the export. This nearest-trend rule is descriptive, not a separately trained or validated multiclass diagnostic model.
Age-Z uses the frozen CN age normalization for the selected model. This exploratory normalization does not establish that disease and aging have been separated. Reference curves and normalization never change with filters or your CSV. Browser and native floating-point calculations can differ slightly.
Private identity-based reference-overlap checking is unavailable in the browser. The export records this as not checked; your own scan identifiers and optional metadata remain only in local memory and your downloaded CSV.
RegAlign publication ↗ · Source code ↗ · Established SPARE framework ↗