AD-TabFusion

Research question: Can longitudinal multimodal clinical tables predict future CN/MCI/AD state and MCI-to-AD progression while preserving strict temporal validation, calibration, and uncertainty reporting?

Core Results

Evidence	Split	Metric	Frozen result
Transition-aware diagnosis model	Locked temporal test	Macro F1	0.881
Transition-aware diagnosis model	Locked temporal test	ROC-AUC OvR	0.958
Source diagnosis + forecast horizon ablation	Validation	Macro F1	0.820
MCI progression, 24 months	Locked temporal test, n=53	ROC-AUC	0.797
MCI progression, 36 months	Locked temporal test, n=44	ROC-AUC	0.804
Isotonic calibration	Locked temporal test	Log Loss	0.352
Selective prediction, 80% target	Locked temporal test	Coverage / Macro F1	81.1% / 0.926

Method

flowchart LR
    A["TADPOLE longitudinal tables"] --> B["Leakage audit"]
    B --> C["Subject-level temporal split"]
    C --> D["Transition and MCI risk models"]
    D --> E["Validation-only calibration and thresholding"]
    E --> F["Locked temporal test"]
    F --> G["Read-only dashboard"]

Loading

Scientific Rigor

• Subject-level isolation prevents the same participant from crossing data splits.
• Future diagnoses and post-index outcomes are excluded from predictors.
• Model, calibrator, and confidence-threshold choices are made on validation data only.
• The dashboard reads frozen CSV/JSON artifacts and never trains, fits, calibrates, or overwrites results.
• D4 is always labeled exploratory post-hoc replay, not independent confirmatory validation.

Run Locally

pip install -r requirements.txt
streamlit run dashboard/app.py

The dashboard supports Chinese and English from the sidebar.

Data Compliance

Source ADNI/TADPOLE data are not redistributed. Public screenshots and dashboard pages contain aggregate metrics only; RID, PTID, and participant-level rows are not displayed. Users must obtain source data through the applicable ADNI/TADPOLE access and data-use process.

Limitations

• This is retrospective research software, not a clinical decision system.
• The 48-month MCI result is small-sample exploratory evidence (n=22).
• PET, CSF, and DTI are sparse and are not complete active modalities in the primary model.
• D4 replay is post-hoc and cannot establish independent external confirmation.

Detailed Reports

• Phase A: data and leakage audit
• Phase B: compact vs full internal study
• Phase C: D3/D4 matching report
• Phase D: final frozen report
• Dashboard redesign report

Presentation materials: 中文项目简介 · English project brief · 中文演示脚本 · result cards