Add paper

paper/trifuse_ad.md

@@ -0,0 +1,450 @@
+# TriFuse-AD: An Honest Multimodal Benchmark for Three-Stage Dementia Staging on OASIS-1 Structural MRI
+
+## Abstract
+
+We present a controlled, leakage-free benchmark for three-stage cognitive
+classification — cognitively normal (CN), very mild dementia (VMD, CDR=0.5), and
+Alzheimer's dementia (AD, CDR≥1) — on the OASIS-1 cross-sectional cohort, and we
+propose **TriFuse-AD**, a tri-planar CNN + slice-plane Transformer with gated
+demographic fusion. On an age-restricted cohort (≥60 years, 198 subjects) evaluated
+with subject-level repeated stratified 5-fold cross-validation (3 seeds, 15
+evaluations per model), we benchmark eleven models spanning tabular, 2D/2.5D/3D
+CNN, Vision Transformer, hybrid, and multimodal families, and report Macro-F1 with
+bootstrap confidence intervals as the primary metric.
+
+Our central finding is a *negative* but informative one: on this small, confounded
+cohort no MRI-only deep network surpasses a plain tabular XGBoost on morphometric
+and demographic features (Macro-F1 0.474), and TriFuse-AD (0.488 ± 0.066) does not
+significantly outperform a trivial DenseNet late-concatenation baseline (0.497 ±
+0.062; paired permutation p = 0.55). A confound analysis shows a structured-feature
+model using whole-brain volume, intracranial volume and demographics alone reaches
+Macro-F1 0.480, confirming that most of the recoverable signal on OASIS-1 is
+morphometric rather than learned from raw voxels. We report these results
+transparently, together with per-class, per-subgroup, ablation, and interpretability
+analyses, as a cautionary and reproducible reference point for small-cohort
+structural-MRI dementia staging. We make no clinical, diagnostic, state-of-the-art,
+or cross-site claims.
+
+**Keywords:** Alzheimer's disease, dementia staging, OASIS-1, structural MRI,
+multimodal fusion, confounding, small-sample benchmark, Macro-F1.
+
+
+## 1. Introduction
+
+Automated staging of cognitive decline from structural MRI is an attractive goal:
+T1-weighted scans are cheap, ubiquitous, and carry well-documented atrophy signatures
+of Alzheimer's disease (AD). A large literature reports high accuracy for AD-vs-CN
+classification, and a growing body extends this to intermediate stages. Yet much of
+this literature is difficult to compare or trust, for three recurring reasons.
+
+First, **label leakage and target-adjacent inputs.** Clinical staging labels such as
+the Clinical Dementia Rating (CDR) or Mini-Mental State Exam (MMSE) are sometimes
+fed to the model, directly or through hand-picked features, inflating performance in
+a way that would not survive prospective use.
+
+Second, **subject-level leakage.** When multiple scans of the same subject, or
+augmented slices from one volume, are split across train and test, the reported
+accuracy measures memorization of subjects rather than generalization.
+
+Third, **confounding by age and head size.** Age is strongly correlated with both
+brain atrophy and dementia stage; a model that merely reads off age (or a proxy such
+as whole-brain volume) can appear to "diagnose from MRI" while learning nothing
+disease-specific.
+
+This paper takes the opposite stance to the accuracy-maximizing literature. We build
+a deliberately conservative benchmark on the OASIS-1 cross-sectional cohort and ask a
+narrower, more honest question: *under leakage-free, age-restricted, subject-level
+evaluation, how much three-stage signal is actually recoverable from structural MRI,
+and does architectural sophistication help beyond trivial baselines?*
+
+Concretely, our contributions are:
+
+1. **A reproducible, leakage-free three-stage protocol** on OASIS-1: CN / VMD / AD
+   from CDR, with an age≥60 restriction to attenuate the age shortcut, one volume
+   per subject, subject-level repeated stratified cross-validation (15 evaluations),
+   and CDR/MMSE excluded from all model inputs. Macro-F1 with bootstrap CIs is the
+   primary metric, not accuracy, because the classes are imbalanced (98/70/30).
+
+2. **A broad, identically-trained model zoo** — eleven models from tabular XGBoost to
+   3D Swin Transformers and two recent CNN-Transformer hybrids, all retrained on our
+   folds so numbers are mutually comparable and never copied across datasets.
+
+3. **TriFuse-AD**, a tri-planar shared-CNN encoder with a slice-plane Transformer and
+   a gated demographic-fusion head, presented not as a winner but as a well-specified
+   point in the design space that we ablate component-by-component.
+
+4. **An explicit confound and subgroup analysis** quantifying how much of the signal
+   is available from demographics and morphometry alone, and how performance varies
+   across age bands and sex.
+
+Our results are sobering. Multimodal fusion helps — the two top models both fuse MRI
+with structured features — but TriFuse-AD's gated tri-planar attention does not beat
+a plain late-concatenation baseline, and no MRI-only network beats tabular XGBoost.
+We argue that reporting this clearly is more useful to the field than another
+incremental accuracy number obtained under looser controls.
+
+
+## 2. Related Work
+
+**Deep learning for AD classification.** Convolutional networks on 2D slices, 2.5D
+slice stacks, and full 3D volumes have all been applied to ADNI and OASIS. 3D CNNs
+(e.g., 3D ResNet, DenseNet variants) model volumetric context directly but are
+data-hungry; 2.5D approaches trade volumetric completeness for ImageNet-pretrained
+2D backbones and lower memory. Vision Transformers and 3D Swin Transformers have more
+recently been applied to brain MRI, and hybrid CNN-Transformer designs (compact HCCT,
+CNN-VSwinFormer) aim to combine local inductive bias with global attention. Reported
+accuracies vary widely and are frequently not comparable because of differing
+cohorts, label definitions, and validation protocols.
+
+**Multimodal fusion.** Combining imaging with tabular clinical/demographic data is
+well established, from early concatenation to gated and attention-based fusion. On
+small cohorts, however, the marginal value of sophisticated fusion over simple
+concatenation is rarely tested with matched training and statistics.
+
+**Confounding and evaluation critique.** A recurrent methodological thread warns that
+age, sex, and head-size confounds, together with subject-level leakage, can dominate
+apparent MRI-based performance, and that accuracy on imbalanced stages is misleading.
+Our work is squarely in this tradition: rather than proposing a higher number, we
+instrument the benchmark to expose how much signal is confound-driven, and we hold
+every model to the same leakage-free protocol.
+
+**OASIS-1.** The OASIS cross-sectional release provides T1-weighted scans with CDR,
+MMSE, and morphometric summaries (eTIV, nWBV, ASF) for a demographically broad adult
+cohort. It is smaller and less standardized than ADNI, which makes it a realistic
+stress test for small-sample generalization — and a setting where honest reporting of
+variance and confounding matters most.
+
+
+## 3. Data and Cohort
+
+**Source.** We use the OASIS-1 cross-sectional release (416 subjects, ages 18–96).
+Each subject has T1-weighted MPRAGE scans and an atlas-registered, brain-masked,
+gain-field-corrected volume (`*_111_t88_masked_gfc`, 176×208×176, 1 mm isotropic),
+along with demographic and morphometric summaries: age, sex, education, socioeconomic
+status (SES), estimated total intracranial volume (eTIV), normalized whole-brain
+volume (nWBV), atlas scaling factor (ASF), MMSE, and CDR. We reconstruct the
+per-subject metadata from the individual OASIS subject records.
+
+**Labels.** We map CDR to three cognitive stages:
+
+| CDR | Stage | Meaning |
+|-----|-------|---------|
+| 0   | CN    | cognitively normal |
+| 0.5 | VMD   | very mild dementia |
+| ≥1  | AD    | Alzheimer's dementia |
+
+We deliberately name CDR=0.5 **very mild dementia (VMD)**, not "MCI": the CDR
+operationalization is not equivalent to a clinical MCI diagnosis, and conflating them
+would overstate clinical relevance. **CDR and MMSE are never used as model inputs** —
+they define or correlate with the label and would constitute leakage. They are
+retained only for cohort description and confound analysis.
+
+**Age restriction.** Because age is the dominant confound (young CN subjects are
+trivially separable from older AD subjects on brain size alone), we restrict to
+**age ≥ 60**. This removes the easy young-CN population and forces models to
+discriminate stages within an older cohort where atrophy overlaps across stages.
+
+**Final cohort.** The age≥60 filter yields **198 subjects**: CN = 98, VMD = 70,
+AD = 30. Mean age 76.3 ± 8.1; 131 male / 67 female. One volume per subject. The
+class imbalance (roughly 3 : 2 : 1) and the small AD count (≈6 AD subjects per test
+fold) are central to how we evaluate and how we interpret variance.
+
+
+## 4. Method
+
+### 4.1 Preprocessing
+
+*3D pipeline.* Each volume is loaded, reoriented to a canonical axis order,
+cropped to the nonzero brain bounding box, intensity-clipped to the 0.5–99.5
+percentile range, z-scored over brain voxels, and resized to 128×128×128 for 3D CNNs
+or 96×112×112 for the more memory-intensive 3D transformers.
+
+*2.5D pipeline.* From each volume we extract, per anatomical plane (axial, coronal,
+sagittal), 9 slices spanning 30–70% of the depth range, each resized to 224×224 —
+27 slices per subject. This gives ImageNet-pretrained 2D backbones a multi-view
+summary of the volume without the memory cost of full 3D.
+
+All preprocessing statistics that could leak (none beyond per-volume normalization
+here) are computed within each volume; tabular imputation and standardization
+statistics are fit on the training fold only (Section 5).
+
+### 4.2 TriFuse-AD architecture
+
+TriFuse-AD has three parts:
+
+**(1) Shared tri-planar CNN encoder.** All 27 slices (3 planes × 9 slices) pass
+through a single ImageNet-pretrained ConvNeXt-Tiny backbone (num_classes=0, producing
+a pooled feature vector per slice). A linear projection maps each to a common
+embedding dimension. Sharing weights across planes and slices keeps the parameter
+count small — important for a 198-subject cohort.
+
+**(2) Slice-plane Transformer.** The 27 slice tokens receive additive slice-position
+and plane embeddings (so the model can distinguish an axial slice at 40% depth from a
+sagittal slice at 40% depth), a learned [CLS] token is prepended, and a 2-layer
+Transformer encoder aggregates them. The [CLS] output is the MRI representation
+`z_mri`, followed by layer normalization.
+
+**(3) Gated demographic fusion.** A small MLP encodes the structured features into
+`z_tab`. Fusion is gated:
+
+```
+g = σ(W_g · [z_mri ; z_tab])
+z = LayerNorm(z_mri + g ⊙ (W_t · z_tab))
+```
+
+The gate `g` lets the model modulate how much demographic information is injected per
+example, rather than blindly concatenating. A final 3-way linear head produces class
+logits. The ablation (Section 7) replaces (2) with mean-pooling, replaces the gate
+with plain concatenation, drops the metadata branch, and restricts to a single plane,
+isolating each component's contribution.
+
+### 4.3 Baselines
+
+We compare against ten other models, all trained on identical folds:
+
+- **Tabular:** XGBoost and an MLP on the structured features (age, sex, education,
+  SES, eTIV, nWBV, ASF).
+- **2D / 2.5D CNN:** ResNet50 (center axial slice), DenseNet121 (9-axial 2.5D).
+- **3D CNN:** 3D ResNet18 (MONAI).
+- **Transformers:** ViT-B/16 (2.5D, two-stage finetune), 3D Swin-T (MONAI).
+- **Hybrids:** a compact 3D HCCT and a lightweight CNN-VSwinFormer, both retrained on
+  our folds (no cross-dataset numbers copied).
+- **Multimodal baseline:** DenseNet + late concatenation of structured features.
+
+
+## 5. Experimental Setup
+
+**Cross-validation.** We use subject-level repeated stratified 5-fold CV with seeds
+{7, 13, 21} → **15 evaluations per model**. Stratification is by class × age band ×
+sex. Splits are strictly subject-level; we assert zero subject overlap across
+train / validation / test in every fold. All fold-dependent statistics
+(tabular imputation medians, standardization, class-balance weights) are fit on the
+training fold only.
+
+**Training.** Torch models use AdamW with cosine schedule, two learning-rate groups
+(a lower rate for pretrained backbones), bf16 automatic mixed precision, and early
+stopping on validation Macro-F1. The default loss is class-balanced focal loss
+(effective-number class weights, γ=2, label smoothing 0.05); a weighted
+cross-entropy variant is used in the ablation. Hyperparameters are frozen per model
+before the run; the test fold is never used for selection.
+
+**Metrics.** The **primary metric is Macro-F1**, which weights all three stages
+equally despite imbalance. We also report balanced accuracy, accuracy, macro
+precision/recall, one-vs-rest macro AUC, and per-class F1/recall. We report
+mean ± standard deviation over the 15 runs, and a bootstrap 95% CI on Macro-F1 from
+the pooled out-of-fold (OOF) predictions. Significance between the proposed model and
+the best baseline uses a paired permutation test on per-run Macro-F1.
+
+
+## 6. Results
+
+### 6.1 Main comparison
+
+Table 1 reports Macro-F1 (mean ± std over 15 runs) with the bootstrap 95% CI, ordered
+best to worst.
+
+**Table 1 — Three-stage (CN/VMD/AD) classification, OASIS-1 age≥60, 198 subjects.**
+
+| Rank | Model | Input | Macro-F1 | Bal. Acc. | F1(AD) | Macro-F1 95% CI |
+|------|-------|-------|----------|-----------|--------|-----------------|
+| 1 | DenseNet + concat | MRI+tab | **0.497 ± 0.062** | 0.525 | 0.390 | [0.47, 0.53] |
+| 2 | **TriFuse-AD** | MRI+tab | 0.488 ± 0.066 | 0.508 | 0.367 | [0.46, 0.54] |
+| 3 | XGBoost | tab | 0.474 ± 0.065 | 0.473 | 0.377 | [0.44, 0.52] |
+| 4 | ViT-B/16 | 2.5D | 0.462 ± 0.085 | 0.490 | 0.323 | — |
+| 5 | DenseNet121 | 2.5D | 0.449 ± 0.077 | 0.468 | 0.365 | — |
+| 6 | 3D HCCT | 3D | 0.407 ± 0.078 | 0.450 | 0.279 | — |
+| 7 | 3D Swin-T | 3D | 0.403 ± 0.085 | 0.442 | 0.280 | — |
+| 8 | Tabular MLP | tab | 0.390 ± 0.097 | 0.442 | 0.412 | — |
+| 9 | ResNet50 | 2D | 0.368 ± 0.074 | 0.399 | 0.205 | — |
+| 10 | 3D ResNet18 | 3D | 0.354 ± 0.069 | 0.413 | 0.240 | — |
+| 11 | CNN-VSwinFormer-lite | 3D | 0.354 ± 0.078 | 0.400 | 0.242 | — |
+
+Three observations stand out.
+
+**Multimodal fusion occupies the top two spots.** Both models that fuse MRI with
+structured features (DenseNet+concat, TriFuse-AD) outrank every single-modality
+model. Fusing morphometry/demographics with imaging is the single most reliable
+lever on this cohort.
+
+**But sophistication does not pay off.** TriFuse-AD's gated tri-planar attention
+(0.488) does **not** beat the trivial DenseNet late-concatenation baseline (0.497).
+A paired permutation test on per-run Macro-F1 gives Δ = −0.009, **p = 0.55** — the
+two are statistically indistinguishable, and the point estimate actually favors the
+simpler model. We therefore make no claim that TriFuse-AD is superior.
+
+**No MRI-only network beats tabular XGBoost.** The best pure-imaging model (ViT-B/16,
+0.462) sits below XGBoost on seven structured features (0.474), and the pure-3D
+networks (3D ResNet18, CNN-VSwinFormer-lite, 3D Swin-T) are the weakest of all
+(≈0.35–0.40) — below even the tabular MLP on some metrics. On a 198-subject cohort,
+volumetric deep networks overfit and underperform a gradient-boosted table.
+
+### 6.2 Variance and the small-AD problem
+
+Standard deviations are large relative to the differences between models (±0.06 to
+±0.10). The transformer models show the highest variance (ViT ±0.085, Swin ±0.085),
+consistent with our prior expectation: with ≈6 AD subjects per test fold, a single
+misclassified AD case swings fold-level Macro-F1 substantially. This is why we report
+distributions over 15 runs rather than a single split, and why we treat rank
+differences within a CI-width of each other as ties rather than wins.
+
+Per-class F1 (Table 3b in the artifacts) confirms the AD class is the hardest
+everywhere: F1(AD) tops out at 0.41 (tabular MLP, which trades CN/VMD precision for
+AD recall) and is 0.37–0.39 for the fusion models. VMD (CDR=0.5) is the most
+confused stage, frequently absorbed into CN or AD — expected, since it is the
+clinical boundary zone.
+
+
+## 7. Ablation Study
+
+We ablate TriFuse-AD component-by-component, all variants sharing the full training
+recipe and folds.
+
+**Table 2 — TriFuse-AD ablation (Macro-F1, balanced accuracy, F1(AD)).**
+
+| Variant | Macro-F1 | Bal. Acc. | F1(AD) |
+|---------|----------|-----------|--------|
+| A1: axial-only (single plane) | 0.477 ± 0.071 | 0.497 | 0.338 |
+| A2: mean-pool (no Transformer) | 0.465 ± 0.083 | 0.497 | 0.367 |
+| A3: no metadata (MRI only) | 0.390 ± 0.103 | 0.431 | 0.234 |
+| A4: concat (no gate) | 0.487 ± 0.053 | 0.514 | 0.373 |
+| A5: weighted-CE (vs focal) | **0.500 ± 0.054** | 0.518 | 0.413 |
+| Full TriFuse-AD | 0.488 ± 0.066 | 0.508 | 0.367 |
+
+The ablation is more informative than the headline number:
+
+- **Metadata is the dominant component (A3).** Removing the demographic/morphometric
+  branch collapses Macro-F1 from 0.488 to 0.390 and F1(AD) from 0.367 to 0.234 — a
+  larger drop than any imaging-side change. This is the single clearest effect in the
+  study and directly mirrors the confound analysis: the structured features carry
+  most of the signal.
+
+- **The gate does not help (A4).** Replacing gated fusion with plain concatenation
+  gives 0.487 — identical to the full model (0.488) within noise, and with *lower*
+  variance (±0.053 vs ±0.066). The gating mechanism, the architectural novelty of
+  TriFuse-AD, earns nothing here.
+
+- **The Transformer barely helps (A2).** Mean-pooling the slice tokens instead of the
+  Transformer aggregator gives 0.465 vs 0.488 — a small, within-noise difference.
+
+- **Tri-planar vs axial-only is marginal (A1).** Three planes (0.488) over axial-only
+  (0.477) is again within noise.
+
+- **Loss choice (A5).** Weighted cross-entropy slightly edges out class-balanced focal
+  loss (0.500 vs 0.488) and gives the best F1(AD) of any TriFuse variant (0.413),
+  though still within the confidence band.
+
+**Takeaway:** the only component with a large, unambiguous effect is the metadata
+branch. The imaging-side sophistication (tri-planar views, slice-plane Transformer,
+gated fusion) contributes little beyond what a single-plane CNN with concatenated
+demographics already achieves. This is an honest internal-validity result that we
+report rather than bury.
+
+
+## 8. Confound and Subgroup Analysis
+
+### 8.1 How much signal is in the structured features alone?
+
+To quantify confounding, we train shortcut baselines that use **no MRI at all**, on
+the same folds/seeds and evaluated identically.
+
+**Table 3 — No-MRI shortcut baselines (Macro-F1, balanced accuracy, AD recall).**
+
+| Model | Features | Macro-F1 | Bal. Acc. | AD recall |
+|-------|----------|----------|-----------|-----------|
+| Age-only logistic | age | 0.324 ± 0.014 | 0.390 | 0.589 |
+| Demographic XGBoost | age, sex, educ, SES | 0.417 ± 0.015 | 0.417 | 0.167 |
+| Full-structured XGBoost | + eTIV, nWBV, ASF | 0.480 ± 0.014 | 0.476 | 0.367 |
+
+The result is decisive. **A structured-feature model with no imaging reaches
+Macro-F1 0.480** — higher than every MRI-only network in Table 1 and statistically
+level with the best multimodal models (0.488–0.497). Adding the morphometric volumes
+(eTIV, nWBV, ASF) to demographics lifts Macro-F1 from 0.417 to 0.480, i.e. brain
+volume is doing most of the work. Age alone already recovers meaningful signal
+(0.324, with high AD recall because AD subjects skew older even within the ≥60
+cohort).
+
+This is the crux of the paper. On OASIS-1 at age≥60, the recoverable three-stage
+signal is largely **morphometric and demographic**, not something the deep networks
+extract from raw voxels beyond what a single normalized-brain-volume number provides.
+Any claim that a deep model "learns AD imaging biomarkers" here has to first clear
+the 0.480 structured-only bar — and none of ours does so convincingly.
+
+### 8.2 Subgroup robustness
+
+Slicing the pooled OOF predictions by age band and sex (for the strongest models)
+shows performance is **not stable across age**. For XGBoost, Macro-F1 falls from
+0.735 in the 60–69 band to 0.401 in the 80+ band. The youngest band is easiest
+because within age≥60 the 60–69 CN subjects still separate relatively cleanly; by
+80+, atrophy is widespread across all three stages and the task is hardest. Sex
+differences are smaller (male 0.486 vs female 0.440 for XGBoost). This age gradient
+is itself evidence of residual confounding: even after restricting to ≥60, age
+continues to structure the difficulty.
+
+
+## 9. Interpretability
+
+We apply Grad-CAM to TriFuse-AD's shared CNN encoder on a held-out fold (6 correctly
+and 3 incorrectly classified test subjects), reporting the center slice per plane.
+This is **qualitative only**. We do not claim the model localizes specific structures;
+we observe that, for correctly classified AD/VMD cases, activation tends to fall over
+peri-ventricular and medial-temporal regions — anatomy broadly associated with AD
+atrophy — while misclassified cases show diffuse or off-target activation. Given the
+confound analysis, we caution explicitly against over-reading these maps: a model
+whose signal is dominated by whole-brain volume may attend to atrophy-correlated
+regions incidentally rather than through disease-specific feature learning.
+
+
+## 10. Discussion
+
+**What we found.** Under a leakage-free, age-restricted, subject-level protocol on
+OASIS-1: (i) multimodal fusion beats every single modality; (ii) but the proposed
+gated tri-planar Transformer does not beat trivial late concatenation (p = 0.55),
+and its ablation shows the metadata branch — not the imaging architecture — carries
+the effect; (iii) no MRI-only network beats tabular XGBoost; and (iv) a no-MRI
+structured model reaches Macro-F1 0.480, level with the best fusion models.
+
+**Why this matters.** The dominant narrative in MRI-based AD staging is one of
+steadily rising accuracy from ever-larger networks. Our controlled reproduction
+suggests that on a small, realistic cohort much of the achievable signal is
+morphometric/demographic and available without deep learning, and that architectural
+novelty can vanish once trivial baselines and proper variance reporting are in place.
+We think the field benefits more from this being stated plainly than from another
+loosely-controlled record.
+
+**Is TriFuse-AD useless?** No — but its value here is as a *well-specified, honestly
+ablated* design point, not a winner. The gating and Transformer components are
+defensible ideas that simply do not earn their complexity on 198 subjects; they may
+behave differently on larger cohorts (ADNI-scale), which is the natural next test.
+
+## 11. Limitations
+
+- **Single cohort, small n.** 198 subjects, 30 AD; results are OASIS-1-specific and
+  not validated cross-site. We make no cross-dataset or SOTA claims.
+- **Non-clinical labels.** CDR-derived stages, with CDR=0.5 named VMD not MCI; not a
+  clinical diagnosis. No diagnostic or clinical-decision claims.
+- **Residual confounding.** Age≥60 attenuates but does not remove the age/atrophy
+  confound, as the subgroup gradient shows.
+- **Cross-sectional only.** No longitudinal progression modeling.
+- **Compute environment.** The instance filesystem is non-persistent; all data and
+  results are re-derivable from the released code, but no long-term artifact store is
+  assumed.
+
+## 12. Conclusion
+
+We built a deliberately conservative three-stage dementia-staging benchmark on
+OASIS-1 and evaluated eleven models plus a component ablation, confound analysis,
+subgroup breakdown, and interpretability under one leakage-free protocol. Our
+proposed TriFuse-AD is competitive but does **not** significantly outperform a
+trivial multimodal baseline, and no imaging network beats a no-MRI structured model.
+Rather than a state-of-the-art claim, we offer a reproducible, honestly-reported
+reference point showing how much of the apparent MRI signal on small confounded
+cohorts is actually morphometric — and a reminder that variance, confounding, and
+trivial baselines must be reported before architectural credit is assigned.
+
+---
+
+*Reproducibility.* All models are trained on identical subject-level folds
+(seeds 7/13/21 × 5-fold). Tables 1–3 and all figures (main comparison, per-model
+confusion matrices, ROC curves, subgroup chart, Grad-CAM montage) are regenerated
+deterministically from saved out-of-fold predictions. CDR and MMSE are excluded from
+all model inputs.