بشب

Prompt

Role: Falsification-driven Principal AI Scientist, Mathematical Systems Architect, and Q1 Peer Reviewer (Nature Mental Health, Biological Psychiatry, NeuroImage, IEEE T-PAMI/T-BME) specializing in multimodal neuroimaging AI (EEG, naturalistic movie fMRI, sMRI) and dimensional psychopathology. Rules: 1. Grounding: If web search is available, verify 2024–2026 HBN/RBC/EEG-Challenge benchmarks; if offline, rely strictly on verified literature and tag uncertain numbers/years with [VERIFY]. Never fabricate citations or numbers. 2. Evidence Stratification: Explicitly separate (a) mathematical/statistical facts derived from specs/logs, (b) verified published literature, and (c) testable hypotheses. 3. Anti-Sycophancy & Anti-Peeking: Treat every module, loss term, preprocessing rule, and metric as a hypothesis to falsify. Outer-test logs are for diagnostic failure analysis only; all Phase A/B decisions must lock strictly to inner splits (`fit / selection / calibration`). --- ### Context & Core Empirical Anomaly We model 4 continuous CBCL bifactor psychopathology dimensions in the Healthy Brain Network (HBN) without DSM labels. Exact codebase specs, tensor shapes, parameter counts, formulas, and 5-fold Kaggle execution logs (Tesla T4, 55.12 min) are detailed below. **Core Anomaly:** 5-fold OOF Multimodal achieves Macro ΔR² = -0.0010 (95%CI = [-0.0030, +0.0008]) vs. pure demographic baseline M0 (+0.0462), underperforming M0 (+0.0462), Deep Neural scalar-calibrated (+0.0461, ΔR² = -0.0001), and Standalone EEG-Only (+0.0464, ΔR² = +0.0002)—even though a 5-fold CV Ridge positive control on the exact same brain features (`E+F+S`) predicts chronological age at R² = +0.7198 (r = +0.8489, MAE = 1.35 yrs). On Fold 0 outer test, positive `r_res = +0.0473` co-occurs with `ΔR² = -0.000085`, and subgroup ΔR² is most negative in highest-QC subsets (`Trimodal = -0.00242`, `QC_Pass = -0.00225`). --- ### 1. Cohort, Targets & Pre-Extracted Phase 7 Inputs - **Cohort (`N = 2,845`):** Filtered from 3,241 Phase 7 subjects (`EEG_REQUIRED_FMRI_OPTIONAL`; age = 10.26 ± 3.27y, 34.8%F). Outer 5-fold CV (569 test / 2,276 train per fold, stratified via 128-trial balanced KMeans-4D + fMRI + sMRI-QC); target-free inner split per fold: `fit = 1,478` (~862 paired), `selection = 342` (~200 paired), `calibration = 456` (~266 paired). `family_id` absent from public manifests. - **Sites & Releases:** `HBNsiteCBIC=868, HBNsiteRU=819, HBNsiteCUNY=350, HBNsiteSI(1.5T)=202, HBNsite_Unassigned=606 (21.3%)`; `R1..R11` (`NC = 348`). Low-symptom `broad_q75` reference: `N = 931` (flagged `cohort_clin__td_anchor`, NOT clinical TD); elevated: `N = 1,046`. - **Targets & Scaling:** 4 Challenge CBCL bifactor scores (`p_factor, attention, internalizing, externalizing`), left floor-censored at `p_factor <= -1.908` (`rbc_mcelroy_native` available on ~1,540). Fold target scaling on `fit`: `tc = nanmedian(y_fit)`, `ts = 0.6*MAD + 0.4*SD`, `y_reg = (y_raw - tc) / ts`. - **Cached Feature Tensors (Phase 7; NO raw 200 Hz waveforms, NO `RestingState / SuS / CCD / SL / SyS`):** 1) **EEG (`N = 2,845`; `2,013` dual-movie):** `ThePresent` (249 bins) & `DespicableMe` (213 bins) @ 0.8s (`valid_ratio>=0.75, bad<=25`). Tensors: `eeg_nodal_biophys [2,100,16]` (5 log10 band powers, 5 scaled rel powers, log1p TBR, custom peak-excluded 2–35 Hz aperiodic slope & intercept, 3 Hjorth), `eeg_asymmetry [2,7,5]` (35), `eeg_log_pec_deg [2,3,100]`, `eeg_log_pec [2,3,100,100]` (theta/alpha/beta orthogonalized PEC), `eeg_temporal_env [2,250,6]` (`log1p(RMS)` detrended via `0.65*(x - uniform_filter1d(15)) + 0.35*(x - mean)` + per-subject Z-score), `eeg_periph_env [2,250,28,2]`, `eeg_dynamic_isc [2,48,6]` (win=15, stride=5 vs. `broad_q75` template, LOO on train anchors), `tangent_eeg_pec [48]` (alpha-PEC SPD log-Euclidean tangent, adaptive shrinkage `rho>=0.2`, `bg_basis=None`, soft-whitened PCA-48). 2) **Movie fMRI (`N = 1,660` complete; `1,106` dual-movie; `1,185` missing; `0/1,431` `movieDM` pairs have verified EEG-fMRI time alignment):** `movieTP` (250 TRs) & `movieDM` (750 TRs) (`max_mean_fd=0.5mm`; Phase 7 compilation required `clean_trs>=200` uniformly across sites, including 1.5T `HBNsiteSI`). Tensors: `tangent_fmri_s100 [48]`, `fmri_glasser_vmhc [2,180]`, `fmri_glasser_nodal [2,360,4]` (log-DC, ALFF, ReHo, VMHC) + `fmri_glasser_summary [2,360,8]`, `fmri_s100_vmhc [2,50]`, `fmri_tian_vmhc [2,25]`, `fmri_nodal_156 [2,156,4]`, `fmri_temp_net_ts_by_task [2,750,14]` (bilateral Yeo-7), `fmri_dynamic_isc_by_task [2,148,14]`. Pooled via `DURATION_WEIGHTED` (3:1 `movieDM:movieTP`). `USE_REST_FC = False`. 3) **sMRI (`1,477/2,216` QC-eligible under `PASS_AND_ARTIFACT`:** `Pass=1,089, Artifact=388, Fail/NoQC=1,032` overlapping fMRI-fail, leaving 336 no-MRI uncategorized; `Trimodal=1,184`; 293 `EEG+sMRI` without fMRI included in `eeg_only=1,185`): 1,885 FreeSurfer (`Glasser-360×4 [ct,sa,gv,wgpct] + Schaefer-100×4 + Subcort logvol 45`). --- ### 2. Fold-Contained Preprocessing (`FoldPreprocessor`, fit on `fit = 1,478`) - **`FoldContainedComBat`:** EB site location/scale preserving `[age, sex, age^2]`. Skips `HBNsite_Unassigned` (21.3%) in `fit` and `transform` (unassigned MRI stays on raw scale while known sites shift to `grand_mean`; EEG mixes sites + `EEG_{release}` batches `R1..R11`). - **sMRI Normative (`1,885 -> 122`):** Median-fill -> `anat_combat` -> sex-specific `SplineTransformer(knots<=4, deg<=2, alpha=10)` on `broad_q75` anchors (251 total: 189M, ~62–95F/fold) with leverage-adjusted residuals `(Y - Y_hat) / max(1 - h_lev, 0.30)` -> `relative_cortical_summary` (global cortical mean subtracted; `hbn_analysis.py` also puts `morph_log_total_area` in `_age_basis`). - **`BlockMaskedNormativePipeline` (`497 = 130 EEG + 245 fMRI + 122 sMRI`):** For EEG & fMRI, fits `SexSpecificBSplineNormativeModel` (`knots=3, deg=2, alpha=10`, 12y hinge, heteroscedastic age-slope) on `broad_q75` anchors (492 EEG, ~300 fMRI) and evaluates in-sample on `fit` anchors (no LOO). Purges `valid_ratio, effective_mean_fd, euler/100` via ridge (`denom = sum(dev^2) + 10`). With `NORMATIVE_DEVIATION_RAW_BLEND_WEIGHT = 0.0`, `StandardScaler` is bypassed for EEG/fMRI (leaving anchor-MAD/SD units) while sMRI passes through `StandardScaler`. - **`FoldBrainPADModel`:** 5-fold inner-OOF `RidgeCV(1..1500)` age from `X_harm` (`[0:130]` EEG, `[0:375]` MM), linearly age-debiased -> `brain_pad_z [B,2]`. - **`FoldConfoundBaselineModel`:** `M0` (8 standardized demo cols: `age_z, age_z^2, age_z^3, hinge(age>12), sex_c` + interactions) via `RidgeCV(15..5000)` + 5-fold inner-OOF. `M2` adds `has_fmri, has_fmri*FD, has_anatomy, has_anatomy*Euler` + site one-hots. --- ### 3. `LALFNetV3` Architecture (2,763,766 Params), Losses & Stacker - **Profile (`reference`):** `VIEW_TOKEN_DIM=80, FUSION_DIM=160, USE_RAW_LABRAM_STREAM=False, USE_REST_FC=False, USE_CONTINUOUS_MANIFOLD_MIXUP=False, EEG_GEOMETRY_ATTENTION=False, EEG_QUERY_INTERACTION=False, EEG_BAND_RELATION_WEIGHT=0, SITE_MMD_WEIGHT=0, GRL_MAX_ALPHA=0`. - **Subsystems:** 1) `eeg_encoder (336,694)`: encodes 2 EEG movies (2-layer `EdgeScaledSignedGraphConv` across 3 PEC bands + `BrainOmniLatentSourceProjector(16)` + `CogRepLKBlock1D(15,3)` on `temp_env+periph_env+dyn_isc`) + global tangent -> `movie_tokens [B,2,80], global_eeg_token [B,80]`. 2) `fmri_encoder (266,626)`: `DURATION_WEIGHTED` (3:1 `DM:TP`), encodes `4S156+Tian50` GNN + tangent, `Glasser360+VMHC`, and `exp(-2*FD)` `750×14` TS + `dyn_isc` -> `fmri_tokens [B,3,80], fmri_pooled [B,80]`. 3) `context_encoder (274,202) + fusion (13,040)`: mean active `norm_vec [497]` + `anatomy [1,885]` (`1885->24->1885` sigmoid gate + `1885->24` bottleneck blended with 4-family split `SA 460, CT 460, GV/WG 920, Subcort 45`) + `brain_pad_z [2]` -> `[B,80]`. 4) `CrossModalLatentTranslator (382,673)`: predicts `synth_fmri_tokens [B,3,80]` + `synth_conf` (`init=0.85`). When `has_fmri==0` (or 12% deferred dropout), outputs `(0.38*synth_conf)*synth_fmri_tokens` and sets `fmri_active_mask = ones(B,3)` (~36% view share in `TransformerU`). 5) `StructureAwareMutualTransformerU (1,087,841)`: 7 views (`[B,7,80] -> [B,7,160]`), DINT self-attn + Sparse MoE + capped pooling (`T=0.55, cap=0.42, prior=0.08`; unused `shape[1]==9` branch) -> L2-normalized `pre_covariate [B,160]`. 6) `PostEmbeddingCovariateFiLM (16,800)`: modulates `pre_covariate` by `[age, age_miss, sex_c, sex_miss, age*sex_c, brain_pad_z[0:2]]` -> `post_covariate [B,160]`. 7) `OrthogonalBifactorRegressionHeads (308,192)`: target cross-attn (`48`) + 4-modality gated bypass on `norm_vec[497]` (`48`) + `post_covariate (160)` + `mod_ctx (4)` = `260` -> 4 `target_trunks` (`Linear+ResidualGLU(64)`). `orth_strength=0.0`; `cross_target_coupler`; `reg_residual_mu = 0.42*sigmoid(logits)*(0.65*r + 0.875*tanh(r/2.5))`; `reg_mu = 1.0*y_confound_scaled + reg_residual_mu`. - **Training & Loss:** - Stage 1 (6 ep, 861 paired, `bs=32`): `VICReg(25*inv+12*var+2*cov) + Translator(10*MSE+8*cos+2*PCC+3*var+2*ssl+1.5*rel)` (`val cos_sim=+0.6236, loss=22.6807`). - Stage 2 (22 ep, 1,478 fit, `bs=32`): `TRUE_FREEZE_UNFREEZE` (ep 1-3 heads/FiLM [392k], ep 4-7 +ctx/trans/U [1.38M], ep 8+ all [2.76M]; `LR_HEAD=1.4e-4, LR_BACKBONE=4.5e-5, WD=0.06, patience=9`). - Stage 2 Loss: `0.5*MSE + 0.35*SmoothL1(res, beta=0.5) + 0.12*res_align + 0.05*BetaNLL(0.5) + 0.02*attn_entropy`, where `res_align = (1-damped_r) + 0.1*(1-ccc) + 0.25*overshoot + 0.02*bias^2`, `desired_sd = 0.15*stopgrad(sd(y_res))`, `overshoot = relu(pred_sd - 1.6*desired_sd)^2`, `damped_r = cov/sqrt(vy*(vp+desired_sd^2))` + `0.02*VICReg (~0.227)` + `0.05*TranslatorMSE (~0.037)` (~25% non-reg loss). Selection uses `crossfit_selection_calibration` (3-fold `fit_centered_residual_calibrator`) on `selection=342`. - **Post-Hoc Stacker & Calibrators (`N_cal = 456`):** - `FoldContainedHybridStackingCalibrator`: fits 5 `_MaskedModalityBlockExpert` (`EEG_Block, Movie_fMRI_Block, Pure_sMRI_Block, Deep_Embedding, RESTFC_Block[inactive]`; inner-CV `abs_r` top-K + 24 SVD, RidgeCV `100..50000` on `fit`), `Deep_Head`, and 5 active `Phase7ResidualExpert`s (`EEG_ThePresent_Stable, EEG_DespicableMe_Stable, EEG_Transition` [VAR(1) on 0.8s 5-lobe RMS], `Movie_fMRI_Dynamics, sMRI_Regional`) trained on raw `extract_feature_bank()` without `NuisanceDesign` residualization (median-imputed via `ObservedScaler(add_masks=False)`). - `PatternRidgeCalibrator` (`N_cal=456`): partitions by exact NaN pattern (`min_pattern=140`, fallback `global_model` with zero-imputation & `use_indicators=False`), fits nonnegative `BVLS (coef>=0)` minimizing `||yc - zc@w||^2 + n*lambda*||w||^2` (`lambdas=(50,15,5,1,0.2)`) + 3-fold `oof_`. - `deep_mm`: 1D `fit_centered_residual_calibrator` (`strength=n*max(r,0)^2`, `ratio=min(0.25, strength/(1+strength)*max(r,0))`, `slope=ratio*sy/sx`) on `N_cal=456`. - Blend (`select_residual_blend(penalty=0.0015)`): grid search over 231 pairs `(s,d) in {0,0.05,...,1}^2, s+d<=1` minimizing `mean((y_res - s*stack_delta - d*deep_delta)^2) + 0.0015*var(y)*(s^2+d^2)` on `N_cal=456`. - **Aux Scripts:** `hbn_baseline_audit.py` builds `D:10, SITE:17, Q:12, MISS:9, E:433, F:354, S:178`, applies `ObservedScaler(add_masks=True)` separately to `D` and `B`, and fits `[dt, gamma*bt]` (`gamma in {0,0.1,1.0}`) under ONE shared Ridge `alpha in {1..10000}`. `Cell 8` (`hbn_thesis_cell8.py`) fits `MissingDiagonalGMM(K=6..12)` on 12 residualized physiology features on fixed split `fit=folds(0,1,2), val=3, test=4` (`paired=[528,170,173]`, `elevated=[135,45,41]`) + `partition_panels` `LogisticRegression` predicting GMM clusters from the same 12 features. --- ### 4. Empirical Kaggle Execution Log (`N = 2,845`) - **Confound M0 (`fit=2,276`):** `p_factor R2=-0.00007 (r=0.0348)`, `attention +0.01454 (r=0.1215)`, `internalizing +0.09724 (r=0.3118)`, `externalizing +0.07660 (r=0.2768)` -> Macro R2 = `+0.0471` (`fit=1,478` -> `+0.0462`); `M2 = +0.0445`. - **`hbn_baseline_audit` (5-fold nested OOF):** - `full (2,845)`: `M0=+0.045272` | `D+S=+0.045817 (+0.000545)` | `D+E=+0.044964` | `D+F=+0.044449` | `D+EF=+0.045189` | `D+EFS=+0.044819` | `D+TECH+MASK=+0.039911`. - `eeg_fmri_paired (1,660)`: `M0=+0.045373` | `D+F=+0.045422` | `D+S=+0.045174` | `D+E=+0.042763` | `D+EFS=+0.043199`. - `trimodal_paired (1,184)`: `M0=+0.051703` | `D+EF=+0.051582` | `D+EFS=+0.051553` | `D+S=+0.050589` | `D+E=+0.049532`. - Brain Age Ridge on `E+F+S`: `R2=+0.7198, r=+0.8489, MAE=1.35y`. - **5-Fold OOF Neural Results:** `Full Multimodal (Blended)`: Macro R2=`+0.0452 (dR2=-0.0010, 95%CI=[-0.0030,+0.0008])`, `r_tot=+0.1844`, `r_res=+0.0175`, `rho_s=+0.1745`, `CCC=+0.0922`, `AUC_Q25v75=0.6241`, `P75_AUC=0.5902` (`p_factor=0.5063` vs `M0=0.5890`/`0.5041`) | `Deep Neural (scalar cal)`: Macro R2=`+0.0461 (dR2=-0.0001)`, `r_tot=+0.1879`, `r_res=+0.0152`, `AUC_Q25v75=0.6257` | `Standalone EEG-Only`: Macro R2=`+0.0464 (dR2=+0.0002)`, `r_tot=+0.1862`, `r_res=+0.0181` | `Counterfactual No-QC sMRI`: Macro R2=`+0.0455, r_res=+0.0215`. - **Subgroups (ΔR² vs M0):** `Trimodal=-0.00242` | `QC_Pass=-0.00225` | `QC_Pass_And_Artifact=-0.00206` | `All_Selected_EEG=-0.00182` | `Strict_LowMotion=-0.00167` | `Dual_Movie_fMRI=-0.00150` | `Complete_EEG_fMRI=-0.00139` | `EEG_Only=-0.00044` | `Fail/NoQC=+0.00020`. - **Training & Calibration Diagnostics:** Early stopping fired BEFORE encoder unfreezing (ep 8) in Fold 1 (`best_ep=4, mm_dR2=0.0`) & Fold 2 (`best_ep=6, mm_dR2=+0.000069`). Fold 0/4: by ep 16–19, `train r_res` reached `+0.201..+0.232` (`dR2=+0.032..+0.043`) while `inner_val raw_deep dR2` crashed to `-0.0218..-0.0261` (`p_factor SD=0.170` vs `confound SD=0.066..0.102`). Fold 0 outer test: despite `r_res=+0.047335`, `mm dR2=-0.000085` (`Deep +0.04530 > M0 +0.04502 > MM +0.04493`, `EEG-Only +0.04596`), blend `mm=[0.15,0,0,0.05]`, `PatternRidge` coefs `Movie_fMRI_Dynamics=4.1971, EEG_DM_Stable=1.4420, Movie_fMRI_Block=1.1104, sMRI_Regional=0.8865`. `Cell 8`: `paired_films K=6` (`ARI=0.245<0.35, sil=0.021`, BH-FDR n.s.) & `elevated K=12` (`N_test=41`, sparse) labeled `exploratory_candidate_only`.