independent grouping units
Post-acquisition analysis
Current workflow state and aggregate evidence
Cohort construction, internal validation, predictive feature analysis, and recovery-index results from the frozen controls-only cohort.
Complete
All registered analysis stages are current.
18 of 18 stages satisfy the manifest, output, and lineage contract. No catalog or result-projection issues are reported.
Open the reviewer summary18of 18
repeated-observation analysis rows
represented in the frozen cohort
of 109 registered groups current
Frozen controls-only analytical cohort
396 rows from 115 animals across 5 experiments. Animals—not rows—are the independent grouping unit.
Analysis domains
Methods and results by analytical stage
Each module links the relevant aggregate results to its validation design, uncertainty specification, and interpretation limits.
8evidence families
Source inventory · Canonicalization · Quality control · Metadata harmonization · Frozen cohort
Data & QC
Post-acquisition assembly of five source experiments, prespecified exclusions, bilateral paw aggregation, phase harmonization, and derivation of the frozen controls-only cohort.
Full-feature nested CV · Statistical evaluationValidation
The ordered Early–Middle–Late outcome was represented by two cumulative binary tasks and evaluated with pooled out-of-fold predictions from animal-grouped nested validation.
Cross-fitted calibration · Prediction sensitivityCalibration
Uncalibrated, logistic, and isotonic probabilities were evaluated using calibrators fitted to grouped inner out-of-fold predictions and applied to untouched outer-test records.
Information sourcesInfo sources
Source-block performance, paired incremental contrasts, Shapley-style contributions, and experiment-fingerprint diagnostics within the validated cohort.
Fold-safe SHAP · Feature ablation · Feature consensusFeature evidence
Native TreeSHAP explanations for held-out records and leave-one-feature-out refits on the same train–test partitions provide complementary predictive evidence.
Fixed-panel sensitivity · Adaptive three-feature procedureThree-feature
The fixed panel derives from 26-feature out-of-fold evidence in this cohort, whereas the adaptive procedure repeats feature selection within each outer-training partition.
Recovery index · GAM surrogateRecovery index
Platt-calibrated probabilities from the three fixed-panel classifiers are combined into a 0–100 expected ordinal class-position score, with observed-phase distributions and surrogate-fidelity diagnostics.
Classifier bundleClassifier
The three fixed-panel frameworks were tuned and refitted on the complete frozen cohort for application, while all displayed performance estimates remain out of fold.
Data-access boundary
Browser-visible data scope
Raw CatWalk workbooks, derived animal-level tables, identifiers, per-record predictions and explanations, model binaries, tuning databases, repository locations, and source fingerprints are intentionally absent from this application.