Public repository milestones and architecture releases.
Multiple public package releases with CPU-verified scientific contracts: CardiAnatomy 0.5.0 (complete CPU anatomy workflows, QC/cache repairs), CardiMech 0.3.0 (mechanics layer, truthful QC, constitutive material-point checks), CardiEP 0.3.0 (EP engine, bounded calibration, cross-platform CI), CardiFlow 0.2.0 (0D numerics, waveform/coupling contracts), CardiAgent 0.4.0, CardiVex 0.4.0, and DCCP 0.4.0 (label-blind evaluation, strict artifact contracts, pinned CPU notebook).
Calibration-grade ECG observation handoffs for CardiEP/CardiInfer: multi-lead median-beat templates, R-relative timing, QRS summaries, residual-noise estimates, lead-quality metadata, SHA-256 provenance, and electrotrace calibration prepare / HeartTwin electrical.prepare_calibration routing. Documentation and release automation centered on the public package. Does not retrain frozen detector evidence.
Public HeartTwin layers for patient-specific mechanics (PV/Windkessel reference + FE plugin boundary), electrophysiology (anisotropic Eikonal, scar-aware propagation, pseudo-ECG, PDE backend gateway), and inference (multimodal parameter estimation, uncertainty, calibration diagnostics). Designed to consume CardiAnatomy geometry and ElectroTrace calibration artifacts.
Major public release on PyPI and GitHub. Highlights: public Python API and electrotrace CLI; detector plugins and provenance manifests; certified WFDB benchmarks (pinned 10.7.0); leakage-safe polarity gates; prospective lead-selector line on EDB, LTAFDB, SVDB, and Zymed LTSTDB; QTDB 105-record tolerance-curve delineation; full 75-record INCART source-cohort characterization; archived validation evidence under validation_reports/experiments/. Python 3.10–3.13. Research software with explicit evidence boundaries.
New HeartTwin anatomy layer: CMR/CT ingestion contracts, segmentation-to-mesh staging, fiber and scar utilities, geometry QC, license-aware external-tool catalog (nnU-Net, biv-me, Meshtool, and related), and readiness gates for EP, mechanics, and flow. AGPL package with CI and CLI (cardianatomy).
Two new stack contracts: CardiFlow for hemodynamics and flow simulation (boundary conditions, CFD/reduced-order backends, flow biomarkers), and CardiTherapy for virtual-intervention experiments (pacing, ablation, pharmacologic, regenerative, device plans with provenance). Fail-closed when numerical backends are missing.
Starvation-rescue lead selector v3 frozen; Zymed LTSTDB validation preregistered with fail-closed evaluator and post-hoc comparator staging. Parallel PhysioNet HTTPS transfer for LTSTDB. Continues leakage-aware ECG phenotyping work on public datasets.
Quality-consensus selector v2, LTAFDB prospective protocol, SVDB annotation-window alignment, label-free feature audits, and post-hoc comparator workflows merged across the ElectroTrace evaluation path.
Public intro cut (~1:20) covering the computational cardiovascular stack , measure, understand, simulate, design , and research programs VB-101, VB-204, and HeartTwin / VB-310. Research-stage framing throughout. Available on the site and on X.
CI test fixes, Ruff cleanup, and removal of residual INCART assignment noise after the polarity-threshold and domain-shift work. Continues the leakage-free evaluation path on PhysioNet data; worst-case external records remain an active research problem, not a solved claim.
Repolarization companion-peak suppression heuristic, config fields, and regression tests merged (detector-fixes line). Aimed at cleaner MEA beat handling for CiPA-aligned research scoring , not a clinical cardiotoxicity product.
Repaired source corruption (literal newlines), stabilized strict typing across supported Python versions, constrained NumPy compatibility, and restored unique factorial run identifiers so experimental-design runs stay reproducible.
Pull request merged: contracts, integrity checks, and validation kit aligned with a protected benchmark workflow. Independent evaluation boundaries documented; scorecard and robustness tests hardened. Research evaluation infrastructure , not a clinical claim.
Docs and release-manifest fixes for integrity boundaries, optional reference maps, and consistent benchmark identities across scorecard fixtures.
Isolated Colab validation path for defensive challenge runs: dependency isolation, pip audit in a test environment, and cleaned runtime notebook flow so validation does not depend on a sticky Colab venv.
Locked eligible cardiac validation cohorts, quarantined ambiguous studies, froze Stage-1 real-data benchmark definition, and hardened GEO sample mapping (schema-driven matrix columns, complete sample contracts, safer per-GSM count assembly). Colab T4 real-data validation notebook linked in-repo.
Hybrid segmentation prefers Cellpose when threshold paths collapse to zero objects (LIVECell). Measured hybrid n=20 Dice ≈0.622 , collapse fixed; pure Cellpose still stronger where it runs cleanly.
Typed multimodal workflow through the Python facade, runnable demo command, full native E2E pipeline tests, CardiBridge consumer results retained and validated, CardiVex result typing in workflow state, and a cross-repository integration matrix. Research orchestration , not a clinical twin.
BBBC038 validation path (mirrors BBBC039; per-nucleus mask decoding), continuing the public microscopy QC validation track.
Public scaffold for the Computational Graft Integration & Risk Engine , layers docs, Layer 2.5 benchmark gate, phenotype maturity axes, locked scientific boundaries.
Benchmark infrastructure layer and public APIs for reproducible defensive challenge runs.
Reproducible Cardiac-Digital-Twin numerical-equivalence CI gate; endocardial propagation regimes aligned with upstream-equivalent behavior.
windowed_std scale estimator, stage-2 feature normalization, and confidence-gated polarity selection on real PhysioNet data (INCART aggregate F1 moved from ~0.47 to ~0.82 across the stage-1/2 sequence).
Surgical dose-response endpoint guards, browser parity tests, and CLI/wheel smoke-test fixes on the MEA scoring pipeline.
Public write-up of MIT-BIH vs INCART performance gap and Stage-1 candidate-generation diagnosis , framed as an open problem, not a solved claim.
Computational Cardiovascular Research Platform , DOI 10.5281/zenodo.22680867. Research framework documentation, not clinical validation.
