# Human Digital Twin — technical case study

A local scientific workbench for whole-body drug transport on a detailed anatomical view, with an optional Pulse 4.4.0 physiology runtime and a preserved set of independent solvers. It is a computational engineering platform. It is not a clinically validated patient model.

## Objective

Show a reproducible path from a scenario document to organ flows, drug amounts, optional tissue fields, and conservation residuals, on a body a person can recognize. Keep the independent cardiovascular and organ research available without presenting its unresolved outputs as the portfolio default.

## Architecture

The browser loads BodyParts3D release 4.0 meshes with Three.js. A scenario is a JSON document. `run_experiment` builds a circulation at each synchronization knot, advances the well-stirred PBPK state with a matrix exponential, and, when configured, advances finite-volume tissue samples that exchange drug with arterial and venous blood. Results are written under `results/workbench/` with a manifest, trajectory, and visualization file. The viewer holds stored samples. It does not integrate again and it does not interpolate.

If the schedule asks for Pulse, `pulse_runtime.run_healthy_adult` advances Pulse 4.4.0 and the coupler maps named vascular inflows onto the Jones organs. Limb beds stay in the residual. Pulse drug substances are not requested. If Pulse cannot start, the run fails and no synthetic exercise trace is inserted.

Static scenarios use the Jones and Rowland-Yeo 2013 flows directly.

## Classification of the work

A. Independently developed implementations. The PBPK state and matrix exponential, the finite-volume tissue solver, the scheduler, the renal and transit updates, the independent heart and lung, the organ glucose pool, and the blood-count baseline are code in `src/hdt`. They implement published forms or stated closures. They are not copies of Pulse.

B. Independently engineered integration. Pulse samples are mapped onto the Jones circulation without adding organs Pulse does not have. Drug mass is owned by the PBPK state, so Pulse clearance and local clearance are not both applied. Spatial samples return venous drug to that same state. One integer clock aligns the knots.

C. Demonstrated improvements. Donor-limited renal exchange matches a small Euler step and refuses a negative store on a long step. Closed-form transit stays non-negative where a 3600 s Euler step does not. The matrix exponential is stable where a 30 s hepatic Euler step is not. Run residuals for the teaching-compound dose close to roundoff on the recorded protocols. `course-correction-1` shows that a 200 mL venous removal lowers independent-engine output, pressure, and filtration together. That last result is an experimental-backend result, not the portfolio default.

D. Not established as novel. RK4, the matrix exponential, donor limiting, finite volumes, Murray’s cube law, and polar meshing are standard. No literature review in this repository claims a new physiological law. The independent heart’s low resting output and weak exercise response are recorded limitations, not tuned away.

E. External pieces. Pulse 4.4.0 (Apache-2.0) supplies whole-body physiology when requested. BodyParts3D 4.0 (CC BY-SA 2.1 Japan) supplies the meshes. Three.js (MIT) draws them. NumPy and SciPy support the arrays.

## Representative checks

The recorded independent comparisons live in `docs/ALGORITHM_RESEARCH_LOG.md` and the JSON reports under `results/independent/`. They include negative results: predictor-corrector exchange, step doubling, conductance controllers that drop mean pressure below 60 mmHg, and a frozen-flow hold whose urine error was 5.1 percent. Those candidates were not adopted as defaults.

## Reproducibility

See `docs/METHODS.md` for the test and server commands. Each workbench run stores `config.json`, `manifest.json`, and `summary.json`. Export from the workbench writes a short manifest of the loaded visualization. It does not contain a filesystem path and it does not rerun the solver.

## Later research

Central-command gains, a diastolic curve identified for the independent circuit, an interstitial refill time, and organ surfaces generated from published lengths rather than BodyParts3D vertices. Those are research items. They are not required to use the portfolio workbench.