A pediatric dose, extrapolated from adult pharmacokinetics — with a cited, graded, auditable rationale.
🔗 Live at paedscale.by-dhruvik.in — built at the 2026 Claude: Life Sciences Hackathon.
PaedScale derives a defensible starting dose for a child from published adult PK, using allometry × organ maturation (Anderson–Holford), and returns a graded A→D recommendation with the reasoning, the guideline concordance, and every uncertainty made explicit — not a black-box number.
Medication errors harm an estimated 1.5 million people a year in the United States alone, and children are disproportionately exposed — because nearly every pediatric dose is a weight-based calculation done by hand, for a body whose organs are still maturing.
Sources: Institute of Medicine, Preventing Medication Errors (2006); Kaushal et al., JAMA 2001;285:2114 (pediatric medication-error rate).
A real graded run: clindamycin, 6 yr — the estimate, the grade, the guideline concordance (1.14×), and the flags, all in one auditable view.
Drug clearance does not scale linearly with body size in early life: the organs that eliminate the drug are still maturing, so scaling an adult dose by weight over-doses the young. Published pediatric labels cover the common drug × age combinations — but they run out for neonates, off-label drugs, organ impairment, and narrow-therapeutic-index agents, where clinicians today extrapolate by hand from sparse references.
PaedScale encodes the distance between the naïve weight-linear line and the true maturation curve, and — the whole point — maps each drug to its elimination pathway so a single deterministic engine extrapolates across the drug space without a hardcoded per-drug lookup table.
- Multi-agent by design — an Opus orchestrator (
agents/agent.py) drives a cheaper Sonnet retrieval subagent (retrieval/) that gathers cited adult PK live from PubMed + openFDA. - Python does the arithmetic; Claude does the judgment — a deterministic engine does the math; the model maps drug → pathway → maturation curve and writes the justification. The mapping across the drug space is what a hardcoded calculator cannot do.
- Cite-or-abstain, enforced in code — the engine raises rather than invent a maturation curve for an unknown pathway, and flags unattributed clearance instead of hiding it. No unsourced PK value drives a confident point estimate — it grades D and abstains instead.
- Organ function, from real inputs — renal impairment derives its clearance modifier from a bedside Schwartz eGFR (serum creatinine + height), not a flat guess; hepatic adjustment is surfaced as a drug-specific note rather than a silent reduction.
- MCP server (
api/mcp_server.py) exposes the same retrieval tools over FastMCP (stdio).
The differentiator isn't a confident number; it's a number that knows its own limits.
| Confident & concordant | Honest under uncertainty |
|---|---|
|
 |
| Clindamycin — clean CYP3A4 pathway, concordant with the guideline (1.14×) → grade A. | Neonatal vancomycin — a narrow-TI renal drug where allometry under-predicts; it flags the discordance, recommends TDM, anchors to the guideline, and caps the grade at C. |
Per elimination pathway:
CL_child = CL_adult × (WT/70)^0.75 × MF(PMA) × OF
MF(PMA) = PMA^H / (TM50^H + PMA^H) # maturation, normalised so adult ≈ 1
WT weight (kg) · PMA postmenstrual age (weeks) · TM50 age at 50% maturation · H Hill coefficient ·
OF organ-function modifier (renal from Schwartz eGFR / hepatic). Vd scales linearly; the dose is solved
by the effect-driving metric (css/auc for maintenance, cmax for peaks, time_mic flagged as a proxy
for β-lactams). Oral doses are corrected by bioavailability F; toxic/effective bounds fire a prominent
safety warning.
Allometric ¾-power scaling and sigmoidal maturation: Anderson & Holford, Annu. Rev. Pharmacol. Toxicol. 2008;48:303. GFR maturation model: Rhodin et al., Pediatr. Nephrol. 2009;24:67. Bedside eGFR: Schwartz et al., J. Am. Soc. Nephrol. 2009;20:629.
| Grade | Meaning |
|---|---|
| A | Passes concordance vs a real published guideline |
| B | Solid PK / guideline-anchored, honestly caveated |
| C | Sparse / uncertain — directional only |
| D | Insufficient data or a safety stop (dose withheld) |
Concordance = the recommended dose ÷ a guideline dose, reported as a ratio inside a strict 0.67×–1.5× band (and a wider clinical 0.5×–2× band).
The deterministic engine is unit-tested and reproducible with no API key
(python3 -m tests.test_pk): maturation monotonicity, cite-or-abstain, safety hard-stops, oral-F,
time>MIC, LRU cache, and the mechanism scorer.
frontend/index.html self-contained UI (form → /calculate/stream → graded result + chat)
backend/ run all commands from here (backend/ is the package root)
engine/ deterministic backbone — no I/O, no LLM
constants.py MATURATION params (TM50/Hill) ONLY. NO per-drug PK.
pk_engine.py allometry × maturation, dose solve, oral-F, safety, Schwartz eGFR → OF
edge_cases.py deterministic flags: prodrug / obesity / protein-binding / illness
pk_cache.py bounded in-process LRU for live dossiers
mechanism_score.py mechanistic-reasoning scorer (6 dimensions)
retrieval/ retrieval subagent + tools (import retrieval → __init__.py)
__init__.py RETRIEVAL SUBAGENT (Sonnet) → cited dossier, cache hit, or abstain
retrieval_tools.py httpx: PubMed E-utilities + openFDA + web_fetch
agents/agent.py Opus ORCHESTRATOR: load_skill → retrieve → compute → edge_cases → grade
api/main.py FastAPI: /, /calculate, /calculate/stream, /chat, /pk, /health
api/mcp_server.py MCP server for the same retrieval tools (FastMCP, stdio)
tests/ test_pk.py (no key) / test_agent.py (key + network)
skills/ lean markdown skills loaded on demand
eval_data/ ANSWER KEYS ONLY — harness, never the product path
No hardcoded per-drug PK in the product path. The agent retrieves adult PK live (PubMed + openFDA),
serves a TTL-bounded cache hit, or abstains (grade D). eval_data/ is harness-only.
Hosted: the live app is at paedscale.by-dhruvik.in. To run locally:
cd backend # backend/ is the package root — run everything from here
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python3 -m tests.test_pk # deterministic core + scorers — NO key needed
cp .env.example .env # ANTHROPIC_API_KEY (+ optional OPENFDA_API_KEY / NCBI_API_KEY)
uvicorn api.main:app --reload --port 8000
# open http://localhost:8000
python3 -m tests.test_agent # end-to-end eval (needs key + network)
python3 -m api.mcp_server # optional: the retrieval MCP server (stdio)/pk and tests/test_pk run without a key. /calculate, the orchestrator and the retrieval subagent
need the key and network (live PubMed/openFDA); expect ~60–95 s/query for the full live path.
From one maturation curve per pathway to a physiologically-based PK (PBPK) "digital twin" of the child — validated against real clinical concentrations, then personalised to the patient's genotype and measured organ function. This is the path from careful extrapolation to individualised pediatric dosing.
PaedScale is a research prototype. Its output is a defensible starting estimate for a qualified clinician — not an autonomous order, and not a validated clinical-decision-support device. Narrow-therapeutic-index drugs require therapeutic drug monitoring.
MIT.