nutramilo-nis 1.1.6

Nutramilo Insulin Score (NIS) — open algorithm for predicting postprandial insulin response from macronutrient composition.

nutramilo-nis

Reference implementation of the Nutramilo Insulin Score (NIS™) — an open-source algorithm designed to estimate the relative insulinogenic potential of a meal from its nutritional composition. NIS is intended for research, educational use, software development, and exploratory nutritional analysis. It should not be interpreted as a clinically validated predictor of postprandial insulin response. ⚠️ Scientific status. NIS v1.1.6 represents an exploratory methodology. The algorithm is intended for comparative analysis of meals and hypothesis generation rather than diagnostic or therapeutic decision-making.

---

✨ Why NIS?

Property	Holt Food Insulin Index (FII)	NIS v1.1.6
In-vivo insulin AUC measurement required	✅ Yes	❌ No — composition only
Suitable for mobile apps & CGM-less use	❌ No	✅ Yes
Frozen, versioned coefficients	❌ No	✅ Yes (Apache 2.0, since v1.0)
Mandatory citation in derivatives	❌ No	✅ Yes (NOTICE clause)
Open dataset (SHA-256-verified)	❌ No	✅ Yes (Zenodo on release)
Out-of-sample evaluation reported	Partial	✅ Yes (n = 25 hold-out)
Externally validated on fresh CGM cohort	n/a	❌ Not yet — see technical documentation

Scientific Status

NIS has been evaluated during development using internal validation procedures and independent test subsets.

The methodology is intended to provide a transparent, reproducible, and open framework for estimating relative meal insulinogenicity from nutritional composition.

Users should interpret outputs as comparative estimates rather than direct physiological measurements.

Future independent validation studies may further clarify the strengths, limitations, and potential applications of the methodology.

📦 Install

pip install nutramilo-nis

Optional extras:

pip install "nutramilo-nis[validation]"   # numpy / pandas / scipy for re-running the regression notebook
pip install "nutramilo-nis[dev]"          # pytest, ruff, coverage

🚀 Quick start — 30 seconds

from nutramilo_nis import compute_nis

# A balanced lunch: salmon + quinoa + broccoli
result = compute_nis(
    carbs_g=35,
    protein_g=30,
    fat_g=15,
    fiber_g=8,
)

print(result.nis_percent)   # → 55.4
print(result.tier)          # → "Medium"
print(result.tier_color)    # → "#F59E0B"
print(result.contributions) # → {"carbs": 26.19, "protein": 11.93, "fat": 10.84, "fiber": -5.49}
print(result.citation)      # → "Inkov I. (2026). NIS …"

JSON-serialisable for APIs:

import json
print(json.dumps(result.to_dict(), indent=2))

🧪 Worked examples

Computed with default mixed source-type and calibration on (slope = 0.7275, intercept = 8.111). See tests/test_nis_v1_1_*.py for the full pinned numerical regression suite.

Plate	C / P / F / Fb (g)	NIS %	Tier
White bread, lean cheese, butter	50 / 12 / 8 / 1	77.3	High
Salmon + quinoa + broccoli	35 / 30 / 15 / 8	55.4	Medium
Plain Greek yogurt + walnuts	8 / 18 / 14 / 2	50.6	Medium
Black bean & spinach bowl	30 / 14 / 6 / 13	44.3	Moderate
Avocado-omelette + side salad	6 / 18 / 22 / 5	42.0	Moderate

📐 Algorithm (5-step pipeline, v1.1.6)

1. Per-1000-kJ normalisation using Atwater factors (17·C + 17·P + 37·F kJ/g).
2. Linear regression layer — independent OLS on the frozen HoltBellBao_v1_frozen_2026.csv cohort (n = 147 foods). Coefficients exposed as NIS_COEFFICIENTS: carbs = +1.61 · protein = +0.66 · fat = +1.20 · fiber = −1.14 (per 1000 kJ). Protein and fat scaled by source-type multipliers (plant/mixed/animal: protein 0.55/1.00/1.30, fat 0.45/1.00/1.25).
3. Pure-fat extrapolation guard — when (net-carb + protein) / 1000 kJ < 5 g the regression output is scaled by 0.15× (energy-normalised, serving- invariant since v1.1.1).
4. Cross-cohort linear calibration — obs ≈ 0.7275 × pct + 8.111 (fit on the n = 63 cross-cohort validation set; opt-out via apply_calibration=False).
5. Optional Holt cross-track + clinical-tier floor, then Final NIS = max(regression_cal, holt, floor), clamped [0, 100].

Note on validation framing. The n = 147 training cohort and the n = 63 cross-cohort calibration set share Holt 1997 foods (derivational overlap, see manuscript §4.3 L1). All inferential claims in v1.1.6 are therefore restricted to the n = 25 strictly out-of-sample subset (Bao 2011, Nilsson 2004, Boirie 1997, Trichopoulou 2003, Sahyoun 2008).

📘 API reference

compute_nis(carbs_g, protein_g, fat_g, fiber_g=0.0, *, insulin_impact_tier="low", il_total_g=None, fat_type="mixed", protein_type="mixed", sugar_g=0.0, apply_calibration=True) → NisResult

Argument	Type	Default	Description
carbs_g	float	—	Total carbohydrate grams (net carb = carbs_g − fiber_g)
protein_g	float	—	Protein grams
fat_g	float	—	Total fat grams
fiber_g	float	0.0	Dietary fibre grams (blunts response)
insulin_impact_tier	Literal["low","medium","high"]	"low"	Clinical floor band (paper §2.4)
il_total_g	Optional[float]	None	Pre-computed Insulin Load grams — activates the Holt cross-track
fat_type	Literal["plant","mixed","animal"]	"mixed"	Fat-source multiplier — plant 0.45, mixed 1.00, animal 1.25 (Holt 1997 / Trichopoulou 2003)
protein_type	Literal["plant","mixed","animal"]	"mixed"	Protein-source multiplier — plant 0.55, mixed 1.00, animal 1.30 (Nilsson 2004 / Boirie 1997)
sugar_g	float	0.0	Total sugars; when supplied, ~50 % is treated as fructose (×0.30 insulinogenicity per Le 2008)
apply_calibration	bool	True	Apply cross-cohort linear calibration (slope 0.7275, intercept 8.111). Set False to recover raw v1.0.1 regression output

NisResult (frozen dataclass)

Field	Type	Description
nis_percent	float	Final NIS, 0–100
tier	str	"Low" / "Moderate" / "Medium" / "High" / "Very high"
tier_color	str	Hex colour for UI (Tailwind-compatible)
piru	float	Predicted Insulin Response Unit (regression layer raw)
regression_pct	float	Pure regression output (0–100)
holt_pct	float	Holt cross-track output (0–100)
tier_floor_pct	float	Clinical-tier floor applied
contributions	dict[str, float]	Per-macro absolute contribution to PIRU
nis_version	str	Frozen version (e.g. "1.0.0")
coefficients_date	str	Date coefficients were regressed (ISO)
citation	str	Required academic citation

Use .to_dict() → JSON-serialisable.

Reproducibility

Artifact	Availability
Source code	Included in this repository
Unit tests	Included in the package
Coefficient definitions	Publicly documented in source code
Example calculations	Included in README and tests
Version history	Available through Git tags and releases

All calculations are fully reproducible from the published source code and released package versions.

™ Trademark Notice

NIS™, Nutramilo Insulin Score™, and Nutramilo™ may be used to identify the project and associated materials.

The Apache 2.0 license applies to the software source code.

Commercial use of names, logos, and branding elements may be subject to separate intellectual property rights where applicable.

⚖️ Disclaimer & regulatory framing

NIS is intended for educational and research purposes. It is not a medical device, does not diagnose any disease, and should not replace professional clinical judgement.

🏗️ Roadmap (semver-frozen)

• v1.x — patch fixes only; coefficients never change.
• v2.0.x — prospective CGM trial re-regression.

| GDPR | No personal data processed by the library itself |

🤝 NIS vs DIL — different questions, different units

NIS and DIL are complementary, not interchangeable:

	NIS (per-meal)	DIL (per-meal & daily)
Question	"How insulinogenic is THIS meal compared to pure glucose?"	"How many grams of carb-equivalent insulin load did I eat today?"
Output	0–100 % (intensity score)	grams of carb-equivalent (load)
Use-case	Compare two meals; rank foods on a phone; coach single-meal swaps	Daily budget tracking; weekly trend monitoring; clinical dosing-adjacent reasoning
Formula	5-step pipeline with source-aware multipliers, pure-fat guard, cross-cohort calibration	0.69 × protein + max(carbs − fiber, 0) per food, summed across the day
Validation	n = 25 OOS Pearson r = 0.83, MAE = 10.2	OLS-stable across Ridge/LASSO; food-level coefficient frozen at 0.69 since 2026-02-10
In this SDK	compute_nis(...)	not bundled — see nutramilo-health package or services/dil_formula.py in the Nutramilo platform

If you need both layers (per-meal NIS for ranking + daily DIL for budget), combine compute_nis here with the open compute_dil formula: DIL_g = max(0, 0.69·protein + max(0, carbs − fiber)).

📜 Licence

Apache 2.0 — see NOTICE for the mandatory citation clause.

💬 Contact

For	Email
Research collaboration	[e-mail]
Trademark / commercial licensing	[e-mail]
Bug reports & feature requests	github.com/nutramilo/nutramilo-nis/issues
General product info	https://nutramilo.com/science