# MyChild Engine — Raw Study Data

> Full validation dataset for the MyChild Engine verification study: 1,747 curated profiles, 70,192 adversarial stress-test profiles, manifest, and SHA-256 checksums. Open data, CC BY-SA 4.0. All profiles are synthetic — no real children.

Canonical HTML page: https://www.mychildapp.in/study-data (request download links by email there, or browse GitHub: https://github.com/hsongra11/mychild-app/tree/main/docs/research)

## What's in the dataset

- **Curated Profiles** (JSON, 1,747 profiles, ~8 MB) — hand-labeled developmental profiles plus promoted high-confidence adversarial profiles. Each simulates a child's milestone responses across 8 domains (1–24 months) with ground-truth labels and the engine's classification.
- **Adversarial Profiles** (JSON.GZ, 70,192 profiles, ~3.7 MB compressed / ~124 MB uncompressed) — profiles designed to break the engine: regression patterns, sparse milestone data, contradictory answers, preterm edge cases. Includes per-profile disagreement analysis and failure mode classification.
- **Data Manifest** (PDF, 17 KB) — dataset descriptions, file sizes, SHA-256 checksums, data provenance.
- **Complete Package** (ZIP, ~3.7 MB) — everything in one download.

## How the data was generated

1. **Synthetic profile generation** — an AI system generates developmental profiles simulating realistic milestone response patterns: a hypothetical child at a specific age (1–24 months) with answers across 131 evidence-weighted questions spanning 8 domains. Includes both on-track children and simulated delays of varying severity.
2. **Ground-truth labeling** — each profile is independently evaluated by an AI clinical evaluator assigning "no concern," "monitor," or "refer," using CDC 2022 / WHO guidelines through a different reasoning pathway than the engine. Not a gold standard — a licensed clinician would be — but consistent and reproducible.
3. **Engine classification** — the engine processes each profile through its deterministic, inspectable rule-based scoring.
4. **Comparison and analysis** — classifications compared to compute Cohen's kappa, sensitivity, specificity; disagreements analyzed into under-flags, over-flags, and threshold boundary cases.
5. **Adversarial stress testing** — increasingly difficult edge cases: regression after meeting milestones, sparse or contradictory data, preterm-adjusted ages, genuinely ambiguous boundaries. Most of the 781 disagreements come from here.

## Data format

```json
{
  "id": "syn-00042",
  "age_months": 14,
  "preterm_weeks": 0,
  "domains": {
    "gross_motor":     { "score": 0.82, "flags": [] },
    "fine_motor":      { "score": 0.65, "flags": ["pincer_grasp_absent"] },
    "communication":   { "score": 0.91, "flags": [] },
    "cognitive":       { "score": 0.78, "flags": [] },
    "social_emotional": { "score": 0.88, "flags": [] },
    "adaptive":        { "score": 0.73, "flags": [] },
    "sensory":         { "score": 0.95, "flags": [] },
    "feeding":         { "score": 0.80, "flags": [] }
  },
  "engine_result": "monitor",
  "ground_truth": "monitor",
  "agreement": true,
  "generation_type": "adversarial"
}
```

- `engine_result` — engine classification: "no_concern", "monitor", or "refer"
- `ground_truth` — AI clinical evaluator's expected label
- `agreement` — whether they match
- `generation_type` — "hand_verified" (Phase 1) or "adversarial" (Phase 2)
- `domains.*.flags` — specific developmental red flags detected

## License & attribution

- **Data:** CC BY-SA 4.0 — use for research, products, publications; attribute and share alike.
- **Engine:** Apache-2.0.
- **Citation:** Songra, H. & Ansari, A. (2026). "MyChild Engine Validation Study: Synthetic Developmental Screening Data." Available at www.mychildapp.in/study-data.

---

Pages: [Home](/index.md) · [About](/about.md) · [Study](/study.md) · [Study Data](/study-data.md) · [Docs](/docs.md) · [RBSK](/rbsk.md) · [Privacy](/privacy.md) · [Terms](/terms.md) · [Contact](/contact.md)