Face → Health

Vision as a longitudinal wellness signal inside AI‑Steve, with strict temporal alignment and evidence‑first modeling.
Overview Pipeline Temporal Alignment ML Direction Links

Why Face → Health Exists

The face is treated as a sensor, not a narrative. Face → Health does not diagnose or infer emotion; it evaluates weak visual signals only through longitudinal correlation. The system supports two aligned modes: retrospective (face today → last night’s sleep) and predictive (face today → tonight’s sleep). Together, they provide both diagnostic context and forward‑looking insight.

Temporal Correctness Face Embeddings Health Linkage Evidence‑First

What’s Implemented in Production

Health Export Faces

Daily face images are detected and stored with strict date normalization. Each image receives a health‑specific vision analysis focused on alertness, tension, hydration, and stress markers.

Dual Embeddings

Every face is embedded twice: CLIP embeddings for similarity search and 128D face embeddings for identity‑specific retrieval.

Health Linkage

A dedicated health record date links every face to Apple Health daily records, powering a materialized view for ML and correlation analysis.

Face to health visual signal

Face images become structured wellness signals tied to daily health data.

Face health ML overview

Embeddings + health metrics flow into interpretable models.

Face health correlation view

Correlation and feature‑importance reporting over time.

Face health capture app

Health Companion Export app (in review). App Store listing placeholder: ShareSteps ID 6738940089.

Temporal Alignment (Critical)

Model A (Retrospective): Face on Day N → sleep from Night N‑1→N (recorded on Day N). Model B (Predictive): Face on Day N → sleep from Night N→N+1 (recorded on Day N+1).

Both modes are supported with explicit alignment. No inferred dates, no silent shifts. This keeps the dataset defensible as it grows.

Machine Learning Direction

Two complementary models run in parallel. Retrospective explains current appearance using last night’s sleep. Predictive estimates tonight’s sleep from today’s face image and same‑day activity metrics. Early models prioritize interpretability and temporal integrity over raw accuracy.

Retrospective (Diagnostic)

Face today → last night’s sleep. Useful for explaining current appearance and recovery signals.

Predictive (Forecast)

Face today → tonight’s sleep. Uses same‑day activity + embeddings with next‑day labels.

Generalization First

Early results show training fit > validation; the goal is clean data and defensible generalization.

Correlation is always reported as correlation. Causation is never implied.