After studying this material, students should be able to:

1. Explain why BMI is insufficient for cardiometabolic risk assessment
2. Understand key body composition biomarkers and their metabolic significance
3. Describe how PhotoScan works and what problem it solves
4. Interpret model performance metrics (AUROC, NRI, MAE)
5. Evaluate the clinical trade-offs between DXA, BIA, and smartphone-based approaches

---

Step-by-Step Teaching Guide

STEP 1: Why BMI Falls Short

The Core Problem

BMI is calculated simply as:

$BMI = \frac{weight(kg)}{height(m)^2}$

The critical limitation: BMI tells you how much mass you carry, but not where it is stored or what type it is

Example: Two people can share identical BMI scores yet have dramatically different metabolic health — one carrying mostly muscle, the other carrying dangerous internal fat

What BMI Misses

---

STEP 2: Understanding Insulin Resistance — The Hidden Threat

What Is Insulin Resistance?

Think of insulin as a key and your cells as locked doors

• Normal state: Insulin key opens cell doors → glucose enters → blood sugar stays balanced
• Insulin resistant state: Keys stop working properly → glucose builds up in blood → pancreas produces more keys → system eventually breaks down

Why It Matters Clinically

Insulin Resistance develops silently
         ↓
Years pass with no obvious symptoms
         ↓
Vascular damage begins
         ↓
Liver function impairs
         ↓
Energy metabolism disrupts
         ↓
Eventually: Type 2 Diabetes diagnosis

Key insight: Damage occurs long before blood sugar reaches diagnostic thresholds

Measuring Insulin Resistance: HOMA-IR

HOMA-IR = Homeostasis Model Assessment for Insulin Resistance

It models the feedback loop between:

• Liver glucose production
• Insulin secretion
• Under fasting/steady-state conditions

---

STEP 3: Body Composition Biomarkers That Actually Matter

Beyond Basic Body Fat Percentage

Body fat percentage (BF%) gives a useful baseline but lacks precision. Two additional ratios provide much deeper clinical insight:

Biomarker 1: Android-to-Gynoid Ratio (A/G Ratio)

Visual concept:

Android (Apple Shape)        Gynoid (Pear Shape)
     ●●●●●                        ●
    ●●●●●●●                      ●●●
   ●●●●●●●●●                    ●●●●●
      |||                       ●●●●●●●
      |||                         |||

• Android fat = stored in trunk/abdomen
• Gynoid fat = stored in hips and thighs
• High A/G ratio = more apple-shaped = higher metabolic risk

Biomarker 2: Visceral-to-Subcutaneous Ratio (V/S Ratio)

Cross-section of abdomen:

    [Skin]
    [Subcutaneous fat layer] ← measured by V/S
    [Muscle wall]
    [Visceral fat] ← surrounds organs, highly metabolic
    [Organs]

Key clinical fact: Elevated A/G ratios AND higher visceral fat mass both strongly correlate with insulin resistance

---

STEP 4: Current Gold Standard and Its Limitations

DXA Scans — The Best Tool We Have

DXA = Dual-Energy X-Ray Absorptiometry

How it works: Two X-ray beams at different energy levels pass through the body, and different tissues absorb energy differently, allowing precise composition mapping

Strengths:

• Extremely precise measurements
• Measures BF%, A/G ratio, V/S ratio accurately
• Clinical gold standard

Critical weaknesses:

The gap: The most accurate tool is the least accessible — exactly when early detection matters most

---

STEP 5: Introducing PhotoScan — The Proposed Solution

Core Concept

PhotoScan is a deep learning framework that extracts body composition metrics from standard 2D smartphone photos

The fundamental insight: Your body's external geometry contains information about internal fat distribution

How PhotoScan Was Built

Phase 1 — Pre-training:

35,000+ UK Biobank participant records
              ↓
    Deep Neural Network learns
    body geometry → composition
    relationships at scale

Phase 2 — Fine-tuning:

677 diverse adults (PhotoBIA cohort)
              ↓
    Model adapts to real-world
    smartphone photo conditions

Phase 3 — Validation:

Independent MetabolicMosaic cohort
              ↓
    Unbiased performance testing
    on completely unseen data

What PhotoScan Outputs

Smartphone Photo Input
        ↓
[Deep Neural Network]
        ↓
┌─────────────────────────┐
│ Body Fat Percentage (BF%)│
│ A/G Ratio               │
│ V/S Ratio               │
└─────────────────────────┘
        ↓
Combined with demographics
        ↓
Insulin Resistance Risk Classification

---

STEP 6: Understanding the Performance Metrics

Metric 1: Mean Absolute Error (MAE)

Simple definition: On average, how far off are the predictions from the true values?

$MAE = \frac{1}{n}\sum|predicted - actual|$

Lower MAE = Better performance

Metric 2: AUROC (Area Under ROC Curve)

Simple definition: How well can the model distinguish between someone WITH insulin resistance versus someone WITHOUT it?

AUROC Scale:
0.5 ──────────────────── 1.0
 |                         |
Random               Perfect
guessing          discrimination

Think of it as: If you randomly picked one insulin-resistant person and one healthy person, AUROC tells you the probability the model correctly identifies which is which

Metric 3: NRI (Net Reclassification Index)

Simple definition: Compared to the baseline model, how many people are now correctly categorized who weren't before?

• Positive NRI = improvement over baseline
• Higher NRI = more people correctly reclassified

---

STEP 7: Comparing All Approaches — The Key Results

Performance Comparison Table

Visualizing the Performance Gap

AUROC Performance:

Demographics  |████████████████████░░░░░░░| 0.692
Demo + BIA    |████████████████████░░░░░░░| 0.692  ← No improvement
Demo+PhotoScan|██████████████████████░░░░░| 0.760  ← Significant jump
Demo + DXA    |███████████████████████░░░░| 0.773  ← Gold standard

                                          Near-DXA accuracy!

Why Did BIA Fail to Improve Results?

This is a critical insight:

BIA provides → BF% only
                  ↓
BF% has LOW feature importance
for insulin resistance prediction
                  ↓
Adding BIA = adding low-value information
                  ↓
No improvement in classification

PhotoScan provides → BF% + A/G ratio + V/S ratio
                              ↓
A/G and V/S have HIGH feature importance
for insulin resistance prediction
                              ↓
Adding PhotoScan = adding high-value information
                              ↓
Significant improvement in classification

The lesson: What you measure matters more than how precisely you measure the wrong thing

---

STEP 8: Synthesizing the Clinical Value Proposition

The Three-Way Trade-off

                    ACCURACY
                       ▲
                       │
                  DXA  ●
                       │
           PhotoScan   ●
                       │
                  BIA  ●
                       │
            Demo only  ●
                       │
─────────────────────────────────────────► ACCESSIBILITY
Low                                    High

Positioning PhotoScan

---

STEP 9: Limitations and Future Directions

Current Limitations to Understand

1. Research prototype status

• Not yet validated for clinical use
• Requires further large-scale testing

2. Cohort composition effects

• MetabolicMosaic had 67% female vs. PhotoBIA's 57% female
• Females generally show lower A/G and V/S ratios due to gynoid fat storage patterns
• This affects prediction error metrics across cohorts

3. Body composition is one piece

• Cardiometabolic health requires multiple signals
• No single metric tells the complete story

The Multimodal Future

Future Integration Vision:

PhotoScan          Wearables           Blood Biomarkers
(Body composition) (Continuous data)   (Clinical labs)
      │                  │                    │
      └──────────────────┴────────────────────┘
                         │
              Holistic Metabolic
              Health Assessment

---

Summary: Key Takeaways

Central insight: The future of metabolic health screening lies not in more expensive clinical tools, but in extracting richer, more clinically meaningful signals from technology people already carry in their pockets