SHAP and Feature Importance: A Shift Toward Biometric Signals

Introduction

Effort-based features used to be at the core of my models.
But today, SHAP analysis showed me something different: biometric signals like heart rate and body temperature mattered more.

---

What I Discovered

• SHAP values pointed to heart_rate, body_temperature, and respiration_rate as more impactful than any effort-related metric.
• This insight shifted my focus — I realized my model wasn’t learning from what I assumed mattered.

---

What I Decided to Build

• heart_rate_delta: Difference between max and min heart rate during sessions.
• temp_peak_ratio: Ratio of intervals where temperature exceeded a defined threshold.
• I’m aiming for features that explain physiological responses, not just mechanical effort.

---

Why This Matters

• SHAP gave me interpretability. It helped me trust my model — and question my assumptions.
• The value of a feature isn’t in how intuitive it sounds, but how much it helps the model generalize.

---

What I’ll Try Next

• Use SHAP values earlier in my pipeline for feature selection.
• Compare models trained with vs. without biometric features.
• Publish a notebook that documents this change and result shift.