This project analyzes SitSense's posture and session datasets to uncover patterns in posture behavior, evaluate posture metrics, and generate actionable insights for improving the product experience.
The analysis includes:
- Data cleaning & preprocessing
- Metrics parsing (7 posture metrics)
- Session duration exploration
- Posture classification behavior
- Correlation analysis of metrics
- Visualization & insight generation
Key steps performed:
- Combined multiple CSV files
- Trimmed whitespace, normalized IDs
- Converted timestamps to datetime format
- Parsed metrics JSON strings into numeric columns
- Dropped invalid or malformed rows
- Validated session durations
- Joined posture + session datasets on
(user_id, session_id)
- Good posture: 70.7%
- Fair posture: 20.7%
- Poor posture: 8.6%
Users mostly maintain “good” posture, but “poor” posture sessions show extremely strong metric deviations.
- Mean duration: 58.1 minutes
- Median duration: 75.6 minutes
- 99th percentile: ~85 minutes
- Long sessions show more posture drift and angle deviations.
1. CVA Angle (Chin–Vertical Angle)
- Strongest indicator of posture
- Lower CVA = significantly poorer posture
- Excellent separation between good → fair → poor
2. Head Forward
- Increases sharply in poor posture
- Strong correlation with neck angle (0.59)
3. Head Tilt
- Mostly near zero
- Larger positive/negative tilt appears in poor posture
4. Neck Angle
- Increases dramatically in poor posture
- One of the top predictors for posture label
5. Neck Turn
- Mostly centered on 0°
- Large rotations appear in poor posture only
6. Shoulder Slope
- Very strong increase in poor posture cases
- Correlates with trunk angle (0.55)
7. Trunk Angle
- Sharp rise in poor posture
- Indicates full upper-body lean, not just head movement
The posture dataset includes seven key angular metrics extracted from SitSense’s posture-tracking pipeline. The correlation matrix reveals how these biomechanical measurements interact with one another and jointly define posture quality (Good, Fair, Poor). Understanding these relationships is essential for:
- Building robust posture scoring rules
- Identifying redundant metrics
- Improving model features for ML posture classification
- Detecting compensatory movements (e.g., neck angle ↑ when head forward ↑)
Below is an in-depth interpretation of the correlations and what they imply about real posture behavior.
This is the most meaningful relationship in the dataset.
As a user’s head moves forward (greater forward drift), the neck angle increases sharply.
This aligns perfectly with real biomechanics:
- When the chin moves forward → the cervical spine must flex → neck angle increases.
- Poor posture amplifies this coupling dramatically.
➡ Interpretation:
These two together measure the classic “tech-neck / forward-head posture,” a major posture risk indicator.
When the shoulders slope downward (slouching), the trunk angle also increases.
This reflects upper-body collapse into forward flexion.
➡ Interpretation:
This pair captures full-torso slouch rather than neck-only posture.
It is highly relevant for long-session fatigue detection.
Slouching shoulders often induce compensatory neck lift or tilt.
➡ Interpretation:
This indicates multi-segment “chain reactions” during poor posture —
when one part deteriorates, adjacent segments adjust unnaturally.
CVA is the strongest overall posture quality indicator.
Higher CVA = upright head.
Lower CVA = forward head posture.
CVA correlates negatively with:
- m_neck_angle (-0.35)
- m_trunk_angle (-0.31)
- m_head_forward (-0.27)
- m_shoulder_slope (-0.29)
➡ Interpretation:
As posture worsens, CVA sharply decreases while all other angles rise.
This validates CVA as a reliable anchor feature for scoring overall posture.
This also proves that:
Good posture = high CVA + low angles across the board
Poor posture = low CVA + spikes in neck, head, trunk, shoulder metrics
Metrics such as:
- m_head_tilt
- m_neck_turn
show very low correlations with the core slouch metrics.
➡ Interpretation:
These represent lateral motion rather than forward/backward posture.
They are useful for detecting:
- head tilt habits,
- asymmetry,
- side-leaning behavior,
but do not directly correspond with forward-slouch patterns.
They are complementary — not redundant.
The correlation structure reveals THREE major posture behavior clusters:
- CVA (negative)
- Head forward
- Neck angle
This cluster defines tech-neck posture, the most common poor posture form.
- Trunk angle
- Shoulder slope
- Neck angle (secondary)
This cluster indicates fatigue-driven collapse seen in long sessions (75+ minutes).
- Head tilt
- Neck turn
Independent of posture quality but useful for ergonomic asymmetry and balance.
-
CVA should be central to posture scoring.
It provides the cleanest separation across posture labels and aligns with ergonomics research. -
Neck Angle + Head Forward = top predictors for poor posture.
These two should trigger early posture warnings. -
Shoulder Slope + Trunk Angle detect prolonged-session slouching.
Useful for session-based posture fatigue indicators. -
Low-correlation metrics add dimensionality.
Tilt/turn can detect habitual asymmetries or workspace layout problems. -
Feature redundancy is low.
All seven metrics provide unique information — beneficial for future ML models.
- Combine CVA + Neck Angle + Head Forward as a composite “Forward Posture Index.”
- Use Shoulder Slope + Trunk Angle for “Torso Collapse Index.”
- Keep tilt/turn as secondary features for classifying posture subtypes.
This will dramatically stabilize classification accuracy and interpretability.
Based on the analysis:
- Prioritize CVA, head-forward, neck angle, shoulder slope for posture scoring
- Implement threshold-based alerts for sharp metric spikes
- Personalize posture scoring per user baseline
- Smooth metrics over time to reduce noise
- Build a simple ML classifier to automate posture classification
- Build a posture classification model (Good/Fair/Poor)
- Add time-series smoothing
- Explore user-level posture trends
- Build a Streamlit dashboard for real-time insights
Niranjan Shankar
UC Davis — Mathematics & Statistics
Aspiring Data Scientist | Actuarial Analyst | ML Enthusiast
LinkedIn: www.linkedin.com/in/neo2604
GitHub: https://github.com/StrangeStorm243-bit