Jaw tension, clenching, and grinding are common but often ignored — especially among students, remote workers, and people under stress.
These habits can lead to headaches, dizziness, neck pain, ear tension, and temporomandibular joint (TMJ) disorders.

According to clinical research, 5–12% of the population experience TMJ-related symptoms, making it one of the most common orofacial pain conditions.
(American Academy of Family Physicians, 2023)

TMJ Coach is a lightweight AI agent that:

• Watches your jaw via webcam
• Detects clenching, lateral shifting, or grinding in real-time
• Gently nudges you to relax
• Learns from your feedback to adapt over time

• Many users clench or grind unconsciously during stressful or focused tasks.
• This can lead to:
  • Jaw pain and stiffness
  • Headaches or dizziness
  • Ear issues or tinnitus
  • Sleep disruption
• Existing solutions (mouth guards, physiotherapy) are reactive, not preventive.

TMJ Coach addresses this by being proactive and personalized as it learns your relaxed state and provides live feedback before symptoms worsen.

1. Calibration – Learns your relaxed baseline for lip gap and mouth width.
2. Detection – Uses your webcam and MediaPipe FaceMesh to track:
   • Lip gap (vertical jaw position)
   • Mouth width (horizontal tension)
   • Jaw offset (left/right deviation)
   • Lateral motion (side-to-side activity)
3. Classification – Uses heuristics + an online SGDClassifier trained from your feedback.
4. Nudging – Gives you a subtle red overlay and a spoken message when clenching is detected.
5. Learning – Asks, “Were you clenching?” — your yes/no feedback improves its model.
6. Adaptivity – Reinforcement learning (RL) adjusts how often and how quickly it reminds you based on your responses.

Webcam Input
     │
     ▼
MediaPipe FaceMesh
     │
     ▼
Feature Extraction
(lip gap, mouth width, jaw offset, lateral activity)
     │
     ├──► Heuristic Clench Detection
     │
     ▼
Online Classifier (SGD)
     │
     ▼
RL Agent (Adaptive Sensitivity)
     │
     ▼
UI + Voice Feedback
(PyQt6 app + pyttsx3 TTS)

Per frame, the model computes the following features:

# Core per-frame feature extraction
lip_gap_ratio = lip_gap / baseline_lip_gap
mouth_width_ratio = mouth_width / baseline_mouth_width
jaw_offset = (chin_x - nose_x) / mouth_width   # normalized left/right shift
lateral_score = compute_lateral_activity(jaw_history, window_seconds=2.0)

These features describe vertical closure and horizontal motion of the jaw, forming a compact 4D feature vector:

features = [lip_gap_ratio, mouth_width_ratio, jaw_offset, lateral_score]

Before the ML model has enough data, the app uses simple heuristics to detect clenching or grinding:

# Heuristic rules
vertical_clench = (lip_gap_ratio < 0.7) and (mouth_width_ratio > 1.05)
lateral_clench  = (lip_gap_ratio < 0.95) and (abs(jaw_offset) > 0.2)
grinding        = (lateral_score > 0.4)

heuristic_clench = vertical_clench or lateral_clench or grinding

Once you start providing YES/NO feedback, the live classifier replaces these heuristics:

# Online model prediction
is_clenched, p_clench = model_predict_is_clenched(features, fallback_is_clenched=heuristic_clench)

When clenching is detected for longer than a threshold (e.g., 5 seconds), the system issues a nudge:

if is_clenched and clenched_duration > threshold:
    speak("Try relaxing your jaw.")
    show_red_overlay()

Then, TMJ Coach asks for feedback:

# Prompt
print("Were you clenching?")
# User clicks:
#   ✓ YES → label as clenching (1)
#   ✗ NO  → label as relaxed (0)

add_training_example(features, label)

The labeled data updates the classifier immediately through continual learning:

clf.partial_fit(X, y)

To avoid over-notifying, the app uses a lightweight Q-learning agent that dynamically adjusts its sensitivity and cooldown:

# Three action modes
param_sets = [
    {"thresh": 3.0, "cooldown": 60.0},   # aggressive
    {"thresh": 5.0, "cooldown": 90.0},   # neutral
    {"thresh": 8.0, "cooldown": 150.0},  # gentle
]

It learns from your behavior and feedback:

# User feedback buttons
👍 Helpful → reward = +1.0
⚠ Too Much → reward = -1.0
🚫 Can’t    → reward = -1.5

Q-learning update rule:

td_target = reward + gamma * max(Q[next_state])
Q[state, action] += alpha * (td_target - Q[state, action])

Over time, the agent converges toward the best balance between helpful and unobtrusive reminders adapting to the user's unique clenching habits.