Scoring Algorithm

Overview

The Mind Measure scoring algorithm fuses multimodal inputs (text, audio, visual) into a single wellness score (0-100). This document describes the V2 scoring algorithm implemented in December 2025.

Score Range

Score	Interpretation
80-100	Excellent wellbeing
60-79	Good wellbeing
40-59	Moderate wellbeing
20-39	Low wellbeing
0-19	Very low wellbeing (clinical attention recommended)

V2 Scoring Weights

Daily Check-ins

Check-ins use text-heavy weighting because natural conversation is the primary signal:

// All modalities available
finalScore = (textScore * 0.70) + (audioScore * 0.15) + (visualScore * 0.15);
 
// Audio only (visual failed)
finalScore = (textScore * 0.80) + (audioScore * 0.20);
 
// Visual only (audio failed)
finalScore = (textScore * 0.80) + (visualScore * 0.20);
 
// Text only (both failed)
finalScore = textScore;

Baseline Assessments

Baselines use clinical-heavy weighting because structured questions provide validated data:

// All modalities available
finalScore = (clinicalScore * 0.70) + (audioScore * 0.15) + (visualScore * 0.15);
 
// Visual failed
finalScore = (clinicalScore * 0.85) + (audioScore * 0.15);
 
// Both failed
finalScore = clinicalScore;

Sanity Floor

To prevent obviously positive check-ins from being dragged down by conservative audio/visual scores:

const avgQuality = (audioConfidence + visualConfidence) / 2;
 
const shouldApplyFloor = 
  textScore >= 75 &&           // High text score
  riskLevel === 'none' &&      // No risk detected
  avgQuality >= 0.6;           // Reasonable signal quality
 
if (shouldApplyFloor && finalScore < 60) {
  finalScore = 60;  // Apply floor
  warnings.push('Sanity floor applied');
}

Rationale: If someone verbally expresses feeling great (text score 85) but wasn't smiling at the camera, the final score shouldn't drop below 60.

Example Calculations

Example 1: Positive Check-in

Input:

Text score: 85 (Bedrock detected positive language)
Audio score: 25 (Voice features scored conservatively)
Visual score: 41 (Limited smiling detected)
Risk level: none
Quality: 0.77

Calculation:

rawScore = (85 × 0.70) + (25 × 0.15) + (41 × 0.15)
         = 59.5 + 3.75 + 6.15
         = 69.4

Sanity floor check: text=85 >= 75 ✓, risk=none ✓, quality=0.77 >= 0.6 ✓
Floor applies but rawScore (69) > 60, so no adjustment needed.

Final Score: 69

Example 2: Sanity Floor Applied

Input:

Text score: 80
Audio score: 15
Visual score: 20
Risk level: none
Quality: 0.65

Calculation:

rawScore = (80 × 0.70) + (15 × 0.15) + (20 × 0.15)
         = 56 + 2.25 + 3
         = 61.25 → 61

Sanity floor check: text=80 >= 75 ✓, risk=none ✓, quality=0.65 >= 0.6 ✓
rawScore (61) > 60, so no floor adjustment.

Final Score: 61

Example 3: No Visual Data

Input:

Text score: 70
Audio score: 30
Visual: not available

Calculation:

rawScore = (70 × 0.80) + (30 × 0.20)
         = 56 + 6
         = 62

Final Score: 62

Individual Modality Scoring

Text Score (Bedrock)

Claude 3 Haiku produces a 0-100 score based on:

Sentiment analysis
Linguistic patterns
Wellness indicators
Risk markers

See Text Analysis for details.

Audio Score

Computed from 10 audio features:

function calculateAudioScore(features: AudioFeatures): number {
  const scores: number[] = [];
  
  // Pitch: optimal around 150-180 Hz
  const pitchScore = 100 - Math.abs(features.meanPitch - 165) / 2;
  scores.push(clamp(pitchScore, 0, 100));
  
  // Speaking rate: optimal around 120-150 wpm
  const rateScore = 100 - Math.abs(features.speakingRate - 135) / 1.5;
  scores.push(clamp(rateScore, 0, 100));
  
  // Jitter: lower is better
  scores.push((1 - features.jitter) * 100);
  
  // Shimmer: lower is better
  scores.push((1 - features.shimmer) * 100);
  
  return average(scores);
}

See Audio Features for details.

Visual Score

Computed from 10 visual features:

function calculateVisualScore(features: VisualFeatures): number {
  const scores: number[] = [];
  
  // Smile frequency
  scores.push(features.smileFrequency * 100);
  
  // Eye contact
  scores.push(features.eyeContact * 100);
  
  // Blink rate: optimal around 15-20 bpm
  const blinkScore = 100 - Math.abs(features.blinkRate - 17) * 3;
  scores.push(clamp(blinkScore, 0, 100));
  
  // Affect: -1 to +1 mapped to 0-100
  scores.push((features.affect + 1) * 50);
  
  // Tension: lower is better
  scores.push((1 - features.facialTension) * 100);
  
  return average(scores);
}

See Visual Features for details.

Confidence and Uncertainty

Modality Confidence

Each modality reports a confidence value (0-1):

Modality	Confidence Source
Text	`1 - uncertainty` from Bedrock
Audio	Quality metric from feature extraction
Visual	Face presence quality from Rekognition

Overall Uncertainty

The final uncertainty is the text analysis uncertainty (most reliable indicator):

result.uncertainty = textResult.uncertainty;

Low uncertainty (< 0.3) indicates high confidence in the score.

Calibration Notes

The V2 weights are initial values that may be tuned based on:

User feedback - Does the score feel accurate?
Longitudinal data - Do scores correlate with self-reported outcomes?
Clinical validation - Correlation with PHQ-9, GAD-7, etc.

Tunable Parameters

Parameter	Current Value	Range	Purpose
Text weight (all)	0.70	0.5-0.9	Text influence when all modalities present
Audio weight	0.15	0.05-0.25	Audio influence
Visual weight	0.15	0.05-0.25	Visual influence
Sanity floor threshold	60	50-70	Minimum score for positive check-ins
Quality threshold	0.6	0.5-0.8	Minimum quality for floor to apply
Text score threshold	75	70-85	Minimum text score for floor to apply

Version History

Version	Date	Changes
V1	Nov 2025	Initial: 60/20/20 weighting
V2	Dec 2024	Text-heavy: 70/15/15 + sanity floor

Last Updated: December 2025
Scoring Version: V2

Text Analysis (Bedrock)Testing Setup Guide