BEAM X1 — Metrics Documentation

requestAnimationFrame — runs as fast as the browser can handle. Typically 30–60 FPS (depends on GPU/CPU). Each frame updates all metrics in the object cur.

Each snapshot contains 82 data points: 52 blendshapes + 30 derived metrics (emotions, attention, gaze, fatigue, headpose, distance, blink, smile, iris, gestures, symmetry, ...). Configuration: CONFIG.SNAPSHOT_INTERVAL = 1000

Estimated face distance from camera in centimeters.

detectDistance(landmarks) v camera.js

Head rotation angles in degrees: Pitch (up/down), Yaw (left/right), Roll (side tilt).

detectHeadPose(landmarks, transformMatrix) v camera.js

(mouthSmileLeft + mouthSmileRight) × 50 — direct conversion from blendshapes, range 0–100%.

Red curve in Emotion timeline chart. Also shown in photo modal: 😊X%.

Hand gestures in camera view. 8 recognized gestures: 👍 Thumbs Up, 👎 Thumbs Down, ✌️ Peace, ☝️ Point, ✋ Open Palm, ✊ Fist, 👌 OK, 🤟 Rock.

MediaPipe Hand Landmarker — 21 bodu ruky. Funkce detectGesture(handLandmarks) compares fingertip positions against joints (tip.y vs joint.y).

Each frame (1x/s) is logged — including "None" (hand not visible). This allows measuring how long the hand was visible. "None" entries are filtered for the report.

Percentage of frames where the face was successfully detected. Low values = face out of frame, covered, or poor lighting.

Bio summary: "Detection" X% (count/total). Displayed as first row.

Presence of other people in frame. Second MediaPipe instance (without blendshapes, numFaces: 3) checks every 5 seconds.

Complete log of all 52 blendshapes (muscle values) from MediaPipe. Every 2 seconds a raw snapshot is stored. Used for future detailed analysis and debugging.

eyeBlinkLeft/Right, eyeLookDownLeft/Right, eyeLookInLeft/Right, eyeLookOutLeft/Right, eyeLookUpLeft/Right, eyeSquintLeft/Right, eyeWideLeft/Right, browDownLeft/Right, browInnerUp, browOuterUpLeft/Right, cheekPuff, cheekSquintLeft/Right, jawForward, jawLeft/Right, jawOpen, mouthClose, mouthDimpleLeft/Right, mouthFrownLeft/Right, mouthFunnel, mouthLeft/Right, mouthLowerDownLeft/Right, mouthPressLeft/Right, mouthPucker, mouthRollLower/Upper, mouthShrugLower/Upper, mouthSmileLeft/Right, mouthStretchLeft/Right, mouthUpperUpLeft/Right, noseSneerLeft/Right, _neutral

Not displayed directly — JSON export contains the complete log. Used for JSON export and future AI analysis.

Ze sessionLog eventu session_started (pole sensors) a dopocitano z dat (fallback).

All charts are SVG s viewBox="0 0 100 100" a preserveAspectRatio="none". Data is mapped to 0–100% axes. Curves use vector-effect="non-scaling-stroke" for consistent stroke width. Question markers are dashed vertical lines. Multi-face detections are red zones.

Dynamic text: "During training we captured X biometric snapshots (Y data points), Z photographs, N hand gestures, and M voice comments. All in real-time." Data points count = snapshots × 82.

Analysis is deterministic — no AI generation, no random elements. Every conclusion follows directly from numerical data. Interpretations are formulated as observations, not judgments. Goal: provide a contextual framework for raw data that is replicable and verifiable.

Complete session JSON export containing: gazeAttention[], fatigue[], emotions[], blinkRate[], distance[], headPose[], gestures[], blendshapeLog[], micVolume[], voiceComments[], videoSteps[], reactionTimes[], presenceChecks[], sessionLog[].

Data is segmented by videoSteps[] (for video scenarios) or answers[] (for classic scenarios). Each segment is defined by time range [startedAt, endedAt] and all biometric data is filtered by timestamp into the respective segment.

Summary number 0–100 expressing overall participant engagement during the session. Displayed as a circular indicator in the report.

0.30×0.694 + 0.15×1.0 + 0.20×0.992 + 0.20×1.0 + 0.15×1.0 = 0.208 + 0.15 + 0.198 + 0.20 + 0.15 = 0.906 → zaokrouhleno na 80
Note: in current implementation, the score is set manually based on expert estimation. The formula above is the proposed automation.

Weights are adjustable per scenario. For classic scenarios, add component correctAnswersNorm. For platforms without voice inputs, adjust interactionNorm to use gestures and clicks.

Horizontal mini-bars showing average gaze attention for each video step. Color: green (>70), orange (60-70), yellow (<60). Allows identifying which segment engaged the participant most and least.

segmentAvg = avg(gazeAttention.score WHERE timestamp BETWEEN step.startedAt AND step.endedAt)
For each videoSteps[] element, all records from gazeAttention[] whose timestamp falls within the time range of that step are filtered.

1. Find segment with max(avg) → "Highest attention at [segment name]"
2. Find segment with min(avg) → "Lowest attention at [segment name]"
3. Pokud max - min > 10 → "This suggests the topic of [max segment] was more engaging than [min segment]"
4. Pokud max - min < 5 → "Attention was evenly distributed across all segments"

Bar mini-chart showing average fatigue per segment. Color: green (<45%), orange (45–54%), yellow (≥55%). Below the chart are segment descriptions. Text: trend identification (rising/falling) and exceptions.

segmentFatigueAvg = avg(fatigue.score WHERE timestamp BETWEEN step.startedAt AND step.endedAt)
Identicky filtr as u pozornosti, pouze z pole fatigue[].

1. Calculate firstHalfAvg (average fatigue of first half of segment) and secondHalfAvg
2. Pokud secondHalfAvg - firstHalfAvg > 10 → "Fatigue gradually increased"
3. For each segment: if segAvg[i] < segAvg[i-1] AND segAvg[i] < segAvg[i+1] → "[Segment] re-energized the participant" (local minimum)
4. Najdi min(segAvg) a max(segAvg) and report both with time
5. Global minimum from entire fatigue[] array → calmness peak (Math.min(...fatigue.map(f => f.score)))

Text description identifying WHICH emotions appeared in WHICH segments. Correlation with gestures and other events.

1. For each non-Focused emotion, find segments where it appears → "Emotion [X] appeared exclusively during [segment]"
2. Pokud gesto (napr. Thumbs Up) ma timestamp v rozsahu stejneho segmentu → "This precisely correlates with gesture [Y]"
3. Pokud vsechny segmenty = 100% Focused → "Entire session ran in Focused mode — stable concentration without emotional fluctuations"
4. Dominantni emoce = ta s highestm celkovym poctem
5. "Positive emotions" = Happy, Surprised; "Neutral" = Focused, Neutral, Relaxed; "Negative" = None (not yet supported)

For each gesture from gestures[] (filtr gesture != 'None') find temporal overlap with segments:
gestureSegment = videoSteps.find(s => gesture.timestamp >= s.startedAt && gesture.timestamp <= s.endedAt)
Pokud segment s gestem = segment s ne-Focused emoci → silna korelace, uvest v reportu.

Breakdown of voice answers (tip), their reaction times, and answer consistency analysis. For closing choice, formulation analysis.

1. reactionTimeVariance < 1000 → "Very similar reaction times indicate consistent and active thinking"
2. avgReactionTime < 2000 → "Fast answers — possibly random guessing or certainty"
3. avgReactionTime 3000–7000 → "Active thinking before answer"
4. avgReactionTime > 10000 → „Dlouhe premysleni — slozita otazka nebo nerozhodnost"
5. closingChoice.detectedKeyword != standard words (biometrics/hardware/adaptivity/no) → natural formulation (e.g. "chci do reportu" instead of "ne") → poznamka o porozumeni kontextu
6. closingChoice.attempts > 0 → „[N]x upozorneni pred spravnou detekci"

Summary of physical metrics: distance (range, drift), head movement (average pitch/yaw), blink rate vs norm. Combined interpretation.

1. Vzdalenost: drift > 0 → „Mirny trend oddalovani — prirozeny projev postupne relaxace"
drift < -5 → "Approaching screen — possible eye fatigue or trying to see better"
|drift| < 2 → „Stabilni pozice po celou dobu"

2. Head movement: avg_pitch < 3 AND avg_yaw < 5 → „Minimalni pohyb hlavy — vysoka vizualni fixace"
avg_pitch > 5 OR avg_yaw > 8 → „Zvyseny pohyb — mozna nepohodli nebo rozptyleni"

3. Blink rate: rate < 10 → "Below norm — combined with stable head indicates high visual fixation"
rate < 10 AND avg_pitch < 3 → "Combination of low blink rate and stable head = strong concentration signal"
Poznamka: "Low blink rate may partly be a detection artifact"

Three paragraphs: (1) Participant type, (2) Highlight — most vs least engaging segment, (3) Content recommendations.

1. Calculate per-segment: gazeAvg, emotionVariety (count of unique emotions != Focused), gestureCount, fatigueAvg
2. Segment s max(gazeAvg) AND max(emotionVariety) AND max(gestureCount) → „nejzajimavejsi"
3. Segment s min(gazeAvg) AND max(fatigueAvg) → „nejmene poutavy"
4. Textova sablona: "[Max segment name] triggered the strongest reaction — highest attention, [specifics]. [Min segment name] was conversely the least engaging."

segmentScore = gazeAvg × (1 - fatigueAvg / 100)

Logika:
• Segment s min(segmentScore) → „Zvazit zkraceni nebo oziveni [segmentu]"
• Segment with max(segmentScore) → "[Segment] serves as the main scenario hook"
• Pokud max(segmentScore) / min(segmentScore) > 1.5 → „Vyrazny rozdil v efektivite mezi segmenty"
• Pokud max(segmentScore) / min(segmentScore) < 1.2 → „Segmenty jsou vyrovnane"

All text outputs in the report are generated from these templates and numerical thresholds. This is not free text — every sentence has a clear data basis. If data does not reach the threshold, the corresponding sentence is not shown in the report.

Detect cognitive load levels based on spontaneous blink suppression. Lower blink rate than normal = higher visual engagement.

Magliacano et al. (2020) — „Eye blink rate increases as a function of cognitive load during an auditory oddball paradigm." Neuroscience Letters, Vol. 736, doi:10.1016/j.neulet.2020.135293. Finding: EBR increases during active non-visual tasks vs. rest.

Holland & Tarlow (1972) — „Blinking and mental load." Psychological Reports, 31, 119–127. Finding: blink rate and cognitive load are inversely proportional during visual tasks.

Frontiers in Human Neuroscience (2017) — „What Does Eye-Blink Rate Variability Dynamics Tell Us About Cognitive Performance?" doi:10.3389/fnhum.2017.00620. Finding: BRV dynamics predict cognitive performance.

Nakano et al. (2019) — „Rapid serial blinks: An index of temporally increased cognitive load." PLOS ONE, doi:10.1371/journal.pone.0225897. Finding: rapid serial blinks (RSB) indicate local cognitive load increase.

Norma: spontaneous EBR in adults = ~15–20 blinku/min (Bentivoglio et al., 1997; Ponder & Kennedy, 1927).

suppression = 1 - (measured_blink_rate / baseline_norm)
Where baseline_norm = 17.5 (midpoint of 15–20 range).
Example: 1 - (5 / 17.5) = 0.71 → 71% suppression = high cognitive load.

For each video segment, count blink events (= momenty kdy blinkRate[i].rate > blinkRate[i-1].rate).
segment_blink_events = count(rate_increases) within [stepStart, stepEnd]
Segments with 0 blink events = maximum visual fixation.

Extract patterns from 52 ARKit blendshapes matching known facial expressions per FACS (Ekman & Friesen, 1978). ARKit blendshapes map to FACS Action Units (AU).

Ekman & Friesen (1978) — Facial Action Coding System (FACS). Standard system for describing facial muscle movements.

ARKit → FACS mapovani (Ozel, 2022, facethefacs.com):
browDownLeft/Right → AU4 (Brow Lowerer) — corrugator supercilii
eyeSquintLeft/Right → AU7 (Lid Tightener) — orbicularis oculi, palpebral
cheekSquintLeft/Right → AU6 (Cheek Raiser) — orbicularis oculi, orbital
mouthSmileLeft/Right → AU12 (Lip Corner Puller) — zygomaticus major
mouthFrownLeft/Right → AU15 (Lip Corner Depressor) — depressor anguli oris
mouthPressLeft/Right → AU24 (Lip Pressor) — orbicularis oris
jawOpen → AU26/27 (Jaw Drop) — masseter; internal pterygoid

Blendshape values are 0.0–1.0. Averaging across a segment gives "basal muscle tone". Maximum shows peak intensity. For identifying brief smiles, it suffices that max(mouthSmile) > 0.5 even if avg < 0.05.

Average attention shows "how much", but standard deviation shows "how dynamically" — higher std = participant actively shifts attention = deeper cognitive processing. Low std = monotonous watching.

std = sqrt( Σ(xi - mean)² / n )
Where xi = individual gaze scores in segment, mean = segment average, n = sample count.

Distinguish low std from concentration (= high avg + low std) vs. low std from disinterest (= low avg + low std). Always report std together with the average.

avg_movement = Σ(|pitch[i]-pitch[i-1]| + |yaw[i]-yaw[i-1]|) / (2 × (n-1))
Measures frame-to-frame "jitter" — how much the head moves on average each frame.

spike = |value[i] - value[i-1]| > threshold
Prah pro yaw: 10°, prah pro pitch: 5°. Spikes indicate sudden head turns — reaction to external stimulus, position change, or surprise.

Verify whether fatigue and attention measure a consistent phenomenon. Expected negative correlation (higher fatigue → lower attention) validates both metrics.

r = Σ((xi - x̄)(yi - ȳ)) / sqrt(Σ(xi - x̄)² × Σ(yi - ȳ)²)
Where xi = gaze score, yi = fatigue score, paired by nearest timestamp (tolerance 2s).

Identify precise moments where the participant lost attention (drop) and where they returned (spike). Enables correlation with external events (segment transition, sound, movement).

drop = score[i-1] - score[i] > threshold
spike = score[i] - score[i-1] > threshold
Default threshold: 15 points. Adjustable per scenario.

Evaluate whether the participant sits in the ergonomic zone and how their distance changes over time (drift = relaxation vs. leaning forward = trying to see better).

CCOHS (Canadian Centre for Occupational Health and Safety) — Resting Point of Accommodation (RPA) ≈ 80 cm. At this distance, eye muscles require no effort to focus.

OSHA eTools Computer Workstations — recommended range 50–100 cm (20–40 inches) from eyes to monitor.

Ergonomic drift interpretation: gradual distancing = relaxation, approaching = trying to see better (possible eye fatigue or small text).

Distinguish active speech from ambient noise. Identify participant voice profile.

In demo session, the speaking detector was inconsistent (detected 2 out of 3 actual voice inputs). Recommendation: compare voiceComments[] (speech-to-text results) with micVolume[].isSpeaking and calibrate the speaking detector threshold.

Automatically generated text profile of the participant based on combination of all metrics. Intended for trainers and training managers.

For each video segment, calculate segmentScore = gazeAvg × (1 - fatigueAvg/100). Segments with lowest score → candidates for shortening or reworking. Segments with highest score → confirm as effective.