FlexTail Sensor Evaluation Playbook

MinkTec GmbH, Braunschweig, Germany

info@minktec.comminktec.com

Introduction

This document illustrates what FlexTail sensor data can be used for in the context of occupational and domestic ergonomics. Each chapter applies a different analytical framework to real recordings, demonstrating how continuous lumbar kinematics translate into actionable ergonomic insight.

The analyses draw on three datasets: a controlled human activity recognition (HAR) study [1] in which ten participants performed everyday household tasks, a longitudinal office-worker dataset capturing spontaneous sitting behaviour across multiple workdays, and cleaning session recorded in a domestic environment.

Several themes emerge across the chapters. Activity type is the primary driver of spinal load: seated tasks such as eating and cutting produce low, stable lumbar angles, while reaching and stooping activities — vacuuming, wiping, loading a dishwasher — generate sustained forward flexion combined with axial rotation, pushing composite risk scores substantially higher. Individual variation within the same activity is also considerable; the scatter plots in the ergonomics chapter make clear that two people performing the same task can produce very different spinal load profiles. In the office dataset, spontaneous position-change rate follows a consistent within-day pattern — higher in the morning, lower across the mid-afternoon — with meaningful differences between individuals. The cleaning analysis demonstrates how a single long session can be decomposed into an ISO 11226 classification, a time-resolved bending frequency profile, and per-bend quality scores, giving a layered picture of cumulative exposure that point-in-time assessments cannot provide.

Together, these examples show that FlexTail data supports analysis at multiple levels of granularity: from population-level activity comparisons down to individual bend events within a single session.

Ergonomics and Occupational Exposure

The analysis in this chapter is reusing the data from a controlled recording study that compared the performance of human activity recognition (HAR) using the FlexTail camera system [1]. Ten participants each performed everyday household activities for one minute per activity, while the FlexTail recorded continuous lumbar spine kinematics.

The activities were chosen deliberately for their familiarity: cutting vegetables, eating at a table, loading and unloading a dishwasher, vacuuming, wiping a surface, and walking. These are tasks that most people perform several times a week, making the data intuitive to interpret and straightforward to relate to one’s own daily movement habits. All activities engage the spine to varying degrees — some through sustained forward flexion, others through rotation or a combination of both.

The risk model applied here is intentionally simple. Two biomechanical planes are assessed independently: the sagittal (lumbar) angle, which corresponds to the degree of lordosis or forward bend at the lower back, and the axial rotation (twist) of the upper body relative to the pelvis, which the FlexTail captures directly. Each plane yields a sub-score on a 0–1 scale, and the two sub-scores are combined into a single composite score. The model makes no claim to clinical completeness; its purpose is to provide a transparent, digestible summary of spinal load that is easy to understand and discuss. Despite its simplicity, it surfaces several interesting differences in how the spine is loaded across activities and across individual participants.

Lumbar–Twist Risk Summary per Activity

Figure 1: Mean lumbar (navy), twist (orange), and composite (lavender) risk scores per activity, averaged across all participants. Individual participant scores are overlaid as strip dots. Horizontal dashed lines at 0.50 and 0.75 mark the medium and high-risk tier boundaries. Scores are on a globally anchored 0–1 scale (0.5 = dataset mean, 1.0 = $+ 2$ SD).

Figure 2: Numeric summary of lumbar–twist risk scores per activity (mean ± SD across all participants). Cell shading denotes tier: green = Low (< 0.50), amber = Medium (0.50–0.75), red = High (≥ 0.75).

Cutting and eating stand out as the lowest-risk activities in the dataset. Both were performed at a table with participants seated, and the sagittal angles recorded reflect a relatively upright posture throughout. The difference between the two activities is nonetheless measurable: during cutting, participants maintained a mean lumbar angle of 0.6°, whereas during eating the mean rose to 10.0° (being slouched forward). The likely explanation is postural: when cutting, participants lean slightly forward and actively engage with the task in front of them, keeping the torso more erect; when eating, they tend to sit more relaxed, with a mild posterior pelvic tilt that increases lumbar flexion. This is a small but consistent effect visible across the participant group.

The remaining activities — vacuuming, wiping, and loading or unloading the dishwasher — require participants to reach forward and downward, producing substantially greater trunk flexion and, in several cases, considerable axial rotation. These demands are directly reflected in their higher lumbar and twist risk scores. Wiping in particular combines both planes: the sustained forward lean elevates the lumbar sub-score, while the lateral arm sweep drives upper-body rotation.

Lumbar vs. Twist Risk per Participant

Figure 3: Lumbar risk score vs. twist risk score, one point per participant, for each of the five risk activities (walking excluded; loading and unloading dishwasher merged). Marker colour denotes composite risk tier (red ≥ 0.75, amber 0.50–0.75, green < 0.50); marker size is proportional to the composite score. Dashed lines at score = 0.50 divide each panel into four quadrants.

Activity Profiles: Lumbar Kinematics Across Everyday Tasks

The HAR study recorded ten participants performing seven household activities for one minute each. This chapter examines the lumbar kinematic signature of each activity in depth — combining angle distributions, temporal patterns, and the joint structure of flexion and rotation to characterise the spinal demands imposed by ordinary daily tasks.

Lumbar Angle Distributions per Activity

Figure 4: Violin plots of lumbar flexion angle (°) per activity, pooled across all ten participants. Each violin shows the full empirical distribution; the embedded box shows the inter-quartile range and median. Positive values denote forward flexion; negative values denote extension. Horizontal dashed line at 20° marks the commonly cited low-risk boundary for lumbar loading.

Eating and cutting produce the narrowest, most upright distributions; both were performed seated at a table. The dishwasher tasks and vacuuming produce right-shifted, wider distributions, reflecting the sustained forward reach those activities require. Walking occupies a distinct regime with near-zero mean flexion and low variance, consistent with its role as a neutral reference activity.

Joint Distribution: Lumbar Flexion and Axial Twist

Figure 5: Joint distribution of lumbar flexion angle (x-axis) and axial twist angle (y-axis) per activity, pooled across all ten participants. Each panel shows up to 1 000 sampled data points (scatter) plus KDE contour lines. Dashed lines at 0° delineate the four posture quadrants. Each activity has a distinct cluster shape that reflects its biomechanical demands.

The joint distributions reveal that activities impose different constraints on the two planes simultaneously. Eating and cutting cluster tightly near the origin — low flexion and near-zero twist. Vacuuming and wiping produce elongated clusters oriented diagonally: forward lean co-occurs with lateral rotation because both tasks involve sweeping arm movements that drive trunk rotation. The dishwasher tasks form a horseshoe-shaped distribution: participants bend forward (increasing flexion) while rotating to reach objects at different depths within the appliance.

Sitting Position Change

Sustained static sitting is a primary ergonomic risk in administrative roles. Beyond average posture scores, however, the frequency and timing of spontaneous position shifts provides additional diagnostic information: workers who move less frequently may accumulate greater cumulative spinal load even if their mean angle appears moderate.

 Sitting Position Change — Detection Algorithm

Intra-day detection (Tier 1). For each recording session, every raw sensor sample is compared against a sticky reference pose that advances only when a position change is detected. The pose is represented as a three-component vector in degrees:

where $𝐿$ is lumbar flexion, $Lat$ is lateral lean, and $𝑆_{approx}$ is sagittal approximation.

The Euclidean distance between the current pose and the reference pose is computed each sample:

A position change is detected when $𝑑 \geq 15 °$ and at least 2 minutes have elapsed since the last detection (minimum position duration holdoff). The holdoff allows the reference pose to track the body during a slow transition while suppressing re-detection of the same shift at 5 Hz. Samples with sensor movement magnitude above threshold are excluded.

Day-level aggregate analysis (Tier 2). For the day-over-day analyses, the windowed change-score formula is applied to daily mean posture vectors from the pre-aggregated statistics:

with $𝜏_{𝐿} = 5 °$ , $𝜏_{𝑆} = 8 °$ , $𝜏_{𝑇} = 10 °$ . A change is detected when $𝐶 > 1.0$ between consecutive workdays, capturing persistent day-to-day postural adaptation.

Time-of-Day Change Patterns

All 31 available recording days were loaded and the sample-by-sample SPC detection was run for each. Change events and their clock-hour were pooled across all days within each participant. The change rate at each clock hour was computed by dividing the event count by the sitting exposure at that hour and scaling to an hourly rate:

This normalisation is essential: not all clock hours are equally represented in the data, and raw counts would be dominated by hours with longer coverage. Hours with fewer than 5 minutes of sitting exposure are excluded.

Figure 6: Estimated sitting position change rate (changes / hour) by clock time, across all office workers. Coloured lines: individual participants (all recording days pooled per participant). Black line + shaded band: group mean ± 95% CI. Grey vertical bands mark typical break and lunch windows. Bottom panel shows data coverage (number of participant-days contributing to each time bin). Only bins with at least 5 minutes of sitting exposure are shown.

Figure Figure 7 complements the rate analysis by showing the distribution of time between consecutive changes at each clock hour. Short intervals reflect frequent position adjustments; long intervals indicate prolonged static sitting.

Figure 7: Inter-change interval (minutes between consecutive sitting position changes) by clock hour of the change event. Violin = pooled distribution across all participants and days; orange dashed line = median; grey dotted line = 5-minute reference. Clipped at 10 min for readability. Minimum 5 observations required per hour. Observation counts shown above each violin.

Cleaning Activity: Joint Posture and Bend Analysis

Cleaning and tidying tasks are among the most biomechanically demanding activities in domestic and care environments: the combination of prolonged trunk flexion, repeated lateral reaches, and high-frequency bending cycles creates cumulative spinal load that is rarely captured by point-in-time assessments. This chapter presents a joint analysis of a 2 hour cleaning recording using two complementary frameworks: ISO 11226 [2] three-tier posture classification applied across the full session, and a bend detection with per-bend quality scoring using the FlexTail lift evaluation engine.

ISO 11226 Background

ISO 11226 (Ergonomics — Evaluation of Static Working Postures) [2] and DIN EN 1005-4 (Safety of Machinery — Human Physical Performance — Part 4: Evaluation of Working Postures and Movements) [3] provide a standardised three-tier framework for classifying trunk posture risk from continuous kinematic recordings. The framework operates in three stages: zone assignment in two independent planes (sagittal flexion and lateral deviation), cross-matrix combination of the two zones, and a pelvic-tilt escalation criterion applied via the 90th-percentile lumbar angle.

Two mechanistically distinct routes can produce a High-risk classification, and they require different interventions:

 Pathway A: Sagittal Mean Escalation

A recording enters the Awkward sagittal zone when the mean trunk flexion reaches ≥ 40°. This pattern reflects prolonged static forward lean — characteristic of desk-based work with a poorly positioned screen or keyboard, or of sustained stooping during cleaning tasks. ISO 11226 flags this because sustained static loading, even at moderate flexion angles, elevates compressive disc load and fatigues lumbar extensor musculature. The relevant interventions are workstation or task-height adjustment, regular postural breaks, and tool ergonomics (e.g. long-handled implements for floor-level cleaning).

 Pathway B: Pelvic-Tilt Escalation

A recording is escalated to High risk when the 90th-percentile lumbar angle reaches ≥ 40° while the mean angle remains in the Neutral or Moderate zone. This pattern indicates that brief, deep forward bends occur repeatedly without substantially elevating the session mean. The relevant injury mechanism is peak disc shear force at L4/L5 and L5/S1: shear force increases approximately quadratically with bending depth beyond 40°, so a small number of deep bends can produce cumulative loading comparable to hours of moderate flexion. Interventions target peak bend depth — raising working heights, coaching bend technique, and limiting floor-level reaches.

The cross-matrix alone cannot distinguish these two pathways. The 90th-percentile lumbar angle is the practical discriminant: high P90 with moderate mean signals Pathway B; high mean directly signals Pathway A. Both are present in the cleaning recording analysed below.

ISO 11226 Session Classification

ISO 11226 and DIN EN 1005-4 classify trunk postures into three risk tiers based on the distribution of sagittal flexion and lateral deviation across the recording. Each sample is independently assigned to a sagittal zone and a lateral zone; their intersection determines the cell risk in the cross-matrix. A pelvic-tilt escalation criterion is then applied via the 90th-percentile lumbar angle.

 ISO 11226 Three-Tier Classification

Each sensor sample is classified into a sagittal zone and a lateral zone independently:

The cross-matrix maps zone combinations to base risk (values show % of recording time):

Session-level risk is the dominant cell (by sample count). Pelvic-tilt escalation: if $𝜃_{lum,P90} \geq 40 °$ , the session risk is incremented by one tier (clamped at High).

Figure Figure 8 shows the cross-matrix for this recording. Each cell displays the percentage of the 124,581 samples that fell into that sagittal × lateral zone combination. The largest cell is Sagittal Moderate × Lateral Neutral, which is the characteristic posture of sustained forward stooping without significant sideways lean — consistent with tasks such as wiping surfaces and pushing a mop. The Sagittal Awkward column (≥40°) accumulates substantial time, driven by deep stooping during floor-level tasks. Very little time is spent in the Sagittal Neutral column, confirming that true upright posture was rare throughout the session.

The session mean sagittal angle of 26.2° places the recording in the Moderate sagittal zone. The mean absolute lateral deviation of 2.4° is Neutral. The 90th-percentile lumbar angle of 34.6° is below the 40° pelvic-tilt escalation threshold. The resulting ISO 11226 classification is Medium risk.

Figure 8: ISO 11226 posture zone cross-matrix for the cleaning session. Each cell shows the percentage of recording time (124,581 samples) spent in that sagittal × lateral zone combination. Cell colour indicates risk tier (green = Low, amber = Medium, red = High). The metrics table summarises the session mean angles, zones, and the final ISO 11226 classification.

Bending Frequency Over Time

Figure Figure 9 shows the number of discrete bending events detected per 5-minute window across the session. Bends are identified by delayed Schmitt trigger applied to the sagittal channel (enter ≥30°, exit <15°, holdoff 1 s). Bar colour indicates whether the window is below the session average (green), close to average (amber), or above average (red).

The distribution of bend frequency is uneven across the session. Several 5-minute windows contain 15 or more bends, while others contain fewer than 5, suggesting alternation between high-intensity cleaning phases and lower-activity periods such as moving between rooms, preparing equipment, or brief rest. This within-session variability is relevant: high-frequency windows represent concentrated spinal loading episodes that contribute disproportionately to cumulative fatigue.

Figure 9: Bending frequency across the cleaning session in 5-minute buckets. Each bar shows the number of Schmitt-trigger bend events starting within that window. Colour: green = below session average, amber = within ±25% of average, red = above average. Dashed line: session mean bends per 5-minute window.

Sagittal Bend Distribution

Figure Figure 10 shows the full distribution of sagittal angles across all samples. The distribution is right-skewed with a long tail extending beyond 90°, consistent with deep stooping during floor-level tasks. Very little time is spent in the Neutral zone below 20°. The mode lies in the 20–40° Moderate band, the characteristic standing-stooped posture during surface cleaning.

Figure 10: Distribution of sagittal flexion across all 124,581 samples. ISO zone boundaries (20° and 40°) shown as coloured overlays. Dashed line: session mean (26.2°). Dotted lines: P10 and P90.

Schmitt-Trigger Bend Detection

The delayed-predicate Schmitt trigger was applied to the sagittal channel to identify discrete bending events. Active phases begin instantly when sagittal flexion exceeds 30° and end only after the signal has remained continuously below 15° for at least 1 second. The holdoff prevents fragmentation of compound movements.

 Bend Detection Parameters

Parameter	Value	Rationale
Enter threshold ( $𝜃_{enter}$ )	30°	Above ISO Neutral; captures meaningful stoops
Exit threshold ( $𝜃_{exit}$ )	15°	Hysteresis margin; avoids fragmentation
Holdoff ( $Δ 𝑡_{low}$ )	1.0 s	Absorbs brief mid-bend pauses
Channel	sagittal	Primary forward flexion channel

Over the recording, 246 bending events were detected. The upper panel of Figure 11 overlays the detected phases on the sagittal time-series; the density of shading confirms that bending is nearly continuous throughout the session. The peak angle distribution (lower left) shows that the majority of bends exceed 40°, placing them in the Awkward zone. The duration distribution (lower right) shows that most bends last 5–30 seconds — sustained stoops rather than brief dips.

Figure 11: Schmitt-trigger bend detection across the full session. **Top:** Sagittal time-series with detected bend phases (orange shading), enter threshold (30°) and exit threshold (15°). 246 events detected. **Bottom left:** Distribution of peak sagittal angles per bend. **Bottom right:** Distribution of bend durations.

 Bend Frequency and Cumulative Exposure

At 246 bends, the session averages approximately 2 bending events per minute — substantially above the ≤1 bend/min threshold cited in manual-handling risk guidelines. With a mean peak angle of 65° and mean duration of 17 s per bend, cumulative disc loading over a full cleaning shift is substantial. If this was a workplace setting, raising working heights and using long-handled tools are the primary ergonomic interventions to reduce both bend frequency and peak angle.

Bend Quality Evaluation

Each detected bend was scored frame-by-frame using the FlexTail lift evaluation engine. Four channels contribute simultaneously — lumbar, sagittal, lateral, and twist — each penalised proportionally to absolute angle in radians. The composite score (0–100, higher = better) is the unweighted mean of the four components, averaged over all frames within the bend phase.

Figure Figure 12 shows the quality results. Panel A plots the composite score for each of the 246 bends sorted in ascending order. Panel B shows mean component scores across all bends. Panel C gives the grade distribution.

Figure 12: Bend quality evaluation across all 246 detected bends. **Left:** Composite score per bend (sorted), coloured by grade. **Centre:** Mean component scores averaged across all bends. **Right:** Grade distribution (Good / Fair / Poor).

References

[1] L. S. T. M. D. Jonas Walkling Arwed Masch, “Wearable Spine Tracker vs.\ Video-Based Pose Estimation for Human Activity Recognition,” Sensors, vol. 25, no. 12, p. 3806, 2025, doi: 10.3390/s25123806.
[2] International Organisation for Standardisation, “ISO 11226: Ergonomics — Evaluation of Static Working Postures,” no. ISO 11226:2000. 2000.
[3] Deutsches Institut für Normung, “DIN EN 1005-4: Safety of Machinery — Human Physical Performance — Part 4: Evaluation of Working Postures and Movements in Relation to Machinery,” no. DIN EN 1005-4:2005. 2005.