PaperlayerView exact revision on arXiv

Source-linked AI summary

Kinematic signatures of impairment: Detecting alcohol intoxication in e-scooter riders using sensor data and machine learning

Rahul Rajendra Pai, Marco Dozza, Alexander Rasch, Ali Mohammadi, Marco Capuccini

arXiv:2609.38276v1cs.LGeess.SPstat.AP

TL;DR

Current e-scooter countermeasures do not continuously assess rider impairment. This study evaluates onboard kinematic detection, finding that an entropy-based classifier distinguishes riding conditions with 85% overall accuracy.

  • Problem

    Current responses to intoxicated e-scooter riding do not continuously assess riders’ motor control, while enforcement and blanket restrictions have practical drawbacks.

    Source: §1 · p. 1

  • Method

    The study tested permutation-entropy features from onboard kinematic signals with a logistic-regression classifier for detecting e-scooter intoxication.

    Source: §2.4.2 · p. 3, §4 · p. 6

  • Results

    85% overall accuracy; the permutation-entropy logistic-regression model achieved an AuROC of 1.00 for sober-versus-high discrimination.

    Source: §4 · p. 6

  • Takeaways & Limitations

    Onboard kinematic monitoring may complement existing countermeasures by monitoring rider control during a ride.

    Source: §5 · p. 10

    Testing was conducted indoors at 7 km/h, below typical real-world riding speeds of 20–25 km/h.

    Source: §4.4 · p. 9

Abstract

from arXiv · show

Alcohol intoxication is a leading contributor to fatal and severe-injured e-scooterist crashes. Current countermeasures, such as temporal restrictions or pre-ride cognitive screening, cannot continuously assess an e-scooterist's physical motor control or impairment in real time. We conducted a controlled experiment in which 25 participants rode an instrumented e-scooter through a test track while sober and at two targeted blood alcohol concentration levels (0.05% and 0.08%). The e-scooter was instrumented with a six-axis inertial measurement unit (IMU), and throttle and brake lever position sensors, all sampled at 100 Hz. Two complementary signal features were computed: normalised permutation entropy, which quantifies temporal complexity, and standard deviation, which quantifies signal amplitude. Repeated measures correlation identified seven kinematic features (all IMU and throttle signals) whose entropy decreased (p < 0.001) while standard deviation increased (p < 0.01) with increasing intoxication, indicating that intoxicated riders shift from continuous, low-amplitude micro-corrections to fewer, high-amplitude reactive corrections. An entropy based multi-class logistic regression classifier, evaluated through leave-one-participant-out cross-validation, achieved 85% overall accuracy and a weighted one-vs-rest area under the receiver operating characteristic curve (AuROC) of 0.94, with a sober-vs-high AuROC of 1.00. Steering rate and lateral acceleration were the most important predictive features, indicating that alcohol induces a distinct collapse in lateral equilibrium during riding. Ultimately, these results demonstrate that onboard kinematic sensing combined with entropy-based signal analysis can reliably distinguish sober from intoxicated e-scooter riding, providing a foundation for automatic intoxication detection systems that preserve mobility for sober riders.

1. Introduction

Alcohol is a major e-scooter safety concern, while current countermeasures do not continuously assess riders’ motor control. This study tests whether onboard kinematics can identify intoxication and characterize associated changes in riding.

1. Introduction

  • Alcohol involvement is reported in e-scooter crashes and fatalities, making intoxication a critical modifiable safety risk.Reported involvement ranges from 12% of crashes in Washington, D.C. to 44% of Swedish e-scooter fatalities.Source: §1 · p. 1
  • Existing responses include BAC limits, reaction tests, and nighttime bans, but enforcement is challenging and blanket bans also restrict sober riders.The passage notes that Sweden has no legal e-scooter BAC limit and that roadside testing is difficult to monitor at scale.Source: §1 · p. 1
  • Prior cycling and e-scooter studies link intoxication to degraded performance, but onboard detection of e-scooter impairment had not been tested.The study tests whether signal complexity decreases and signal amplitude increases with intoxication, consistent with a shift from continuous micro-corrections to fewer, larger reactive corrections.Source: §1 · pp. 1–2
  • The study investigates whether onboard kinematic signals can detect intoxication through a controlled test-track experiment and a classification framework.Its stated objective also includes characterizing the kinematic changes associated with intoxication.Source: §1 · p. 2

2. Method

The study retained 25 participants and used a sensor-equipped commercial e-scooter on an indoor track designed to elicit distinct riding kinematics.

2.1. Participants and ethics

  • The final sample comprised 25 participants after eight of 33 participants were excluded because of data loss.The retained group included 19 men and 6 women, with a mean age of 26.5 ± 4.6 years.Source: §2.1 · p. 2
  • Participants were screened for health, alcohol-use, and medication-related eligibility, and the study used ethics approval, informed consent, and supervised post-trial discharge.The alcohol-use screening required regular but moderate consumption and excluded AUDIT scores above 14.Source: §2.1 · p. 2

2.2.1. Instrumented e-scooter

  • A commercial e-scooter was limited to 7 km/h and instrumented with a six-axis IMU and continuous brake and throttle position sensors.Signals were logged synchronously at 100 Hz; the IMU measured three-axis acceleration and angular velocity.Source: §2.2.1 · p. 2

2.2.2. Test track

  • The indoor test track combined a straight path, a progressively tighter cone slalom, a braking zone, and a figure-eight manoeuvre.The track controlled for weather and variable road friction.Source: §2.2.2 · pp. 2–3

2.3. Experimental procedure

Participants practised while sober, then completed trials across sober, low-intoxication, and high-intoxication conditions in a single session. Alcohol doses targeted BAC levels of 0.05% and 0.08%, with achieved means below those targets.

2.3.1. Familiarisation phase

  • Participants practised the scooter and track while sober until they felt comfortable, to mitigate learning effects.Source: §2.3.1 · p. 3

2.3.2. Experiment protocol

  • Each participant completed two trials in each of three sequential conditions: sober, low intoxication, and high intoxication.The dataset contained 141 valid trials: 45 sober, 50 low, and 46 high.Source: §2.3.2 · p. 3

2.3.3. Intoxication protocol

  • Participants selected a spirit and consumed an individually calculated volume, while BAC was measured before trials without adjusting for target overshoots or undershoots.Target BAC levels were 0.05% and 0.08%; achieved means were 0.035 ± 0.008% for Low and 0.065 ± 0.008% for High.Source: §2.3.3 · p. 3

2.4. Data analysis

The analysis processed sensor signals into centred entropy and amplitude features, tested their association with intoxication, and evaluated classifiers using participant-held-out cross-validation.

2.4.1. Pre-processing

  • The analysis defined trial boundaries using vehicle speed and synchronized optical tracking, filtered IMU signals at 10 Hz, and linearly interpolated missing data.Throttle and brake position signals were not filtered.Source: §2.4.1 · p. 3

2.4.2. Feature extraction and within-subject centring

  • Normalised permutation entropy measured ordinal-pattern complexity across nine signals, while standard deviation measured signal amplitude.Entropy used embedding dimension m = 5 and delay τ = 5 samples; within-subject centring subtracted the participant’s mean across other trials, excluding the trial being centred.Source: §2.4.2 · p. 3

2.4.3. Feature selection

  • Repeated-measures correlations retained seven IMU and throttle features whose PE decreased (p < 0.001) and SD increased (p < 0.01) with intoxication; brake signals were not significant.Holm–Bonferroni correction was applied separately to the PE and SD test families.Source: §2.4.3 · p. 4

2.4.4. Classification model and baselines

  • The primary classifier was L2-regularised multi-class logistic regression trained on seven centred PE features, compared with entropy-threshold, RBF-SVM, and SD-based logistic regression baselines.The models compared the discriminative value of temporal complexity and signal amplitude.Source: §2.4.4 · pp. 4–5

2.4.5. Cross-validation and evaluation metrics

  • Leave-one-participant-out cross-validation held out all trials from one participant per fold and evaluated concatenated out-of-fold predictions.The evaluation included accuracy, macro-averaged precision, recall, and F1-score, plus one-vs-rest AuROC and sober-versus-high AuROC.Source: §2.4.5 · p. 5

2.4.6. Feature importance

  • Feature importance was estimated by averaging logistic-regression coefficients across 25 folds and summing their absolute class-specific means.Source: §2.4.6 · p. 5

3. Results

The results show that permutation entropy decreases while signal amplitude increases with intoxication, and that entropy-based classification distinguishes intoxication conditions. Steering rate and lateral acceleration are the strongest predictive features.

3.1. Feature selection

  • All IMU signals and throttle showed decreasing PE (p < 0.001) and increasing SD (p < 0.01) with intoxication; brake signals were not significant.Fig. 2 shows WSC-PE distributions for nine signals by condition, with selected features trending downward from Sober to High.Source: §3.1 · p. 5, Figure 2 · p. 6

3.2. Classification performance

  • 85% overall accuracy and a weighted one-vs-rest AuROC of 0.94 were achieved by the PE-based logistic regression, with sober-vs-high AuROC of 1.00.The SVM achieved 86% accuracy and comparable per-class performance; machine-learning models outperformed the 74% sum-entropy threshold baseline.Source: §3.2 · p. 5

3.3. Feature importance

  • Steering rate (ry; Σ|β| = 3.21) and lateral acceleration (ax; Σ|β| = 3.03) ranked as the dominant predictive features.Roll rate (rz; 1.93) and throttle (1.76) followed; vertical and longitudinal acceleration contributed least.Source: §3.3 · p. 6

4. Discussion

The study reports that permutation entropy and logistic regression can distinguish intoxication from onboard e-scooter kinematics. The discussion interprets the signal changes, identifies lateral-control features as especially informative, and outlines limits on generalisation and deployment.

4. Discussion

  • 85% overall accuracy and a sober-vs-high AuROC of 1.00 were achieved by the permutation-entropy logistic regression classifier.The authors conclude that intoxicated riding has kinematic signatures distinct enough for automated detection.Source: §4 · p. 6

4.1. Rider impairment characterisation during intoxicated riding

  • As intoxication increased, riders shifted from continuous, low-amplitude corrections to fewer, high-amplitude reactive corrections as signal entropy fell and standard deviation rose.The entropy decline indicates more predictable movement timing, not smaller steering movements.Source: §4.1 · pp. 6–7

4.2. Directional importance of kinematic features

  • Steering rate and lateral acceleration were more important for detection than vertical or longitudinal dynamics, while similar speed profiles may persist as lateral control deteriorates.The discussion links this pattern to the continuous balance and steering adjustments required for lateral stability.Source: §4.2 · p. 8

4.3. Model performance and practical implications

  • 85% vs 86% accuracy and 0.94 vs 0.94 AuROC show that the PE-based logistic regression performed comparably to the more complex SVM.The authors suggest added model complexity does not meaningfully improve detection and note the linear model is easier to interpret and deploy on embedded hardware.Source: §4.3 · p. 8
  • 74% accuracy and a sober-vs-high AuROC of 0.99 were achieved by the simple sum-entropy threshold model, while temporal complexity carried most three-class discriminative information.The logistic regression primarily improved discrimination of the intermediate Low class.Source: §4.3 · p. 8
  • 61% accuracy and a weighted AuROC of 0.79 for WSC-SD classification were below PE-based performance, although amplitude measures achieved a sober-vs-high AuROC of 0.90.The authors state that complementing PE with amplitude features may further improve performance.Source: §4.3 · p. 8
  • Achieved BAC levels were below targets, so classifier performance at 0.08% or higher remains an empirical question.Mean levels were 0.035 ± 0.008% for Low and 0.065 ± 0.008% for High, against targets of 0.05% and 0.08%.Source: §4.3 · p. 9

4.4. Limitations and future research

  • Testing was limited to 7 km/h indoors, and fixed ascending intoxication order did not fully control fatigue or practice effects; eight of 33 participants were excluded after data loss.Real-world outdoor riding can occur at 20–25 km/h; the authors recommend operational-speed studies and note the attrition reduced statistical power.Source: §4.4 · p. 9

5. Conclusion

The study characterizes intoxicated riding as a collapse of lateral dynamic equilibrium, marked by fewer, higher-amplitude corrections concentrated in steering and roll kinematics. It proposes continuous kinematic monitoring as a complement to existing countermeasures.

5. Conclusion

  • Intoxicated riders shift from adaptive high-frequency micro-corrections to fewer, high-amplitude reactive corrections.The conclusion characterizes this change as a collapse of lateral dynamic equilibrium.Source: §5 · p. 9
  • The change is concentrated in steering and roll kinematics and is measurable from onboard sensors for automated detection.Source: §5 · p. 10
  • Continuous motor monitoring is proposed as a complement to pre-ride screening and nighttime restrictions, with intervention focused on risky riding behaviour.The framework is intended to preserve shared e-scooter accessibility while concentrating intervention where risk is greatest.Source: §5 · p. 10

CRediT authorship contribution statement

The authors report distinct contributions across methodology, analysis, investigation, software, supervision, funding, and writing.

CRediT authorship contribution statement

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

The authors used large language models for grammar correction and language review, then reviewed and edited the work themselves.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

Declaration of competing interests

The disclosure reports relationships with Voi Technology AB for two authors and no known competing interests or relevant relationships for the others.

Declaration of competing interests

Appendix A. Alcohol volume estimation

Appendix A describes estimating beverage volume using target and previous measured alcohol concentrations, alongside absorption and elimination assumptions. It calculates the Widmark factor from participant anthropometry and specifies additional alcohol and distribution parameters.

Appendix A. Alcohol volume estimation

  • The required-volume estimate uses target BAC (0.05 or 0.08), previous measured BAC, an elimination rate of 0.15 per hour, and a 15-minute absorption time.Source: Appendix A · p. 10
  • The calculation specifies a 15% resorption deficit, ethanol density of 0.789 g/mL, beverage alcohol content of 0.40 or 0.35, and distribution volume as weight multiplied by r.Source: Appendix A · pp. 10–11
  • The Widmark factor r is calculated using anthropometric equations, with weight in kilograms and height in centimetres.Source: Appendix A · p. 11

Related papers

Loading 2609.38276v1…