Machine Learning-Based Stroke Segmentation in Kayaking Using

Integrated IMU and EMG Data

abor Nagy

, P

eter Katona

, Levente Gannorouwa

and L

aszl

o Grand

azm

any P

eter Catholic University, Faculty of Information Technology and Bionics, 40 Pr

ater Str, Budapest, Hungary

Hungarian University of Sports Science, 42-48 Alkot

as Str, Budapest, Hungary

Keywords:

Kayaking, Electromyography, Inertial Measurement Unit, Boat Acceleration, Machine Learning Models,

Ensemble Learning.

Abstract:

Accurate classiﬁcation of stroke side in rowing motions is essential for performance monitoring and injury

prevention. This study evaluates three machine learning models — Naive Bayes (NB), Logistic Regression

(LR), and Gradient Boosting Decision Trees (GBDT) — using biomechanical and electromyographic (EMG)

features. A core set of 25 features was identiﬁed, with normalized joint coordinates and latissimus dorsi

EMG activity among the most inﬂuential. The NB model achieved 92.21% cross-validation accuracy using

only three coordinate-based features, while the full feature set improved accuracy modestly by 1.94%. The

LR model attained 94.48% accuracy, slightly outperforming NB. The GBDT model achieved the highest

accuracy with 96.18% on the test set, alongside the lowest mean absolute stroke onset detection error of 24.6

± 51.6 ms, corresponding to just 4.5% of average stroke duration. Classiﬁcation accuracy remained stable

across stroke paces. A strong negative correlation (R = -0.935) between classiﬁcation accuracy and onset

detection error was observed across subjects, indicating that poorer spatial classiﬁcation corresponds with

greater temporal imprecision. Signiﬁcant inter-subject variability was found, with accuracy ranging from

91.89% to 98.9%, likely reﬂecting individual differences in stroke technique and muscle activation patterns. A

core set of biomechanical features were identiﬁed, such as normalized joint coordinates of th eulnar styloid and

right olecranon, latissimus dorsi EMG activity among the most inﬂuential, vertical pelvis lateral bending and

bilateral shoulder ﬂexion. Tempo-based relative time averages of these features reveal clear phase-dependent

patterns that contribute strongly to model decision-making. These results demonstrate that accurate stroke

side classiﬁcation can be achieved using a relatively small set of biomechanical features, with GBDT models

providing superior performance.

1 INTRODUCTION

Measuring kayaking technique is inherently complex,

requiring the integration of biomechanical, kinematic,

and neuromuscular analyses. Recent advancements

have expanded the use of technologies such as motion

capture systems, surface electromyography (sEMG),

and machine learning (ML) models to assess perfor-

mance more accurately. Kinematic analysis helps

quantify joint and segment movement, while sEMG

offers insight into muscle coordination and fatigue

during paddling, particularly in key upper-body mus-

cles such as the latissimus dorsi, triceps brachii, and

anterior deltoid (Lauder and Kemecsey, 1999).

Studies such as Garnier et al. have demonstrated

that combining sEMG with kinematic data allows for

a more complete understanding of how fatigue im-

pacts paddling performance (Garnier et al., 2023).

Similarly, McDonnell et al. proposed reﬁned stroke

models to address inconsistencies in stroke descrip-

tions, further enhanced by tools like the e-Kayak

system for performance feedback (McDonnell et al.,

2012; Bonaiuto et al., 2020). Other research has

emphasized the value of three-dimensional analyses

and recognized inherent movement asymmetries in

kayaking that affect force transmission and efﬁciency

(Li, 2017; Vasiljev et al., 2024).

Importantly, multiple studies have reported sig-

niﬁcant differences between on-water and ergometer-

based kinematics. Klitgard et al. found variations in

joint movement and increased thoracolumbar motion

during on-water trials, while Harbour et al. identiﬁed

large discrepancies in timing and force parameters

across environments (Klitgaard et al., 2021; Harbour

256

Nagy, G., Katona, P., Gannorouwa, L. and Grand, L.

Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data.

DOI: 10.5220/0013781200003988

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2025), pages 256-264

ISBN: 978-989-758-771-9; ISSN: 2184-3201

et al., 2021). These ﬁndings suggest that ergometer-

based assessments may not fully capture the biome-

chanical realities of on-water performance.

Despite the widespread use of sEMG and motion

tracking, our review identiﬁed a lack of studies that

combine these modalities in on-water kayaking sce-

narios. This represents a key research gap in under-

standing how muscle activation and movement inter-

act under real-world conditions.

In parallel, the application of machine learning in

sports biomechanics has rapidly progressed. IMU-

based data capture combined with ML models has

shown strong potential in real-time movement clas-

siﬁcation, including in kayaking (Kranzinger et al.,

2023; Liu et al., 2021). For example, Liu et al.

achieved 98% accuracy in segmenting kayak stroke

phases using ML validated against video analysis.

Similar methodologies have also been successfully

applied to gesture recognition tasks using sEMG data,

despite challenges such as electrode shift (Kim et al.,

2021).

In this study, we investigate whether sEMG and

motion capture data can be used to develop an accu-

rate ML model to classify left and right kayak strokes.

We also examine the role of input features and ex-

plore between-subject variability in classiﬁcation ac-

curacy. Our hypothesis is that some athletes exhibit

more consistent and repeatable stroke patterns, lead-

ing to higher classiﬁcation accuracy, while others dis-

play more variation, resulting in reduced model per-

formance. Since classiﬁcation models perform best

with well-separated and consistent input patterns, this

work aims to better understand how stroke consis-

tency affects model outcomes and which features are

most inﬂuential in side prediction and also provide a

basis for further analysis.

2 METHODS

2.1 Participants and Environment

The study involved nine male participants, aged 17

to 21, all of whom were experienced kayakers with

notable athletic achievements in the sport. The ex-

periments were conducted over two consecutive days

at the Kolonics Gy

orgy Watersport Center, located

on the Danube River. These trials took place during

mid-summer, under favorable weather conditions. It

is noteworthy that the section of the Danube River

where the study was conducted is regulated by the

Kvassay Dam, which ensures modest presence of nat-

ural water currents in the area.

2.2 Motion Capture System

A wireless inertial motion capture system (Xsens

MVN Link, Xsens Technologies BV, Enschede,

Netherlands) and its software (Xsens MVN Studio

version 4.2.4, Xsens Technologies BV, Enschede,

Netherlands) was utilized to record on water kayaker

motion at a sampling rate of 60 Hz. The measure-

ment system consists of small (47 × 30 × 13 mm, 16

g) box shaped measurement units that are placed on

both shoulders, upper arms, forearms, hands, thighs,

shanks and feet, with one additional sensor placed on

the head, chest and on the sacrum of the athlete with

Velcro straps.

The measurement units communicate with a base

station called Awinda, that are on a boat 5-20 me-

ters away from the athlete. The technology provider

claims 50 meter open air wireless range, however in

our studies it was signiﬁcantly less than that, which

resulted in signal loss. For better reception a bigger

antenna was used in the base station, that somewhat

improved signal stability.

Before each recording session a strict IMU unit

calibration procedure provided by the manufacturer

was strictly followed to avoid electromagnetic ﬁelds.

The inertial motion capture systems’ data fusion

algorithm employed an extended Kalman ﬁlter to pro-

vide accurate positional readings of body segments

and joint angles. In this study we use features from

the positional coordinates and joint angles. The

XSense biomechanical model consists of 23 segments

linked through 22 joints. (Technologies, 2025)

2.3 EMG Signal Acquisition

Cometa Miniwave wireless sensors with a WavePlus

base station were used to acquire EMG signals and

was also used as the boat and paddle accelerometers.

The manufacturer claims 40 meters of indoor range.

Tested in an outside environment, it is notable that the

basestaion was far more reliable than the IMU mea-

surement base station in terms of signal acquisition.

Sensors are able to collect EMG data at 2 kHz and

also employ an accelerometer that is capable to trans-

fer 3-axis accelerometer data at a sampling rate of 140

Hz.

During the exercise sessions, surface electromyo-

graphy (EMG) signals were collected from various

muscles on both the right (abbreviated as ”R”) and

left (abbreviated as ”L”) sides of the participants. In

total 14 sensors were used to collect data from the fol-

lowing speciﬁc muscles:

1. Latissimus dorsi

2. Trapezius (middle ﬁbers)

Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data

257

3. Deltoideus Anterior (anterior deltoids)

4. Pectoralis Major

5. Obliquus Externus Abdominis

6. Rectus Femoris

7. Flexor Carpi Radialis

These muscles were strategically chosen to capture a

comprehensive representation of the muscular activity

involved in the kayaking exercise.

In accordance with the international protocol

recommendations established by SENIAM (Surface

EMG for Non-Invasive Assessment of Muscles), we

positioned the surface electrodes on the muscles,

speciﬁcally targeting the area between the tendon

and the motor point. This standardized placement

aimed to ensure consistency and reliability in cap-

turing electromyographic signals, facilitating accurate

assessment of muscle activity during the kayaking

motions under investigation. Adhering to SENIAM

guidelines enhances the comparability of our ﬁndings

with broader scientiﬁc literature and contributes to

the overall validity of our electromyographic analy-

ses. (Stegeman and Hermens, 2007)

The boat-mounted accelerometer was positioned

directly behind the kayaker, aligned with the longitu-

dinal axis of the kayak. From this setup, we utilized

acceleration data from the axis parallel to the boat’s

heading to capture forward motion dynamics. Sim-

ilarly, the paddle accelerometer was mounted at the

midpoint of the paddle shaft, with the relevant axis

also aligned parallel to the paddle. This conﬁgura-

tion enabled consistent detection of stroke direction

and intensity. The parallel axis readings from both

devices were essential for accurately deriving stroke

side and synchronizing movement phases.

Both the Awinda and WavePlus base stations sup-

port hardware synchronization, which we used dur-

ing the recording process. This hardware sync en-

sured that both systems began recording at the same

time, with their internal clocks aligned. As a result,

the timestamps recorded by each system were already

synchronized when saved to the computer, enabling

accurate temporal alignment of data across devices.

The recording was initiated via the WavePlus data ac-

quisition software, which triggered the start of data

collection in Xsens as well.

2.4 Protocol

Participants were instructed to paddle at a ﬁxed stroke

rate for multiple, short, sustainable recording ses-

sions. Stroke rates began at 80 strokes per minute

(SPM), representing a low-to-intermediate training

load, and increased with 10 SPM incrementally up to

120 SPM, which is considered highly demanding for

the athletes. The duration of each recording session

was adjusted based on the target stroke rate, allowing

for the collection of approximately 60 stroke cycles

per session. Speciﬁcally, the session lasted 45 sec-

onds at 80 SPM, 40 seconds at 90 and 100 SPM, 35

seconds at 110 SPM, and 30 seconds at 120 SPM. No-

tably, the 30-second duration at 120 SPM was physi-

cally demanding for some athletes, given the high in-

tensity of the effort. A rest period of 60 to 120 sec-

onds was provided between recording sessions to al-

low for adequate recovery. The protocol ensured that

we can gather samples from low-intermediate training

scenarios to very demanding intensities. Some ath-

letes performed two runs successively.

An on-land reference trial was conducted to de-

termine the maximum EMG values required for sub-

sequent signal normalization. During this trial, the

athlete was instructed to apply maximal force to the

kayak paddle while it was held stationary by a helper.

This was performed separately for the left and right

arms, resulting in isometric contractions that engaged

the relevant muscle groups. The resulting EMG max-

ima provided a consistent and reproducible reference

point for normalizing muscle activity recorded during

the on-water paddling sessions.

3 SIGNAL PROCESSING

3.1 IMU Readings

All IMU-derived body segment positions were ex-

ported in C3D format for further analysis. Due to the

absence of a ﬁxed track, the kayak could not maintain

a perfectly straight trajectory during recordings. As

a result, the heading and tilt of the athletes were nor-

malized to account for variations in orientation. The

center of the pelvis was deﬁned as the origin, and all

positional data were translated accordingly.

Motion capture systems are inherently subject to

positional drift, a phenomenon that is further ampli-

ﬁed in on-water conditions due to the kayak’s con-

tinuously shifting heading. To mitigate this drift,

the biomechanical model was dynamically rotated.

Speciﬁcally, the model was aligned such that the

vector connecting the midpoint between the athlete’s

heels and the midpoint between the ischial tuberosi-

ties (sit bones) was made parallel to the kayak’s head-

ing, which in this context is considered the X-axis

of the coordinate system. The Z-axis was deﬁned as

pointing upward, and the Y-axis pointed leftward rel-

ative to the athlete. The biomechanical model of the

athlete was rotated within the XY plane such that the

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

258

line connecting the seventh cervical vertebra (C7) and

the midpoint of the ischial tuberosities was aligned

parallel to the Y-axis.

To account for inter-individual differences in body

dimensions, per-athlete z-score standardization was

applied to the body part coordinate data. These stan-

dardized coordinates were then used as features in the

subsequent analysis, ensuring comparability across

athletes. Joint angles were also normalized on a per-

athlete base with z-score standardization.

3.2 EMG, Boat and Paddle

Accelerometer Preprocessing

To enhance EMG signal quality and minimize the im-

pact of various artifacts, a series of standard prepro-

cessing techniques were applied. A bandpass ﬁlter

was implemented with cutoff frequencies set at 20

Hz (low) and 400 Hz (high) to remove movement

artifacts and high-frequency electrical noise, respec-

tively. The ﬁlter was designed as a second-order But-

terworth ﬁlter to retain the physiologically relevant

frequency content of the EMG signals (Raez et al.,

2006).

Following ﬁltering, full-wave rectiﬁcation was

performed using a centered moving window of 100

ms. The rectiﬁed signal was then smoothed using a

Savitzky–Golay ﬁlter (window size: 100 ms; polyno-

mial order: 1) to further enhance signal clarity while

preserving temporal features.

Finally, all EMG signals were normalized to the

maximum EMG amplitude recorded during each ath-

lete’s respective on-land reference trial. This normal-

ization allowed for inter-participant comparisons and

reduced variability due to individual differences in

signal amplitude.

Both boat and paddle accelerometer readings in

all axes were preprocessed and smoothed with a Sav-

itzky–Golay ﬁlter (window size: 50 ms, polynomial

order: 2). High frequency noise does not interfere

with resulting measurements.

To ensure uniformity across recordings and facil-

itate subsequent analysis, all sensor data were resam-

pled to a standardized frequency of 1 kHz. This re-

sampling was carried out using linear interpolation,

which estimates intermediate data points to produce

evenly spaced time series, thereby allowing consistent

comparison and processing across all runs.

3.3 Outlier Filtering

Outlier strokes were removed from the analysis using

the interquartile range (IQR) method applied to sev-

eral key stroke metrics. A multiplier of 2 was used to

deﬁne the lower and upper bounds for outlier detec-

tion, identifying strokes with values lying outside this

range as outliers. The measurements considered in-

cluded stroke duration (time taken to complete a sin-

gle stroke), tempo duration (time taken to complete

a full left-right stroke cycle), air phase duration, and

the durations of the catch, pull, and recovery phases.

Out of 1509 recorded strokes across all athlete runs,

125 strokes were identiﬁed as outliers and excluded,

resulting in 1384 strokes retained for analysis.

4 MODELING

4.1 Target Variable

The target variable was computed using synchronized

data from both the boat and paddle accelerometers.

Stroke segmentation was performed using the boat

accelerometer’s x-axis — aligned with the kayak’s

heading — by identifying zero-crossings in the for-

ward acceleration signal, which marked the beginning

of new stroke segments. To determine stroke side, we

used the paddle accelerometer axis that was parallel to

the paddle shaft. For each segmented stroke, we cal-

culated the mean acceleration along this axis. Based

on the measurement orientation, strokes with negative

mean acceleration values were labeled as left-handed,

while those with positive values were labeled as right-

handed.

4.2 Training and Test Dataset

Preparation

A total of 74 runs were collected from nine athletes

across varying intensities. Of these, 44 runs were used

for model training, feature selection, and parameter

optimization, while the remaining 30 runs were re-

served for ﬁnal testing. Each subset maintained a rep-

resentative distribution of stroke paces, ranging from

80 to 120 strokes per minute, ensuring consistent pro-

portions across all pace classes. Additionally, data

from all athletes were included in both training and

testing sets to preserve subject diversity. This bal-

anced sampling approach ensured that model evalu-

ation reﬂected a comprehensive range of performance

conditions.

We employed a 6-fold grouped cross-validation

across 18 available training runs. Grouping ensured

that all samples from the same run (i.e., from the

same subject and session) remained entirely within ei-

ther the training, validation or test set for each fold,

thereby preventing data leakage. This approach is

Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data

259

particularly appropriate when dealing with repeated

measures or correlated observations — such as multi-

ple recordings from the same athlete — as it reduces

the risk of overﬁtting to individual-speciﬁc patterns.

Furthermore, non-overlapping runs of varied intensi-

ties were selected to enhance independence between

training and validation subsets, improving generaliza-

tion to unseen conditions.

We have chosen accuracy as the error metric for

all because of interpretability. We also employed

the error metric mean absolute stroke on-set differ-

ence (MAEOSD): that is the mean absolute difference

measured in samples between the true onset of a suc-

cessive stroke and the predicted onset of the stroke.

4.3 Models

We compare results of 3 classiﬁers: Gaussian Naive

Bayes, Logistic Regression, Gradient Boosting Deci-

sion Tree.

Gaussian Naive Bayes (NB) is a probabilistic clas-

siﬁer based on Bayes’ Theorem, assuming indepen-

dence between features and that they follow a normal

(Gaussian) distribution. It is simple, fast, and works

well with small datasets or when the independence as-

sumption roughly holds.

Logistic Regression (LR) is a linear model used

for binary classiﬁcation, which estimates the proba-

bility of a class label using the logistic (sigmoid) func-

tion. It assumes a linear relationship between input

features and the log-odds of the outcome, and it’s ef-

fective for well-separated classes with linearly sepa-

rable boundaries.

Gradient Boosting Decision Trees (GBDT) is an

ensemble method that builds a series of decision

trees sequentially, where each new tree corrects errors

made by the previous ones. It combines the strengths

of multiple weak learners to produce a strong classi-

ﬁer, often achieving high accuracy on complex, non-

linear datasets. GBDT is more computationally inten-

sive but generally offers superior performance com-

pared to simpler models. (Chen and Guestrin, 2016)

A Gaussian Naive Bayes classiﬁer was employed

as a baseline to benchmark the performance of more

complex models. Performance metrics are reported as

the average classiﬁcation accuracy across six cross-

validation folds and on the held-out test set.

The classiﬁers operated on a per-sample basis, as-

signing a conﬁdence score to each time point, indicat-

ing the likelihood of it corresponding to a left or right-

hand stroke. Training was also conducted at the per-

sample level, enabling the model to learn ﬁne-grained

temporal patterns speciﬁc to stroke side. Accordingly,

the accuracy metric can be interpreted as the propor-

tion of time during which the classiﬁer correctly iden-

tiﬁed the stroke side.

4.4 Feature Selection

To reduce dimensionality and improve performance

feature selection techniques were implemented: for

Naive Bayes and Logistic Regression forward selec-

tion and backward elimination were both used. These

stepwise feature selection techniques commonly used

in machine learning. Forward selection begins with

an empty model and iteratively adds the feature that

improves the error metric (accuracy in this case). In

contrast, backward elimination starts with the full set

of features and progressively removes the least signif-

icant ones, continuing until no further improvement in

model performance is observed. We used backward

elimination on the feature set obtained with forward

selection to reduce the feature space. Once no further

improvement was observed, Greedy Backward Elim-

ination was applied to remove any features whose ex-

clusion resulted in improved accuracy. This process

of forward selection followed by backward elimina-

tion continued iteratively until no additional modiﬁ-

cations improved model performance.

4.5 Feature Importance

Following feature selection, permutation importance

was employed to further evaluate and interpret feature

contributions. This model-agnostic technique esti-

mates feature importance by randomly permuting fea-

ture values and measuring the resulting drop in per-

formance. Features causing greater reductions in ac-

curacy were deemed more important. This provided

a robust and interpretable measure of feature rele-

vance, especially beneﬁcial for understanding com-

plex or non-linear models, but also helps with simple

models that do not provide an intrinsic feature impor-

tance such as Naive Bayes. Features with negative

permutation importance scores were discarded.

5 RESULTS

5.1 Naive Bayes

The ﬁnal selected feature set contained 25 features

and achieved a cross-validation accuracy of 94.71%

± 0.01 and 94.48% on the test set. Among all

features, the most inﬂuential were the normalized

Z-coordinates of the left olecranon and both left

and right ulnar styloid markers. Remarkably, using

just these three features, the Naive Bayes classiﬁer

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

260

achieved 92.21% accuracy in cross-validation. The

additional 23 features contributed a modest 2.5% im-

provement, suggesting that the model performs effec-

tively even with a very minimal feature set. This ﬁnd-

ing has practical implications for applications where

data collection capabilities are limited. This ﬁnding

shows that using only joint coordinates would be suf-

ﬁcient to classify stroke sides.

5.2 Logistic Regression

The ﬁnal model contained 28 features the validation

set accuracy was 94.38% ± 0.005, slightly worse than

the NB classiﬁer and achieved 94.53% on the test

set. The model used features from all three fea-

ture sets. The ﬁnal model parameters used were L2-

penalty, C=100.0 (inverse of regularization strength,

smaller values specify stronger regularization). Inter-

estingly the top-5 most important features would pro-

duce a 93.46% ± 0.01 accuracy, and the top-3 would

only achieve 88.72% ± 0.02. The top 5 features in-

clude T8-right upper arm lateral bending and verti-

cal pelvis axial bending, along with normalized EMG

readings of the latissimus dorsi and the right olecra-

non Z-coordinate.

5.3 Gradient Boosting Decision Tree

We ﬁrst trained a base model of GBDT with all fea-

tures and selected the top-50 most important features,

after that we gradually decreased the number of fea-

tures until accuracy improved. The ﬁnal model used

33 features. After hyperparameter tuning, the best

validation accuracy was achieved with: 15 leaves, a

learning rate of 0.050, colsample = 0.75, and subsam-

ple = 0.50. The best model achieved 96.02%± 0.01

on the validation set and 96.18% on the test set. The

top-5 most important features produce a model with

94.83% ± 0.01 accuracy. The used features are left

ulnar stiloid Z-coordinate and right top of hand Z-

coordinate, with left and right latissimus dorsi and left

shoulder abduction adduction joint angles.

5.4 Feature Importances

A core set of features (Figure 1) was identiﬁed as im-

portant across all three models, including the normal-

ized left ulnar styloid Z coordinate, right olecranon

Z coordinate, bilateral latissimus dorsi EMG activity,

vertical pelvis axial bending, left ﬂexor carpi radialis

EMG activity, bilateral shoulder ﬂexion/extension,

and bilateral knee abduction/adduction. When ex-

panding the analysis to features selected by at least

two models, additional relevant variables include

T8–right upper arm lateral bending, right ulnar styloid

(particularly important for the Naive Bayes model),

and left shoulder abduction/adduction, along with

several lower-importance features. Notably, in the

best-performing GBDT model, the right top of hand

Z coordinate had high relative importance, despite not

being selected by the other models.

The identiﬁed feature set reveals a clear anatomi-

cal and functional grouping relevant to kayaking per-

formance. Upper limb features, including wrist and

elbow positions, shoulder joint angles, and upper-

body EMG activity, contribute primarily to paddle

control and propulsion. Trunk features, such as verti-

cal pelvis axial bending and upper spinal lateral bend-

ing, reﬂect core stability and rotational control dur-

ing the stroke. Lower limb features, notably knee ab-

duction/adduction, indicate the role of leg position-

ing in anchoring and force transfer. These ﬁndings

highlight the interdependence of upper body mechan-

ics, trunk stability, and lower limb engagement in

accurately distinguishing stroke sides using machine

learning models.

Figure 1: List of most important features by model.

In terms of feature utilization, both NB and LR

models rely on a smaller but more discriminative fea-

ture set, while GBDT makes use of a broader range of

features—likely due to its use of random resampling

during training. Interestingly, the NB model showed

strong performance using only coordinate-based fea-

tures, suggesting that, in resource-constrained scenar-

ios, coordinate data alone may sufﬁce for side classiﬁ-

cation. GBDT achieved the highest overall accuracy,

beneﬁting from its ability to model feature interac-

tions, though this advantage resulted in only a 2.5%

improvement. While NB slightly outperformed LR in

the validation set and underperformed it in the test set,

the marginal accuracy gain does not justify its higher

complexity.

Tempo-based relative time averages — calculated

across all available data (training and test sets com-

bined) — are shown in Figure 2 for the ﬁve most im-

portant features: normalized left ulnar styloid Z coor-

Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data

261

dinate, right olecranon Z coordinate, bilateral latis-

simus dorsi EMG activity, and vertical pelvis axial

bending. The blue vertical line indicates the midpoint

of the tempo, separating right and left hand strokes.

These features consistently ranked highest in impor-

tance across all classiﬁers, reﬂecting their reliability

and discriminative power for stroke side classiﬁca-

tion.

The temporal proﬁles of these features reveal dis-

tinct and interpretable biomechanical patterns. For

example, the bilateral latissimus dorsi muscles show

unilateral activation that aligns with the stroke side,

being active only during the corresponding stroke

phase. Shoulder ﬂexion/extension patterns also show

a clear alternation across the tempo midpoint, with the

left shoulder exhibiting positive values during right-

hand strokes and negative values during left-hand

strokes. Notably, the normalized Z coordinates of

the left ulnar styloid and right olecranon cross over

at the stroke transition. The right olecranon Z coor-

dinate remains below 0.25 during right strokes and

rises above 0.25 during left strokes, while the left ul-

nar styloid shows the inverse pattern. These coordi-

nated shifts across features illustrate a synchronized

change in posture and muscle engagement, reinforc-

ing their signiﬁcance in stroke classiﬁcation. These

proﬁles may be used as templates for stroke separa-

tion and to identify systematic errors in athletes.

Figure 2: Tempo relative time feature averages.

5.5 Statistical Analysis of Errors

Model performances are summarized in Table 1.

To evaluate the differences in classiﬁcation accuracy

among the three models—Logistic Regression (LR),

Naive Bayes (NB), and Gradient Boosting Decision

Trees (GBDT)—we conducted statistical tests on the

cross-validation results. Normality of the pairwise

differences in model accuracies was assessed using

the Shapiro-Wilk test. The results indicated no sig-

niﬁcant deviation from normality for any model pair:

LR vs. NB (W=0.922, p=0.543), NB vs. GBDT

(W=0.899, p=0.404), and LR vs. GBDT (W=0.999,

p=0.9998). These results justify the use of parametric

tests for further comparisons.

Three paired-sample t-tests were performed to

compare model accuracies. There was no statistically

signiﬁcant difference between LR and NB (t=-0.964,

p=0.390). However, GBDT showed signiﬁcantly

higher accuracy compared to both LR (t=13.816,

p=0.0001) and NB (t=7.564, p=0.0016). These ﬁnd-

ings suggest that GBDT outperforms both LR and NB

in classiﬁcation accuracy under the tested conditions.

Table 1: Stroke side classiﬁcation accuracy of machine

learning models on validation and test set (%).

Accuracy

Model Val. set Test set

NB 94.71 94.48

LR 94.38 94.53

GBDT 96.02 96.18

Table 2 summarizes the MAEOSD of the classi-

ﬁers in milliseconds. As expected, the GBDT model

achieved the best performance, with a mean absolute

onset difference of 24.6 ± 51.6 ms. Considering that

the average stroke duration in the test set is 639.6

± 88.4 ms, this corresponds to an average onset de-

tection error of only 4.5%, highlighting the GBDT

model’s suitability for real-time classiﬁcation scenar-

ios. The results show that the GBDT model achieves

tolerable stroke onset duration error.

Table 2: Mean absolute onset differences on the test set

(ms).

Model MAEOSD

NB 36.5 ± 56.4

LR 35.5 ± 49.7

GBDT 24.6 ± 51.6

Per-sample predictions were also aggregated at

the stroke level to evaluate whether any entire strokes

were misclassiﬁed. On the validation set, 14 mis-

classiﬁed strokes were observed for both the NB and

LR models, while the GBDT model misclassiﬁed 12

strokes. Notably, no misclassiﬁed strokes were found

in the test set for any of the models. These results

indicate that all three models achieve high aggregate

accuracy at the stroke level.

Figure 3 illustrates the classiﬁcation accuracy of

left and right hand strokes across the normalized

stroke cycle for the GBDT model. The two curves are

nearly identical, indicating symmetrical model per-

formance between stroke sides. As anticipated, clas-

siﬁcation accuracy declines in the ﬁrst and last 10%

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

262

of the stroke cycle — regions associated with greater

motion variability and transitional dynamics. This ob-

servation aligns with earlier ﬁndings on stroke onset

detection error, which is most pronounced at stroke

boundaries. In contrast, the central 80% of the stroke

is consistently classiﬁed with high accuracy, suggest-

ing that mid-stroke movement patterns are more sta-

ble and distinguishable.

Figure 3: Relative time accuracy for left and right hand

strokes of GBDT model.

An increase in stroke pace does not appear to sig-

niﬁcantly impact the overall classiﬁcation efﬁciency

of the models. Interestingly, the 90 strokes per minute

(SPM) condition yields the highest average stroke on-

set detection error and the lowest classiﬁcation ac-

curacy among all tested stroke rates for the best-

performing GBDT model. This decrease in accuracy

is primarily due to an outlier on-water measurement

test that yielded only 88% accuracy. When this outlier

is excluded, stroke pace does not have a statistically

signiﬁcant effect on classiﬁcation accuracy, indicat-

ing that the model generalizes well across a range

of movement speeds and model performance is sta-

ble and robust with respect to variations in stroke fre-

quency.

Figure 4 presents the per-subject classiﬁcation ac-

curacy (x-axis) plotted against the mean absolute on-

set detection error (y-axis). A very strong nega-

tive correlation is observed between these two met-

rics (Pearson’s R=-0.9346). The ﬁtted regression line

has a slope of -8.55, meaning that for every 1% in-

crease in classiﬁcation accuracy, the mean absolute

onset detection error decreases by approximately 8.55

milliseconds. This result suggests that improvements

in classiﬁcation performance are directly associated

with greater temporal precision in detecting stroke on-

set.

In practical terms, subjects for whom the model

struggles to correctly classify stroke sides also tend

to show larger deviations in onset timing. This high-

lights the interdependence of spatial and temporal as-

Figure 4: Per-subject accuracy and mean absolute onset dif-

ference (outlier removed).

pects of stroke recognition, and suggests that efforts

to improve classiﬁcation accuracy may also yield ben-

eﬁts in onset timing accuracy.

This relationship is further supported by the trends

observed in Figure 3, which shows classiﬁcation ac-

curacy across the normalized time course of a stroke.

Accuracy is lowest at the earliest (0–10%) and latest

(90–100%) segments of a stroke, where motion tends

to be more variable and transitions are more subtle.

These boundary regions likely contribute dispropor-

tionately to onset detection error, reinforcing the con-

nection between reduced accuracy and timing impre-

cision.

The primary factor affecting classiﬁcation error

appears to be inter-subject variability, with accuracy

rates differing signiﬁcantly between athletes, ranging

from 91.89% to 98.9%, when the outlier is removed.

This variability likely reﬂects individual differ-

ences in technique and style concentrated in the tran-

sitional period between two strokes, which warrants

further investigation in future research.

We hypothesize that subjects exhibiting lower ac-

curacy rates demonstrate greater variability in the fea-

tures utilized by the GBDT model. Such variabil-

ity may arise from inconsistent stroke techniques, a

wider range of hand movement patterns, or variations

in muscle activation timing and intensity, all of which

can challenge the model’s ability to generalize and

accurately classify stroke sides. Conversely, athletes

with a more stable and consistent style are likely to

produce more homogeneous training data, resulting in

higher classiﬁcation accuracy. Furthermore, we pro-

pose that athletes with higher classiﬁcation accuracy

may display a reduced number of distinct muscle ac-

tivation sequences, potentially due to less variability

in technique or lower muscle fatigue. These insights

emphasize the importance of individual biomechani-

cal factors in classiﬁer performance and highlight av-

enues for personalized model optimization.

Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data

263

6 CONCLUSION

This study evaluated and compared the performance

of three machine learning models — Naive Bayes

(NB), Logistic Regression (LR), and Gradient Boost-

ing Decision Trees (GBDT) — for classifying stroke

sides based on biomechanical and electromyographic

features. A core set of features, primarily normal-

ized joint coordinate data and Latissimus Dorsi EMG

activity, was found to be highly informative across

all models. Remarkably, the NB classiﬁer achieved

strong accuracy using only a minimal feature set con-

sisting of three coordinate-based variables, highlight-

ing its potential utility in scenarios with limited data

acquisition capabilities.

Among the models, GBDT demonstrated the

highest classiﬁcation accuracy (96.69% on the test

set) and the lowest stroke onset detection error (24.6 ±

51.6 ms), indicating its suitability for real-time appli-

cations despite increased model complexity. Logistic

Regression offered a modest improvement over NB

but with only marginal gains relative to the additional

complexity involved. The strong negative correlation

between classiﬁcation accuracy and onset detection

error suggests that improvements in spatial classiﬁ-

cation directly enhance temporal precision.

Stroke pace and side had minimal effect on model

performance, whereas signiﬁcant inter-subject vari-

ability was observed. This variability likely reﬂects

individual differences in stroke technique and mus-

cle activation patterns, which inﬂuence classiﬁer ac-

curacy. Athletes with more consistent movement

patterns exhibited higher accuracy, underscoring the

importance of personalized biomechanical factors in

model generalization.

Importantly, since the most critical features in-

clude joint coordinates, these results suggest that pose

estimation models applied to video feeds could pro-

vide a practical and non-invasive means of acquiring

the necessary data for stroke side classiﬁcation. This

opens the possibility of implementing the classiﬁca-

tion pipeline in real-world settings without the need

for extensive sensor instrumentation.

Finally, we plan to test our hypotheses regarding

muscle activation order and per-subject variability in

future research to better understand their impact on

classiﬁcation performance and to improve model per-

sonalization.

Overall, these ﬁndings demonstrate that effective

stroke side classiﬁcation can be achieved using a rela-

tively small set of biomechanical features, with gradi-

ent boosting methods providing the best performance.

Future work should explore subject-speciﬁc adapta-

tions, further investigate biomechanical sources of

inter-subject variability, and evaluate the integration

of video-based pose estimation for broader applica-

bility.

REFERENCES

Bonaiuto, V., Gatta, G., Romagnoli, C., Boatto, P., Lanotte,

N., and Annino, G. (2020). A pilot study on the e-

kayak system: A wireless daq suited for performance

analysis in ﬂatwater sprint kayaks.

Chen, T. and Guestrin, C. (2016). Xgboost.

Garnier, Y. M., Hilt, P. M., Sirandre, C., Ballay, Y., Lep-

ers, R., and Paizis, C. (2023). Quantifying paddling

kinematics through muscle activation and whole body

coordination during maximal sprints of different du-

rations on a kayak ergometer: A pilot study. Interna-

tional Journal of Environmental Research and Public

Health, 20(3).

Harbour, E., McAlpine, P., and Neville, J. (2021). Pad-

dle mechanics differ between on-water and ergometer

sprint kayaking.

Kim, J., Koo, B., Nam, Y., and Kim, Y. (2021). semg-based

hand posture recognition considering electrode shift,

feature vectors, and posture groups.

Klitgaard, K. K., Hauge, C., Oliveira, A. S., and Heinen,

F. (2021). A kinematic comparison of on-ergometer

and on-water kayaking. European Journal of Sport

Science, 21(10):1375–1384. PMID: 33001757.

Kranzinger, C., Bernhart, S., Kremser, W., Venek, V.,

Rieser, H., Mayr, S., and Kranzinger, S. (2023). Clas-

siﬁcation of human motion data based on inertial mea-

surement units in sports: A scoping review.

Lauder, M. and Kemecsey, I. (1999). Kayak technique di-

agnosis and remedies, part two. Canoe Focus, pages

18–19.

Li, M. (2017). The progress of biomechanical researches in

kayaking.

Liu, L., Wang, H., Qiu, S., Zhang, Y., and Hao, Z.-D.

(2021). Paddle stroke analysis for kayakers using

wearable technologies.

McDonnell, L. K., Hume, P., and Nolte, V. (2012). An

observational model for biomechanical assessment of

sprint kayaking technique.

Raez, M. B., Hussain, M. S., and Mohd-Yasin, F. (2006).

Techniques of emg signal analysis: detection, pro-

cessing, classiﬁcation and applications. Biological

Procedures Online, 8:11–35. Epub 2006 Mar 23.

Erratum in: Biol Proced Online. 2006;8:163. doi:

10.1251/bpo124.

Stegeman, D. and Hermens, H. (2007). Standards for suface

electromyography: The european project surface emg

for non-invasive assessment of muscles (seniam). 1.

Technologies, M. (2025). Xsens biomechanical model.

Vasiljev, R., Vasiljev, I., Voronin, D. I., and Zhigalina, A.

(2024). Features of pressure asymmetry manifestation

on the seat of kayakers on a rowing ergometer.

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

264