Smartphone Inertial Sensors in Gait Analysis: A Comparison with a

Commercial Device

Marco Oliveira

1 a

, William Fr

ohlich

1 b

, Rafael Baptista

1 c

, Sandro Rigo

2 d

and C

esar Marcon

1 e

School of Technology, Pontiﬁcal Catholic University of Rio Grande do Sul (PUCRS), Porto Alegre, Brazil

Universidade do Vale do Rio dos Sinos (UNISINOS), S

ao Leopoldo, Brazil

Keywords:

Gait, Inertial Sensors, Biomechanics, Smartphone, Wearable Sensors.

Abstract:

Human gait analysis is a crucial tool in healthcare, providing valuable insights into an individual’s well-being,

as various disorders and diseases can be detected through changes in walking patterns. This study aims to

validate the gait sensing results obtained from a smartphone, an easily accessible and portable device, by

comparing them with equivalent data from the G-Walk, a widely used commercial equipment. The goal is to

assess the applicability and accuracy of the solution with the support of healthcare professionals, ensuring its

effectiveness in clinical settings.

1 INTRODUCTION

This paper presents a detailed investigation of using

inertial sensors embedded in smartphones, such as ac-

celerometers and gyroscopes, for human gait analysis.

The study compares the sensing results obtained from

a smartphone with those obtained from a commercial

device, speciﬁcally the G-Walk (BTS Bioengineering,

2024) system, a widely recognized tool for biome-

chanics assessments. Our validation goal is to deter-

mine whether smartphone data can be utilized to de-

velop an algorithm capable of analyzing gait patterns,

serving as a reliable substitute for established tools.

G-Walk is used in clinical settings to evaluate

movement and gait, serving as a benchmark for re-

liability in motion analysis. The validation devel-

oped during this paper focuses on assessing the com-

parability of data collected by the smartphone with

G-Walk measurements across various parameters. If

smartphones provide sufﬁciently accurate data for the

future development of algorithms designed for gait

pattern recognition, they could enable routine mon-

itoring of gait-related health conditions in everyday

settings, transforming how gait analysis is performed

and expanding its applications in healthcare.

https://orcid.org/0009-0003-7625-0735

https://orcid.org/0000-0003-3551-2623

https://orcid.org/0000-0003-1937-6393

https://orcid.org/0000-0001-8140-5621

https://orcid.org/0000-0002-7811-7896

2 RELATED WORK

Gait analysis is a valuable tool in healthcare, tradi-

tionally relying on methods such as optical motion

capture systems, force platforms, and pressure sen-

sors. While these methods are considered highly ac-

curate, they are often expensive and require special-

ized setups that normally involve complex data pro-

cessing, limiting their use outside clinical environ-

ments (Hausdorff and Alexander, 2005).

Wearable technologies and smartphones, how-

ever, typically feature embedded Inertial Mea-

surement Units (IMUs)—accelerometers and gyro-

scopes—which serve as affordable and portable al-

ternatives for gait analysis. Their accessibility allows

for broader use beyond clinical settings. Research has

demonstrated that smartphone accelerometer and gy-

roscope data can yield gait analysis results compara-

ble to those produced by traditional methods (Caro-

Alvaro et al., 2024).

The quality of data from smartphone sensors can

signiﬁcantly impact the accuracy of gait analysis. To

mitigate this, preprocessing techniques such as noise

ﬁltering, signal normalization, and feature extraction

are applied to make raw sensor data more usable.

Standard ﬁltering methods, like the Kalman ﬁlter,

help reduce noise (Kim et al., 2018), while frequency-

domain techniques such as the Fourier Transform

aid in identifying gait patterns within acceleration

data (Shi et al., 2023). Additionally, studies have

shown the effectiveness of preprocessing across var-

ious walking conditions and sensor placements (Ar-

shad et al., 2022) (Ramli et al., 2024).

Oliveira, M., Fröhlich, W., Baptista, R., Rigo, S. and Marcon, C.

Smartphone Inertial Sensors in Gait Analysis: A Comparison with a Commercial Device.

DOI: 10.5220/0013172700003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 595-602

ISBN: 978-989-758-731-3; ISSN: 2184-4305

595

Machine learning techniques have further en-

hanced gait analysis by enabling the automated de-

tection of gait events. Commonly used algorithms

in this ﬁeld include Convolutional Neural Networks

(CNNs) and Long Short-Term Memory (LSTM) net-

works, which are particularly well-suited for analyz-

ing time-series sensor data. LSTM networks, in par-

ticular, can outperform traditional techniques in rec-

ognizing gait phases (Huang et al., 2021) (Marimon

et al., 2024), and combining CNNs with LSTM net-

works has proven effective in classifying complex gait

patterns (Das et al., 2022).

While techniques used in G-Walk (BTS Bioengi-

neering, 2024) offer highly detailed data using spe-

cialized hardware, smartphone-based gait analysis

aims to replicate this precision at a lower cost. Pre-

vious studies have found that calibrated smartphone

sensors can closely match the performance of widely

used commercial equipment when used under con-

trolled conditions (Lopez-Nava et al., 2020). How-

ever, several challenges remain for smartphone-based

gait analysis to achieve the accuracy of traditional

methods. Factors such as variations in sensor qual-

ity across smartphone models, noise from different

sensor attachment methods, and environmental inﬂu-

ences like walking speed or uneven surfaces can af-

fect the results. Ongoing validation, calibration, and

algorithm improvements are necessary to ensure the

reliability of smartphone-based gait analysis in real-

world scenarios (Kocuvan et al., 2023).

This study builds on previous research by com-

paring smartphone sensor data with the G-Walk sys-

tem to assess the viability of smartphones as data-

gathering tools for gait monitoring.

3 MATERIALS AND METHODS

This section outlines the data collection, processing,

and analysis procedure to ensure the study follows a

robust methodological framework. All essential ele-

ments were carefully addressed to enable the valida-

tion of the smartphone-based method as a reliable al-

ternative to established systems used in gait analysis.

3.1 Sensors Speciﬁcations

Comparing the hardware speciﬁcations of the sensors

was essential for an initial compatibility evaluation,

ensuring that both devices have a similar capacity for

capturing data. This subsection presents the main

speciﬁcations of the sensors.

The ’device info’ function in Phyphox (Staacks

et al., 2018) was used to discover the hardware con-

tained in the smartphone. The main sensor is the

LSM6DSO from STMicroelectronics. Its speciﬁca-

tions are as follows (STMicroelectronics, 2024):

• Accelerometer: 16-bit/axis triaxial accelerome-

ter with a full-scale range of ±2, ±4, ±8, ±16 g

• Gyroscope: 16-bit/axis triaxial gyroscope with

a full-scale range of ±125, ±250, ±500, ±1000,

±2000 °/s

The G-Walk has three main sensors, but since this

paper does not address the usage of the magnetome-

ter, the speciﬁcations of the two sensors used are as

follows (BTS Bioengineering, 2024):

• Accelerometer: 16-bit/axis triaxial accelerome-

ter with a full-scale range of ±2, ±4, ±8, ±16 g

• Gyroscope: 16-bit/axis triaxial gyroscope with a

full-scale range of ±250, ±500, ±1000, ±2000 °/s

An initial evaluation shows that both devices

have identical accelerometer hardware speciﬁcations.

However, the main difference for the gyroscope is that

the LSM6DSO is better suited for measuring low-

speed rotations, as it offers an additional range of

±125 degrees per second.

3.2 Data Gathering

The initial step involved collecting raw gait data from

our reference commercial system using inertial sen-

sors from both the smartphone and G-Walk.

The smartphone application Phyphox (Staacks

et al., 2018) captures linear acceleration and angu-

lar velocity. Phyphox is an open-source tool devel-

oped by RWTH Aachen University, allowing access

to all available sensor data on smartphones. The pri-

mary sensors used in this study include accelerome-

ters, which measure linear acceleration (in m/s²), and

gyroscopes, which measure angular velocity (in °/s).

Figure 1: Methodology used in the trial.

Figure 1 shows that both devices—the smartphone

and G-Walk—were mounted on the subject’s waist

HEALTHINF 2025 - 18th International Conference on Health Informatics

596

using a running belt. This placement was selected

because it is the standard position for using the G-

Walk system, allowing for direct comparisons. Ad-

ditionally, it ensures consistency in the data collec-

tion process, as this location effectively captures hip

motion. Both devices recorded data simultaneously

while walking to minimize variations caused by sep-

arate data collection sessions.

The trial consisted of a single subject walking reg-

ularly along a straight path. It lasted approximately 40

seconds, during which 60 steps were recorded. This

number was determined empirically by analyzing a

video recorded during data collection. The video was

also used for time synchronization purposes.

The data was saved in a structured format. Phy-

phox generated a spreadsheet containing the smart-

phone data, and the G-Walk system provided a text

ﬁle with the recorded signals and a report summariz-

ing the gait analysis.

3.3 Preprocessing

Several steps were taken to prepare the data for fur-

ther analysis, primarily because the data from Phy-

phox and G-Walk were in different formats. The ﬁrst

step was to ensure compatibility between the datasets

by renaming columns, converting units, and ensuring

consistent time references.

The initial timestamps were discrepant since the

recordings were manually started on both devices. An

offset correction was applied to align the timestamps,

ensuring both datasets had a common structure for

more straightforward analysis. Additionally, an ex-

tra offset adjustment was made to ensure that the zero

level of each axis was consistent across both datasets.

Following the data cleaning and synchronization,

some features were calculated from the raw signals.

The columns below were added to both datasets:

• Absolute acceleration - calculated as the magni-

tude of the acceleration vector, combining the X,

Y, and Z axes;

• Absolute angular velocity - similar to absolute

acceleration, but for the gyroscope data;

• Roll, Pitch, and Yaw - obtained by integrating the

angular velocity over time, allowing the analysis

of rotational motion.

3.4 Visual Comparison

Before conducting a more advanced statistical analy-

sis, a visual comparison was used to evaluate the col-

lected data preliminary. This approach provided an

intuitive way to assess the similarity between the sig-

nals captured from the smartphone and those from the

G-Walk system.

The primary method involved creating ﬁgures to

overlay the smartphone and G-Walk time-series data:

acceleration and angular velocity. Speciﬁc data seg-

ments were enlarged to focus on individual gait cycles

to facilitate this comparison. These zoomed-in ﬁgures

provided a closer look at the alignment of peaks and

the shape of the waveforms, helping to identify mi-

nor differences that might not be as evident in full-

length plots. The zoomed-in ﬁgure for the accelerom-

eter data is presented in Figure 2, while the zoomed-in

ﬁgure for the gyroscope data is shown in Figure 3.

Figure 2: Accelerometer time series zoomed-in.

Figure 3: Gyroscope time series zoomed-in.

3.5 Statistical/Hypothesis Test

This section outlines the steps taken for data prepro-

cessing and the statistical methods used to compare

the datasets collected by the wearable device and the

smartphone.

3.5.1 Data Preprocessing

Initially, the raw data underwent preprocessing to en-

sure consistency and comparability between the de-

vices. Invalid values recorded during data collection

were ﬁrst removed. Next, interpolation techniques

were applied to address missing data (NaN values)

and ﬁll gaps in the time series. A low-pass Butter-

worth ﬁlter was then used to eliminate high-frequency

noise that could distort the signal and affect the qual-

ity of subsequent analyses. The StandardScaler nor-

malization technique was applied to standardize the

datasets, ensuring that both were on the same scale,

which is critical for meaningful comparisons.

The two devices had different sampling rates, so

the datasets were synchronized using a temporal ap-

proximation technique. This alignment was essential

to ensure that data points from both devices corre-

sponded to the same events in time, allowing for an

accurate comparison between the specialized sensors

of the wearable device and the general-purpose IMU

sensors of the smartphone.

Smartphone Inertial Sensors in Gait Analysis: A Comparison with a Commercial Device

597

3.5.2 Correlation and Cross-Correlation

We applied correlation techniques to assess the simi-

larity between the data from the two devices. Pearson

correlation was used to measure the linear relation-

ship between the datasets. Pearson is ideal when the

data follows a normal distribution and a linear rela-

tionship between variables is assumed. However, this

method is sensitive to outliers, which can dispropor-

tionately affect the correlation coefﬁcient, potentially

leading to skewed results.

In addition to Pearson, we applied Spearman cor-

relation to evaluate cases where the relationship be-

tween variables might not be strictly linear. Spear-

man is a non-parametric measure of rank correlation,

assessing how well the relationship between two vari-

ables can be described using a monotonic function.

This method does not assume normality or linearity;

it is less sensitive to outliers, making it suitable for

scenarios where the data might contain anomalies or

the relationship between the datasets is non-linear.

We captured linear and non-linear relationships be-

tween the two data sources by combining Pearson and

Spearman correlations.

We also applied cross-correlation to analyze the

temporal alignment between the signals. This tech-

nique identiﬁes the time shift that maximizes the sim-

ilarity between the two signals, enabling us to detect

potential delays or desynchronization between the de-

vices. This is particularly relevant for cyclic events

such as gait, where slight timing discrepancies could

affect the data interpretation.

3.5.3 Bland-Altman Analysis

We employed the Bland-Altman analysis to assess

the agreement between the wearable and smartphone

measurements. This method plots the difference be-

tween the two measurements against their mean, al-

lowing for identifying biases and limits of agreement.

The Bland-Altman plot is handy for detecting system-

atic differences (bias) between methods, something

that correlation analysis alone may fail to reveal. For

instance, even when two methods show high correla-

tion, they may still exhibit a consistent bias, evident

in the Bland-Altman plot but not in the correlation co-

efﬁcient.

3.5.4 Error Analysis: RMSE and MAE

In addition to these qualitative assessments, we quan-

tiﬁed the magnitude of the errors between the two

devices using Root Mean Square Error (RMSE) and

Mean Absolute Error (MAE). RMSE emphasizes

more signiﬁcant deviations by taking the square root

of the average of the squared differences between the

two sets of measurements, making it more sensitive

to substantial errors. In contrast, MAE provides a

straightforward interpretation of the average absolute

error, offering a more robust measure in the presence

of outliers. These error metrics are crucial for evalu-

ating the overall accuracy of the devices in capturing

similar measurements.

3.5.5 Lin’s Concordance Correlation Coefﬁcient

To provide a more robust evaluation of concordance

between the devices, we also computed Lin’s Con-

cordance Correlation Coefﬁcient (CCC). This metric

combines measures of correlation and agreement, as-

sessing how closely the measurements from the two

devices align in trend and magnitude. Lin’s CCC is

beneﬁcial when the goal is to determine if the two

methods are correlated and if their values are similar

in absolute terms.

3.5.6 Fast Fourier Transform Analysis

Lastly, we performed a Fast Fourier Transform (FFT)

analysis to compare the frequency content of the sig-

nals from both devices. FFT decomposes the time-

domain signals into frequency components, enabling

us to assess whether the devices captured the same

dominant frequencies in the gait cycle. Frequencies

related to step cadence and stride patterns are critical

for gait analysis. Differences in the frequency com-

ponents between the two datasets may indicate that

the devices are not capturing the movement dynamics

with the same precision, which could be necessary de-

pending on the intended application of the data.

4 RESULTS AND DISCUSSION

4.1 Correlation and Cross-Correlation

When analyzing Pearson and Spearman correlations

between wearable sensors and smartphones, we ob-

serve varying levels of correlation across acceleration

variables (acc), gyroscope variables (gyro), and rota-

tion angles (roll, pitch, yaw).

From an initial analysis using Pearson correla-

tions, as seen in Figure 4, which measures linear re-

lationships between variables, we ﬁnd high correla-

tions for most acceleration and gyroscope axes. For

instance, the absolute acceleration (acc abs) shows a

strong correlation of 0.931, suggesting that wearable

devices and smartphones similarly capture overall ac-

celeration magnitude. Gyroscope values exhibit even

HEALTHINF 2025 - 18th International Conference on Health Informatics

598

higher correlations, such as gyro x (0.951), gyro z

(0.911), and absolute gyroscope (gyro abs, 0.944), in-

dicating substantial agreement between the sensors in

measuring rotational movement. The highest correla-

tions are found in roll (0.99996) and yaw (0.99976),

reﬂecting a near-perfect agreement in these rotational

angles. Yet, a fundamental exception is the low cor-

relation in the Y-axis, particularly for the gyroscope,

where gyro y shows a correlation of just 0.165. This

suggests a signiﬁcant discrepancy between the wear-

able and smartphone measurements on the Y-axis for

gyroscope data. The accelerometer on the Y-axis also

has the lowest correlation among the acceleration axes

but performs signiﬁcantly better than gyro y, with a

Pearson correlation coefﬁcient of 0.779.

Figure 4: Pearson correlation between wearable sensor and

smartphone data across acceleration, gyroscope, and rota-

tion variables.

Figure 5 also displays Spearman correlations,

measuring monotonic relationships and generally

conﬁrming the patterns observed with Pearson corre-

lations, albeit with some differences.

Figure 5: Spearman correlation between wearable sensor

and smartphone data across acceleration, gyroscope, and ro-

tation variables.

As expected, the Spearman correlations for ac-

celeration are also high, such as acc x (0.858) and

acc abs (0.928), indicating that the overall trend be-

tween the two devices remains strong even when

considering non-linear relationships. However, for

the gyroscope, Spearman correlations are slightly

lower, especially for gyro x (0.884) and gyro abs

(0.835), suggesting that the devices may detect rota-

tions differently, particularly when considering non-

linear movements.

We observed a consistent pattern in the cross-

correlation analysis across all data points analyzed.

There was either a low negative or low positive cor-

relation for negative lag values, with values gradually

increasing. However, the correlation remained stable

at the observed levels for positive lag values, as shown

in Figure 6.

Figure 6: Cross-correlation analysis for gyro x (left) and

gyro z (right), showing correlation patterns across lag val-

ues.

4.2 Bland-Altman Analysis

Figure 7 displays the results for acc

using the Bland-

Altman technique, which assesses the agreement be-

tween two measurement methods: a wearable device

and a smartphone. The X-axis represents the average

of the measurements between the two methods, while

the Y-axis shows the difference between these mea-

surements.

Figure 7: Bland-Altman plot comparing wearable sensor

and smartphone measurements for acc x.

The central gray line indicates the mean of the

differences, which is slightly below -1. This sug-

gests that the wearable device tends to register slightly

lower values than the smartphone, revealing a slight

Smartphone Inertial Sensors in Gait Analysis: A Comparison with a Commercial Device

599

negative bias, implying that one device may slightly

underestimate the values.

The red dashed lines mark the limits of agree-

ment, which range approximately from -3 to +1. Most

data points fall within these limits, indicating good

agreement between the methods for most of the data.

However, a few points fall outside these limits, repre-

senting outliers where the devices show more signiﬁ-

cant disagreement. Hence, both methods demonstrate

good agreement on the acc x axis, with a slight nega-

tive bias and relatively consistent differences through-

out the measurements.

In contrast, Figure 8 analyzes gyro y using the

Bland-Altman agreement technique. The central gray

line represents the mean of the differences, which is

around 0 but shows a slight upward shift. This indi-

cates a small positive bias, suggesting that the wear-

able device records slightly higher values than the

smartphone.

Figure 8: Bland-Altman plot comparing wearable sensor

and smartphone measurements for gyro y.

The red dashed lines mark the limits of agreement,

encompassing most of the data points. These lim-

its, approximately between -40 and +40, represent the

range where most differences are concentrated, with

a few outliers falling outside these limits.

Furthermore, the data suggests a proportional

bias. The agreement between the two methods is

stronger for smaller measurements, but the differ-

ences also tend to grow as the average values in-

crease. This divergence is particularly noticeable dur-

ing faster rotations or larger movements along the

gyro y axis, indicating a potential limitation of the

wearable device in capturing such movements.

4.3 Error Analysis: MSE, RMSE, and

MAE

Table 1 presents the error metrics analyzed in this

study, comparing the results from data obtained using

a wearable sensor and a smartphone. The table high-

lights three key error metrics: Mean Absolute Error

(MAE), Mean Squared Error (MSE), and Root Mean

Squared Error (RMSE).

Table 1: Error metrics (MAE, MSE, RMSE) comparing

wearable sensor and smartphone data across various vari-

ables.

Variable MAE MSE RMSE

acc x 0.416 0.286 0.535

acc y 0.531 0.441 0.664

acc z 0.455 0.346 0.588

acc abs 0.248 0.137 0.370

gyro x 0.237 0.097 0.312

gyro y 1.005 1.668 1.291

gyro z 0.335 0.177 0.421

gyro abs 0.251 0.111 0.333

pitch 1.544 3.349 1.830

roll 0.007 7.712e-05 0.008

yaw 0.018 0.001 0.021

rpy abs 0.004 3.199e-05 0.005

When analyzing the accelerometer variables, we

observe that the absolute acceleration (acc abs) has

the lowest errors across all metrics, with an MAE of

0.248, MSE of 0.137, and RMSE of 0.370. This in-

dicates that the overall magnitude of acceleration is

well-estimated, demonstrating a solid agreement be-

tween the sensors.

Among the individual accelerometer axes, the

acc x axis performs slightly better with an MAE of

0.416, followed by the acc z axis with an MAE of

0.455. The acc y axis shows the highest error among

the accelerometer axes, with an MAE of 0.531, MSE

of 0.441, and RMSE of 0.664, indicating lower preci-

sion on this axis.

For the gyroscope variables, the gyro x axis stands

out with the lowest errors, showing an MAE of 0.237,

MSE of 0.097, and RMSE of 0.312, followed closely

by gyro abs with similar values. In contrast, the

gyro y axis presents signiﬁcantly higher errors, with

an MAE of 1.005, MSE of 1.668, and RMSE of 1.291,

indicating considerable inaccuracy in gyroscope mea-

surements on this axis. This discrepancy reﬂects a

more signiﬁcant variation between the wearable and

smartphone sensors and aligns with the previously an-

alyzed correlation values.

The rotation angles (pitch, roll, yaw) also exhibit

error variability. The roll angle shows the lowest er-

HEALTHINF 2025 - 18th International Conference on Health Informatics

600

rors, with minimal values (MAE of 0.007), suggest-

ing almost no discrepancy. The yaw angle performs

slightly worse, though still very accurate, with an

MAE of 0.018. On the other hand, the pitch angle has

the highest errors among the rotation variables, with

an MAE of 1.544, MSE of 3.349, and RMSE of 1.830,

indicating more incredible difﬁculty in estimating this

angle.

Overall, the results show strong agreement be-

tween the sensors for most variables, with notable ex-

ceptions for the gyro

axis and the pitch angle, which

present higher error levels.

4.4 Lin’s Concordance Correlation

Figure 9 depicts the results of the Concordance Cor-

relation Coefﬁcient (CCC) for absolute acceleration

(acc abs). We have the actual and predicted values

on the X and Y axes. Each blue dot corresponds to a

pair of actual and predicted values from the dataset,

while the red dashed line represents the identity line,

which indicates perfect agreement.

Figure 9: Lin’s CCC for absolute acceleration (acc abs),

showing agreement between wearable sensor and smart-

phone measurements.

We can discern a strong positive correlation be-

tween the actual and predicted values. Most points

are clustered around the red identity line, indicating a

good ﬁt. However, some variations are present, par-

ticularly at the higher and lower extremes, suggest-

ing potential limitations in extreme conditions. The

spread of the points suggests a certain degree of vari-

ability. Yet, the strong clustering near the identity line

reassures us about the model’s accuracy in capturing

the overall trend, suggesting that the model or sensors

perform well but may show limitations in extreme

conditions, where predictions deviate slightly from

the actual values. Nonetheless, the overall alignment

with the identity line indicates good agreement be-

tween the actual and predicted measurements. How-

ever, Figure 10 shows a strong correlation between the

points. Still, they show a slight offset below the red

identity line, indicating that the data does not exhibit a

perfect concordance correlation, therefore, there may

be variations in magnitude compared to absolute ac-

celeration, demonstrating a closer correlation.

Figure 10: Lin’s CCC for acc z, showing slight deviations

from the identity line.

4.5 Fast Fourier Transform Analysis

Figure 11 shows the ﬁrst analysis with the FFT ap-

plied to acc

. We observed a very sharp peak around

0 Hz in magnitude and frequency, indicating a dom-

inant frequency component near zero and predomi-

nantly low-frequency components, with minimal vari-

ation in magnitude outside the central peak. This sug-

gests that the accelerometer’s X-axis data contains a

signiﬁcant DC component or a slow-moving trend.

The accelerometer on the X-axis may be sensing min-

imal or constant motion over time, leading to this

dominant low-frequency component. The sharp peak

near 0 Hz implies that both datasets recorded mini-

mal acceleration on the X-axis, indicating little to no

signiﬁcant movement or only slow changes.

Figure 12 displays the second signal analyzed -

gyro

y. This ﬁgure shows more distributed peaks cen-

tered around 0 Hz, with visible spikes across the fre-

quency range. We identiﬁed higher frequencies, indi-

cating more variability or dynamic activity in the gy-

roscope’s Y-axis data. The signal’s amplitude is also

more spread, suggesting that the gyroscope detected

rotational movements oscillating at different frequen-

cies. This could result from rotational motion or ori-

entation changes, causing the gyroscope’s Y-axis to

pick up varying frequencies. Both devices recorded

rotational movements on the Y-axis but with different

frequency components.

Smartphone Inertial Sensors in Gait Analysis: A Comparison with a Commercial Device

601

Figure 11: FFT analysis of acc x from wearable sensor and

smartphone data, showing the frequency spectrum.

Figure 12: FFT analysis of gyro y from wearable sensor

and smartphone data, showing dynamic movement across

frequencies.

5 CONCLUSION

Linear correlation provides an overview of relation-

ships between signals but lacks precision in evalu-

ating agreement. In contrast, Bland-Altman plots

and Lin’s Concordance Correlation Coefﬁcient of-

fer a more robust assessment, essential for validating

methods in clinical and biomechanical studies. Error

metrics like RMSE and MAE quantify discrepancies,

while FFT conﬁrms whether devices capture critical

motion frequencies similarly.

Our ﬁndings demonstrate that smartphones with

inertial sensors are highly promising for gait anal-

ysis. They showed strong correlations with the

G-Walk system for variables like absolute accel-

eration (acc abs) and gyroscope data (gyro abs).

Bland-Altman analysis revealed satisfactory agree-

ment overall, though notable discrepancies in the gy-

roscope Y-axis (gyro y) highlight limitations in cap-

turing speciﬁc rotational movements. Error metrics

conﬁrmed these results, with minimal errors in abso-

lute acceleration and larger discrepancies in gyro y.

FFT analysis validated the smartphone’s ability to

represent essential gait cycle phases accurately.

This study establishes smartphones as accessible

alternatives for gait analysis, though improving Y-

axis gyroscope accuracy is critical for complete in-

terchangeability. Future work should focus on devel-

oping algorithms to analyze accelerometer and gyro-

scope data, identify gait phases, and generate detailed

reports for healthcare professionals.

REFERENCES

Arshad, M., Jamsrandorj, A., Kim, J., and Mun, K.-R.

(2022). Gait events prediction using hybrid cnn-rnn-

based deep learning models through a single waist-

worn wearable sensor. Sensors, 22(21):8226.

BTS Bioengineering (2024). G-walk: Wearable inertial sen-

sor for motion analysis.

Caro-Alvaro, S., Garcia-Lopez, E., et al. (2024). Gesture-

based interactions: Integrating accelerometer and gy-

roscope sensors in the use of mobile apps. Sensors,

24(3):1004.

Das, S., Meher, S., and Sahoo, U. K. (2022). A uni-

ﬁed local–global feature extraction network for hu-

man gait recognition using smartphone sensors. Sen-

sors, 22(11):3968.

Hausdorff, J. and Alexander, N. (2005). Gait disorders eval-

uation and management. CRC Press.

Huang, H., Zhou, P., Li, Y., and Sun, F. (2021). A

lightweight attention-based cnn model for efﬁcient

gait recognition with wearable imu sensors. Sensors,

21(8):2866.

Kim, Y., Bang, H., et al. (2018). Introduction to kalman

ﬁlter and its applications. Introduction and Implemen-

tations of the Kalman Filter, 1:1–16.

Kocuvan, P., Hrasti

c, A., Kareska, A., and Gams, M.

(2023). Predicting a fall based on gait anomaly de-

tection: a comparative study of wrist-worn three-axis

and mobile phone-based accelerometer sensors. Sen-

sors, 23(19):8294.

Lopez-Nava, I. H., Valent

ın-Coronado, L. M., Garcia-

Constantino, M., and Favela, J. (2020). Gait ac-

tivity classiﬁcation on unbalanced data from inertial

sensors using shallow and deep learning. Sensors,

20(17):4756.

Marimon, X., Mengual, I., et al. (2024). Kinematic analysis

of human gait in healthy young adults using imu sen-

sors: exploring relevant machine learning features for

clinical applications. Bioengineering, 11(2):105.

Ramli, A., Liu, X., , et al. (2024). Gait characterization in

duchenne muscular dystrophy (dmd) using a single-

sensor accelerometer: classical machine learning and

deep learning approaches. Sensors, 24(4):1123.

Shi, L.-F., Liu, Z.-Y., Zhou, K.-J., Shi, Y., and Jing, X.

(2023). Novel deep learning network for gait recog-

nition using multimodal inertial sensors. Sensors,

23(2):849.

Staacks, S., H

utz, S., Heinke, H., and Stampfer, C. (2018).

Advanced tools for smartphone-based experiments:

phyphox. Physics education, 53(4):045009.

STMicroelectronics (2024). Lsm6dso. Acessado em: 20 de

outubro de 2024.

HEALTHINF 2025 - 18th International Conference on Health Informatics

602