Proceedings of the 10th World Congress on Electrical Engineering and Computer Systems and Sciences (EECSS'24)

Barcelona, Spain – August 19 - 21, 2024

Comparison of Short Fast Fourier Transform and Continuous Wavelet

Abstract – Neurodegenerative diseases (NDD) are a heterogeneous group of complex diseases characterized by neuronal
loss and progressive degeneration of different areas of the nervous system. Gait analysis presents an early recognition system
for NDD which is important to increase the patient’s awareness of their health conditions. However, it is very difficult to
identify and formulate suitable digital biomarkers from the data collected from gait experiments such as stride interval and
swing. The objective of this paper is to compare the result of Short – Time Fourier Transform (STFT) and Continuous
Wavelet Transform (CWT) on the collected stride interval of healthy young people and healthy old people. In this paper,
STFT and CWT are performed on the collected stride interval and from the result of the STFT and CWT, further features
are extracted like instantaneous RMS and maximum RMS value. STFT is performed on the collected stride interval from a
window length of 64 to 512 while CWT is performed on the collected stride interval from the scale of 128 to 2048. The
processing time of the STFT and CWT with varied window lengths and scales respectively are collected. Besides, the actual
maximum time from the time – frequency plot derived from STFT and CWT is also collected. Both STFT and CWT show
that the young group has a higher maximum RMS, an indication of higher stride interval than the old group and higher
variance, an indication of higher gait complexity. The suitable window lengths for STFT in analyzing the stride interval are
64 and 128 while the scale for CWT should be set to the lowest scale. Overall, STFT with a window length of 64 and 128 is
better in analyzing the stride interval due to low processing time at the expense of slightly less accurate time – frequency
representation.

Keywords – Neurodegenerative Diseases (NDD), Gait Analysis, Digital Biomarkers, Short – Time Fourier Transform
(STFT), Continuous Wavelet Transform (CWT)

1. Introduction
Gait analysis is important in diagnosing and rehabilitating gait abnormalities, as well as understanding the physiological
changes associated with aging and treating injuries. Additionally, gait analysis can be used to improve the quality of life for
patients with physical and neurological pathologies, as it helps in understanding the inner workings of the human nervous
system [1], [2]. From 1990 to 2016, the number of people with Parkinson's Disease (PD) doubled to more than 6 million,
with over 200,000 PD-related deaths globally which shows that the study of gait analysis is important [3]. One of the methods
for gait analysis is analyzing the stride interval. Stride interval refers to the time between successive heel strikes of the same
foot [4], [5]. From the literature review, many features are being used to formulate the digital biomarkers from the stride
interval such as time feature, time-frequency feature, non–linear feature, and entropy [6]–[8]. There is no best feature for the
data analysis as it depends on the type of cases and applications. Examples of time-frequency features include STFT and
CWT. Each time-frequency feature has its advantages and disadvantages. STFT is suitable for analyzing non-stationary
signals but uses fixed window lengths for all frequencies. Wavelet transform is sensitive to noise and has high computational
complexity, but its upside is that it uses frequency – a dependent window [9].

ICBES 123-1
2. Method
The stride interval of five healthy old people with a mean age of 74.6 and five healthy young people with a mean
age of 24.4 are collected from the Gait in Aging and Disease Database in Physionet [10]. The healthy old people and
healthy young people were asked to walk in a roughly circular path for 15 minutes. STFT and CWT are performed on
the stride interval to identify the technique that gives the best visualization of the time-frequency plot.

The use of STFT is restrained by the uncertainty principle of time-frequency where a signal is divided into short
pieces. The time resolution and frequency resolution in STFT are uniform because STFT has the same observation time–
width in all frequency ranges.

When 𝑥𝑥(𝑛𝑛) is an input signal, the time-axis sampling function of STFT can be denoted by equation (1).

𝑥𝑥(𝑛𝑛) ⊗ 𝑊𝑊𝑛𝑛 (𝑛𝑛 − 𝑘𝑘. 𝑆𝑆𝑛𝑛 ) (1)

Equation (1) is transformed into Equation (2) after a Fourier Transform is applied.

𝑋𝑋(𝑘𝑘, 𝑖𝑖𝑖𝑖) = [∑ +∞ +∞
𝑛𝑛=−∞ 𝑥𝑥(𝑛𝑛). exp(−𝑗𝑗𝑗𝑗𝑗𝑗) ]. [∑𝑛𝑛=−∞ 𝑊𝑊𝑛𝑛 (𝑛𝑛 − 𝑘𝑘. 𝑆𝑆𝑛𝑛 ). exp(−𝑗𝑗𝑗𝑗𝑗𝑗)] (2)

Based on Equation (1) and (2), 𝑊𝑊𝑛𝑛 is the window function, N is the FFT number, 𝑆𝑆𝑛𝑛 is the time interval shift and 𝑘𝑘
is an integer [11].

In CWT, the observation time-width changes according to the frequency. When the frequency is low, the
observation time-width is lengthened which results in a small frequency resolution. When the frequency is high, the time
resolution becomes small which results in a large frequency resolution. The time resolution and the frequency resolution
of the wavelet transform are not uniform.

The CWT of a discrete-time signal with uniform time intervals is determined by calculating the inner product with
a scaled and shifted mother wavelet as described in Equation (3)[12].
𝑁𝑁−1
�(𝑛𝑛′ − 𝑛𝑛)𝛿𝛿𝑡𝑡 �
𝑊𝑊𝑛𝑛 (𝑠𝑠) = � 𝑥𝑥𝑛𝑛′ 𝜓𝜓 ∗ � � (3)
′
𝑠𝑠
𝑛𝑛 =0

Based on Equation (3), discrete-time signal is denoted as 𝑥𝑥𝑛𝑛 , uniform time intervals is denoted as 𝛿𝛿𝑡𝑡 , inner product
is denoted as 𝑥𝑥𝑛𝑛 and a scaled and shifted mother wavelet is denoted as ψ. 𝑛𝑛′ and 𝑛𝑛 are sequence positions.

ICBES 123-2
3. Results & Discussion

To ensure the result of STFT has a good time – frequency resolution and can derive the instantaneous RMS value
accurately, it is important to select the appropriate window length. The sampling frequency,𝑓𝑓𝑠𝑠 for STFT in this project is
defined by the reciprocal of the difference of consecutive time as shown in Equation (4).

𝑛𝑛 − 1
𝑓𝑓𝑠𝑠 = (4)
∑𝑛𝑛−1
𝑖𝑖=1 (𝑥𝑥𝑖𝑖+1 − 𝑥𝑥𝑖𝑖 )

Based on Equation (4), 𝑛𝑛 is the number of data points, 𝑥𝑥 is the data.

Table 1 shows the time – frequency plot of STFT (hanning window), derived instantaneous RMS of STFT and recorded
processing time with varied window lengths from Old Subject 1. The window length is adjusted from 64 to 512. When the
window length is increased from 64 to 512, the derived instantaneous RMS of STFT converges slowly from a maximally
flat amplitude to a single peak. It can also be seen that when the window length increases from 128, the maximum x – value
(time) has increased and deviated significantly from the original maximum x – value which is 869.79s. Table 2 shows the
actual maximum x – value (time) that is derived from the result of STFT of varied window lengths from Old Subject 1. There
is no noticeable change in recorded processing time when the window length is increased.

Table 1 The Time – Frequency Plot of STFT (hanning window), Derived Instantaneous RMS of STFT and Recorded
Processing Time with Varied Window Length from Old Subject 1
Window Time – frequency Plot of STFT Derived Instantaneous RMS of STFT Recorded Processing
Length Time [sec]
64 0.0656

128 0.0645

ICBES 123-3
 256 0.0704

512 0.0611

Table 2 The Actual Maximum Time Derived from STFT of Varied Window Length from Old Subject 1
Window Length Actual Maximum Time [sec]
64 857.546
128 857.546
256 923.511
512 1055.441

Since the result of STFT of the window length above 128 does not come out with the accurate span of time in time –
frequency plot, the window lengths of 64 and 128 will be chosen for further analysis because the maximum time derived
from STFT is nearest to 869.79s. Figure 1 shows the comparison of the boxplot of maximum RMS in STFT versus the old
group and young group with window lengths of 64 and 128. There is no difference in the overall pattern and trend of the
boxplot in window lengths of 64 and 128. The difference lies in the magnitude of the maximum RMS. The higher the window
length, the smaller the magnitude of the maximum RMS. Therefore, any of the two window lengths, 64 and 128 can be used
to analyze the stride interval in this project with almost the same processing time.

ICBES 123-4
 (a) (b)
Figure 1 The Box Plot of Maximum RMS in STFT versus Old Group and Young Group with (a) Window Length of 64 and
(b) Window Length of 128

Table 3 shows the time – frequency plot of CWT, Derived Instantaneous RMS of CWT and recorded processing time
with varied scales from 128 to 2048 from Old Subject 1. When the scale increases, the derived instantaneous RMS of CWT
converges slowly from multiple peaks to a single peak. Also, the recorded processing time increases when the scale increases.

Table 3: The Time – Frequency Plot of CWT, Derived Instantaneous of CWT and Recorded Processing Time with
Varied Scale from Old Subject 1
Scale Time – frequency Plot of CWT Derived Instantaneous RMS of Recorded Processing Time
CWT [sec]
128 0.695

256 0.825

ICBES 123-5
 512 1.107

1024 2.188

2048 6.409

Based on Table 4, the actual maximum time derived from CWT of varied scale from 128 to 2048 from Old Subject 1 is
all the same which is 869.03s which is nearer to 869.79s than STFT. It shows that CWT produces slightly more accurate
time – frequency visualization than STFT.
Table 4 The Actual Maximum Time Derived from CWT of Varied Scale from Old Subject 1
Scale Actual Maximum Time [sec]
128 869.043
256 869.043
512 869.043
1024 869.043
2048 869.043

ICBES 123-6
 Figure 2 shows the comparison of the boxplot of maximum RMS in CWT versus the old group and young group with a
scale of 128 and 2048. The variance of maximum RMS in CWT of the old group is higher on the scale of 2048 than in 128.
Other than that, there are no differences in the overall pattern and trend on a scale of 128 and 2048. The higher the scale, the
higher the magnitude of maximum RMS in CWT. It is better to use the lower scale for CWT because the processing time
increases with the scale and there are no significant differences between lower scale and higher scale in terms of overall
pattern and trend.

(a) (b)
Figure 2 The Box Plot of Maximum RMS in CWT versus Old and Young Group on a scale of (a) 128 and (b) 2048

From the result of STFT and CWT, the young group has a higher maximum RMS than the old group. The higher
maximum RMS indicates that the stride interval of the young group is larger than the old group. In STFT, the young group
has a higher variance of maximum RMS than the old group at both window lengths of 64 and 128. In CWT, the young group
has a higher variance at a low scale and has almost the same variance as the old group at a high scale. The higher variance
of the stride interval indicates higher complexity in the gait of the young group. The complexity of the gait of the young
group had been studied extensively using Detrended Fluctuation Analysis (DFA) in the past that showed gait complexity
increased during young adulthood and declined with aging [13]–[15].

STFT and CWT are both capable of distinguishing the gait of old people and young people using the feature, maximum
RMS value. The suitable window lengths for STFT in analyzing the stride interval are 64 and 128 while for CWT, the scale
should be set to the lowest which is 128. However, in terms of processing time, STFT can be performed in a significantly
shorter time than CWT. CWT at any scale produces a slightly more accurate time – frequency representation than STFT.

Overall, STFT with window lengths of 64 and 128 is better in analyzing the stride interval due to low processing time at
the expense of slightly less accurate time – frequency representation.

4. Conclusion
The result of STFT and CWT, both show that the young group has a higher maximum RMS value than the old group
which indicates a higher stride interval. The result of STFT shows that the young group has a higher variance of maximum
RMS than the old group at both window lengths of 64 and 128. The result of CWT shows that the young group has a higher
variance of maximum RMS at a low scale and almost the same variance as the old group at a high scale. The suitable window
lengths for STFT in analyzing the stride interval are 64 and 128 while for the CWT, the lowest scale which is 128 should be
chosen to reduce the processing time. Overall, STFT with window lengths of 64 and 128 is better in analyzing the stride
interval due to low processing time at the expense of slightly less accurate time – frequency representation.

ICBES 123-7
Acknowledgements
The authors acknowledge the support and funding provided by the Ministry of Higher Education (MOHE) of Malaysia
and Universiti Teknikal Malaysia Melaka (UTeM) through the Fundamental Research Grant Scheme (FRGS), No:
FRGS/1/2023/SKK06/UTEM/02/1.

References
[1] P. Ren, S. Tang, F. Fang, L. Luo, Lei. Xu, M. L. Bringas-Vega, D. Yao, K. M. Kendrick and P. A. Valdes-Sosa, “Gait
Rhythm Fluctuation Analysis for Neurodegenerative Diseases by Empirical Mode Decomposition,” IEEE Trans.
Biomed. Eng., vol. 64, no. 1, pp. 52–60, 2017, doi: 10.1109/TBME.2016.2536438.
[2] G. Cicirelli, D. Impedovo, V. Dentamaro, R. Marani, G. Pirlo, and T. R. D’Orazio, “Human Gait Analysis in
Neurodegenerative Diseases: A Review,” IEEE J. Biomed. Heal. Informatics, vol. 26, no. 1, pp. 229–242, 2022, doi:
10.1109/JBHI.2021.3092875.
[3] K. R. Chaudhuri, J. P. Azulay, P. Odin, S. Lindvall, J. Domingos, A. Alobaidi, P. L. Kandukuri, V. S. Chaudhari, J.
C. Parra, T. Yamazaki, J. Oddsdottir, J. Wright, P. Martinez-Martin, “Economic Burden of Parkinson’s Disease: A
Multinational, Real-World, Cost-of-Illness Study,” Drugs - Real World Outcomes, no. 0123456789, 2024, doi:
10.1007/s40801-023-00410-1.
[4] M. Akhtaruzzaman, A. A. Shafie, and M. R. Khan, “GAIT ANALYSIS: SYSTEMS, TECHNOLOGIES, and
IMPORTANCE,” J. Mech. Med. Biol., vol. 16, no. 7, pp. 1–45, 2016, doi: 10.1142/S0219519416300039.
[5] K. Okamoto, I. Obayashi, H. Kokubu, K. Senda, K. Tsuchiya, and S. Aoi, “Contribution of Phase Resetting to
Statistical Persistence in Stride Intervals: A Modeling Study,” Front. Neural Circuits, vol. 16, no. June, pp. 1–11,
2022, doi: 10.3389/fncir.2022.836121.
[6] V. T. Sai Sandeep Raju and M. Belwal, “Driver Drowsiness Detection,” Lect. Notes Data Eng. Commun. Technol.,
vol. 58, pp. 975–983, 2021, doi: 10.1007/978-981-15-9647-6_77.
[7] S. M. Cai, P. L. Zhou, H. J. Yang, T. Zhou, B. H. Wang, and F. C. Zhao, “Diffusion entropy analysis on the stride
interval fluctuation of human gait,” Phys. A Stat. Mech. its Appl., vol. 375, no. 2, pp. 687–692, 2007, doi:
10.1016/j.physa.2006.10.027.
[8] V. Alcan, “Nonlinear Analysis of Stride Interval Time Series in Gait Maturation Using Distribution Entropy,” Irbm,
vol. 43, no. 4, pp. 309–316, 2022, doi: 10.1016/j.irbm.2021.02.001.
[9] E. F. Shair, S. A. Ahmad, M. H. Marhaban, S. B. M. Tamrin, and A. R. Abdullah, “EMG processing based measures
of fatigue assessment during manual lifting,” Biomed Res. Int., vol. 2017, 2017, doi: 10.1155/2017/3937254.
[10] A. Goldberger, L. Amaral, L. Glass, J. Hausdorff, P. C. Ivanov, R. Mark, and H. E. Stanley, “PhysioBank,
PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals,” Circulation
[Online], vol. 101, no. 23, pp. e215–e220, 2000.
[11] T. Baba, “Time-Frequency Analysis Using Short Time Fourier Transform,” Open Acoust. J., vol. 5, no. 1, pp. 32–38,
2012, doi: 10.2174/1874837601205010032.
[12] N. Ji, H. Zhou, K. Guo, O. W. Samuel, Z. Huang, L. Xu, G. Li, “Appropriate mother wavelets for continuous gait
event detection based on time-frequency analysis for hemiplegic and healthy individuals,” Sensors (Switzerland), vol.
19, no. 16, pp. 1–18, 2019, doi: 10.3390/s19163462.
[13] K. Jordan, J. H. Challis, and K. M. Newell, “Long range correlations in the stride interval of running,” Gait Posture,
vol. 24, no. 1, pp. 120–125, 2006, doi: 10.1016/j.gaitpost.2005.08.003.
[14] J. M. Hausdorff, S. L. Mitchell, R. Firtion, C. K. Peng, M. E. Cudkowicz, J. W. Wei, A. L. Goldberger, “Altered
fractal dynamics of gait: Reduced stride-interval correlations with aging and Huntington’s disease,” J. Appl. Physiol.,
vol. 82, no. 1, pp. 262–269, 1997, doi: 10.1152/jappl.1997.82.1.262.
[15] S. Yano, L. Dimalanta, Y. Suzuki, and T. Nomura, “Fluctuation of stride time intervals during walking with
smartphone,” 2019 IEEE 1st Glob. Conf. Life Sci. Technol. LifeTech 2019, no. LifeTech, pp. 296–297, 2019, doi:
10.1109/LifeTech.2019.8884072.

ICBES 123-8