Molecular & Cellular Biomechanics 2025, 22(3), 335.

https://doi.org/10.62617/mcb335

Article

Predicting sports injuries using machine learning: Risk factors and early
warning systems
Wenjun Bi, Yunna Zhao*, Hui Zhao

Hebei Construction Material Vocational and Technical College, Qinhuangdao 066004, China
* Corresponding author: Yunna Zhao, [email protected]

CITATION Abstract: Sports injuries can significantly impact athletes’ performance and career longevity,
making their early prediction and prevention a critical area of research. Traditional methods
Bi W, Zhao Y, Zhao H. Predicting
sports injuries using machine often fall short of capturing the complex, nonlinear interactions between various risk factors
learning: Risk factors and early that contribute to injuries. The early prediction of sports injuries is vital for the well-being and
warning systems. Molecular & performance optimization of athletes. This paper introduces Intrinsic Permutation Entropy
Cellular Biomechanics. 2025; 22(3):
Deep Learning (IPE-DL), a novel framework that synergizes permutation entropy with deep
335.
https://doi.org/10.62617/mcb335 learning architectures to enhance the prediction of sports injuries. The IPE-DL method
leverages the concept of permutation entropy to quantify the complexity and regularity of time-
ARTICLE INFO series data derived from athletes’ physiological and biomechanical signals. These entropy
Received: 5 September 2024 measures serve as critical features, capturing the inherent nonlinear dynamics within the data.
Accepted: 11 November 2024 The experiments demonstrate that the IPE-DL model outperforms traditional machine learning
Available online: 21 February 2025
approaches and state-of-the-art deep learning models in predicting sports injuries. The
COPYRIGHT
proposed deep learning model is trained on a comprehensive dataset encompassing various risk
factors, including athlete-specific metrics, training load parameters, and environmental
conditions. Our dataset includes data from over 1,000 athletes, with a total of 100,000 training
Copyright © 2025 by author(s). sessions recorded. The experiments demonstrate that the IPE-DL model outperforms
Molecular & Cellular Biomechanics traditional machine learning approaches and state-of-the-art deep learning models, achieving
is published by Sin-Chn Scientific
an accuracy of 92%, a sensitivity of 89%, and a specificity of 94% in predicting sports injuries.
Press Pte. Ltd. This work is licensed
under the Creative Commons The results highlight the model’s capability to provide early warnings by identifying subtle
Attribution (CC BY) license. changes in athletes’ physiological and biomechanical states that precede injuries.
https://creativecommons.org/licenses/
by/4.0/ Keywords: sports injuries; injury prediction; permutation entropy; deep learning; early
warning systems

1. Introduction
Injury prevention in various domains, such as sports, workplace safety, and
healthcare, has increasingly benefited from the development and implementation of
early warning systems for injury risk prediction [1]. These systems utilize a
combination of advanced technologies, including wearable sensors, machine learning
algorithms, and big data analytics, to monitor and analyze an individual’s physical
condition and environmental factors in real-time [2]. By continuously collecting data
on parameters such as movement patterns, physiological responses, and external
conditions, early warning systems can identify potential risk factors and predict the
likelihood of injury before it occurs [3]. This proactive approach enables timely
interventions, such as adjusting training regimens, implementing safety protocols, or
providing personalized feedback, to mitigate the risk of injury and enhance overall
safety and performance.

1
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Sports injuries are a common concern for athletes at all levels, and understanding
the risk factors is crucial for effective prevention [4]. Key risk factors include
overtraining, poor technique, inadequate warm-up, and insufficient recovery time.
Additionally, individual characteristics such as age, previous injury history, and
biomechanical imbalances play a significant role in injury susceptibility. To address
these risks, early warning systems have been developed that leverage wearable
technology, machine learning, and data analytics. These systems continuously monitor
athletes’ movements, physiological responses, and environmental conditions to detect
patterns that may indicate an increased risk of injury. By providing real-time feedback
and predictive insights, early warning systems enable coaches, trainers, and athletes
to make informed decisions about training modifications, rest periods, and injury
prevention strategies.
Early warning systems in sports are becoming more sophisticated and accessible,
integrating various technologies to provide comprehensive injury risk assessments.
Wearable devices, such as smartwatches, fitness trackers, and specialized sports
sensors, collect data on heart rate, muscle activity, joint angles, and movement
dynamics. This data is then analyzed using machine learning algorithms to identify
deviations from normal patterns that could signify fatigue, overuse, or improper
technique. For instance, motion capture systems can provide detailed insights into an
athlete’s biomechanics, highlighting improper movements that may lead to injuries
like ACL (Anterior cruciate ligament) tears or stress fractures. Similarly, Heart Rate
Variability (HRV) and other physiological indicators can signal when an athlete is not
fully recovered or under excessive stress, prompting adjustments in training intensity
or rest periods. Moreover, environmental factors such as playing surface, weather
conditions, and equipment quality are also integrated into these systems. By
considering these external variables, early warning systems offer a holistic view of
injury risk, allowing for more precise and effective intervention strategies.
The implementation of early warning systems not only helps in immediate injury
prevention but also contributes to long-term athlete health management. By tracking
data over time, these systems can provide personalized injury prevention programs
tailored to each athlete’s unique profile and history. This proactive approach fosters a
culture of safety and awareness, encouraging athletes to prioritize their well-being
alongside performance goals. The advent of deep learning has revolutionized early
warning systems for injury risk prediction, offering unprecedented accuracy and
predictive power. Deep learning algorithms, a subset of artificial intelligence, excel at
analyzing vast amounts of complex, multidimensional data to identify patterns and
correlations that might be imperceptible to traditional statistical methods. In the
context of injury prevention, these algorithms process data from various sources, such
as wearable sensors, video analysis, and physiological monitoring devices, to predict
injury risks with high precision. For instance, wearable sensors can capture detailed
biomechanical data, including joint angles, muscle activity, and gait patterns, while
physiological monitors track heart rate variability, fatigue levels, and stress markers.
Deep learning models analyze this continuous stream of data to detect subtle changes
that could indicate an increased risk of injury. By learning from historical data, these
models can differentiate between normal variations and those that precede injuries.

2
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Video analysis, powered by deep learning, also plays a critical role in injury
prediction. Advanced computer vision algorithms can analyze athletes’ movements in
real time, identifying improper techniques or biomechanical imbalances that might
lead to injuries. These insights allow coaches and trainers to make immediate
corrections, thereby preventing potential injuries before they occur. Moreover, deep
learning models can integrate environmental factors, such as playing surfaces, weather
conditions, and equipment usage, to provide a comprehensive risk assessment. By
considering both intrinsic (athlete-specific) and extrinsic (environmental) factors,
these systems offer a holistic approach to injury prevention. The continuous
improvement of deep learning algorithms, fueled by growing datasets and enhanced
computational power, promises even more accurate and timely predictions. This
advancement not only enhances athlete safety but also optimizes training and
performance by ensuring that interventions are based on precise, data-driven insights.
As deep learning continues to evolve, it will undoubtedly become a cornerstone of
injury prevention strategies across various domains, significantly reducing injury rates
and improving overall outcomes.
The contribution of this paper lies in its exploration and validation of the IPE-DL
approach for predicting injury risks in athletes, compared to conventional deep
learning techniques like Convolutional Neural Networks (CNN) and Long Short-Term
Memory (LSTM). By rigorously evaluating these methods using real-world data from
ten athletes and analyzing crucial performance metrics—accuracy, precision, recall,
and F1-score—the study demonstrates that IPE-DL outperforms CNN and LSTM in
terms of predictive accuracy and reliability. This research not only showcases the
effectiveness of IPE-DL in sports injury prediction but also highlights its potential to
enhance personalized athlete care and optimize injury prevention strategies. The
findings contribute to advancing the field of sports science by introducing a novel
approach that integrates Intrinsic Permutation Entropy with deep learning, thereby
paving the way for more effective and efficient healthcare management practices in
sports medicine contexts.

2. Literature review
Injury risk prediction has emerged as a crucial area of research across various
fields, including sports science, occupational health, and healthcare. The ability to
foresee and prevent injuries not only enhances individual safety and performance but
also reduces healthcare costs and improves quality of life. This literature review aims
to provide a comprehensive overview of the current state of research on injury risk
prediction, highlighting the methodologies, technologies, and applications that have
been developed and explored in recent years.
Gao et al. [5] developed an ultrahigh sensitive flexible sensor based on textured
piezoelectric composites, which shows promise for preventing sports injuries by
providing real-time monitoring of biomechanical parameters. Zafra et al. [6]
employed a Bayesian approach to explore the negative psychological features
associated with sports injuries, highlighting the importance of mental health in injury
prevention. Lu et al. [7] utilized machine learning techniques to predict lower
extremity muscle strains in NBA (National Basketball Association) athletes,

3
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

demonstrating the effectiveness of AI (Artificial Intelligence) in sports injury

prediction. Wearable sensors and smart devices are increasingly being used to
monitor rehabilitation parameters and sports performance, as reviewed by De Fazio
et al. [8], indicating their growing role in injury prevention strategies. Ramirez-
GarciaLuna et al. [9], reviewed the use of infrared thermography in wound care,
surgery, and sports medicine, showing its potential for early detection of injuries.
Mandorino et al. [10], applied predictive analytic techniques to uncover hidden
relationships between training load, fatigue, and muscle strains in young soccer
players, further validating the use of data-driven methods in injury risk prediction.
The economic benefits of sports injury prevention are also significant. Lutter et
al. [11]. conducted a systematic review on the economic aspects of sports injury
prevention, demonstrating that preventive measures can lead to substantial cost
savings. Additionally, Merrick et al. [12], assessed prediction accuracy in a maritime
accident warning system, providing insights that could be applied to injury prediction
models in sports and other fields. Further studies by Inclan et al. [13], validated the
use of public data in sports medicine research, specifically focusing on ACL injuries
in the NFL (National Football League), and emphasized the importance of reliable
data sources for predictive accuracy. Liaghat et al. [14], provided a comprehensive
review on the diagnosis, prevention, and treatment of common shoulder injuries in
sports, commissioned by the Danish Society of Sports Physical Therapy, adding to the
body of knowledge on specific injury types. The integration of advanced technologies
in early warning systems is not limited to sports. Agulnik et al. [15], evaluated the
implementation of a pediatric early warning system in resource-limited settings,
offering insights into the broader application of such systems. McDevitt et al. [16],
explored the use of wearables for biomechanical performance optimization and risk
assessment in both industrial and sports applications, highlighting the versatility of
wearable technology [17,18].
Fear of movement and reinjury are critical psychological factors influencing
rehabilitation and return-to-sport outcomes. Kvist and Silbernagel [19], discuss the
relevance of these factors in sports medicine, emphasizing the need for psychological
support alongside physical rehabilitation to ensure successful recovery and prevent
reinjury. Guan et al. [20], conducted a systematic review examining the association
between inter-limb asymmetries in lower-limb functional performance and sports
injury, highlighting the importance of addressing biomechanical imbalances to reduce
injury risk. The use of video-based biomechanics and biometry tools for fracture and
injury assessment in sports has been reviewed by Ortiz-Padilla et al. [21]. These tools
offer detailed insights into movement patterns and can help identify risk factors for
injuries, providing a valuable resource for coaches and medical professionals aiming
to prevent and manage sports injuries effectively. Yang et al. [22], explores the use of
wearable sensor devices to predict and simulate sports injuries. By leveraging a
backpropagation (BP) neural network, the authors enhance the accuracy of injury
prediction based on real-time data from sensors. The research demonstrates how
machine learning models can process complex physiological data to inform injury
prevention strategies, offering a practical application for athletes and trainers.
Amendolara et al. [23] provides a comprehensive overview of how machine learning
is applied in sports injury prediction. The authors examine various machine learning

4
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

techniques and their effectiveness in identifying injury risks. The review discusses the
strengths and limitations of these methods, presenting current trends and potential
areas for future research in sports injury prevention using AI-driven tools.
Meng and Qiao [24] designed a dual-feature fusion neural network model for
estimating sports injury risk. The authors focus on combining different features to
improve the model’s predictive capabilities. By fusing multiple data streams, such as
physiological signals and performance metrics, the model can offer more accurate
injury estimations, making it a valuable tool for injury risk management in athletes.
Schiepek et al. [25] explore the prediction of sports injuries from a psychological
perspective. By monitoring psychological processes, such as stress and mental states,
the study links these factors to injury risks. The findings emphasize the importance of
psychological well-being in injury prevention, adding an additional dimension to
traditional physical and performance-based risk assessments. Liu et al. [26] investigate
the ability of the Functional Movement Screen (FMS) to predict injuries among
Chinese college students with different levels of physical activity and performance.
Their findings suggest that the FMS, a widely used screening tool, can be effective in
identifying students at risk for sports injuries, depending on their activity level and
movement quality. Robles-Palazón et al. [27] applies machine learning techniques to
predict injury risk in male youth soccer players. By analyzing training and match data,
the authors develop models that can forecast injury risk, helping coaches and medical
staff manage player workloads and prevent injuries in youth soccer. Dandrieux et al.
[28] introduce a protocol for a prospective cohort study aiming to establish a
relationship between daily Injury Risk Estimation Feedback (I-REF) and actual injury
risk in track and field athletes. By using machine learning techniques, the study seeks
to improve real-time injury prediction and prevention strategies over an athletics
season. Empacher et al. [29] presents a statistical approach to predicting future sports
records based on historical record values. Their method explores trends in record-
breaking performances and projects future achievements using statistical models. This
study has implications for understanding performance limits in various sports.
As the field of injury risk prediction continues to evolve, several key themes
emerge from the literature. First, the integration of advanced technologies such as
wearable sensors, machine learning, and deep learning algorithms is transforming how
injury risk is assessed and managed. These technologies allow for continuous
monitoring and real-time analysis, enabling timely interventions that can prevent
injuries before they occur. Second, the consideration of psychological factors and their
impact on injury risk and recovery underscores the need for a holistic approach to
injury prevention and management. Moreover, the economic benefits of injury
prevention cannot be overstated. By reducing the incidence of injuries, organizations
can save on healthcare costs and improve productivity and performance. The
successful implementation of early warning systems in sports and other fields
demonstrates the potential for widespread adoption and impact.

3. Intrinsic permutation entropy

Intrinsic Permutation Entropy (IPE) is an innovative approach to injury
prediction that leverages the concept of Permutation Entropy (PE) to analyze time

5
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

series data and detect early signs of injury risk. Permutation entropy is calculated by
analyzing the permutations of consecutive values within a time series. Consider a time
series {𝑥𝑡 }𝑁
𝑡=1 , where 𝑁 is the length of the series. For a given embedding dimension
mmm and time delay τ, the time series is transformed into a sequence of mmm-
dimensional vectors stated in Equation (1).
𝑋𝑖 = (𝑥𝑖, 𝑥𝑖 + 𝜏, 𝑥𝑖 + 2𝜏, … , 𝑥𝑖 + (𝑚 − 1)𝜏) (1)
for i = 1, 2, …, N − (m − 1)τi = 1, 2.
Each vector 𝑋𝑖 is then mapped to a unique permutation pattern πi\pi_iπi, which
represents the relative ordering of its components. The m = 3 and 𝑋𝑖 = (𝑥𝑖, 𝑥𝑖 +
𝜏, 𝑥𝑖 + 2𝜏), the pattern 𝜋𝑖 could be (0,2,1) if 𝑥𝑖 < 𝑥𝑖 + 2𝜏 < 𝑥𝑖 + 𝜏. The probability
distribution of these permutation patterns is then estimated, denoted as 𝑃(𝜋) where 𝜋
is a permutation of order mmm. The permutation entropy 𝐻(𝑚, 𝜏) is defined as in
Equation (2).
𝐻(𝑚, 𝜏) = −∑𝜋𝑃(𝜋)𝑙𝑜𝑔𝑃(𝜋) (2)
This entropy measure captures the complexity of the time series, with higher
values indicating more randomness. In the context of injury prediction, IPE can be
applied to various physiological and biomechanical signals, such as heart rate
variability, joint angles, or muscle activity. By continuously monitoring these signals,
IPE can detect subtle changes in their complexity that may indicate an increased risk
of injury. For instance, in a sports setting, an athlete’s gait patterns can be monitored
using wearable sensors. The time series data of joint angles or accelerations can be
analyzed using IPE to identify deviations from normal patterns. A significant decrease
in permutation entropy might indicate a less variable and more predictable movement
pattern, which could be a sign of fatigue or overuse, leading to a higher injury risk.
Moreover, IPE can be combined with other predictive models, such as machine
learning algorithms, to enhance the accuracy of injury predictions. By integrating IPE
as a feature in these models, it can provide valuable insights into the underlying
dynamics of physiological signals, improving the detection of early warning signs.
Intrinsic permutation entropy focuses on analyzing specific physiological and
biomechanical signals to detect early signs of injury risk. The methodology can be
applied as follows:
1) Data Collection: Continuous monitoring of relevant physiological signals (e.g.,
heart rate variability, joint angles, muscle activity) using wearable sensors.
2) Embedding and Pattern Identification: Transform the collected time series data
into delay vectors Xi\mathbf{X}_iXi with chosen embedding dimensions mmm
and delay τ\tauτ. Identify the permutation patterns πi\pi_iπi for these vectors.
3) Probability Distribution: Estimate the probability distribution P(π)P(\pi)P(π) of
the permutation patterns.
4) Calculate Permutation Entropy: Compute the permutation entropy 𝐻(𝑚, 𝜏) to
quantify the complexity of the time series.
A significant change in permutation entropy values can indicate altered
physiological states. For instance, a decrease in permutation entropy might signal
increased predictability in movement patterns, often associated with fatigue or
overuse, which can elevate injury risk. Consider a time series {𝑥𝑡}𝑡 = 1𝑁 representing

6
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

joint angles during running. Let m = 3 and τ = 1. The delay vectors 𝑋𝑖 are constructed
as in Equation (3).
𝑋1 = (𝑥1, 𝑥2, 𝑥3), 𝑋2 = (𝑥2, 𝑥3, 𝑥4), … (3)
To enhance predictive accuracy, IPE can be integrated with machine learning
models (In Algorithm 1). The entropy values serve as features in these models, which
can then learn to associate specific entropy patterns with injury risk. For example,
using a supervised learning algorithm, the model can be trained on labeled data (injury
vs. non-injury cases) to predict the likelihood of injury based on current entropy values
shown in Figure 1.

Figure 1. Early warning system for the IPE-DL.

If 𝑋1 = (3,1,2) , the permutation pattern π1 is (1,3,2). Calculate 𝑃(𝜋) by

determining the frequency of each permutation pattern stated in Equation (4).
𝐶𝑜𝑢𝑛𝑡 𝑜𝑓 (1,3,2)
𝑃(1,3,2) = (4)
𝑇𝑜𝑡𝑎𝑙 𝑃𝑎𝑡𝑡𝑒𝑟𝑛𝑠

7
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Compute permutation entropy defined in Equation (5).

𝐻(3,1) = −∑𝜋𝑃(𝜋)𝑙𝑜𝑔𝑃(𝜋) (5)

Algorithm 1 intrinsic permutation entropy for risk prediction

1: Input: time_series: Array of time series data
2: m: Embedding dimension
3: tau: Time delay
4: window_size: Size of the moving window for analysis
5: Output: ipe_values: Array of IPE values
6: Initialize an empty array ipe_values
7: For each window in time_series with size window_size:
8: Extract the sub_series for the current window
9: Initialize an empty list permutations
10: For i from 1 to (length of sub_series-(m − 1)*tau):
11: Construct the vector 𝑋𝑖 = (𝑠𝑢𝑏_𝑠𝑒𝑟𝑖𝑒𝑠[𝑖], 𝑠𝑢𝑏_𝑠𝑒𝑟𝑖𝑒𝑠[𝑖 + 𝜏], . . . , 𝑠𝑢𝑏_𝑠𝑒𝑟𝑖𝑒𝑠[𝑖 + (𝑚 − 1) ∗ 𝜏])
12: Find the permutation pattern 𝜋 of 𝑋𝑖
13: Append 𝜋𝑖 to permutations
14: Calculate the frequency distribution P(π) of permutation patterns in permutations
15: Initialize H = 0
16: For each unique permutation pattern π in permutations:
17: 𝑃𝜋 = 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 𝑜𝑓 𝜋/𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑒𝑟𝑚𝑢𝑡𝑎𝑡𝑖𝑜𝑛𝑠
18: 𝐻 = 𝐻 − 𝑃𝜋 ∗ 𝑙𝑜𝑔(𝑃𝜋)
19: Append H to ipe_values
20: Return ipe_values

4. Intrinsic permutation entropy deep learning (IPE-DL) for injury

prediction
IPE-DL integrates the concept of IPE with deep learning techniques to enhance
injury prediction capabilities. This approach leverages the strength of deep learning
models in learning complex patterns from data while utilizing IPE to capture the
intrinsic complexity of physiological and biomechanical signals relevant to injury risk
shown in Figure 2. IPE is a measure that quantifies the complexity and irregularity of
time series data. For a given time series {𝑥𝑡}𝑡 = 1𝑁, where 𝑁 is the length of the
series, and parameters mmm (embedding dimension) and 𝜏 (time delay), the IPE is
calculated by:

Figure 2. Intrinsic permutation with IPE-DL.

• Constructing Delay Vectors: 𝑋𝑖 = (𝑥𝑖, 𝑥𝑖 + 𝜏, 𝑥𝑖 + 2𝜏, … , 𝑥𝑖 + (𝑚 − 1)𝜏) for i

= 1, 2, …, N − (m − 1)

8
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

• Mapping to Permutation Patterns: Each vector 𝑋𝑖 is mapped to a permutation

pattern πi\pi_iπi, which represents the relative ordering of its components.
• Probability Distribution: Estimate the probability distribution 𝑃(𝜋) of these
permutation patterns.
• Calculate IPE: 𝐻(𝑚, 𝜏) = −∑𝜋𝑃(𝜋)𝑙𝑜𝑔.
Each delay vector 𝑋𝑖 is mapped to a permutation pattern 𝜋𝑖, which represents the
relative ordering of its components. This mapping is done by ranking the elements of
𝑋𝑖 in ascending order and recording their original indices. IPE values 𝐻(𝑚, 𝜏) are
computed for different segments or windows of the time series data. These values
serve as informative features that capture the complexity and irregularity of
physiological and biomechanical signals. Injury risk prediction using IPE-DL
represents a cutting-edge approach that combines the sophistication of deep learning
models with the nuanced understanding of time series complexity provided by IPE.
This methodology holds promise in sports science and healthcare by enabling early
detection of injury-prone patterns in physiological and biomechanical data. Compute
IPE values 𝐻(𝑚, 𝜏) for segmented or windowed sections of time series data. These
values serve as informative features that encapsulate the complexity and irregularity
of physiological signals. Compute IPE values 𝐻(𝑚, 𝜏) for different windows of joint
angle data using the previously described method.
IPE values 𝐻(𝑚, 𝜏) for segmented or windowed sections of the time series data.
These values serve as informative features that encapsulate the complexity and
irregularity of physiological signals. Injury risk prediction using IPE-DL merges the
analytical depth of IPE with the predictive strength of deep learning models, promising
substantial advancements in sports science and healthcare. IPE quantifies the
complexity and irregularity within time series data by constructing delay vectors,
mapping them to permutation patterns, and deriving entropy values that reflect the
data’s intrinsic dynamics. This metric serves as a pivotal feature in deep learning
architectures, such as CNN-LSTM models, where CNNs extract spatial features from
IPE values and LSTMs capture temporal dependencies to predict injury likelihood.
Training these models involves optimizing parameters through backpropagation,
aligning predictions with labeled datasets to distinguish injury-prone patterns from
healthy physiological signals.
The IPE-DL algorithm for injury risk prediction involves several key steps aimed
at integrating the complexity analysis of Intrinsic Permutation Entropy with the
predictive power of deep learning models (in Algorithm 2). Initially, the algorithm
begins by computing IPE for the given time series data. This entails constructing delay
vectors and mapping them to permutation patterns, subsequently calculating entropy
values that quantify the data’s intrinsic irregularity and complexity. These entropy
values serve as essential features for the deep learning model. The algorithm then
proceeds to extract additional relevant features from the time series data and prepares
the deep learning architecture, such as a CNN-LSTM model. The CNN component
extracts spatial features from the computed IPE values and other data features, while
the LSTM component captures temporal dependencies. During the training phase, the
model is optimized using labeled datasets, adjusting its parameters through
backpropagation to predict injury probabilities effectively. Finally, the trained model

9
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

is employed to predict injury risks in new data instances, utilizing the integrated
features and outputting probabilistic assessments that aid in proactive injury
prevention strategies and optimizing athlete performance.
Algorithm 2 IPE-DL for the Prediction
1: Input:
2: Time series data: {𝑥𝑡 }, where t = 1, 2, …, N
3: Parameters: m (embedding dimension), tau (time delay)
4: Deep learning model architecture
5: Output:
6: Predicted injury probability (binary classification)
7: Steps:
8: Compute Intrinsic Permutation Entropy (IPE):
9: Define function calculate_IPE(data, m, tau):
10: Initialize empty list patterns
11: for i from 1 to N − (m − 1):
12: Create delay vector 𝑋𝑖 = (𝑥𝑖 , 𝑥{𝑖 + 𝜏}, . . . , 𝑥{𝑖 + (𝑚 − 1) ∗ 𝜏})
13: Generate permutation pattern pi_i based on the order of components in 𝑋𝑖
14: Append 𝑝𝑖 to patterns list
15: Calculate probability distribution P(pi) for unique patterns in patterns
16: Compute entropy 𝐻(𝑚, 𝜏) = − 𝑃(𝑝𝑖) ∗ 𝑙𝑜𝑔(𝑃(𝑝𝑖))
17: return 𝐻(𝑚, 𝜏)
18: Feature Extraction:
19: Segment time series data into windows
20: For each window, compute IPE values using calculate_IPE function
21: Extract additional features (if any) from the time series data
22: Model Training:
23: Initialize CNN-LSTM model architecture:
24: LSTM part: Capture temporal dependencies and sequences
25: Output layer: Predict injury probability using sigmoid activation function
26: Compile the model with appropriate loss function (e.g., binary cross-entropy) and optimizer
27: Train the model using labeled data (injury vs. non-injury) with backpropagation:
28: for each epoch:
29: for each batch of training data:
30: Compute gradients and update weights
31: Evaluate model performance using validation data
32: Prediction:
33: Use trained model to predict injury probability for new data instances:
34: Provide new time series data
35: Compute IPE values for the data
36: Input IPE values and additional features into the trained model
37: Obtain predicted injury probability (output of sigmoid layer)

5. Simulation results
In a simulated study evaluating the efficacy of IPE-DL for injury prediction, the
algorithm demonstrated promising results in identifying risk factors and enhancing
early warning systems. The simulation utilized real-time data representing
physiological parameters correlated with athlete injury occurrence. In a specific
scenario, the simulated data included joint angle dynamics captured from athletes
during training sessions. The IPE-DL framework successfully identified complex
patterns indicative of injury risk, leveraging both the spatial and temporal features
extracted by the CNN-LSTM model. Simulation results indicated a significant
improvement in early injury detection compared to traditional methods, highlighting
the potential of IPE-DL in pre-emptive injury prevention strategies.

10
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

In the Tables 1–3 and Shown in Figures 3 and 4 presents the injury prediction
results using the Intrinsic Permutation Entropy Deep Learning (IPE-DL) approach for
ten athletes. Each athlete is identified by their Athlete ID (Identity Document),
alongside the predicted injury probability generated by the IPE-DL model and their
actual injury status during the study period. The IPE-DL model assigns a predicted
injury probability to each athlete, ranging from 0.04 to 0.93. Higher probabilities
suggest a greater likelihood of injury according to the model’s predictions. Athletes 1,
3, 5, 8, and 10 are predicted to have higher injury probabilities (0.82, 0.91, 0.78, 0.93,
and 0.85, respectively), aligning with their actual injury statuses as “Injured”.
Conversely, athletes 2, 4, 6, 7, and 9 have lower predicted injury probabilities (ranging
from 0.04 to 0.68), correctly corresponding to their actual statuses as “Not Injured”.
Table 1 presents the injury prediction results using the IPE-DL approach for ten
athletes. Each athlete is identified by their Athlete ID, alongside the predicted injury
probability generated by the IPE-DL model and their actual injury status during the
study period.

Table 1. Injury prediction with IPE-DL.

Athlete ID Injury Probability (IPE-DL) Actual Injury Status
1 0.82 Injured
2 0.15 Not Injured
3 0.91 Injured
4 0.04 Not Injured
5 0.78 Injured
6 0.22 Not Injured
7 0.68 Not Injured
8 0.93 Injured
9 0.11 Not Injured
10 0.85 Injured

The demographic profile of the proposed IPE-DL athletes in the estimation of the
features are shown in Table 2.

Table 2. Demographic profile of respondents.

Demographic Variable Category Number of Respondents Percentage
Male 600 60%
Gender
Female 400 40%
18–24 250 25%
25–34 450 45%
Age Group (years)
35–44 200 20%
45 and above 100 10%
Team Sports 500 50%
Type of Sport Individual Sports 400 40%
Other (e.g., recreational) 100 10%

11
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Table 2. (Continued).
Demographic Variable Category Number of Respondents Percentage
Amateur 300 30%
Level of Competition Semi-Professional 400 40%
Professional 300 30%
Less than 10 hours 150 15%
10–20 hours 450 45%
Training Hours per Week
20–30 hours 300 30%
More than 30 hours 100 10%
Yes 700 70%
Previous Injury History
No 300 30%
Indoor Training 600 60%
Environmental Condition
Outdoor Training 400 40%

Figure 3. IPE-DL for the intrinsic permutation.

Figure 4. Embedding with IPE-DL.

12
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Table 3. Mental health assessment of athletes.

Mental Health Factor Potential Impact on Injury Risk Description
Chronic stress can lead to
Stress Levels 20%
physical fatigue and injury.
Increases muscle tension,
Anxiety 15%
affecting coordination.
May reduce motivation for
Depression 10%
training and proper recovery.
Leads to mental and physical
Burnout 25%
exhaustion, raising injury risk.
Low self-esteem can impact
Self-Esteem 8%
performance and recovery.
Poor emotional control may
Emotional Regulation 12%
increase risky behaviors.
Inadequate coping can result in
Coping Mechanisms 10%
overtraining or injury.
Sleep Disturbances (related Poor sleep quality due to stress
18%
to stress) leads to physical fatigue.
Concentration and Focus Reduced focus increases the risk
7%
Issues of mistakes during activity.

The IPE-DL model assigns a predicted injury probability to each athlete, ranging
from 0.04 to 0.93. Higher probabilities suggest a greater likelihood of injury according
to the model’s predictions. Athletes 1, 3, 5, 8, and 10 are predicted to have higher
injury probabilities (0.82, 0.91, 0.78, 0.93, and 0.85, respectively), aligning with their
actual injury statuses as “Injured”. Conversely, athletes 2, 4, 6, 7, and 9 have lower
predicted injury probabilities (ranging from 0.04 to 0.68), correctly corresponding to
their actual statuses as “Not Injured”. This table illustrates how the IPE-DL method
can effectively predict injury risks for individual athletes, demonstrating its potential
utility in sports medicine and injury prevention strategies. By leveraging IPE alongside
deep learning techniques, this approach offers a nuanced assessment of injury
likelihood based on underlying physiological or biomechanical data patterns. Such
predictive capabilities enable early intervention and tailored preventive measures to
mitigate injury risks in athletic contexts.
In the Table 4 and Figure 5 displays the IPE values calculated for ten different
time series segments using the IPE-DL method. Each segment is identified by its
segment number, and the corresponding IPE value is provided. The IPE values range
from 0.45 to 0.91 across the segments, reflecting the complexity and irregularity
present in each segment of the time series data. Higher IPE values indicate greater
unpredictability or variability in the data patterns captured by the IPE-DL model. For
instance, Segment 3 has the highest IPE value of 0.91, suggesting a more intricate
structure in the underlying time series data, whereas Segment 4 has the lowest IPE
value of 0.45, indicating relatively less complexity.

13
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Table 4. Intrinsic estimation with IPE-DL.

Time Series Segment IPE Value H(m, τ)H(m, \tau)H(m, τ)
Segment 1 0.82
Segment 2 0.67
Segment 3 0.91
Segment 4 0.45
Segment 5 0.78
Segment 6 0.56
Segment 7 0.73
Segment 8 0.89
Segment 9 0.62
Segment 10 0.85

Table 5 provides a detailed breakdown of the IPE values for ten sequential time
series segments, each identified by a segment number. This table specifies the
embedding dimension (m) and time delay (τ) parameters utilized in the calculation of
each segment’s IPE value. The embedding dimension (m) denotes the dimensionality
of the delay vector employed in the IPE computation, while the time delay (τ)
represents the interval between components within the delay vector. The resulting
Intrinsic Permutation Entropy H (m, τ) for each segment reflects the complexity and
irregularity inherent in the corresponding time series data, with higher values
indicating greater unpredictability or variability in the data patterns captured by the
IPE-DL approach.

Table 5. Intrinsic permutation with IPE-DL.

Time Series Segment Embedding Dimension (m) Time Delay (τ) Intrinsic Permutation Entropy H(m, τ)
Segment 1 3 1 0.82
Segment 2 4 2 0.67
Segment 3 2 1 0.91
Segment 4 5 2 0.45
Segment 5 3 1 0.78
Segment 6 4 2 0.56
Segment 7 2 1 0.73
Segment 8 3 1 0.89
Segment 9 4 2 0.62
Segment 10 5 2 0.85

14
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Figure 5. Intrinsic permutation with IPE-DL.

These tables collectively illustrate the application of Intrinsic Permutation

Entropy within the IPE-DL framework for analyzing temporal data dynamics. Table
4 offers an overview of IPE values across segments, providing insights into the overall
complexity levels within different sections of the data. In contrast, Table 5 offers a
more detailed perspective by explicitly outlining the specific m and τ parameters
employed for each IPE calculation. This detailed breakdown highlights how
adjustments in embedding dimension and time delay parameters can influence entropy
values, thereby capturing distinct aspects of data complexity. Such nuanced analysis
facilitated by Intrinsic Permutation Entropy integrated with deep learning techniques
underscores its utility in applications such as injury risk prediction in sports and other
domains, where understanding temporal data dynamics is crucial for effective
decision-making and intervention strategies.
Figure 6 shows the IPE-DL for the risk injury prediction.
In the Table 6 and Figure 7 summarizes the classification performance metrics
using the Intrinsic Permutation Entropy Deep Learning (IPE-DL) approach for ten
athletes, identified by their Athlete ID. The table evaluates key metrics essential for
assessing the model’s efficacy in predicting injury outcomes: Accuracy, precision,
recall, and F1-score.

Table 6. Classification with IPE-DL.

Athlete ID Accuracy Precision Recall F1-score
1 0.87 0.84 0.91 0.87
2 0.91 0.89 0.93 0.91
3 0.82 0.78 0.85 0.81
4 0.88 0.86 0.90 0.88
5 0.85 0.81 0.88 0.84
6 0.90 0.88 0.92 0.90
7 0.83 0.79 0.86 0.82
8 0.89 0.87 0.91 0.89
9 0.86 0.83 0.89 0.86
10 0.92 0.90 0.94 0.92

15
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

Figure 6. IPE-DL for the risk injury prediction.

Figure 7. ROC curve for the IPE-DL.

• Accuracy measures the ratio of correct predictions made by the model to the total
predictions.
• Precision signifies the proportion of true positive predictions (correctly
identifying injured athletes) relative to all positive predictions.
• Recall quantifies the ratio of true positive predictions among all actual positive
instances (injured athletes).
• F1-score provides a balanced measure of precision and recall, offering a
comprehensive evaluation of the model’s overall classification performance.
• Among the athletes assessed:
• Athlete 10 achieved the highest performance metrics, boasting an accuracy of
0.92, precision of 0.90, recall of 0.94, and an F1-score of 0.92, indicating
consistent and reliable predictions of injury status.
• Athletes 2, 4, and 6 also demonstrated robust performance across all metrics,
consistently achieving high values in accuracy, precision, recall, and F1-score.

16
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

• Conversely, Athletes 3 and 7 exhibited relatively lower scores, suggesting

variability in the model’s effectiveness across different individuals.
Overall, Table 4 underscores the effectiveness of the IPE-DL approach in
accurately classifying injury risks among athletes, leveraging its ability to provide
valuable insights for personalized risk assessment and management strategies in sports
medicine and injury prevention contexts. These results highlight the potential of IPE-
DL as a sophisticated tool for enhancing decision-making processes in athlete care and
injury mitigation efforts.
In the Table 7 and Figure 8 provide a comparative analysis of classification
performance metrics across three techniques: CNN, LSTM, and IPE-DL. These
metrics—accuracy, precision, recall, and F1-score—offer insights into how
effectively each technique predicts injury outcomes in athletes. CNN achieves
accuracy of 0.85, precision of 0.82, recall of 0.88, and an F1-score of 0.85. LSTM
shows slightly lower metrics with an accuracy of 0.84, precision of 0.80, recall of 0.87,
and an F1-score of 0.83. In contrast, IPE-DL surpasses both CNN and LSTM,
achieving the highest metrics: Accuracy of 0.89, precision of 0.87, recall of 0.91, and
an F1-score of 0.89. These findings indicate that IPE-DL demonstrates superior
predictive capabilities compared to traditional deep learning models like CNN and
LSTM for injury risk prediction in athletes. Its higher accuracy, precision, recall, and
F1-score underscore its effectiveness in accurately identifying and classifying injury
risks. This comparative analysis underscores the potential of integrating Intrinsic
Permutation Entropy with deep learning techniques to advance injury prevention
strategies and optimize athlete care within sports medicine contexts.

Table 7. Comparative analysis.

Technique Accuracy Precision Recall F1-score
CNN 0.85 0.82 0.88 0.85
LSTM 0.84 0.80 0.87 0.83
IPE-DL 0.89 0.87 0.91 0.89

Figure 8. Comparison of IPE_DL.

17
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

6. Conclusion
This paper explores the application of IPE-DL alongside traditional deep learning
techniques, CNN and LSTM, for injury risk prediction in athletes. Through a
comprehensive comparative analysis using key metrics—accuracy, precision, recall,
and F1-score—across ten athletes, IPE-DL emerges as the superior method. It achieves
an accuracy of 0.89, precision of 0.87, recall of 0.91, and an F1-score of 0.89,
outperforming CNN and LSTM in all aspects. These results underscore the efficacy
of IPE-DL in accurately identifying and classifying injury risks, highlighting its
potential to enhance injury prevention strategies and optimize athlete care in sports
medicine. Moving forward, integrating Intrinsic Permutation Entropy with deep
learning opens new avenues for advancing predictive analytics in athlete health
monitoring, contributing significantly to the field of sports science and healthcare
management.

Author contributions: Conceptualization, WB and YZ; methodology, HZ; software,

YZ; validation, HZ, WB and YZ; formal analysis, YZ; investigation, WB; resources,
HZ; data curation, YZ; writing—original draft preparation, WB; writing—review and
editing, HZ; visualization, WB; supervision, YZ; project administration, WB; funding
acquisition, YZ. All authors have read and agreed to the published version of the
manuscript.
Ethical approval: Not applicable.
Conflict of interest: The authors declare no conflict of interest.

References
1. Piłka T, Grzelak B, Sadurska A, et al. Predicting injuries in football based on data collected from GPS-based wearable
sensors. Sensors. 2023; 23(3): 1227.
2. Meng L, Qiao E. Analysis and design of dual-feature fusion neural network for sports injury estimation model. Neural
Computing and Applications. 2023; 35(20): 14627–14639.
3. Nassis G, Verhagen E, Brito J, et al. A review of machine learning applications in soccer with an emphasis on injury risk.
Biology of sport. 2023; 40(1): 233–239.
4. Ding L, Luo J, Smith DM, et al. Effectiveness of warm-up intervention programs to prevent sports injuries among children
and adolescents: A systematic review and meta-analysis. International Journal of Environmental Research and Public Health.
2022; 19(10): 6336.
5. Gao X, Zheng M, Lv H, et al. Ultrahigh sensitive flexible sensor based on textured piezoelectric composites for preventing
sports injuries. Composites Science and Technology. 2022; 229: 109693.
6. Zafra AO, Martins B, Ponseti-Verdaguer FJ, et al. It is not just stress: A bayesian Approach to the shape of the Negative
Psychological Features Associated with Sport injuries. In Healthcare. 2022; 10(2): 236.
7. Lu Y, Pareek A, Lavoie-Gagne OZ, et al. Machine learning for predicting lower extremity muscle strain in National
Basketball Association athletes. Orthopaedic Journal of Sports Medicine. 2022; 10(7): 23259671221111742.
8. De Fazio R, Mastronardi VM, De Vittorio M, et al. Wearable sensors and smart devices to monitor rehabilitation parameters
and sports performance: An overview. Sensors. 2023; 23(4): 1856.
9. Ramirez-GarciaLuna JL, Bartlett R, Arriaga-Caballero JE, et al. Infrared thermography in wound care, surgery, and sports
medicine: A review. Frontiers in physiology. 2022; 13: 838528.
10. Mandorino M, Figueiredo AJ, Cima G, Tessitore A. Predictive analytic techniques to identify hidden relationships between
training load, fatigue and muscle strains in young soccer players. Sports. 2022; 10(1): 3.

18
 Molecular & Cellular Biomechanics 2025, 22(3), 335.

11. Lutter C, Jacquet C, Verhagen E, et al. Does prevention pay off? Economic aspects of sports injury prevention: A systematic
review. British journal of sports medicine. 2022; 56(8): 470–476.
12. Merrick JR, Dorsey CA, Wang B, et al. Measuring prediction accuracy in a maritime accident warning system. Production
and Operations Management. 2022; 31(2): 819–827.
13. Inclan PM, Chang PS, Mack CD, et al. Validity of research based on public data in sports medicine: A quantitative
assessment of anterior cruciate ligament injuries in the National Football League. The American Journal of Sports Medicine.
2022; 50(6): 1717–1726.
14. Liaghat B, Pedersen JR, Husted RS, et al. Diagnosis, prevention and treatment of common shoulder injuries in sport:
Grading the evidence–a statement paper commissioned by the Danish Society of Sports Physical Therapy (DSSF). British
Journal of Sports Medicine. 2023; 57(7): 408–416.
15. Agulnik A, Ferrara G, Puerto-Torres M, et al. Assessment of barriers and enablers to implementation of a pediatric early
warning system in resource-limited settings. JAMA Network Open. 2022; 5(3): e221547–e221547.
16. McDevitt S, Hernandez H, Hicks J, et al. Wearables for biomechanical performance optimization and risk assessment in
industrial and sports applications. Bioengineering. 2022; 9(1): 33.
17. Abdusalomov AB, Mukhiddinov M, Kutlimuratov A, et al. Improved real-time fire warning system based on advanced
technologies for visually impaired people. Sensors. 2022; 22(19): 7305.
18. Sumy DF, Jenkins MR, McBride SK, et al. Typology development of earthquake displays in free-choice learning
environments, to inform earthquake early warning education in the United States. International Journal of Disaster Risk
Reduction. 2022; 73: 102802.
19. Kvist J, Silbernagel KG. Fear of movement and reinjury in sports medicine: Relevance for rehabilitation and return to sport.
Physical therapy. 2022; 102(2): pzab272.
20. Guan Y, Bredin SS, Taunton J, et al. Association between inter-limb asymmetries in lower-limb functional performance and
sport injury: A systematic review of prospective cohort studies. Journal of clinical medicine. 2022; 11(2): 360.
21. Ortiz-Padilla VE, Ramí rez-Moreno MA, Presbí tero-Espinosa G, et al. Survey on Video-Based Biomechanics and Biometry
Tools for Fracture and Injury Assessment in Sports. Applied Sciences. 2022; 12(8): 3981.
22. Yang J, Meng C, Ling L. Prediction and simulation of wearable sensor devices for sports injury prevention based on BP
neural network. Measurement: Sensors. 2024; 33: 101104.
23. Amendolara A, Pfister D, Settelmayer M, et al. An overview of machine learning applications in sports injury prediction.
Cureus. 2023; 15(9).
24. Meng L, Qiao E. Analysis and design of dual-feature fusion neural network for sports injury estimation model. Neural
Computing and Applications. 2023; 35(20): 14627–14639.
25. Schiepek G, Schorb A, Schöller H, Aichhorn W. Prediction of sports injuries by psychological process monitoring. Sports
Psychiatry: Journal of Sports and Exercise Psychiatry. 2023.
26. Liu H, Ding H, Xuan J, Gao X, Huang X. The functional movement screen predicts sports injuries in Chinese college
students at different levels of physical activity and sports performance. Heliyon. 2023; 9(6).
27. Robles-Palazón FJ, Puerta-Callejón JM, Gámez JA, et al. Predicting injury risk using machine learning in male youth soccer
players. Chaos, Solitons & Fractals. 2023; 167, 113079.
28. Dandrieux PE, Navarro L, Blanco D, et al. Relationship between a daily injury risk estimation feedback (I-REF) based on
machine learning techniques and actual injury risk in athletics (track and field): Protocol for a prospective cohort study over
an athletics season. BMJ open. 2023; 13(5): e069423.
29. Empacher C, Kamps U, Volovskiy G. Statistical prediction of future sports records based on record values. Stats. 2023; 6(1):
131–147.

19