*Result*: Beyond self-reports after anterior cruciate ligament injury - machine learning methods for classifying and identifying movement patterns related to fear of re-injury.
Original Publication: [London, England] : Published by E. & F.N. Spon on behalf of the Society of Sports Sciences, [c1983-
*Further Information*
*Anterior cruciate ligament (ACL) tears are prevalent career-ending sports injuries. A barrier to successful return to activity is fear of re-injury. Evaluating psychological readiness is however limited to insufficient self-reported assessments. We developed machine learning models using biomechanical data from standardized rebound side hops (SRSH) to objectively classify fear levels post-ACL reconstruction (ACLR) and identify key biomechanical variables. Sixty individuals with ACLR and 47 controls performed up to 10 side hops per leg. Kinematic and kinetic data were collected using motion capture and force platforms. ACLR participants were classified (Tampa Scale for Kinesiophobia-17) as HIGH-FEAR (n = 32) or LOW-FEAR (n = 28). Analyses involved 1D convolutional neural networks (1D CNN) and logistic regression. Integrated gradients identified influential movement variables. The 1-D CNN distinguished HIGH-FEAR versus LOW-FEAR ACLR individuals in agreement with Tampa Scale scores, achieving a mean accuracy of 75.6% (F₁ Score = 0.76, Matthews Correlation Coefficient = 0.52), which was 8.6% better than logistic regression. Influential variables included trunk tilt, hip flexion/extension, and ankle supination/pronation. Machine learning from biomechanics can identify movement linked to fear of re-injury post-ACLR, potentially informing personalised rehabilitation to mitigate fear and enhance recovery.*
AN0190859090;5bv01feb.26;2026Jan16.02:52;v2.2.500
Beyond self-reports after anterior cruciate ligament injury – machine learning methods for classifying and identifying movement patterns related to fear of re-injury
Anterior cruciate ligament (ACL) tears are prevalent career-ending sports injuries. A barrier to successful return to activity is fear of re-injury. Evaluating psychological readiness is however limited to insufficient self-reported assessments. We developed machine learning models using biomechanical data from standardized rebound side hops (SRSH) to objectively classify fear levels post-ACL reconstruction (ACLR) and identify key biomechanical variables. Sixty individuals with ACLR and 47 controls performed up to 10 side hops per leg. Kinematic and kinetic data were collected using motion capture and force platforms. ACLR participants were classified (Tampa Scale for Kinesiophobia-17) as HIGH-FEAR (n = 32) or LOW-FEAR (n = 28). Analyses involved 1D convolutional neural networks (1D CNN) and logistic regression. Integrated gradients identified influential movement variables. The 1-D CNN distinguished HIGH-FEAR versus LOW-FEAR ACLR individuals in agreement with Tampa Scale scores, achieving a mean accuracy of 75.6% (F₁ Score = 0.76, Matthews Correlation Coefficient = 0.52), which was 8.6% better than logistic regression. Influential variables included trunk tilt, hip flexion/extension, and ankle supination/pronation. Machine learning from biomechanics can identify movement linked to fear of re-injury post-ACLR, potentially informing personalised rehabilitation to mitigate fear and enhance recovery.
Keywords: Artificial intelligence; biomechanics; kinesiophobia; knee; machine learning integration; rehabilitation
Introduction
Anterior cruciate ligament (ACL) ruptures are highly prevalent and debilitating knee injuries, with approximately 70% occurring during sports or physical activities, especially those involving sudden stops, jumps, or directional changes (Griffin et al., [17]). Standard treatment approaches for ACL injuries are conservative management or surgical reconstruction (ACLR) (Rodriguez et al., [47]). Regardless of treatment, extensive rehabilitation is imperative to restore knee function, strength, and stability (Wu et al., [55]). Despite this, many individuals still fail to return to pre-injury activity levels (McPherson et al., [38]), often due to psychological factors (Flanigan et al., [14]; Hart et al., [21]; Tripp et al., [52]).
Fear of re-injury represents a common psychological factor that negatively impacts the individual and their return to pre-injury activity levels despite a successful physical recovery (Kvist et al., [29]). Observed avoidance behaviours include stiffened movement patterns during jump landings (Trigsted et al., [51]). We have earlier reported greater lower limb muscle co-activation during a standardized rebound side hop (SRSH) for those who self-reported high versus those who indicated low fear of injury within 2 years of ACLR (Markstrom et al., [36]). In the same study, we also found greater hip and knee flexion at initial contact of the landing for both high and low fear ACLR groups compared with asymptomatic controls. Such behaviours may potentially increase the risk of re-injury and promote long-term disability. There is however no gold standard for assessing fear of re-injury after ACL rupture.
Current methods used to assess fear of re-injury in the field of sports medicine and rehabilitation include patient-reported outcome measures, interviews, clinical evaluations, and some biomechanical studies of movement patterns and muscle activity (Filbay & Kvist, [13]; Kvist et al., [29]; Little et al., [33]; Meierbachtol et al., [39]). One of the most frequently used questionnaires for this purpose is the Tampa Scale for Kinesiophobia-17 (TSK-17) (Miller et al., [40]). The TSK-17 was originally developed to assess fear of movement-induced pain among patients with chronic low back pain but has been implemented also among ACL injured populations (Kvist et al., [30]; Luc-Harkey et al., [34]; Trigsted et al., [51]). To classify low and high fear subgroups in individuals with ACLR, we previously (Markström et al., [37]) used the statement from the TSK-17 'I am afraid that I might injure myself accidentally'. Subjective self-reported measures can nevertheless be prone to biases like social desirability, recall inaccuracies, and lack of self-awareness, potentially leading to misrepresentation of true fear levels (French et al., [15]; Grinberg, [18]; Huang et al., [23]). For example, an athlete eager to return to sport may underreport fear to appear more confident, while another might overreport due to anxiety, consciously or unconsciously. Consequently, there is a need for objective assessment methods that can detect subtle biomechanical indicators of fear. Analysing movement patterns through biomechanical data may provide deeper insights into how fear manifests physically, offering a more comprehensive understanding that complements self-reported measures.
Objective assessment and quantification of fear of re-injury following ACLR could inform rehabilitation and improve outcomes to facilitate successful return to sport (Hsu et al., [22]). Machine learning (ML) has emerged as a powerful tool in this domain, for instance when evaluating the risk of ACL injuries based on kinematic and kinetic data from physical activities. Early studies employed algorithms like support vector machines (SVMs) (Taborri et al., [50]), and convolutional neural networks (CNNs) (Jauhiainen et al., [24]) to analyse biomechanical data from various tasks such as jumps. More recent advancements utilise multi-modal deep learning approaches a subset of ML, which leverage both imaging and non-imaging data to predict future ACL injury risk (Han et al., [20]). Utilising diverse data sources in conjunction with deep learning has the potential for improved prediction accuracy. Additionally, ML techniques have identified specific gait biomechanical parameters associated with ACL injuries (Kokkotis et al., [27]). However, traditional feature-based ML methods often rely on manually selected variables, potentially discarding valuable information contained within the raw movement data. This limitation was recently highlighted when random forest analysis was used to identify key biomechanical factors related to knee joint kinematics and kinetics (Jauhiainen et al., [24]). Similarly, CNNs outperformed random forests in predicting joint kinematics, kinetics, and muscle forces, emphasising the importance of feature selection for optimal performance (Moghadam et al., [41]). In parallel, hybrid feature – time-series neural networks have been applied to predict ACL forces after reconstruction with resistive bracing, demonstrating that biomechanical loading patterns can be captured directly from time-series signals (Li et al., [31]). Separate work has developed ML models to predict ACLR failure risk using large-scale clinical and imaging variables, indicating the growing adoption of data-driven prognosis in ACL care (Alaiti et al., [3]).
The primary aim of this study was to develop and evaluate an ML model using kinematic and kinetic time series data from SRSH landings to objectively classify high or low fear of re-injury among individuals with ACL injury and in relation to non-injured controls. We hypothesised that the model would categorise these two fear levels more accurately than logistic regression (LR). In addition, we wanted to identify the biomechanical variables that were most influential for fear classification using Explainable Artificial Intelligence (XAI) techniques.
Methods
Study design
This was a cross-sectional observational study with data collection performed in a movement laboratory. The study was approved by the regional ethical review board in Umeå, Sweden (Dnr. 2015/67–31). All participants received a verbal and written description of the study and provided their written informed consent before participating. The study was conducted following the ethical principles stated in the Declaration of Helsinki.
Participants
Participants were recruited from orthopaedic and sports medicine clinics, local sports clubs and via advertisements at the university campus. Eligibility criteria for all participants were: aged 17–34 years, able to perform maximal single-leg forward hops, no major injuries other than ACL rupture and concomitant meniscal tear for the ACLR participants and no other conditions that could influence hop performance.
Sixty individuals who had undergone ACLR within the previous two years and 47 age and sex-matched asymptomatic controls representing a range of physical activity levels (25 non-athletes and 22 athletes) participated. All ACLR participants had undergone unilateral ACLR with an ipsilateral hamstring autograft.
Eight of the ACLR participants and 25 controls performed testing on at least two occasions and thus data from two test sessions were included in the analyses for these participants. See Table 1 for participant characteristics.
Table 1. Participant characteristics by group for participants with anterior cruciate ligament reconstruction (ACLR) and asymptomatic controls.
1 <sups>a</sups>Eight ACLR participants and 25 controls have two test sessions included in the main data analysis but only data related to their first test session have been included to calculate group characteristics.
Study procedures
Patient-reported outcome measures
All participants completed the following patient-reported outcome measures before hop testing: 2000 International Knee Documentation Committee subjective knee form (IKDC), Knee injury and Osteoarthritis Outcome Score (KOOS), Lysholm Scale, Tegner Activity Scale (current). All ACLR participants additionally completed the following: TSK-17, Tegner Activity Scale (pre-injury).
Classification of fear of Re-injury
In line with a previous study (Markstrom et al., [36]), participants were divided into HIGH-FEAR and LOW-FEAR groups based on their responses to statement 9 of the TSK-17 questionnaire: 'I am afraid that I might injure myself accidentally'. The HIGH-FEAR group consisted of 32 participants who answered 'Agree' or 'Strongly agree'. The LOW-FEAR group consisted of 28 participants who answered, 'Strongly disagree' or 'Disagree'. All control participants were classified by default as having no fear of re-injury without completing the TSK-17.
Hop testing
Participants performed the SRSH test, which has previously been shown to reliably evaluate landing mechanics for ACLR and controls (Markström et al., [37]). The SRSH was also used in the above mentioned study to discriminate between individuals with ACLR who have high or low fear, as well as asymptomatic controls based on biomechanical variables during landings (Markstrom et al., [36]). Testing was conducted at the U-Motion Laboratory, Umeå University, Sweden.
Briefly, two white tape strips approximately 30 cm in length and 4 cm in width were affixed parallel to each other on separate force plates with their outer edges at an individually adjusted distance equivalent to 25% body height of each participant. Participants stood barefoot on the testing leg with both tape strips lateral to the body and the foot positioned against the nearest tape strip. A 25 cm rope with knots on either side was held with the hands behind the back. Participants hopped laterally to the furthest side of the lateral tape strip and rebounded back as quickly as possible to the start position (Figure 1). The final landing was required to be held for 3 s. Two familiarisation repetitions were performed on each leg before 10 successive repetitions were performed on each leg in an alternating order regardless of whether the hops were successful. The start leg was randomised. Unsuccessful hops, which were not included in the analyses, were those with significant adjustments of the foot upon landing, loss of balance within 3 seconds of landing, release of the rope, and/or excessively short or long hop distance.
Graph: Figure 1. Workflow for 1D CNN model analysis. 1. Recording of side hop tests. 2. Data processing. 3. A 1D CNN model classifies the data to identify the level of fear of re-injury. 4. Integrated Gradients (IG) to highlight biomechanical variables which are most influential for the decision.
Biomechanical assessment
Kinematics and kinetics
Kinematic and kinetic data were collected using an infrared 8-camera 3D motion capture system (240 Hz, Oqus 300, Qualisys AB, Sweden) and two time-synchronised force platforms (1200 Hz, Kistler 9286AA, Switzerland). Fifty-six passive retroreflective markers were attached to anatomical landmarks on the skin using double-sided adhesive tape according to an established full-body marker set protocol (Markström et al., [37]). The raw motion capture data were recorded and pre-processed in Qualisys Track Manager software (v2019.3, Qualisys). In cases of missing marker data, gap filling was implemented via cubic spline interpolation. All interpolated segments were then reviewed manually by examining each marker's 3D trajectory and its motion relative to adjacent markers to identify tracking artifacts; trials with abnormal interpolations were excluded.
Data processing
Further processing of marker trajectories and force plate data was performed in Visual3D (Visual3D v5, C-Motion Inc., Germantown, MD, USA). Marker trajectories, kinematics and kinetics were filtered using a critically damped zero-lag digital low-pass filter with a 15 Hz cut-off frequency. A kinematic model defined trunk, pelvis, thigh, shank, and foot segments. Joint kinematics and kinetics were derived by a Cardan XYZ rotation sequence, commonly used for describing joint rotations in biomechanics (Wu et al., [54]). Hip joint centres were determined using a functional joint method, while knee and ankle joint centres were calculated based on markers placed over femoral epicondyles and malleoli, respectively (Schwartz & Rozumalski, [48]).
Hop segmentation
The biomechanical recordings were segmented into discrete hop trials, yielding a total of 1 714 approved hops across all participants (HIGH-FEAR = 272, LOW-FEAR = 255, Controls = 1 187). Every participant contributed at least five valid hops from the analysed leg, ensuring robust within-subject representation for subsequent modelling. A complete overview of participants, hop counts, and the resulting train/validation/test partitions is provided in Table 2.
Table 2. Composition of the study dataset, feature matrix, and data partitions.
5 †Control trials are appended to the LOWFEAR class during training only (dataaugmentation strategy; see Supplementary S1).
Handling dependency of repeated hops
To ensure independence between training and testing phases and prevent data leakage, a Leave-One Participant Out Cross-Validation (LOPCV) strategy was employed (Cheng et al., [8]; Molinaro et al., [42]). In each iteration, all hops from a single participant were held out as the test set, while the remaining hops of the other participants were used to train the model. This process was repeated for each participant, allowing the model to be tested on every participant's data once, while being trained on the others. Since there were 60 participants, the LOPCV strategy resulted in a total of 60 training/testing combinations, with each participant's data serving as the test set once. An 80/20 split within the training set (at the participant level) was used for hyperparameter tuning via a validation set. By structuring the data in this way, we ensured that data from a single participant never appeared in both the training and testing phases within a single iteration, effectively addressing the issue of intra-individual correlations and repeated measures. This approach rigorously evaluates the model's ability to generalise to new participants, even in the absence of a fixed test set.
Data extraction
Outcome variables were trunk, pelvis, hip, knee and ankle angles, as well as hip, knee and ankle moments of force, measured in all three directions: medio-lateral, anterior-posterior and inferior-superior. To enable comparisons and analyses between trials and participants, the landing phase was aligned across trials using a time normalisation procedure with a 3rd order polynomial in Visual3D. Each landing (medial and lateral) was defined by the time interval from initial contact to maximum knee flexion and was time-normalised to 101 data points, ensuring consistent temporal aspects of the biomechanical waveforms. The resulting integrated dataset thus consisted of 24 time-series variables across 101 normalised time points for both medial and lateral landings, for each individual and trial.
Data augmentation and standardization
To enhance the ability of the model to learn diverse movement patterns and biomechanical variations, data from controls were integrated into the LOW-FEAR category. This not only increased dataset size but also introduced greater variability within the training environment (details in Supplementary Materials S1). The time series were standardised using min-max normalisation, with reference values derived from the minimum and maximum values observed in the control group data.
Artificial Intelligence methods
Classification
After a thorough evaluation of classification methods (see Supplementary material S1), a 1D convolutional neural network (CNN) model was chosen as the main classification method. As a baseline comparison, we also performed logistic regression. Models were trained on medial landings and lateral landings separately and on the combination of both.
For the 1D CNN, the HyperBand algorithm in Keras Tuner (Molinaro et al., [42]) guided the selection of optimal model architecture and hyperparameters. An overview of the full analysis pipeline, from recording and preprocessing to classification and explainability, is provided in Figure 1. To further illustrate the analytic procedure, a flow chart of the LOPOCV framework is presented in Figure 2, complemented by the corresponding pseudocode in Figure 3. These depictions clarify the sequential structure of the analysis, including data partitioning, model training, prediction, and metric aggregation. At the model level, the detailed 1D CNN architecture, showing the sequence of convolutional, pooling, and dense layers, is reported in Figure 4. A fixed model architecture and set of hyperparameters were used for the main experiment (see Supplementary Material S1).
Graph: Figure 2. Flow chart of the Leave-one-participant-out cross-validation (LOPOCV) pipeline. The chart illustrates sequential steps including data collection, preprocessing, group assignment, model training and evaluation, and performance aggregation. Each iteration holds out one ACLR participant for testing, while the remaining participants are used for training. Metrics are stored per Fold and aggregated across iterations.
Graph: Figure 3. Pseudocode of the proposed Leave-one-participant-out cross-validation (LOPCV) pipeline.
Graph: Figure 4. Visual representation of a deep 1D CNN with regularization and pooling layers.
Throughout the training phase, we closely monitored the 1D CNN for overfitting indicators – such as diverging training and validation curves – and adjusted the hyperparameters of the model accordingly. Specifically, we fine-tuned the critical hyperparameters in relation to our evaluation metrics. To ensure reliable results, we conducted the classification process three times, following recommendations from prior studies on the importance of multiple runs, and presented the average results (Bajgar et al., [4]; Dushatskiy et al., [11]; Gundersen et al., [19]).
Logistic regression
For baseline comparisons, we performed LR with Lasso regularisation. As input, we used biomechanical data where each time series was represented with its average value. The LR analysis was performed using the package glmnet in R (version 4.1.3) (Friedman et al., [16]) using default settings and weights. The larger group (LOW-FEAR including controls) had a weight of one and the smaller group (HIGH-FEAR) had a weight corresponding to the ratio between the group sizes. A threshold of 0.5 was used for the classification of hops, i.e., a hop was classified as HIGH-FEAR if the predicted probability was > 0.5, and LOW-FEAR otherwise.
In our preliminary analyses, we compared average and peak values of biomechanical time series as input features. Models using average values consistently performed better, likely because they capture more consistent movement patterns related to fear levels, while peak values may represent isolated moments not reflective of sustained biomechanical adaptations associated with fear of re-injury.
Majority voting
To classify participants according to fear level at the participant level, we employed a majority vote based on the classifications of individual hops. Participants were classified as HIGH-FEAR if at least 50% of their hops were classified as such. Otherwise, they were classified as LOW-FEAR. This method ensures that the participant's overall classification reflects most of their hop classifications within a single run.
Model performance evaluation
Performance was evaluated using the LOPCV strategy described earlier. The 1D CNN model was implemented with three independent runs to assess the variability due to random weight initialisation, data shuffling, and other stochastic processes inherent in training neural networks (Dodge et al., [9]; Fernández-Delgado et al., [12]; Pineau et al., [44]; Reimers & Gurevych, [45], [46]). These stochastic factors can lead to different model performances even when trained on the same data (Bengio et al., [5]; Pineau et al., [44]; Zhang et al., [57]). To mitigate this variability and obtain robust estimates of model performance, we averaged the results over multiple runs.
We chose to run the model three times to balance the need for assessing variability with the constraints of computational resources and time, as each run was computationally intensive. For each run, we recorded performance metrics at the participant level, including accuracy, sensitivity (recall), specificity, F1 Score, and Matthews Correlation Coefficient (MCC). We calculated the mean and range (minimum and maximum values) of these metrics across the runs to provide comprehensive estimates of the model's performance and to understand the impact of random initialisation and other stochastic factors on the model's stability. Given the small number of runs (n = 3), reporting the range provides a clearer representation of variability without assuming symmetrical distribution around the mean.
While scalar performance metrics were averaged across runs, confusion matrices are specific to each run and cannot be meaningfully averaged. Therefore, we present the confusion matrices for each run in Supplementary Material S4 to provide detailed insights into the model's classification performance.
We fine-tuned the critical hyperparameters of learning rate, early stopping patience, and reduced learning rate patience based on performance indicators such as loss, validation loss, accuracy, and validation accuracy. This iterative refinement was conducted within each cross-validation cycle to minimise overfitting and enhance model generalisability. To optimise the classification performance of the model, we also evaluated various thresholds for assigning class labels based on the confidence scores of the model. Consequently, a threshold of 0.5 was chosen as it yielded the highest accuracy.
Addressing class imbalance
Due to a slight class imbalance between the HIGH-FEAR and LOW-FEAR groups in our dataset, relying solely on accuracy could be misleading. Therefore, we included additional metrics like the F1 Score and MCC to provide a more nuanced assessment of the model's performance. The F1 Score balances precision and recall, which is crucial when the costs of false positives and false negatives are different. The MCC considers all four confusion matrix categories and is particularly robust for evaluating performance on imbalanced datasets, producing a score between −1 and +1. A score of + 1 represents a perfect prediction, 0 represents a random prediction, and −1 represents a complete disagreement between prediction and observation.
Identification of influential variables
To identify influential variables for fear classification, Integrated Gradients (IG) (Sundararajan et al., [49]) was chosen as the primary Explainable Artificial Intelligence (XAI) method due to its compatibility with both the data structure and model architecture. A relative importance score was calculated for each variable based on its contribution to the model output. For detailed information on both the selection of IG and the method itself, see Supplementary Materials S2.
We followed a detailed four-step process for calculating IG scores. First, we computed the absolute IG values for each 'hop' (time point) for every participant, creating an IG matrix that showed how each biomechanical variable, over measured landing phase interval, influenced the decisions of the model at the individual level. Second, to generalise these findings beyond individual differences, we averaged these IG matrices across all participants, offering a broader view of the influence of each variable. Third, to account for the variability in deep learning model initialisation, we performed the classification process three times, as mentioned in the model performance evaluation section. By averaging the IG matrices from each run for every participant, we minimised the influence of initial model variability. Fourth, we averaged the rows of these averaged IG matrices from the last step to rank the biomechanical variables by their overall influence in the classification decisions.
Implementation
The 1D CNN model was implemented in Keras with TensorFlow 2.15 under Python 3.11.4. Training and evaluation were conducted on a workstation equipped with an Intel Core i9 CPU, 64GB RAM, and an NVIDIA GeForce RTX 4090 Graphics Processing Unit (GPU). Across all 60 cross-validation folds, total training time was approximately 17 minutes, with each fold requiring 15.6 ± 3.7 seconds (median 25.1, max 23.4). Peak GPU memory usage reached 23 GB. Inference, executed on the CPU, averaged 83 ± 25 milliseconds per participant.
Results
The classification accuracies and sensitivities of the 1D CNN and LR models across different hop types are presented in Table 3. The specific results are presented and discussed in detail below.
Table 3. Performance of Machine learning models in classifying HIGH-FEAR and LOW-FEAR groups using biomechanical variables from lateral, medial, as well as from the combination of lateral-medial hop trials.
Participant level performance of 1D CNN and LR models
At the participant level, the classification accuracy and sensitivity differed between the model and type of input data. The mean accuracy for the 1D CNN model was 75.6% (range: 73.1%–78.1%), compared to the LR model's accuracy of 67%. The same was seen for class sensitivity, with the 1D CNN model achieving 71.9% (range: 68.8%-75.0%) sensitivity compared to 69% for the LR model.
Considering the class imbalance, the F1 Score and MCC provided additional insights into the model's performance. The mean F1 Score was highest for the 1D CNN model (mean F1 = 0.76, range: 0.73–0.78) for lateral-medial hops, while the LR model rendered a slightly lower score (F1 = 0.69 for medial and lateral-medial). The mean MCC was also higher for the 1D CNN model (mean MCC = 0.52, range: 0.47–0.57) compared to the LR model (MCC = 0.33 for medial and lateral-medial) (see Table 2).
Detailed confusion matrices for each run of the 1D CNN model are provided in Supplementary Material S4. These matrices illustrate the consistency of the model's classification performance across different runs.
Hop-level performance of 1D CNN and LR models
The hop level accuracy and sensitivity were slightly lower compared to the participant level and again depended on hop direction and model. The 1D CNN model reached an overall slightly higher classification mean accuracy (53%-65.3%) compared to the LR model (58%-61%). The 1D CNN model achieved a higher mean F1 Score for lateral-medial hops (mean F1 = 0.63, range: 0.60–0.67), while the LR model had its highest F1 Score for Medial hops (F1 = 0.66). The MCC values further highlighted these differences, with the 1D CNN model having a mean MCC of 0.31 (range: 0.27–0.35) for lateral-medial hops.
Notably, the 1D CNN model showed better performance when predicting the LOW-FEAR group (specificity range: 55.3%–75.7%) compared with the HIGH-FEAR group (sensitivity range: 50.7%–58.3%). In contrast, the LR model better predicted the HIGH-FEAR group (sensitivity: 60%–72%), than the LOW-FEAR group (52%-54%). See Table 2 for detailed results.
Identification of influential biomechanical variables for fear classification
The top 10 most influential biomechanical variables across hops for the classification using our 1D CNN model are illustrated in Figure 5. For medial hops, the most important variable was the degree of trunk anterior/posterior tilt and for lateral hops it was ankle supination/pronation. For detailed quantification of variatable
Graph: Figure 5. Most influential biomechanical variables for classification of HIGH-FEAR and LOW-FEAR hops, as identified by integrated gradient analysis.
influence and additional visualisations, please see Supplementary S3 for associated heatmaps.
Discussion
This study addresses the critical challenge of identifying fear of re-injury following ACL injury by introducing a novel approach for objective assessment using ML based on movement behaviour during challenging side hop landings, which is a common clinical assessment in the context. Our 1D CNN model achieved a promising accuracy of 75.6% (range: 73.1%–78.1%) and high mean sensitivity and specificity of 71.9% and 79.8%, respectively, in identifying individuals with high fear of re-injury.
By analysing each run separately and reporting the mean and range of the performance metrics, we ensured a transparent evaluation of the model's robustness. The observed variability highlights the importance of conducting multiple runs to fully understand the model's performance and reliability. This variability may be attributed to the stochastic nature of neural network training and the sensitivity of the model to weight initialisation and optimisation processes.
Notably, the inclusion of additional performance metrics like the F1 Score and MCC was crucial due to the slight class imbalance in our dataset. The high F1 Score (mean 0.76, range: 0.73–0.78) indicated that the model maintained a good balance between precision and recall. The MCC (mean 0.52, range: 0.47–0.57) further affirmed the model's robustness in handling imbalanced data, demonstrating strong predictive capability.
Leveraging ML and biomechanical data analysis, our findings thus demonstrate the potential of a 1D CNN model to accurately classify fear of re-injury in individuals with ACLR based solely on biomechanical data. Notably, the 1D CNN model showed higher accuracy compared to the LR model in classifying hop consistency for combined lateral-medial hops and for medial hops, while the LR model had higher accuracy for lateral hops, suggesting potential differences in biomechanical patterns associated with fear of re-injury across hop directions. While both models achieved good accuracy in classifying hops according to fear, further analysis revealed variations in their agreement with actual classifications across different hop types (lateral, medial, combined). This suggests the need for further investigation into how fear of re-injury might influence biomechanics differently depending on, for example, movement type and hop direction.
Key biomechanical predictors and their implications
We identified the most influential biomechanical predictors for classification of fear of re-injury: the degree of trunk anterior/posterior tilt, hip flexion/extension, and ankle supination/pronation. These predictors corroborate our prior research indicating that individuals with heightened fear adopt protective movement strategies to enhance knee stability and mitigate injury risk through, e.g., stiffening the joints by muscle co-contractions (Markstrom et al., [36]). Such kinematic patterns may serve as preventive measures to lessen the strain on the ACL and minimise the likelihood of re-injury. This interpretation underscores the influence of psychological factors on biomechanical outcomes following ACLR, reinforced by evidence indicating that greater fear of re-injury may also lead to adaptative movement mechanics aimed at protecting the knee (Markstrom et al., [36]). This also aligns well with prior observations of stiffened jump-landing mechanics and muscle activation in females with high fear of re-injury following ACLR (Trigsted et al., [51]).
Influence of hop direction
A similar presence of the top-ranked variables was observed across medial and lateral landings, but their ordering differed by direction (Figure 5). Trunk anterior – posterior tilt ranked highest during medial landings, whereas ankle supination/pronation and hip flexion/extension were more prominent during lateral landings. This indicates that the models capture multi-joint patterns linked to fear in both directions. Performance reflected these directional differences. At the hop level, the 1D-CNN performed best when combining directions (lateral-medial), while lateral hops yielded the highest specificity for this model; the LR model performed best for medial hops (see Table 3). At the participant level, combining directions also produced the strongest 1D-CNN results (Table 3). Together, these findings suggest that medial and lateral hops emphasise partly different biomechanical features relevant to fear classification.
Side-hop tasks stress frontal – transverse control relevant to return-to-sport (Dos Santos, [10]) and are targeted in ACLR retraining frameworks (Buckthorpe, [7]). Fear has been linked to stiffer side-hop mechanics and altered gluteal activation (Markstrom et al., [36]). Our SRSH protocol uses a lateral hop followed immediately by a medial rebound within the same trial. In our data, the leading variables differed by direction: ankle supination/pronation was more prominent in lateral hops, whereas trunk anterior – posterior tilt led in medial hops (Figure 4). Performance reflected this: at the hop level the 1D-CNN was best when combining directions (Lateral – Medial), with lateral hops yielding the highest specificity; the LR model performed best for medial hops. At the participant level, combining directions also produced the strongest 1D-CNN results (Table 3). These findings indicate that the two directions provide complementary information: lateral hops probe ankle/frontal-plane control, while medial hops emphasise trunk – hip strategies consistent with the clinical emphasis on frontal/transverse-plane demands and hip – trunk control in ACLR screening and retraining (Buckthorpe, [7]; Dos Santos, [10]).
Bridging the gap: Machine learning for objective assessment of fear of Re-injury based on mov...
It is well established that fear of re-injury plays a significant role in rehabilitation and for return to sport following ACL injury (Kvist et al., [30]; Webster & Feller, [53]). However, a critical knowledge gap exists in the application of objective methods to identify fear of re-injury and its consequences. Some biomechanical studies have found evidence that fear of re-injury influences movement behaviour following ACL injury (Khojah et al., [26]; Little et al., [32]; Trigsted et al., [51]). However, to our knowledge, no research has investigated the potential of ML for identifying fear of re-injury based on biomechanical data.
Our work bridges this gap by developing an ML model to objectively classify movement strategies according to fear levels based on biomechanical data. This innovative approach aligns with findings suggesting that individuals with high fear of re-injury exhibit altered movement mechanics, potentially as a subconscious strategy to protect the injured knee (Trigsted et al., [51]). However, while our results are promising, it is important to contextualise what they mean in the multifaceted context of fear assessment. Fear, as a psychological construct, can manifest differently across individuals, and behaviours do not always align with self-reported fear levels due to ingrained movement habits or strategies and awareness. For instance, an individual may report no fear but still behave cautiously as if they were fearful, displaying a movement pattern protective of their knees, or vice versa. This divergence may explain why, even with high accuracy, there remains variability unaccounted for by the model.
Evidence from non-ACL musculoskeletal applications shows that biomechanical time-series can encode clinically meaningful states, for example in scapular-stabilisation assessment (Mabrouk et al., [35]), postpartum low back pain trunk motion (Abdel Hady & Abd El-Hafeez, [2]), and pelvic-tilt estimation in pelvic-floor dysfunction (Abdel Hady & Abd El-Hafeez, [1]). Our findings extend this line of work by showing that hop biomechanics can also capture a psychological state fear of re-injury after ACL reconstruction, a target not addressed in prior applications.
Limitations
Our study has some limitations that warrant consideration. The sample size, while adequate for initial exploration, could be increased to improve the robustness of the model and provide a more comprehensive evaluation of model performance. The range of activity levels among our participants is, however, a strength which allows for greater generalisation of our results compared with a more homogenous sample. Our analysis focused on kinematic and kinetic data from a side hop task, but fear-related movement compensations may manifest differently during other activities such as cutting or pivoting. Additionally, our previous research has found different muscle activity patterns depending on fear status (Markstrom et al., [36]; Trigsted et al., [51]). Adding relevant muscle activity data to the model may thus further enhance its performance. Another limitation may be the use of one question from the TSK for group labelling. While self-reported outcomes are convenient and widely used, this approach might introduce potential biases or discrepancies between self-reported fear and its manifestation in movement behaviour. Future studies could explore incorporating more comprehensive psychological assessments, muscle activation data, qualitative data, or expert clinical evaluations to refine the labelling process and potentially improve the accuracy of the model.
Additionally, the choice of machine learning models presents limitations. The 1D CNN model, while effective in capturing spatial-temporal dependencies in time-series data, requires substantial computational resources. Due to these computational constraints, we limited our analysis to three runs. This limitation might restrict the assessment of variability due to random initialisation. However, the small ranges observed in performance metrics suggest that three runs were sufficient to demonstrate model stability. Future work with larger datasets and greater computational resources could consider increasing the number of runs to further validate model robustness. Other ML methods and data preprocessing techniques are also likely to yield different results.
Clinical implications and future research
Successful classification of how fear of re-injury influences biomechanics could facilitate a more personalised rehabilitation programme. This holistic approach, addressing both physical and psychological aspects, aligns with the established importance of comprehensive recovery strategies for optimal outcomes (Brinlee et al., [6]; Jenkins et al., [25]; Kotsifaki et al., [28]; Paster et al., [43]; Yao et al., [56]; Ziab et al., [58]). Our study represents an advancement in the objective, data-driven assessment of fear of re-injury based on movement strategies after ACL injury. Identification of individuals who exhibit movement patterns indicative of heightened fear of re-injury can eventually facilitate the implementation of targeted interventions such as education about fear of re-injury and the rehabilitation process, graded exposure exercises to increase movement confidence, and cognitive-behavioural therapy to address negative thought patterns and anxiety associated with fear of re-injury.
Further research in this area is important, and future endeavours should explore the generalisability of these findings across diverse populations and functional tasks. The incorporation of electromyographic data or musculoskeletal modelling could provide further insights into the neuromuscular manifestations of fear, potentially enhancing the ability of the model to predict the impact of fear of re-injury on movement patterns. Longitudinal studies tracking the evolution of fear-related movement patterns throughout rehabilitation could elucidate the temporal dynamics of fear of re-injury and inform optimal intervention timings. Additionally, investigations into the relationship between objective assessments of fear of re-injury and long-term outcomes, such as return to sports, re-injury rates and functional performance, would ascertain the clinical utility of these ML models.
The clinical translation of our model depends on modest hardware and software footprint. Training was performed once on a high-performance workstation. At the point of care, classification of a new patient's hop trials runs in < 1 s on a standard CPU and does not require a dedicated GPU. All software required is open-source and license-free. Accordingly, day-to-day use including preprocessing of motion-capture output runs on desktops or laptops already present in biomechanics or physiotherapy clinics. In practice, the only ongoing expense is the motion capture system itself; the model converts those recordings into decision support within the existing infrastructure. This study is a research evaluation; clinical deployment will depend on institutional approval, regulatory compliance where applicable, and site-level verification and training.
In summary, explainable ML applied to hop-biomechanics can objectively characterise fear of re-injury after ACL reconstruction; the next priorities are prospective, multi-site validation and integration with neuromuscular measures (e.g., EMG) to enable clinical translation.
Conclusion
Our study highlights the potential of ML for objective assessment of psychological factors based on movement behaviour following ACL injury. By achieving 75.6% classification accuracy, an 8.6% improvement over logistic regression, our 1D CNN model successfully distinguished individuals with HIGH-FEAR and LOW-FEAR of reinjury based merely on biomechanical data. The significance of this improvement lies in demonstrating that biomechanical data alone can capture psychological readiness post-ACL injury, highlighting the potential of ML models as an objective, scalable complement to traditional self-reported measures in clinical practice. This work underscores the intricate interplay between psychological state and physical movement. Future related research should focus on incorporating neuromuscular data, such as EMG, and conducting longitudinal studies to track fear-related movement patterns throughout the rehabilitation process or evaluation of fear-related targeted interventions.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Data availability statement
The dataset generated and analysed during the current study is not publicly available due to privacy and ethical restrictions concerning participant data but is available from the corresponding author on reasonable request. Researchers wishing to replicate the analysis workflow without personal data can obtain an
Ethics approval
The Regional Ethical Review Board in Umeå approved the study (ref no. 2015/67–31).
Supplementary Information
Supplemental data for this article can be accessed online at https://doi.org/10.1080/02640414.2025.2578584
References
1 Abdel Hady, D. A., & Abd El-Hafeez, T. (2023, October 20). Predicting female pelvic tilt and lumbar angle using machine learning in case of urinary incontinence and sexual dysfunction. Scientific Reports, 13 (1), 17940. https://doi.org/10.1038/s41598-023-44964-0
2 Abdel Hady, D., & Abd El-Hafeez, T. (2024, August 12). Utilizing machine learning to analyze trunk movement patterns in women with postpartum low back pain. Scientific Reports, 14 (1), 18726. https://doi.org/10.1038/s41598-024-68798-6
3 Alaiti, R. K., Vallio, C. S., da Silva, A. G. M., Gobbi, R. G., Pecora, J. R., & Helito, C. P. (2025, March). Predicting ACL reconstruction failure with Machine learning: Development of Machine learning prediction models. Orthopaedic Journal of Sports Medicine, 13 (3), 23259671251324519. https://doi.org/10.1177/23259671251324519
4 Bajgar, O., Kadlec, R., & Kleindienst, J. (2018). A Boo(n) for evaluating architecture performance. [Paper Presentation]. Proceedings of the 35th International Conference on Machine Learning, Proceedings of Machine Learning Research, Stockholm, Sweden, 10–15 July (pp. 334–343).
5 Bengio, Y., Goodfellow, I., & Courville, A. (2016). Deep networks (in deep learning). MIT Press.
6 Brinlee, A. W., Dickenson, S. B., Hunter-Giordano, A., & Snyder-Mackler, L. (2022). ACL reconstruction rehabilitation: Clinical data, biologic healing, and criterion-based milestones to inform a return-to-sport guideline. Sports Health A Multidisciplinary Approach, 14 (5), 770 – 779. https://doi.org/10.1177/19417381211056873
7 Buckthorpe, M. (2021, August). Recommendations for movement Re-training after ACL reconstruction. Sports Medicine, 51 (8), 1601 – 1618. https://doi.org/10.1007/s40279-021-01454-5
8 Cheng, J., Dekkers, J. C. M., & Fernando, R. L. (2021, September). Cross-validation of best linear unbiased predictions of breeding values using an efficient leave-one-out strategy. Journal of Animal Breeding & Genetics, 138 (5), 519 – 527. https://doi.org/10.1111/jbg.12545
9 Dodge, J., Ilharco, G., Schwartz, R., Farhadi, A., Hajishirzi, H., & Smith, N. (2020). Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv: 2002 [Online]. https://arxiv.org/pdf/2002.06305
Dos Santos, T. (2020). Biomechanical determinants of injury risk and performance during change of direction: Implications for screening and intervention. Ph.D. School of health and society, University of Salford (United Kingdom).
Dushatskiy, A., Alderliesten, T., & Bosman, P. A. N. (2022). Heed the noise in performance evaluations in neural architecture search. [Paper Presentation]. Proceedings of the Genetic and Evolutionary Computation Conference Companion, Boston, Massachusetts.
Fernández-Delgado, C. E. M., Barro, S., & Amorim, D. (2014). Do we need hundreds of classifiers to solve real world classification problems? Journal of Machine Learning Research, 15, 3133 – 3181. https://dl.acm.org/doi/10.5555/2627435.2697065
Filbay, S., & Kvist, J. (2021). Fear of reinjury following surgical and nonsurgical management of anterior cruciate ligament injury: An exploratory analysis of the NACOX multicenter longitudinal cohort study. Physical Therapy, 102 (2). https://doi.org/10.1093/ptj/pzab273
Flanigan, D. C., Everhart, J. S., Pedroza, A., Smith, T., & Kaeding, C. C. (2013, August). Fear of reinjury (kinesiophobia) and persistent knee symptoms are common factors for lack of return to sport after anterior cruciate ligament reconstruction. Arthroscopy, 29 (8), 1322 – 1329. https://doi.org/10.1016/j.arthro.2013.05.015
French, D. J., France, C. R., Vigneau, F., French, J. A., & Evans, R. T. (2007, January). Fear of movement/(re)injury in chronic pain: A psychometric assessment of the original English version of the Tampa scale for kinesiophobia (TSK). Pain, 127 (1–2), 42 – 51. https://doi.org/10.1016/j.pain.2006.07.016
Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software, 33 (1). https://doi.org/10.18637/jss.v033.i01
Griffin, L. Y., Agel, J., Albohm, M. J., Arendt, E. A., Dick, R. W., Garrett, W. E., Garrick, J. G., Hewett, T. E., Huston, L., Ireland, M. L., Johnson, R. J., Kibler, W. B., Lephart, S., Lewis, J. L., Lindenfeld, T. N., Mandelbaum, B. R., Marchak, P., Teitz, C. C., & Wojtys, E. M. (2000). Noncontact anterior cruciate ligament injuries: Risk factors and prevention strategies. Journal of the American Academy of Orthopaedic Surgeons, 8 (3), 141 – 150. https://doi.org/10.5435/00124635-200005000-00001
Grinberg, A. (2024). Reinjury fear and anxiety following anterior cruciate ligament injury. Let's get our constructs Straight! JOSPT Open, 2 (1), 4 – 7. https://doi.org/10.2519/josptopen.2023.0807
Gundersen, O. E., Shamsaliei, S., Kjaernli, H. S., & Langseth, H. (2023). On reporting robust and trustworthy conclusions from model comparison studies involving neural networks and randomness. Proceedings of the 1st Acm Conference on Reproducibility and Replicability, Rep'23, 37 – 61. https://doi.org/10.1145/3589806.3600044
Han, M. (2023, July 27). LigaNET: A multi-modal deep learning approach to predict the risk of subsequent anterior cruciate ligament injury after surgery. medRxiv. https://doi.org/10.1101/2023.07.25.23293102
Hart, H. F., Culvenor, A. G., Guermazi, A., & Crossley, K. M. (2020, January). Worse knee confidence, fear of movement, psychological readiness to return-to-sport and pain are associated with worse function after ACL reconstruction. Physical Therapy in Sport, 41, 1 – 8. https://doi.org/10.1016/j.ptsp.2019.10.006
Hsu, C.-J., Meierbachtol, A., George, S. Z., & Chmielewski, T. L. (2017). Fear of reinjury in athletes. Sports Health A Multidisciplinary Approach, 9 (2), 162 – 167. https://doi.org/10.1177/1941738116666813
Huang, H., Nagao, M., Arita, H., Shiozawa, J., Nishio, H., Kobayashi, Y., Kaneko, H., Nagayama, M., Saita, Y., Ishijima, M., Takazawa, Y., Ikeda, H., & Kaneko, K. (2019, September 11). Reproducibility, responsiveness and validation of the Tampa Scale for kinesiophobia in patients with ACL injuries. Health and Quality of Life Outcomes, 17 (1), 150. https://doi.org/10.1186/s12955-019-1217-7
Jauhiainen, S., Kauppi, J. P., Krosshaug, T., Bahr, R., Bartsch, J., & Ayramo, S. (2022, September). Predicting ACL injury using Machine learning on data from an extensive screening test battery of 880 female elite athletes. American Journal of Sports Medicine, 50 (11), 2917 – 2924. https://doi.org/10.1177/03635465221112095
Jenkins, S. M., Guzman, A., Gardner, B. B., Bryant, S. A., Del Sol, S. R., McGahan, P., & Chen, J. (2022, June). Rehabilitation after anterior cruciate ligament injury: Review of Current literature and recommendations. Current reviews in musculoskeletal medicine, 15 (3), 170 – 179. https://doi.org/10.1007/s12178-022-09752-9
Khojah, O., Khojah, O., Dabroom, A., Alghamdi, A., Alshibely, A., Algarni, H., Almadani, M., Alswat, A., & Alshehri, M. (2021). Factors associated with fear of re-injury in physically active individuals after an anterior cruciate ligament reconstruction. Journal of Musculoskeletal Surgery and Research, 5 (1), 30. https://doi.org/10.4103/jmsr.jmsr_113_20
Kokkotis, C., Moustakidis, S., Tsatalas, T., Ntakolia, C., Chalatsis, G., Konstadakos, S., Hantes, M. E., Giakas, G., & Tsaopoulos, D. (2022, April 22). Leveraging explainable machine learning to identify gait biomechanical parameters associated with anterior cruciate ligament injury. Scientific Reports, 12 (1), 6647. https://doi.org/10.1038/s41598-022-10666-2
Kotsifaki, R., Korakakis, V., King, E., Barbosa, O., Maree, D., Pantouveris, M., Bjerregaard, A., Luomajoki, J., Wilhelmsen, J., & Whiteley, R. (2023, May). Aspetar clinical practice guideline on rehabilitation after anterior cruciate ligament reconstruction. British Journal of Sports Medicine, 57 (9), 500 – 514. https://doi.org/10.1136/bjsports-2022-106158
Kvist, J., Bengtsson, J., & Lundqvist, C. (2023). The experience and influence of fear after anterior cruciate ligament reconstruction: An interview study with young athletes. BMC Sports Science, Medicine and Rehabilitation, 15 (1). https://doi.org/10.1186/s13102-023-00659-7
Kvist, J., Ek, A., Sporrstedt, K., & Good, L. (2005, July). Fear of re-injury: A hindrance for returning to sports after anterior cruciate ligament reconstruction. Knee Surgery, Sports Traumatology, Arthroscopy, 13 (5), 393 – 397. https://doi.org/10.1007/s00167-004-0591-8
Li, D., Li, H., & Hang, Y. (2025). Hybrid feature-time series neural network for predicting ACL forces in martial artists with resistive braces after reconstruction. Frontiers in Bioengineering and Biotechnology, 13, 1579472. https://doi.org/10.3389/fbioe.2025.1579472
Little, C., Lavender, A. P., Starcevich, C., Mesagno, C., Mitchell, T., Whiteley, R., Bakhshayesh, H., & Beales, D. (2022, November 18). An investigation of the nature of fear within ACL-Injured subjects when exposed to provocative videos: A concurrent qualitative and quantitative study. Sports (Basel), 10 (11), 183. https://doi.org/10.3390/sports10110183
Little, C., Lavender, A. P., Starcevich, C., Mesagno, C., Mitchell, T., Whiteley, R., Bakhshayesh, H., & Beales, D. (2023). Understanding fear after an anterior cruciate ligament injury: A qualitative thematic analysis using the common-sense model. International Journal of Environmental Research and Public Health, 20 (4), 2920. https://doi.org/10.3390/ijerph20042920
Luc-Harkey, B. A., Franz, J. R., Losina, E., & Pietrosimone, B. (2018, July). Association between kinesiophobia and walking gait characteristics in physically active individuals with anterior cruciate ligament reconstruction. Gait Posture, 64, 220 – 225. https://doi.org/10.1016/j.gaitpost.2018.06.029
Mabrouk, O. M., Hady, D. A. A., & Abd El-Hafeez, T. (2024, November 18). Machine learning insights into scapular stabilization for alleviating shoulder pain in college students. Scientific Reports, 14 (1), 28430. https://doi.org/10.1038/s41598-024-79191-8
Markstrom, J. L., Grinberg, A., & Hager, C. K. (2022). Fear of reinjury following anterior cruciate ligament reconstruction is manifested in muscle activation patterns of single-leg side-hop landings. Physical Therapy, 102 (2), pzab218. https://doi.org/10.1093/ptj/pzab218
Markström, J. L., Schelin, L., & Häger, C. K. (2021). A novel standardised side hop test reliably evaluates landing mechanics for anterior cruciate ligament reconstructed persons and controls. Sports Biomechanics, 20 (2), 213 – 229. https://doi.org/10.1080/14763141.2018.1538385
McPherson, A. L., Feller, J. A., Hewett, T. E., & Webster, K. E. (2019, March). Psychological readiness to return to sport is associated with second anterior cruciate ligament injuries. American Journal of Sports Medicine, 47 (4), 857 – 862. https://doi.org/10.1177/0363546518825258
Meierbachtol, A., Obermeier, M., Yungtum, W., Bottoms, J., Paur, E., Nelson, B. J., Tompkins, M., Russell, H. C., & Chmielewski, T. L. (2020). Injury-related fears during the return-to-sport phase of ACL reconstruction rehabilitation. Orthopaedic Journal of Sports Medicine, 8 (3), 232596712090938. https://doi.org/10.1177/2325967120909385
Miller, R. P., Kori, S. H., & Todd, D. D. (1991). The Tampa Scale: A measure of Kinisophobia. Clinical Journal of Pain, 7 (1), 51. https://doi.org/10.1097/00002508-199103000-00053
Moghadam, S. M., Yeung, T., & Choisne, J. (2023, March 28). A comparison of machine learning models' accuracy in predicting lower-limb joints' kinematics, kinetics, and muscle forces from wearable sensors. Scientific Reports, 13 (1), 5046. https://doi.org/10.1038/s41598-023-31906-z
Molinaro, A. M., Simon, R., & Pfeiffer, R. M. (2005, August 1). Prediction error estimation: A comparison of resampling methods. Bioinformatics, 21 (15), 3301 – 3307. https://doi.org/10.1093/bioinformatics/bti499
Paster, E., Sayeg, A., Armistead, S., & Feldman, M. D. (2022, January). Rehabilitation using a systematic and holistic approach for the injured Athlete returning to sport. Arthroscopy, Sports Medicine, and Rehabilitation, 4 (1), e215 – e219. https://doi.org/10.1016/j.asmr.2021.09.036
Pineau, J., Vincent-Lamarre, P., Sinha, K., Larivière, V., Beygelzimer, A., d'Alché-Buc, F., Fox, E., & Larochelle, H. (2021). "Improving reproducibility in machine learning research" (in English). Journal of Machine Learning Research, 22 164, 1 – 20. https://www.jmlr.org/papers/volume22/20-303/20-303.pdf
Reimers, N., & Gurevych, I. Reporting Score distributions makes a difference: Performance study of LSTM-networks for sequence tagging. https://doi.org/10.48550/arXiv.1707.09861/
Reimers, N., & Gurevych, I. (2017). Reporting Score distributions makes a difference: Performance study of LSTM-networks for sequence tagging. [Paper Presentation]. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Copenhagen, Denmark (pp. 338 – 348). https://doi.org/10.48550/arXiv.1707.09861
Rodriguez, K., Soni, M., Joshi, P. K., Patel, S. C., Shreya, D., Zamora, D. I., Patel, G. S., Grossmann, I., & Sange, I. (2021). Anterior cruciate ligament injury: Conservative versus surgical treatment. Cureus. https://doi.org/10.7759/cureus.20206
Schwartz, M. H., & Rozumalski, A. (2005, January). A new method for estimating joint parameters from motion data. Journal of Biomechanics, 38 (1), 107 – 116. https://doi.org/10.1016/j.jbiomech.2004.03.009
Sundararajan, M., Taly, A., & Yan, Q. (2017, July 17). Axiomatic attribution for deep networks. [Paper Presentation]. International conference on machine learning, PMLR, Sydney, Australia.
Taborri, J., Molinaro, L., Santospagnuolo, A., Vetrano, M., Vulpiani, M. C., & Rossi, S. (2021, April 30). A Machine-learning approach to measure the anterior cruciate ligament injury risk in female basketball players. Sensors (Basel), 21 (9), 3141. https://doi.org/10.3390/s21093141
Trigsted, S. M., Cook, D. B., Pickett, K. A., Cadmus-Bertram, L., Dunn, W. R., & Bell, D. R. (2018). Greater fear of reinjury is related to stiffened jump-landing biomechanics and muscle activation in women after ACL reconstruction. Knee Surgery, Sports Traumatology, Arthroscopy, 26 (12), 3682 – 3689. https://doi.org/10.1007/s00167-018-4950-2
Tripp, D. A., Stanish, W., Ebel-Lam, A., Brewer, B. W., & Birchard, J. (2011). Fear of reinjury, negative affect, and catastrophizing predicting return to sport in recreational athletes with anterior cruciate ligament injuries at 1 year postsurgery. Sport, Exercise, & Performance Psychology, 1 (S), 38 – 48. https://doi.org/10.1037/2157-3905.1.S.38
Webster, K. E., & Feller, J. A. (2018). Psychological factors influencing return to sport after anterior cruciate ligament reconstruction. In Return to play in football (pp. 73 – 83).
Wu, G. (2002, April). ISB recommendation on definitions of joint coordinate system of various joints for the reporting of human joint motion-part I: Ankle, hip, and spine. International society of biomechanics. Journal of Biomechanics, 35 (4), 543 – 548. https://doi.org/10.1016/s0021-9290(01)00222-6
Wu, J., Kator, J. L., Zarro, M., & Leong, N. L. (2022). Rehabilitation principles to consider for anterior cruciate ligament repair. Sports Health, 14 (3), 424 – 432. https://doi.org/10.1177/19417381211032949
Yao, S., Yung, P. S. H., & Lui, P. P. Y. (2021). Tackling the challenges of graft healing after anterior cruciate ligament reconstruction-thinking from the endpoint. Frontiers in Bioengineering and Biotechnology, 9, 756930. https://doi.org/10.3389/fbioe.2021.756930
Zhang, C., Bengio, S., Hardt, M., Recht, B., & Vinyals, O. (2021). Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64 (3), 107 – 115. https://doi.org/10.1145/3446776
Ziab, H., Mazbouh, R., Mahmoodi, R., Gebran, T., Hussein, S., Wehbe, M., Kanj, A., Taha, M., & Reza Hadian, M. (2020). The effectiveness of mental imagery exercises as a neurocognitive-based approach after ACLR: A systematic review. Physical Medicine and Rehabilitation Research, 5 (2). https://doi.org/10.15761/pmrr.1000216
By Abdolamir Karbalaie; Andrew Strong; Tomas Nordström; Lina Schelin; Jonas Selling; Helena Grip; Kalle Prorok and Charlotte K. Häger
Reported by Author; Author; Author; Author; Author; Author; Author; Author