Wearable sensor technology and machine learning for anterior cruciate ligament injury diagnosis: an exploratory feasibility study

Abstract

Objectives

To evaluate the feasibility of using wearable inertial measurement units (IMUs; small body-worn sensors that capture linear acceleration and angular velocity) combined with machine learning (ML) to identify anterior cruciate ligament (ACL) injuries during clinical knee joint laxity assessments.

Design

Prospective exploratory feasibility study using a case–control design. A feed-forward neural network classifier was trained on time-normalised IMU signals recorded during Lachman and Anterior Drawer tests to distinguish ACL-injured from healthy knees. Model performance was assessed using 10-fold participant-level cross-validation and reported at repetition-wise and subject-wise levels.

Setting

Biomechanics laboratory within a secondary care setting in London, UK. Recruitment occurred through an acute soft-tissue injury management clinic.

Participants

50 participants were recruited: 26 healthy controls and 24 individuals with an MRI-confirmed ACL injury (partial or complete) sustained within the current injury episode. Healthy controls contributed 52 uninjured legs and ACL-injured participants contributed 25 injured and 23 contralateral uninjured legs (including 1 bilateral injury). Inclusion criteria for the injured group were acute knee injury, MRI-confirmed ACL tear and ability to bear weight. Controls were ≥18 years with no history of knee ligament injury. Exclusion criteria included age <18 or >45, chronic joint conditions, non-weightbearing status, pregnancy, allergy to adhesives or inability to provide informed consent. All participants completed the study.

Interventions

Not applicable. Participants underwent standardised clinical Lachman and Anterior Drawer tests performed by a specialist physiotherapist while IMUs were mounted on the femur and tibia using a custom three-dimensional-printed rig.

Primary and secondary outcome measures

The primary outcome was diagnostic accuracy of an ML model classifying ACL injury status (injured vs healthy) based on IMU-derived linear acceleration and angular velocity (three axes per sensor). Secondary outcomes included sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), area under the receiver operating characteristic curve (AUROC) and F1 score. Planned model assessments (repetition-wise and subject-wise) were completed as intended. Hyperparameter adjustments were made post hoc to address overfitting when expanding to the full dataset.

Results

Across the full dataset (n=50), the comparative leg model (method 2) achieved a subject-wise diagnostic accuracy of 81%, sensitivity 83%, specificity 79%, PPV 80% and NPV 83% using 10-fold participant-level cross-validation. Repetition-wise accuracy was 71%. Discriminative performance was moderate (AUROC 0.773), with an F1 score of 0.889. The individual leg model (method 1) showed lower performance (subject-wise accuracy 70%; sensitivity 43%; specificity 81%).

Conclusions

This study supports the feasibility of wearable sensors to be used in the clinical assessment of ACL injury to assist in capturing movement features associated with joint laxity. More work should be done to increase generalisability and validate findings further.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• A standardised clinical protocol was used for all participants, with tests performed by the same specialist physiotherapist to minimise inter-examiner variability.

• Inertial measurement unit placement was controlled using a custom three-dimensional printed rig designed to reduce soft-tissue artefact and improve data consistency.

• K-fold cross-validation was employed to evaluate model robustness but the modest sample size limited performance generalisability.

• The physiotherapist conducting index tests was not blinded to clinical status, introducing potential assessment bias.

• Only anterior laxity tests were examined, restricting the biomechanical inputs available for model development.

Background

Wearable sensors are increasingly used to track and assess biomechanics; their wireless, unobtrusive and portable nature has led to them being favourable in providing biofeedback for performance optimisation.¹ Their use within musculoskeletal (MSK) diagnostics, however, is much less established. Wearable sensors, particularly inertial measurement units (IMUs), have revolutionised movement tracking in sports science, rehabilitation and injury prevention.¹ While extensively used for performance monitoring, their diagnostic potential in MSK injuries remains underexplored.²,⁶ In anterior cruciate ligament (ACL) injuries, IMUs could provide objective laxity measures, reducing reliance on clinician-dependent manual laxity examinations. Therefore, the aim of this study was to explore whether IMUs combined with machine learning (ML) could identify biomechanical markers of anterior knee laxity associated with ACL tears during standard clinical tests.

ACL injuries commonly present to accident and emergency and similar acute-care settings in the UK, where diagnosis is frequently delayed or missed. Delays to definitive diagnosis and specialist review remain common, particularly when patients are not assessed in dedicated knee clinics or by clinicians with relevant MSK expertise.^{7 8} In this context, expertise in acute knee injury refers to clinicians with specific training and regular clinical experience in acute knee assessment, such as specialist knee physiotherapists, sports medicine clinicians or orthopaedic clinicians with focused MSK practice. Despite their widespread use, clinical tests such as the Lachman and Anterior Drawer remain inherently subjective, with diagnostic performance highly dependent on examiner experience, injury acuity and patient-specific factors including pain, swelling and muscle guarding.^{7 9 10} In acute and non-specialist settings, diagnosis at initial presentation has been reported to be as low as 14%, with delays of up to 21 months before confirmation, often deferred until specialist review or MRI.^{11 12} Reported sensitivity and specificity of clinical laxity tests vary widely in these contexts (ranging from approximately 9.5%–100%), highlighting the limitations of relying solely on manual grading.⁹ This diagnostic uncertainty underpins the need for objective methods capable of quantifying knee laxity during standard clinical assessments, particularly tools that can be deployed early in the care pathway without altering the examination itself.

The use of IMUs to measure knee instability has started to be investigated; however, this has primarily focused on the pivot shift test. This test involves manipulating the knee to sublux and then reduce it, generating a distinct acceleration-deceleration pattern detected by IMU components.^{4 5} Unlike the pivot shift test, which requires anaesthesia and is primarily used postsurgery, the Lachman and Anterior Drawer tests are widely used in acute ACL injury assessment.^{23 13},¹⁵ These manual tests rely on subjective grading, creating an opportunity for IMU-based quantification to improve diagnostic accuracy.

The wider potential of wearable technology to quantify joint laxity and aid in ACL tear diagnosis has not been extensively explored. While limiting factors such as pain, effusion and muscle guarding remain obstacles, objective quantification of these intricate movements may help identify subtle changes in laxity and reduce reliance on subjective interpretation. The aim of this study was not to produce a direct quantitative measurement of laxity in millimetres, but to determine whether IMU-derived motion patterns during standard laxity tests could act as surrogate markers of laxity and be used by a ML model to classify ACL injury status thereby enhancing diagnosis speed and accuracy in acute settings and facilitating earlier referrals to appropriate management paradigms. While IMU-based assessment does not shorten the clinical testing manoeuvre itself, it has the potential to accelerate decision-making by providing immediate quantitative outputs that do not require manual grading. We hypothesise that wearable sensors will prove useful as a diagnostic indicator of ACL injuries.

This study explores the use of wearable sensors combined with ML to support ACL injury assessment during standard clinical laxity tests. It is presented as a feasibility study to evaluate the practical implementation and preliminary diagnostic performance of this approach prior to larger-scale validation.

Methods

Equity, diversity and inclusion

This study was designed and conducted in accordance with principles of equity, diversity and inclusion. Recruitment was open to all eligible individuals regardless of sex, race, ethnicity or socioeconomic status. The research team included members from diverse clinical, technical and research backgrounds. Analyses did not explore subgroup differences by demographic characteristics due to sample size constraints; future work will address performance across demographic strata.

Participants and recruitment

The study included 26 healthy controls and 24 individuals with ACL injuries (total n=50). ACL-injured participants were recruited prospectively from a secondary-care soft-tissue injury clinic using convenience sampling. Healthy controls were recruited via institutional advertising. Healthy controls contributed 52 uninjured legs, while ACL-injured participants contributed 25 injured legs and 23 contralateral uninjured legs (one participant had bilateral ACL injuries). Of the injured cohort, 19 had complete ACL ruptures and 6 had partial ACL tears confirmed by MRI.

Eligibility criteria

Eligibility criteria were defined a priori. Participants in the injured group were required to present with an acute knee injury, have an MRI-confirmed ACL tear (partial or complete), and be able to bear weight at the time of assessment. Healthy control participants were eligible if they were aged 18 years or older and reported no previous history of knee ligament injury. Exclusion criteria for all participants included age under 18 or over 45 years, presence of chronic joint disease, non-weightbearing status, pregnancy, known allergy to adhesive materials used for sensor attachment or insufficient English proficiency to provide informed consent. No adverse events occurred during clinical testing or MRI assessment.

Sample size calculation

Sample size calculations were based on effect sizes from a prior ACL gait study, indicating that 44 participants would be required to achieve 80% power at α=0.05. To accommodate expected heterogeneity in acute injuries and potential data loss, the target sample size was increased by approximately 10%–50% participants. Ultimately, 51 participants were recruited and included in the analysis.

Equipment

Biomechanical assessments were performed in a laboratory environment using Blue Trident IMUs (IMUs; Vicon, Oxford, UK) and an optical motion capture (OMC) system (Vicon, Oxford, UK). IMUs captured linear acceleration and angular velocity along three axes (x, y, z) for both the femur and tibia. The OMC system was used solely to assist with repetition segmentation and time normalisation and was not used as an input to the ML models.

Sensor placement and setup

Two IMUs were used for each participant:

• one mounted on the distal femur, and

• one mounted on the proximal tibial shaft.

Rather than being placed directly on the skin, the IMUs were fixed using a custom-built, three-dimensional-printed assessment rig designed to anchor over bony prominences of the leg and to reduce soft-tissue artefact (figure 1). The rig ensured consistent sensor orientation and positioning across participants.

Figure 1. Setup of three-dimensional-printed assessment rig.

Clinical testing protocol

All clinical tests were performed by a Health Care Professions Council (HCPC) registered physiotherapist, a UK Bachelor of Science degree in Physiotherapy, 5 years of postqualification clinical experience in acute knee assessment, and a PhD in wearable sensors and applied biomechanics in ACL injuries. Participants lay supine on a plinth while the physiotherapist performed standardised Lachman and Anterior Drawer tests on each eligible leg.

For each participant:

1. The IMU rig was secured to the femur and tibia.

2. The Lachman test was performed multiple times on each test leg.

3. The Anterior Drawer test was then performed multiple times on each test leg.

The physiotherapist performing the index tests was aware of participants’ clinical history and MRI findings. MRI assessments were conducted as part of routine clinical care and were interpreted independently, without access to IMU data or ML outputs.

Data processing and segmentation

Raw IMU data from the femoral and tibial sensors were recorded continuously during each testing session. OMC markers were used to identify the start and end points of each Lachman and Anterior Drawer test repetition. These timepoints were used to segment the IMU signals into individual repetitions.

Each repetition was time-normalised to 0%–100% of the test movement to enable comparison across participants and repetitions. For every repetition, each IMU provided six time-series signals (three axes of linear acceleration and three axes of angular velocity), resulting in a 12×101 matrix (12 channels by 101 time-normalised points) per repetition.

Model architecture

A feed-forward neural network ML framework was designed to classify ACL injury status (injured vs healthy) using IMU-derived surrogate markers of knee laxity. Inputs to the model were the time-normalised linear acceleration and angular velocity signals from the femoral and tibial IMUs recorded during the Lachman and Anterior Drawer tests. Models were trained using a batch size of 1024 samples and a learning rate of 0.01 for 30 training epochs, which represent standard training parameters for networks of this size. Two modelling approaches were evaluated. Method 1 (individual leg analysis) employed three hidden layers with 64, 8 and 8 neurons, respectively, and used selected accelerometer and gyroscope axes (0, 1, 3, 4, 5, 6, 7 and 11). Method 2 (comparative leg analysis) employed two hidden layers with 256 and 16 neurons and used a reduced set of axes (3, 5, 7, 8 and 11), reflecting features most informative for interlimb comparison.

The architectural configurations and input axes for each method were identified empirically through experimentation to optimise diagnostic performance while minimising overfitting.

Two modelling strategies were explored:

1. ML method 1: Individual leg analysis (figure 2)

Figure 2. Representing the neural network of methodology 1. The first approach to process the inertial measurement unit data to predict ACL injuries. Each black circle is a neuron of the neural network model. Blue and red lines are the network parameters. ACL, anterior cruciate ligament; IMU, inertial measurement unit.

This approach used a dual neural network architecture. One network processed the IMU data from a single leg, while an additional component incorporated contextual information about leg side (right or left) Error! Reference source not found. The model aimed to characterise the dynamics of each leg independently, without explicitly modelling side-to-side differences within participants.

• 2 1 ML method 2: comparative leg analysis (figure 3)

Figure 3. Figure representing the neural network of methodology 2. The second approach to process the inertial measurement unit data to predict ACL injuries. Each black circle is a neuron of the neural network model. Blue and red lines are the network parameters. (IMU_f^ant–Femoral IMU during anterior draw test, IMU_f^lach–Femoral IMU during Lachman test, IMU_t^ant–Tibial IMU during anterior drawer test, IMU_t^lach–Tibial IMU during Lachman test). ACL, anterior cruciate ligament; IMU, inertial measurement unit; l, left; r, right.

This approach combined IMU data from both legs for each participant to capture interleg differences. A single neural network model processed this combined dataset, allowing the model to leverage comparisons between injured and non-injured legs (or between both legs where both were healthy or injured) to detect patterns associated with ACL injury. Error! Reference source not found.

For both methods, the input axes and specific sensor channels included were selected based on their relevance to flexion-extension and anteroposterior knee dynamics.

Model training and validation

To minimise data leakage and assess generalisability, model training and validation were conducted using 10-fold cross-validation at the participant level. In each fold, participants were partitioned such that:

• nine-tenths of the sample were used for training, and

• one-tenth were held out for testing.

This procedure was repeated ten times so that each participant appeared in the test set exactly once. All repetitions from a given participant were assigned to the same fold to prevent overlap between training and test sets.

The neural networks were trained using a binary cross-entropy loss function, with ACL-injured legs labelled as 1 and healthy control legs as 0. A gradient-based optimisation algorithm was used to update network weights via backpropagation. Hyperparameters (including batch size, hidden layer sizes, dropout rate, learning rate, step size, and number of epochs) were tuned through a series of experiments to identify settings that improved performance while reducing overfitting.

Two levels of prediction were generated:

• Repetition-wise classification: the model produced a prediction for each individual test repetition.

• Subject-wise classification: predictions from all repetitions belonging to a given participant were aggregated, and the most frequent predicted class was taken as the final classification for that participant.

Subject-wise classification was considered the primary and clinically meaningful outcome, as diagnostic decisions are made at the participant level rather than on individual test repetitions; repetition-wise analysis was included to characterise variability across repeated manoeuvres and to provide contextual insight into model behaviour.

For each participant, a minimum of five repetitions of both the Lachman and Anterior Drawer tests were performed per test leg. The number of repetitions was not fixed and could vary between participants, with additional repetitions undertaken at the discretion of the assessing physiotherapist where initial manoeuvres were limited by muscle guarding, patient apprehension or equivocal joint feel. This approach reflects routine clinical practice in acute knee assessment and was appropriate for this feasibility study. The reported model outputs were derived from combined IMU signals recorded during both the Lachman and Anterior Drawer tests; test-specific performance was not analysed separately in this feasibility study.

Performance metrics

Model performance was assessed using standard diagnostic performance metrics:

• Accuracy

• Sensitivity

• Specificity

• Positive predictive value (PPV)

• Negative predictive value (NPV)

• Area under the receiver operating characteristic curve (AUROC)

• F1 score.

These metrics were calculated for both repetition-wise and subject-wise classification. Receiver operating characteristic (ROC) curves were generated, and AUROC values were used to summarise overall discriminative ability. The F1 score was used to characterise the balance between precision and recall, given the clinical importance of minimising both false positives and false negatives. All reported performance values are presented in the Results.

Results

Demographic information is outlined in table 1 and population separation is explained in figure 4.

Table 1. Demographics table of included participants.

Characteristic	ACL-injured (n=25)	Healthy controls (n=26)
Age (mean±SD)	27.6±5.4 years	28.2±6.1 years
Sex (M/F)	14/11	13/13
Height (mean±SD)	174.5±8.3 cm	172.1±7.6 cm
Weight (mean±SD)	72.4±10.1 kg	70.6±9.8 kg
BMI (mean±SD)	23.8±2.7	23.9±2.5
Time since injury	Median 3 weeks	–

ACL, anterior cruciate ligament; BMI, body mass index.

Figure 4. Flow of participants. ACL, anterior cruciate ligament.

Two modelling approaches were evaluated using 10-fold participant-level cross-validation:

1. Method 1: individual leg analysis.

2. Method 2: comparative leg analysis.

Method 1: individual leg analysis

Method 1 evaluated each leg independently, without incorporating interlimb differences.

Using the full dataset, method 1 achieved:

• Repetition-wise accuracy: 65%

• Subject-wise accuracy: 70%

• Subject-wise sensitivity: 43%

• Subject-wise specificity: 81%

• Subject-wise PPV: 48%

• Subject-wise NPV: 77%

This represents the baseline performance of single-limb IMU classification during laxity testing.

Method 2: comparative leg analysis

Method 2 combined IMU data from both legs for each participant and incorporated interlimb comparison.

Using the full dataset, method 2 achieved:

• Repetition-wise accuracy: 71%

• Subject-wise accuracy: 81%

• Subject-wise sensitivity: 83%

• Subject-wise specificity: 79%

• Subject-wise PPV: 80%

• Subject-wise NPV: 83%

Diagnostic performance improved across multiple metrics compared with method 1, indicating that bilateral comparison provides additional discriminatory value for detecting ACL injury.

Discriminative performance

Across the full dataset, the optimised model achieved:

AUROC: 0.773

F1 score: 0.889

These results reflect moderate discriminatory ability and a balanced relationship between precision and recall (figure 5).

Figure 5. Receiver operating characteristic (ROC) curve for participant-level predictions. The area under the curve (AUC) was 0.77, indicating moderate diagnostic accuracy.

Model comparison summary

Method 2 consistently outperformed method 1, suggesting that interlimb biomechanical differences captured by IMUs enhance classification performance. Method 2 therefore represents the preferred framework for future model refinement and clinical validation.

Discussion

Diagnosing ACL injuries remains challenging due to variability in joint assessment interpretation, impacting clinical decision-making across healthcare settings.¹⁶,¹⁸ This study highlights the potential of wearable devices and ML models in improving diagnostic accuracy by providing objective, data-driven assessments to complement traditional clinical examinations. Even among experienced clinicians, diagnostic accuracy in ACL assessments varies, with sensitivity and specificity averaging 81%.¹⁸ However, initial assessments are often conducted by non-specialists or emergency physicians, where sensitivity can be as low as 25.9%, leading to frequent missed or delayed diagnoses.^{16 17} These disparities underscore the need for objective tools to support early and accurate diagnosis.

Wearables are well-established in performance monitoring and biomechanics¹⁹; they offer the potential to perform real-time knee dynamics assessment, enhancing clinical utility.^{1 9} This study demonstrates their potential in ACL diagnostics, with knee-mounted sensors processing IMU data in real time to assist clinical decision-making. The aim of this approach is not to replace diagnostic imaging, but to support earlier and more confident clinical assessment in acute settings. By providing an objective measure of joint behaviour during standard laxity tests, wearable-assisted assessment may help prioritise referrals for imaging and reduce delays or unnecessary investigations where clinical findings are equivocal. Integrating such technology could improve diagnostic confidence, reduce reliance on imaging, and facilitate rapid assessments in acute settings.²⁰,²²

Two modelling strategies were evaluated using participant-level cross-validation on the full dataset. The individual leg analysis (method 1), which classified each limb independently, demonstrated moderate diagnostic performance, achieving a subject-wise accuracy of 70% with high specificity but limited sensitivity. In contrast, the comparative leg analysis (method 2), which incorporated bilateral limb data to capture interlimb differences, substantially improved performance across multiple diagnostic metrics. This approach achieved a subject-wise accuracy of 81%, with balanced sensitivity (83%) and specificity (79%), alongside an AUROC of 0.77. These findings indicate that interlimb biomechanical comparison provides meaningful additional discriminatory value and represents a more robust framework for wearable-based ACL injury classification.^{23 24}

The diagnostic performance achieved using bilateral IMU data is comparable to reported accuracy of specialist-performed clinical laxity tests under optimal conditions, while operating independently of examiner grading. Although specialist assessment may achieve higher accuracy in controlled settings, such performance is not consistently replicated in emergency or non-specialist contexts, where diagnostic accuracy is substantially lower (sensitivity as low as 25%–30%).^{16 17} However, all clinical tests in this study were performed by a specialist physiotherapist, so these findings should not be assumed to generalise to non-specialist clinicians or frontline emergency settings. This approach may be most relevant where variability lies in interpretation of laxity rather than execution of the clinical manoeuvre itself, although IMU-based assessment would not remove limitations related to pain, swelling, guarding or suboptimal test performance.

ROC analysis demonstrated moderate discriminatory capability (AUROC 0.77), and the high F1 score (0.889) indicates a favourable balance between sensitivity and precision. Together, these metrics suggest that bilateral IMU-based modelling captures clinically relevant information during manual laxity testing and supports further investigation of wearable technologies as objective decision-support tools in acute knee injury assessment.

While the model’s current accuracy is based on a relatively small dataset, further expansion and increased model sophistication could improve diagnostic performance. Future improvements could include incorporating additional inputs such as gait assessments and clinical history. Larger datasets and more complex models may better define the clinical role of wearables in ACL assessment.²⁵ Future work should focus on increasing dataset size and incorporating additional biomechanical features to enhance robustness.²⁶

While the model’s current accuracy is based on a relatively small dataset, further expansion and increased model sophistication could improve diagnostic performance. Further improvements to the model could include incorporating additional inputs such as gait assessments and clinical history. A larger dataset and more complex models may better align wearables with clinical demands, supporting their integration into healthcare pathways. These advancements could streamline workflows, reduce unnecessary imaging and optimise resource allocation for specialist referrals.^{23 26 27} Future work should focus on increasing dataset size and incorporating additional biomechanical features to enhance robustness.

Clinical implications

Integrating IMU with OMC (VICON) data could further enhance biomechanical analysis^{28 29}; however, the set-up requirements for motion capture limit its feasibility in most clinical environments. Despite the benefits of enriching injury assessment with wearable technology, challenges remain regarding data privacy, clinical integration and implementation costs, and pilot studies assessing real-world usability and regulatory considerations will be essential. Sensor placement requires only a few minutes and could be performed by support staff rather than the assessing clinician, minimising workflow disruption, although it does introduce an additional step. Any added set-up time may be justified if wearable-derived outputs provide useful objective information alongside routine clinical examination. ML models are increasingly being explored within clinical workflows, and similar approaches in ACL diagnostics may have future utility in supporting interpretation, providing feedback during testing and monitoring change over time, although these applications were not evaluated in the present study.

Limitations

This study is limited by its modest sample size, which restricts generalisability despite the use of participant-level cross-validation. Model performance may also be influenced by variability in test execution, sensor noise and patient-specific factors such as pain, swelling and muscle guarding. Additionally, the model primarily focused on anterior knee laxity. As the IMU data were collected during the same manual laxity tests, the sensor outputs remain dependent on the quality of the underlying clinical manoeuvre and therefore do not eliminate limitations related to examiner technique or patient presentation. Future studies should incorporate rotational, valgus and varus assessments to broaden diagnostic capabilities. Integrating OMC data and clinician input could improve generalisability. While hyperparameter optimisation improved performance, further refinement and exploration of alternative ML architectures are necessary for clinical scalability.

Conclusions

This feasibility study demonstrated that IMU-derived data can be collected during manual laxity tests and used to classify ACL injury with moderate diagnostic performance. These findings support the feasibility of wearable sensors as an adjunct to specialist clinical assessment for capturing movement features associated with joint laxity. While the present results do not establish generalisability beyond this setting, they provide an initial basis for further investigation of IMU-based assessment within ACL diagnostic pathways. Larger studies are needed to validate these findings and assess performance across different clinical settings and levels of examiner expertise.

Data availability statement

Data are available on reasonable request.