ACLART

Teaching a knee to predict its own future.

The objective of the project is to use AI machine learning with clinical data and in silico data to predict revision probability and to enhance pre and post-operative management of ACLR. By harnessing the power of clinical and in silico data-driven insights and AI driven analyses, the project holds the potential to significantly improve patient outcomes and reshape the future of orthopaedic surgery.

The aim of this application is to present an innovative project that intersects Anterior Cruciate Ligament Reconstruction (ACLR) with artificial intelligence (AI) machine learning and in silico simulation for proposing a predictive model for ACLR clinical outcomes. By integrating patient’s clinical and specific surgery data in machine learning models/algorithms and computational simulation, it is possible to deliver insights to improve patient selection, surgical decision-making, pre and post-operative management to enhance the efficiency of ACLR procedures.

The inclusion of AI-driven in silico simulation into the ACLR management is an innovative element of the design approach, as it offers additional possibilities in the overall analysis of all different domains of ACLR surgery and particularly decision-making that is not possible solely with machine learning driven clinical data per se, surgical planning, especially for complex surgeries that pose difficult decisions to the surgical team. In this way, computational finite element results, including those of published studies, will be integrated and driven by machine learning models and algorithms, aiming to correlate the outputs of AI via clinical data and via biomechanical in silico factors.

The methods considered for the project are:

• Data collection from electronic health records and institutional databases that include
  patient pre-operative characteristics, surgical techniques, and post-operative outcomes;
• Use of imaging analysis to assess for each patient ACL injury severity, anatomical factors, and graft viability for ACLR;
• Develop and use of machine learning algorithms to predict ACLR outcomes based on clinical data;
• Biomechanical in vitro and computational simulation (finite element analysis) of ACLR procedures and evaluate the effects of surgical decisions (grafts, tunnel orientation and tibia and femur fixation devices) on knee biomechanics and stability;
• Validation of AI-driven in silico analysis techniques using independent datasets and clinical validation studies to assess prediction accuracy, generalizability, and clinical utility.

Expected outcomes include:

• Development of AI-driven in silico analysis techniques for predicting ACLR outcomes and optimizing surgical planning;
• Enhancement of patient selection criteria and personalized treatment strategies based
  on AI-generated insights and predictive models;
• Improvement in surgical decision-making, intra-operative guidance, and post-operative management through AI-driven computational simulation and predictive analytics;
• Validation of AI-driven in silico analysis techniques through clinical studies and real- world implementation in orthopaedic practice;
• Contribution to the advancement of AI-driven orthopaedic surgery and optimization of ACLR clinical outcomes.

The following research questions are to be investigated:

• How effectively can machine learning algorithms predict post-operative clinical outcomes of ACLR based on preoperative patient data?
• What are the most significant predictors of successful ACLR outcomes identified through AI-driven analysis?
• How does the performance of AI-driven prediction models compare to traditional regression-based methods in forecasting post-ACLR outcomes?
• Can machine learning algorithms accurately stratify patients into risk categories for complications or poor outcomes after ACL surgery
• How do different surgical techniques or graft choices impact the predictive accuracy of AI-driven in silico models for ACLR outcomes?
• Can AI-driven in silico analysis help personalize rehabilitation protocols following ACL surgery to optimize patient recovery and functional outcomes?
• How generalizable are AI-driven in silico prediction models for ACLR outcomes across different patient populations or surgical practices
• What are the potential limitations or biases inherent in AI-driven in silico predictions for ACLR outcomes, and how can these be mitigated
• How does the integration of longitudinal data, including preoperative, intraoperative,
  and postoperative variables, enhance the predictive power of AI models for ACLR outcomes?

To achieve the objectives, the project includes 5 tasks and 4 research teams (surgical medical team, AI machine learning team, in vitro biomodelling team and in silico biomodelling team. The research team includes senior researchers in the fields of medicine, physics, biomechanics and machine learning and master and doctoral students.

Objectives

To predict ACLR outcomes is challenging and machine learning has the potential to
improve our predictive capability. The project represents an innovative and novel approach to estimate subjective failure risk at a patient-specific level when discussing outcome expectations preoperatively, as well as enhancing surgical decision-making and optimizing patient care in orthopaedic surgery. By leveraging AI technologies and computational modelling, the project aims to improve surgery efficiency and personalize ACLR procedures for benefiting patients and advancing the field of orthopaedic surgery in the inevitable AI environment that cannot be ignored.

The objectives are in line with the questions previously presented. However, by introducing in silico simulations it is possible to complement and widen the objectives of the project. Therefore, the complementary objectives include those that will result from the development and implementation of AI in ACLR. The AI algorithms/models and finite element models will enhance ACLR management. The tool as result of the AI-driven finite element analysis can be used to analyse patient-specific characteristics, injury severity, and other relevant biomechanical factors (graft selection, tunnel orientation, tibia and femur fixation devices) and presurgery conditions for surgery (BTB, 4 ST/G, All-inside, other) decision-making and surgical planning to improve post-operative outcomes to minimise the risk of complications.

• The complementary objectives are:
  Improve patient outcomes and long-term prognosis;
• Refine surgical techniques and personalize rehabilitation protocols;
• Reduce complications and failures by identifying potential risk factors of ACLR
  procedures;
• Build a platform for ongoing research and innovation of ACLR by enabling large-scale
  data analysis and modelling with AI techniques;
• Gain deeper insight into the biomechanical factors influencing ACLR outcomes through
  AI-driven simulations;
• Contribute to the advancement of knowledge in ACLR orthopaedic surgery.

Due to the nature of the project, ethics, patient privacy, and algorithmic transparency will be considered to ensure the responsible and beneficial application of AI-driven clinical data and silico analysis in clinical practice.
In a future perspective, the objective is to integrate AI tools into clinical practice to
streamline workflow processes related to ACLR, such as preoperative planning,
postoperative monitoring, and outcome assessment, as well as to foster collaboration among researchers, health professionals, and other stakeholders to leverage collective expertise and data resources for continuous improvement in ACLR treatment strategies.
To enhance post-operative management and rehabilitation it is necessary to monitor patient progress, predict recovery trajectories, and identify early warning signs of complications to facilitate timely interventions. Ultimately, improve patient outcomes is an objective to be pursued to improve the quality of life and patient satisfaction following ACLR surgery.
The contribution to scientific knowledge and advancements in the field of orthopaedic surgery is an objective that is achieved by disseminating research findings, publishing articles, and presenting study results at scientific conferences, as well as sharing results
with the scientific community and healthcare professionals. A website of the project is assumed to play an important role in the dissemination of the project.

Research plan and methods

The research plan considers five main tasks that, at different stages, intersect, allowing information and project results to flow into subsequent phases. For the preparation of this proposal, a comprehensive literature review was conducted, covering topics such as
ACLR outcomes, clinical data analysis, finite element method analysis, and artificial intelligence (AI) applications in orthopaedic surgery. As previously stated, the study will follow the TRIPOD guidelines and the Guidelines for Developing and Reporting Machine Learning Models in Biomedical Research.

The first phase involves clinical data collection, including both pre-operative and post-operative data from multiple sources. These include historical patient records (teamclinical data and orthopaedic registers from Norway, Denmark, and Sweden), surgicalreports, follow-up assessments, and imaging data such as MRI scans. The use of datafrom multiple healthcare institutions ensures diversity and enhances the generalizability of the study findings.

This is followed by data pre-processing, aimed at eliminating outliers and standardizing the dataset to ensure consistency and reliability. Data mining techniques will be applied to extract meaningful patterns and insights, enabling the identification of key factors
influencing ACLR outcomes.

The next stage focuses on the development, training, and evaluation of machine learning models to predict ACLR outcomes based on the collected data. Various algorithms will be employed, including logistic regression, support vector machines, random forests, and
artificial neural networks. Tools such as Python, TensorFlow, and Keras will support the implementation. Cross-validation techniques, such as k-fold and leave-one-out validation, will be used to assess model generalisation and mitigate overfitting.

In parallel, finite element analysis (FEA) data will be collected from peer-reviewed
publications to support in silico model validation and integration with machine learning approaches. The results obtained from these computational models will be validated against experimental in vitro data.
Subsequently, clinical data and in silico data will be integrated within the AI models to improve prediction accuracy and robustness. Model training and validation will be performed using independent datasets or external validation with published data to ensure reliability. Model interpretability will be addressed using approaches such as
decision trees or linear models, facilitating clinical decision-making.

The interpretation and visualization of results will be carried out to support
understanding and enable meaningful insights from the data.

Finally, communication, promotion, and dissemination activities will be undertaken. Research findings will be published in peer-reviewed journals, contributing to the scientific community. Presentations at national and international conferences, as well as
participation in scientific and professional meetings, will allow for knowledge sharing and feedback. Additional dissemination strategies include the development of educational materials such as workshops, webinars, online seminars, and focus groups, aimed at supporting the training of healthcare professionals in the application of AI-driven in silico analysis in ACLR.

Task description and expected results

T1 - Project management, communication, promotion and dissemination
The project management activities that implies demonstration of impact, dissemination and valuation of results has been describe elsewhere in this application. Even though, the target audience include stakeholders such as healthcare professionals, researchers, orthopaedic surgeons, patients and industry partners, who would benefit from the project's outcomes. A comprehensive project management approach will be implemented for promotion and dissemination aiming to effectively communicate research findings, raise awareness about the importance of AI-driven in silico analysis in ACLR, and facilitate knowledge exchange among stakeholders to improve patient outcomes and advance orthopaedic surgery.

In the start of the project, the Portuguese Society of Orthopaedics and Traumatology
(SPOT) will be contacted to promote a national seminar to present the ACLART - Artificial intelligence-driven in silico simulation for prediction of anterior cruciate ligament reconstruction outcomes project. With this it is our intention to included other
orthopaedic surgeons in the process and establish a protocol for research collaboration.
More specific divulgation actions will be made for orthopaedic surgeons of the
Biomechanics and of the Knee Section of SPOT. It is our intention to have surgeons to
make registrations of their ACLR according to specific data information for the machine learning model/algorithms development and testing.

Regarding project management, online collaboration platforms such as Microsoft Teams or Google Workspace will be used for team communication, file sharing, and project management, as well as tools such as Jira, Trello, or Asana for task tracking, milestone management and project coordination. A website of the project will be implemented as well as dissemination activities in social networks.

T2 - Anterior cruciate ligament reconstruction surgery data collection (database) and statistical analysis
The objective of this task is to collect clinical data of ACLR with information regarding
patient specific factors (gender, age, BMI), pre-operative conditions, mechanism of
injury, previous surgeries, ACL injury severity, TALS preoperative, anterior knee pain, location of pain, MRI, pre-surgery KOOS QOL score, pre-surgery KOOS sports score, days/months lesion to surgery and details (graft choice, tunnel orientation, femur fixation device), sports aetiology, right dominant limb, anterior knee pain duration,
hypoesthesia, location, rehabilitation protocols, Knee Walking Test (+), post-operative outcomes like Tegner Activity Level Scale (TALS), Lysholm Knee Scoring Scale (LKSS) and International Knee Documentation Committee-Subjective Knee Form (IKDC-SKF) from electronic health records, institutional databases and international orthopaedic registers.
All researchers of the surgical medical team have extensive experience surgery activity
in ACLR and their data records will be collected from Centro Hospitalar de Vila Nova de Gaia-Espinho, Hospital São Francisco–Porto and Hospitais da Universidade de Coimbra, as well as the inclusion of data of new surgeries made during the duration of the project.
More than a thousand of records of the team are available for the project. These records
include pre-operative MRI scans and other imaging modalities to assess ACL injury
characteristics, concomitant injuries, and anatomical factors relevant to surgical
planning and outcome assessment.

Systematic reviews of results published in orthopaedic knee registers will be performed.
Inclusion criteria will consider application of AI anywhere along the spectrum of predicting, diagnosing, and managing ACL injuries. Since the Portuguese Orthopaedic
Registers lack of suitable information, data will be collected from Swedish, Norwegian
and Danish orthopaedic registers. Clinical data will also be collected from data bases like Pubmed, EBSCO, B-On PEDRo, EBM reviews, Medline, Amed, Embas, Scopus, Google Scholar and Web of Science. The PRISMA of study of inclusions and exclusions will include identification, screening and data eligibility. In this procedure, completeness of the data will be confirmed to identify inconsistencies, errors, or missing values.

Concerning the treatment of the ACLR data disposable of the surgical medical team, it
will be conducted to ensure data integrity and privacy following data protection
regulations and institutional guidelines. The statistical analysis will include comparative,
regression, survival, and sensitivity analyses. The comparative analysis will allow analysis of clinical variables of different patient groups (e.g., graft types, surgical techniques) and identify potential predictors of surgical outcomes, namely probability of revisions. Regression analysis will allow to determine the association between independent variables (patient characteristics, surgical factors) and dependent variables (surgical success, complication rates). A survival analysis will allow to evaluate time-to- event outcomes; such as graft failure or revision surgery rates. Sensitivity analysis are of great importance in statistical analysis to assess the robustness of statistical findings and evaluate the impact of potential confounding variables or modelling assumptions. The complication of all information aiming to generate evidence-based insights into ACLR surgery outcomes, will be the data to design and develop the machine learning algorithms.

T3 - Artificial intelligence machine learning applied to ACLR clinical-surgery data
The machine learning analysis will adhere to both the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis (TRIPOD) guidelines and the Guidelines for Developing and Reporting ML Models in Biomedical Research.

The methodology for developing and evaluating machine learning models for predicting subjective failure risk following ACLR involves a systematic approach integrating various regression algorithms. The methodology will include model training and testing.
As for the model training, ACL and patient data will be the input data for the machine learning algorithm. The methods will include all patients with 2-year follow-up and that have complete clinical data records.

The KOOS QOL, as our primary regression target, can be one of the scores related to the probability of subjective failure 2 years following primary surgery. However, other
subjective scores such as TALS postoperative, LKSS and IKDC-SKF will be analysed, and possibly also predicted in a multivariable regression approach. It is expected that machine learning analysis of different data set origins can predict subjective failure risk following ACLR with adequate accuracy.

To ensure optimal modelling results, a comprehensive pre-processing methodology is required. A combination of data cleaning tasks, such as missing value imputation and
value replacements, coupled with feature engineering will comprise the first step. This
can later be followed by data transformations, such as normalization and one hot encoding. Finally, feature selection and dimensionality reduction techniques can be employed to improve generalization, reduce noise and computational effort.

Subsequently, a diverse set of regression models, including k-Nearest Neighbours,
Decision Tree, Gaussian Process, Support Vector Regression (SVR), Kernel Ridge
Regression, Linear Regression, Stochastic Gradient Descent Regression (SGD), Bayesian Ridge Regression, Lasso Lars Regression, Multi-Task Elastic Net Regression, Ridge Regression (Ridge), and Multi-Layer Perceptron Regression (MLP), is selected for evaluation. Grid search coupled with cross-validation techniques might be used to optimize hyperparameters for each model, ensuring robustness and generalization.
Following hyperparameter tuning, models are trained on the training dataset and
evaluated using appropriate performance metrics, such as mean squared error (MSE),
mean absolute error (MAE), R-squared (R2) score, and root mean squared error (RMSE).

The best-performing model is selected based on cross-validation results and further trained on the entire dataset. These metrics provide valuable insights into the predictive accuracy and generalization capability of each model, aiding in the selection of the most suitable algorithm for predicting subjective failure risk following ACLR.

Finally, the validated model, capable of predicting subjective failure risk following ACLR with adequate accuracy, is deployed to create an easy-to-use in-clinic calculator for point-of-care risk stratification, facilitating preoperative outcome discussions with patients. This calculator can be used to estimate subjective failure risk at a patient-specific level when discussing outcome expectations preoperatively.

T4 - Artificial intelligence machine learning applied to in silico biomechanical ACRL
simulation
For the biomechanical study (in vitro and in silico simulation), focus will be on
mechanical characteristics of ACL composition and types of grafts, surgical preplanning and techniques, factors that can play an important role in the preparation of tibial and femoral tunnels (orientation and fixation devices).

The in vitro and in silico biomechanical modelling encompasses sub-tasks. An
experimental knee replica will be designed and setup to simulate ACLR and results to be
compared with equivalent in silico models for numerical validation. These models will
be assembled with synthetic composite bones, ligament and meniscus replicas using
synthetic materials. INESCTEC and Universidade de Aveiro teams have these types of experimental devices/materials that have been used in other related knee biomechanics studies. For scientific confidence, it is demanding that the finite element models replicate with significant accuracy the experimental results.

For the in silico biomechanical modelling, several important aspects of simulations are
to be critically analysed. In fact, the input parameters of the finite element analysis are
crucial to obtain reliable output data. Computational models of knee joint and ACLR will be built using multibody dynamics software to correctly determine physiological loading conditions, including gait patterns, joint forces, and muscle activations, to simulate in vivo biomechanics of the knee joint. A concern with finite element analysis has to do with the material properties since all structures involved in the knee are complex material structures, inhomogeneous, anisotropic and of non-linear behaviour. Therefore, different type of material constitutive laws (compression-tension non linearity, spring, elastic, hyperelastic, porohyperelastic, and fibril-reinforced porohyperelastic) will be tested. Of these, graft mechanics is a relevant input for adequate numerical simulations and finite element analysis. For this purpose, mechanical properties (stiffness, strength,
and failure modes) of ACL grafts will be collected from literature. Results (experimental and numerical) to be obtained will be crossed and compared with those disposable by literature review concerning finite element analysis of ACLR.

The confidence of the models built to perform finite element simulations and analysis are based on sensitivity analysis to generate sound scientific results. This analysis is demanding to assess the impact of key parameters, such as graft type, size, and orientation, on biomechanical outcomes and identify factors influencing surgical success to be correlated with AI-machine learning clinical outcomes. In a very similar way as described in task 3 - Artificial intelligence machine learning applied to ACLR clinical surgery data, machine learning of the in silico results is performed integrating data to enhance predictive accuracy and clinical relevance and to analyse the influence of the biomechanical results, identify patterns, and predict ACLR outcomes based on biomechanical parameters and surgical variables. The AI-driven in silico analysis models using experimental and computational data will be made to ensure consistency and accuracy in predicting ACLR clinical outcomes.

T5 - Clinical ACLR outcome predictions (revision probability) by artificial intelligence machine learning
The final task consists in the analyses and outcome results obtain from the machine
learning based on clinical and surgery data and in silico data. The correlation analysis
including relevant features from the integrated dataset including both clinical and biomechanical variables will allow to identify associations and potential predictors for ACLR, like review probabilities. These can be of different nature and will be tested to obtain significant clinical meaning, or minimal clinically important difference.

Some important factors affecting the surgical outcome of ACLR include graft selection, tunnel placement, initial graft tension, graft fixation, graft tunnel motion and healing and machine learning can give insights how these affect outcomes and how to enhance surgical decision-making and optimize patient care. For this analysis, objective and subjective scores will be considered and analysed such as the TALS postoperative, LKSS and IKDC-SKF that will be intersected with stress/strain patterns and other biomechanical variables.

Due to the duration of the project, clinical correlation, evaluation and testing will be
made from all patients that undergo ACLR made by the medical team surgery and by
other surgeons of other Portuguese health institutions with whom we expect to make
research protocols aiming the objectives of the project. Only 2 years’ post-operative
results will be considered and included in the machine learning model. It is expected to
have a significant number of ACLR to be analysed within this project.

A long way is expected before the predicative models can be used in real clinical surgery practice. It will be necessary to deploy machine learning models into more statistical significant clinical practice or decision support systems to assist healthcare professionals in surgical planning, patient selection, and post-operative management. The integration
with clinical workflow is a crucial step to streamline processes and enhance efficiency.
Monitoring of the models integration is necessary to interactively and continuously
monitor the performance of created models in real surgery scenarios and rehabilitation
protocols to refine models based on feedback and new data and adapt to evolving clinical needs and patient outcomes and evolving clinical insights to improve accuracy and performance.

It is intended to improve testing and implement effectively the model predictions in the near future in real in vivo scenarios of ACLR by the orthopaedic surgeons of the project and other orthopaedic surgeons. Based on the project results, it is possible to establish surgical guidance to provide insights and recommendations derived from biomechanical modelling and support decision by integrating it with machine learning analysis to assist clinicians in personalized treatment planning, especially for complex ACLR.

The AI machine learning models to be developed will allow to predict ACLR clinical
outcomes like probability of revision for individual patients based on their pre-operative characteristics and biomechanical parameters. These predicted outcomes can be used in several distinct valuable ways, namely, but not all:

• to stratify patients into risk categories to facilitate personalized treatment planning and patient counselling;
• provide decision support tools integrating clinical-biomechanical correlation and predictive models to assist orthopaedic surgeons in optimizing surgical decisions.

 

Créditos