Big data and artificial intelligence in Orthopaedics: Trends and future directions

Predictive analytics in Orthopaedics

Predictive analytics leverages on rich datasets and progressive ML methods to forecast and predict downstream clinical outcomes. Illustrating how this trend may improve clinical care is developing. Domb et al. analysed the pre-operative factors of a cohort of 2415 hip arthroscopy patients via Cox proportional hazards and Fine-Gray models. Thereafter, an institution-specific, ML based prognostic model with a high C index for predicting survivorship and repeat surgeries for patients undergoing hip arthroscopy was created. This model was then adapted into a web based tool to assist surgeons with shared decision making. ²⁴ Abraham et al. applied XGBoost, an open source software library for ML on the 2014 to 2019 National Surgical Quality Improvement Program database of aseptic revision total hip and knee arthroplasty. This was a large database of over 40,000 patients across a wide range of clinical settings with preoperative laboratory values. XBoost was able to handle missing data, reduce model complexity and assess nuanced and non-liner relationships between variables. This generated a model capable of predicting 30-days mortality, cardiovascular events and respiratory complications for aseptic revision total joint arthroplasty, with a C statistics ranging from 0.78 to 0.88. Similarly, this work led to a freely accessible web calculator for assessing preoperative risk. ²⁵ Ghadirinejad et al. recently performed a large multi-study review, modelling postoperative clinical outcomes for hip and knee arthroplasty using supervised machine learning, reliably forecasting outcomes such as readmission, complications, length of stay and patient-reported outcome measures (PROMs). ²⁶ Ghadirinejad et al. were the first to develop preoperative algorithms to predict postoperative opioid use after THA, using multimodal ML algorithms such as stochastic gradient boosting, random forest, support vector machine, neural networks and elastic-net penalized logistic regression.

Predictive analytics based on new ML techniques are able to achieve better accuracy thus aiding Orthopaedic surgeons at determining what is best for patients, optimising resource usage and minimizing surgically related complications. The capability to accurately predict resource utilization, and patient outcomes also provides immense value such as right-siting of care, healthcare policy planning and cost estimation to institutions, governments and third party payors. However, current predictive models rely on retrospective datasets with internal validation only, impeding scalability. While registry-based studies improve sample size and representativeness, external validation across healthcare systems remains inconsistent.^24,26

Imaging-based analytics

Imaging-based AI applications represent one of the most mature domains of big-data analytics in Orthopaedic Surgery, largely due to the availability of standardised imaging modalities. CNNs and DL models may be applied to radiographs, magnetic resonance imaging (MRI) and computerised tomography (CT) scans for diagnostic tasks. Longo et al. performed a recent systematic review summarising over 11 million imaging studies, concluding that AI algorithms which incorporate CNNs consistently achieved high performance on segmentation, detection and classification across multiple imaging modalities. ²⁷ These models have demonstrated performance comparable to or exceeding radiologists in specific, well-defined imaging tasks under controlled conditions, although large scale generalisability across institutions and clinical settings remain limited. For plain radiographs of the wrist, CNN models such as VGG-16, ResNet-50 and GoogLeNet achieved high accuracy in identifying fractures, achieving rates of up to 93% accuracy. In a 2023 systematic review, Dankelman et al. first noted that CNNs applied to fracture recognition and classification on CT scans could aid surgeons by improving diagnostic accuracy, reducing both time to diagnosis and the number of missed diagnoses consistently. ²⁸ This effect is further amplified in cases where there is high inter-observer variability such as when CT scans are interpreted by more than one professional. Despite the improving diagnostic accuracy of these models, Groot et al. demonstrated that many of the current studies are predominantly retrospective, single-centre datasets with limited external validation. ²⁹ Oeding et al. further described that the reliance on retrospectively curated datasets with homogenous imaging protocols, consistent hardware and expert-level annotations result in failure of the models to capture the variability encountered in routine clinical practice. ³⁰ This includes differences in radiographic positioning, exposure parameters, implant designs and reporting standards. Consequently, models demonstrating excellent internal accuracy still exhibit degraded performances when applied to external datasets.^30,31 Collectively, despite promising results, the translational readiness of Orthopaedic imaging based AI remains constrained. Addressing the external validation gap will require purposeful integration of multicentre datasets, registry-based imaging repositories and prospective evaluation frameworks that ensure algorithmic performance is reproducible, generalisable and clinically meaningful.

Robotics and intraoperative data

Robotic-assisted surgeries in Orthopaedics have become more prevalent, especially in joint arthroplasty. Robotic-assisted TKA (RATKA) has demonstrated improved component alignment, reduced hip-knee-ankle (HKA) outliers and reduced coronal plane outliers compared to conventional TKA. ³² Robotic and navigational systems in arthroplasty necessarily require the acquisition of various anatomical landmark data which are usually stored data. This provides a huge anatomical database that can be further researched beyond aiding the performance of the surgery. The current capabilities for ML and DL provides the catalyst for research using this huge minefield of data. As the ideal targets for TKA and THA are constantly changing, AI applications to existing navigational and robotic acquired data can be directed at integrating alignment accuracy and outcome metrics to derive patient specific personalised targets. MacDessi et al. first described the Coronal Plane Alignment of the Knee (CPAK) as a means to examine soft tissue parameters and balancing associated with TKA, comparing kinematic and conventional TKA alignment strategies. ³³ Since then, the use of RATKA has allowed surgeons to accurately plan surgical alignment strategies in order to find the right alignment strategy for the right patient, even in the South-East Asian population.^34–37 Moreover, RATKA has provided surgeons with more real time high-frequency data such as joint angles, alignment parameters, and resection margins, enabling correlation with clinical outcome measures and final implant positioning.

Methodological and translational limitations

A primary limitation in the current developments of Orthopaedic AI is the prevalence of retrospective, single-institution datasets used for model development.^29,31 While these cohorts demonstrate high internal validity by capitalizing on consistent imaging protocols and standardized surgical workflows, they lack the demographic and technical diversity necessary for robust clinical application. Consequently, models developed in these environments often encounter performance degradation when used across a multitude of different clinical settings, driven by domain shift and inter-observer annotation variability. Literature reviews indicate a systemic deficiency in rigorous testing, with only a small fraction of medical imaging AI models undergoing independent external validation.^9,31 Within the Orthopaedic Surgery domain, this representation is sparse. The emerging trend of “negative findings” in external validation studies highlights a critical disconnect between successful proof-of-concept research and the realization of clinically deployable, generalizable diagnostic tools. Addressing this gap requires a transition toward multi-center prospective validation to ensure AI systems remain resilient to the nuances of global clinical practice.

Algorithmic bias

The key driver for progress in AI is data-driven algorithmic computation. Many successful AI models use adaptive AI, which is designed to continuously learn and improve over time by incorporating new data and information, as compared to previous static AI models which are trained on a fixed dataset. ³⁸ These adaptive AI models are able to modify their outputs based on feedback and evolving environments, allowing the models to constantly adapt and stay accurate. However, if these systems are trained on non-representative cohorts and datasets, outcomes and findings can become biased. A 2024 study done by Siddiqui et al. demonstrated that AI models trained using knee and hip X-ray data from Caucasian males performed sub-optimally on females and non-Caucasian populations, leading to misdiagnosis and unideal treatment recommendations. ³⁹ Unfortunately, AI models trained on registry data may reflect biases such as demographic, socioeconomic and institutional biases if key structural frameworks are not put in place before the promulgation of AI models across registries. Kurmis and Ianuzio first noted in 2022 on the integration of electronic health records, imaging, registry data and intraoperative data that due to varying operating and software platforms used and the overly heterogenous datasets generated, data analysis was difficult. ⁴⁰ While AI models trained on registry data offers larger and more robust datasets, the lack of global consistency, the perennial problem of compliance and subsequent generalizability of results will continue to pose a significant hurdle for AI and Orthopaedics.

Regulatory constraints and data privacy

One of the greatest challenges in the foreseeable future for AI and its’ incorporation into Orthopaedics is the regulatory constraints with respect to the safety and equitable integration of data into models. There are divergent regulatory philosophies employed by the United States, European Union and China, as described by Tang et al. ⁴¹ The United States adopts a more flexible and market-oriented approach, while the European Union enforces rigorous data protection measures as per the General Data Protection Regulation, and finally China employs a more comprehensive and process-driven framework heavily backed by real world data validation and algorithm traceability. Even locally, Singapore enforces a slightly different privacy policy following the Personal Data Protection Act (PDPA), which is more business-friendly and pragmatic, emphasising on consent and sectoral flexibility. For Orthopaedic surgeons operating across these different landscapes, differing priorities will be placed with regard to the use of AI in patient care. In the United States, there is priority in clinical performance over interpretability, whereas the European Union demands explainability and defensibility. Ultimately, clinical decisions such as those involving implant selection and fracture classification must both be defensible and transparent yet be made in a highly accurate, precise and reproducible way.

Data deployment

At the core of AI is data, the foundational building block of all algorithms and models. As more emerging technologies in Orthopaedics present new avenues for data collection, Orthopaedic surgeons need to work with data scientists, regulatory authorities and researchers to develop standardised frameworks and workflows to optimise data analysis. In order to reduce algorithmic biases, registries need to be aggregated appropriately using methods such as standardized coding systems, common data dictionaries and harmonized outcome measures, allowing more care to be taken in order to address biases that might arise from registry based data. Chen et al. proposed techniques to mitigate algorithmic bias in a three pronged approach with pre-processing steps that improve subgroup balance through data blinding, augmentation or reweighting, in-processing steps that impose fairness-oriented optimisation constraints and post-processing strategies that recalibrate predictions to improve equity across subgroups. ⁴² This also highlights that AI programmes and technologies, and the subsequent downstream use cases need to be clinician-informed. These AI based programs need to be explainable and surgeon-validated, allowing for seamless integration into clinical workflows and offer clinical utility to the Orthopaedic surgeon. Once algorithms are matured, there needs to be large scale, multi-institutional and international clinical trials to rigorously validate these AI algorithms across different patient populations and clinical settings. These efforts will be instrumental in enhancing generalisability and allow Orthopaedic surgeons to understand the clinical relevance of these tools. With the generation of multiple matured AI systems, Orthopaedic surgeons should then work to establish standardised data protocols and algorithmic transparency to facilitate integration into existing workflows, however diverse the target patient population is. Once these workflows are promulgated, processes will need to have a quality control system. Lee et al. proposed an auditing framework for registry data in a public hospital, improving data quality and ensuring the presence of integrated feedback loops, further providing internal consistency and validation to the collected data. ⁴³

Clinician driven cost-benefit modelling

Research must quantify AI’s return on investments for it to be adopted. AI requires an increasingly sophisticated technological infrastructure along with hiring of experts with subject matter expertise. An AI based project will require multidisciplinary teams involving clinical scientists, data scientists, process and innovation managers. These infrastructure and manpower demands will lead to a significant increase in resource expenditure which must compete with a myriad of other demands. Returns in clinical efficiency, cost efficiency, cost savings and improved patient outcomes must then be used to justify this added expense. Crucially, these returns may not be immediately apparent and a balance will have to sought between increasing expenses and real returns. Hassan et al. described a process of technical evaluation, followed by clinical validation and finally economic validation by institutions and governing bodies as a means of providing adopters with the confidence in the effectiveness and accuracy of AI models. This confidence will then translate into AI use cases that are scalable and equitable. ⁸

Governance

Governance is a key enabler in the successful integration of AI, facilitating progress from conceptualisation to clinical implementation. Hassan et al. recently introduced a new AI governance framework with the goal of equipping healthcare organisations with a means to implement and adopt AI systems. ⁴⁴ The authors proposed an AI framework based on trust, with the three pillars of Ethics, Bias/Fairness and Transparency and Explainability, allowing for an adoption-centred governance framework with clearly defined responsibilities for organisations to consider. Ultimately, a good governance system within a healthcare institution should aim to ensure that AI is deployed safely, ethically and accurately. It should also aim to protect patient safety and foster trust, while still allowing innovation alongside rapid technological advancement. A good governance system should also uphold key ethical principles, maintaining equity, justice, accountability and informed consent. A robust governance structure should therefore comprise of a multidisciplinary team with sufficiently varied expertise and viewpoints with members holding clearly defined responsibilities. Clearly defined responsibilities ensure subsequent transparency, accountability and improves trust related concerns.^5,6,45,46 A governance body also has to responsible for maintaining high standards of equitable data collection, data quality and analysis whilst being adaptive and capable and adaptive at responding to rapid technological advancements and quick changes in regulatory landscapes. ⁴⁷ As this is a nascent field, insight and direction from professional associations, thought leaders and even international collaborative think tanks will continue to shape the trajectory and regulatory frameworks. Overall, governing bodies will take the lead in being the regulatory watchdog that are responsible for AI oversight and approvals, establishing clear and robust regulatory and legal frameworks that clarify development requirements, clinical and technical responsibilities and the bounds of acceptable AI uses in Orthopaedics.^1,4,48 Government and institutional entities will also play critical roles beyond just oversight, ensuring that funding is secured, fostering integration between the healthcare and data science communities, aligning institutional and national goals as well as maintaining long term performance and public health related targets.^6,47