Beyond the algorithms: Challenges in predicting knee osteoarthritis pain from a machine learning point of view

Abstract

Background:

Knee osteoarthritis (OA) presents complex pain experiences often poorly correlated with structural imaging findings. Machine learning (ML) and deep learning (DL) hold promise for predicting and identifying both current and future pain and personalizing interventions, but translating these insights into clinical practice remains challenging.

Methods:

This study highlights key challenges in applying ML and DL to predict knee OA pain, with a focus on the complex relationship between imaging features and subjective pain experiences. While not based on a systematic search or formal data extraction, the review aims to provide an informed overview of current issues and research directions.

Results:

Key challenges include the limited availability of longitudinal data, small and imbalanced datasets, and the difficulty of aligning multimodal data over time. Existing pain assessment tools often fail to capture the dynamic and multidimensional nature of pain. In addition, many studies underrepresent diverse populations and neglect psychosocial factors, which weakens model robustness and limits generalizability.

Conclusions:

Future research should integrate longitudinal and real-time data (e.g., wearables) and promote diverse, collaborative datasets to bridge the gap between ML-based predictions and clinical applicability in knee OA pain management.

1. Introduction

Knee osteoarthritis (OA) is characterized by persistent pain, a hallmark symptom that often drives patients to seek medical attention [1]. This pain is typically described as a deep, aching sensation, exacerbated by activity and alleviated by rest [2]. The experience of pain can lead to reduced physical activity, creating a cycle of muscle weakness and functional decline [3]. Research shows a strong correlation between knee pain severity and the degree of functional impairment, significantly impacting overall quality of life [4].

The impact of knee OA extends beyond physical symptoms, profoundly affecting emotional and social well-being [5]. Patients often experience feelings of frustration and isolation, reducing their participation in social activities due to pain and functional limitations [5]. This highlights the necessity for a comprehensive treatment approach that addresses different aspects of living with knee OA [4], [6].

Pain in knee OA involves overlapping nociceptive, nociplastic, and potentially neuropathic mechanisms [17], [18], [19]. Nociceptive pain, traditionally considered dominant in OA, arises from mechanical stress and tissue damage [20], [21], [22]. Inflammatory processes involving macrophages contribute via cytokines, neuropeptides (e.g., NGF), growth factors, and proteinases. These mediators also interact with nociceptors and may drive peripheral sensitization through effects on the dorsal root ganglia [23], [24]. Nociplastic pain, in contrast, reflects altered central pain processing without clear evidence of tissue damage or nerve lesions [18], [23], [25]. It is often associated with hyperalgesia and diffuse, persistent pain [18]. While some OA patients show signs of central sensitization, its relationship with disease severity remains unclear, and many improve after joint replacement—indicating a continued role for peripheral drivers [18], [25]. Neuropathic pain, defined as pain from a lesion or disease of the somatosensory system, remains less clearly supported in OA [19], [23]. Direct evidence of nerve damage is limited, and neuropathic-like symptoms – often assessed via tools like PainDETECT [26] – may reflect comorbidities (e.g., diabetes, radiculopathy) rather than OA pathology itself [18], [23].

Managing knee OA pain requires a multifaceted approach, incorporating pharmacological interventions, physical therapy, and lifestyle modifications. Nonsteroidal anti-inflammatory drugs (NSAIDs) are frequently prescribed to alleviate pain and reduce inflammation, but their long-term use can result in adverse effects [27]. Viscosupplementation, which involves injecting hyaluronic acid into the knee joint, offers mixed results and some cases has been found to provide symptomatic relief, particularly for those with mild to moderate knee OA [28]. Exercise therapy may also improve joint function, decreases pain, and enhances overall quality of life [29].

In recent years, predicting pain and pain trajectories in knee OA has gained increasing attention as a strategy to enhance patient management and treatment outcomes (Table 1). For example, Previtali et al. [7] conducted a systematic review synthesizing evidence from seven studies involving 7,747 patients. They reported that pain was unstable in nearly half of patients in the short term, but stable in 85% over the mid-term, with 8% experiencing pain reduction and 7% worsening. Predictors of severe or worsening pain included low education, comorbidities, and depression, whereas factors such as age, alcohol use, smoking, coping strategies, and medications had no significant impact.

Table 1.

A list of recent studies examining pain trajectories in knee osteoarthritis (knee OA).

Authors	Year	Title
Previtali et al [7]	2020	Pain trajectories in knee osteoarthritis — a systematic review and best evidence synthesis on pain predictors
Neelapa et al. [8]	2024	Exploring different models of pain phenotypes and their association with pain worsening in people with early knee osteoarthritis: the MOST cohort study
Altamirano et al. [9]	2023	Identifying multivariate disease trajectories and potential phenotypes of early knee osteoarthritis in the CHECK cohort
Collins et al. [10]	2014	Trajectories and risk profiles of pain in persons with radiographic, symptomatic knee osteoarthritis: data from the osteoarthritis initiative
Nicholls et al.[11]	2014	Pain trajectory groups in persons with, or at high risk of, knee osteoarthritis: findings from the Knee Clinical Assessment Study and the Osteoarthritis Initiative
Bastick et al.[12]	2016	Defining knee pain trajectories in early symptomatic knee osteoarthritis in primary care: 5-year results from a nationwide prospective cohort study (CHECK)
Halilaj et al. [13]	2018	Modeling and predicting osteoarthritis progression: data from the osteoarthritis initiative
Pan et al.[14]	2018	Predictors of pain severity trajectory in older adults: a 10.7-year follow-up study
Wesseling et al.[15]	2015	Identifying trajectories of pain severity in early symptomatic knee osteoarthritis: a 5-year followup of the Cohort Hip and Cohort Knee (CHECK) Study
Wiexzorek et al. [16]	2020	What have we learned from trajectory analysis of clinical outcomes in knee and hip osteoarthritis before surgery?

Traditionally, the clinical identification of OA-related pain types – nociceptive, inflammatory, and neuropathic – has relied on a combination of patient history, physical examination, standardized pain questionnaires (e.g., Visual Analog Scale (VAS) [30] and the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) [31], and imaging modalities that detect structural and inflammatory features such as cartilage degradation, bone marrow lesions, and synovitis. Prediction of future pain progression has often focused on demographic and clinical risk factors – such as age, body mass index, comorbidities, and baseline pain severity – alongside imaging biomarkers. These data have typically been analyzed using statistical approaches such as regression models, survival analyses, and latent class growth modeling.

While traditional assessments and statistical models have provided valuable insights at the population level, they often fall short in accounting for the complexity and individualized nature of pain experiences, as well as the multifaceted interactions among biological, psychological, and environmental factors. Pain assessment remains inherently subjective and vulnerable to inter-clinician variability. In contrast, identifying patient-specific pain trajectories offers a route to more personalized and effective management strategies, particularly for individuals with persistent pain who may require more targeted interventions [32]. Moreover, a deeper understanding of these trajectories can empower patients to engage more actively in self-management and treatment adherence [33].

In this context, machine learning (ML) techniques have emerged as promising tools for advancing pain prediction in knee OA (Table 3). Traditional models are often constrained by assumptions of linearity and their limited ability to handle high-dimensional data. ML methods, by contrast, can model non-linear, complex interactions among numerous variables and integrate diverse data modalities – such as clinical, imaging, psychological, and genetic information – into unified predictive frameworks [34], [35]. This capability allows for the identification of latent patterns and relationships that may remain undetected through conventional statistical approaches. For instance, Ma et al. [36] identified key OA-related genes using ML techniques, while Chan et al. [37] used ML to uncover interactions among risk factors contributing to the onset and deterioration of knee OA. More recently, Castagno et al. [38] developed an AutoML tool for predicting rapid knee OA progression using a combination of clinical, radiographic, MRI, and biochemical data.

Nonetheless, despite these advantages, significant challenges remain in building clinically useful ML models for OA pain. The heterogeneous nature of knee OA, the subjectivity and variability in pain reporting, and limitations in data availability and representativeness all pose barriers to developing accurate and generalizable models. Continued refinement and critical evaluation of ML approaches are needed to realize their full clinical potential.

This review aims to explore these challenges and assess the current potential of ML in identifying and predicting pain trajectories in knee OA. Our narrative review is based on a non-systematic literature search, supplemented by our perspectives on the key obstacles and opportunities in the field.

While many of the limitations discussed here are shared with broader epidemiological research in OA, our focus is on how these challenges specifically affect ML model development and application. A detailed technical overview of ML methodologies is beyond the scope of this manuscript. Readers interested in foundational background on ML in healthcare are referred to comprehensive references including [39], [40], [41], [42].

2. Methods

This work is a narrative review aimed at identifying and discussing recent trends, opportunities, and challenges in applying ML to predict pain aspect of knee OA. While we did not conduct a formal systematic review, we followed a structured approach to gather and evaluate relevant literature. We primarily searched for articles using Google Scholar, focusing on studies published after 2010, to ensure the inclusion of recent advances in the field.

Our search strategy included a combination of keywords such as pain, osteoarthritis, knee, machine learning, artificial intelligence, imaging biomarkers, prediction, and using various permutations of these terms. In addition to keyword-based searches, we also examined reference lists and forward citations of key studies identified during the initial search to capture additional relevant papers.

Although the inclusion of studies was not based on predefined systematic criteria, we prioritized high-quality publications based on our own expert judgment. We included studies that were published in reputable journals and conferences, demonstrated methodological rigor, and contributed meaningful insights to the prediction of OA-related pain using ML.

We included studies using a wide range of data modalities, including imaging data (e.g., radiographs, MRI), non-imaging clinical data (e.g., demographics, psychosocial variables, and pain questionnaires), and multimodal datasets that combine both. While our primary emphasis is on the prediction of both future pain and pain trajectories over time, we also included recent studies that apply ML to predict current pain levels in OA patients, recognizing their relevance to understanding pain mechanisms and informing future trajectory modeling.

3. Results

Given the central role of imaging in ML-based knee OA research, this section first focuses on studies that explore the relationship between imaging biomarkers and OA-related pain. We then discuss challenges in predicting pain from imaging data and summarize recent ML approaches, highlighting representative studies and the types of data used.

3.1. Pain and structural findings on imaging

As imaging serves as the primary data modality for novel machine learning models in knee OA research, understanding the relationship between structural findings on imaging and pain is crucial for developing reliable models. Table 2 lists recent studies examining this relationship. For previous studies, we refer readers to the reviews by Barr et al. [43] and Culvenor et al. [44].

In recent years, the identification of biomarkers associated with inflammation and cartilage degradation has gained traction, with studies linking elevated inflammatory markers to increased pain [21]. Advanced imaging techniques, like MRI, have shown promise in assessing cartilage health and predicting pain progression by identifying early structural changes [45], [46]. For example, studies has demonstrated that the presence of BMLs on MRI is significantly associated with pain severity in knee OA patients [35], [47]. Another study [48] reported that meniscal extrusion, full thickness cartilage loss, and osteophytes are associated with more knee pain. Furthermore, the detection of synovitis through MRI has been linked to increased pain levels [49]. Macri et al. [50] also reported that MRI defined osteophytes associated with anterior knee pain.

On the other hand, there are multiple studies showing that structural and symptomatic progression do not always coincide [47], [51], [52], [53], [54], [55], [56]. Thus, there is a discrepancy between structural changes and pain [20], [54], [56]. For instance, some studies show that the presence of pain does not always correlate with the severity of radiographic findings [55], [56]. Patients can experience significant pain and disability despite having minimal radiographic changes, indicating that factors beyond structural damage, such as psychological distress, central sensitization, and being female, may play a critical role in the pain experience [56]. This complexity is compounded by the limitations of radiographic imaging, which often fails to detect the full extent of disease progression. This inconsistency raises questions about the reliance on imaging as a primary tool for assessing pain and may lead to misinterpretations of a patient’s condition. Understanding these dynamics is essential for developing effective models which could be used to shape treatment strategies that address both the physical and psychological aspects of knee OA.

Table 2.

A list of recent studies examining the relationship between knee osteoarthritis (knee OA) pain and structural changes using imaging techniques.

Authors	Year	Key Findings	Imaging modality
Lanois et al. [57]	2024	Frequent AKP at baseline was associated with worsening cartilage damage in the lateral PF joint over 2 years.	MRI
Lin et al. [58]	2023	MRI-based radiomics nomogram achieves a moderately predictive efficacy and accuracy in differentiating the improvement in knee pain.	MRI
Ghouri et al. [59]	2022	Uncertain association between meniscal pathology and knee pain progression, pain severity and pain incidence.	Multi-modal
Collins et al. [10]	2021	Reported that while certain imaging features correlated with pain, the overall relationship was not consistent across populations.	MRI
Cai et al. [60]	2021	No association between change in bone marrow lesion size and pain	MRI
Zhang et al. [61]	2021	Noted variability in pain responses; some patients with minimal structural changes experienced significant pain. MRI findings predictive of pain trajectories; specific structural changes linked to pain severity.	MRI
Wirth et al.	2016	Cartilage loss and BMLs predictive of worsening pain over time.	MRI
Zhao et al. [35]	2024	Machine learning identified cartilage thickness and BMLs as significant predictors of pain severity	MRI
Macri et al. [50]	2022	MRI-defined osteoarthritis features associated with anterior knee pain; highlights importance of imaging.	MRI
Bowes et al. [62]	2021	3D femur bone shape associated with risk of current and future pain.	MRI
Lee et al. [63]	2021	Deep learning based features extracted from MRI associated with the presence of chronic knee pain and prediction of incident knee pain.	MRI
Kraus et al. [64]	2024	Modest association between radiographic subchondral trabecular bone texture and pain.	Radiography

3.2. Machine learning in predicting OA pain

Various machine learning based studies have explored factors influencing pain progression in OA patients, aiming to identify distinct pain trajectories and their predictors (Table 3, [65], [66]).

Zhao et al. [35] demonstrated that machine learning models can identify key structural contributors to knee pain severity – such as cartilage loss, bone marrow lesions, osteophytes, effusion-synovitis, and meniscal damage – by directly analyzing imaging data. Similarly, Chang et al. [67] used deep learning on sagittal MRI slices to predict bilateral knee pain, achieving higher AUCs than traditional ML on posteroanterior and lateral knee X-rays. Bayramoglu et al. [68] developed a deep learning model to detect painful knees directly from radiographs. The results showed that X-ray has significant added value in pain detection compared to basic model (age, sex, BMI). These findings underscore the potential of machine learning to enhance diagnostic accuracy and pain prediction in OA.

Chen et al. [69] employed automated machine learning techniques to predict knee pain progression and functional decline demonstrating the potential for machine learning to integrate various clinical measures for forecasting outcomes. Additionally, the work of Morales et al. [70] demonstrated that data-driven learned quantitative MRI biomarkers could significantly correlate with chronic pain, further supporting the notion that machine learning can uncover critical associations between imaging features and pain outcomes.

Despite the promising advancements, challenges remain in the application of machine learning to predict knee OA pain. There is a clear discrepancy in the research literature which highlights the need for further refinement of predictive models and the potential for integrating diverse data types, including patient-reported outcomes and functional assessments, to improve pain prediction.

In conclusion, the use of machine learning to predict osteoarthritis pain represents a significant advancement in the field, and machine learning models can provide nuanced insights into pain mechanisms and predictors. However, ongoing research is essential to address existing challenges and refine these predictive models, ultimately leading to improved patient outcomes and personalized treatment strategies.

Table 3.

A list of recent studies that examine the relationship between knee osteoarthritis (knee OA) pain and structural changes using machine learning/deep learning.

Authors	Year	Title	Input modality	Pain scope
Bowes et al. [62]	2021	Machine-learning, MRI bone shape and important clinical outcomes in osteoarthritis: data from the Osteoarthritis Initiative	MRI	Current and Future
Hirvasniemi et al. [71]	2023	An image analysis challenge to predict incident symptomatic radiographic knee osteoarthritis from MRI and X-ray images	Multimodal	Future
Guan et al. [72]	2022	Deep learning model for predicting pain progression	Radiographs	Future
Chang et al. [67]	2018	Predicting bilateral knee pain from MR imaging using deep learning	MRI	Frequent
Chang et al. [73]	2020	A deep learning approach to distinguish painful knees from nonpainful contralateral ones	MRI	Frequent
Bayramoglu et al. [68]	2024	Detecting painful knees from lateral and frontal X-rays using deep learning	Radiographs	Current
Zhao et al. [35]	2024	Identifying significant structural factors associated with knee pain severity in patients with osteoarthritis using machine learning	MRI	Current
Hu et al. [74]	2023	DeepKOA: a deep-learning model for predicting progression in knee osteoarthritis.	MRI	Future
Schiratti et al. [34]	2021	A deep learning method for predicting knee osteoarthritis radiographic progression from MRI	MRI	Future
Castagno et al. [75]	2024	Developed machine learning model to predict rapid pain progressors	Multimodal	Future
Kraus et al. [64]	2024	Predictive models of radiographic progression and pain progression in patients with knee osteoarthritis: data from the FNIH OA biomarkers consortium project	Radiographs	Future
Almhdies et al. [76]	2024	Performance of Radiological and Biochemical Biomarkers in Predicting Radio-Symptomatic Knee Osteoarthritis Progression	Radiographs	Future
Lee et al.[63]	2021	Deep learning based features extracted from MRI associated with the presence of chronic knee pain and prediction of incident knee pain.	MRI	Current and Trajectory
Minciullo et al. [77]	2018	Combination of lateral and PA view radiographs to study development of knee OA and associated pain	Radiographs	Future
Chen et al.[69]	2023	Automated machine learning-based prediction of the progression of knee pain, functional decline, and incidence of knee osteoarthritis in individuals at high risk of knee osteoarthritis: Data from the osteoarthritis initiative study	Clinical	Future
Hill et al. [78]	2023	An Algorithmic Approach to Understanding Osteoarthritic Knee Pain	Radiographs	Current
Chan et al. [37]	2021	A machine learning-based approach to decipher multi-etiology of knee osteoarthritis onset and deterioration	Clinical	Future
Van Helvoort et al. [79]	2023	Predicted and actual 2-year structural and pain progression in the IMI-APPROACH knee osteoarthritis cohort	Radiographs	Future
Widera et al.[80]	2020	Multi-classifier prediction of knee osteoarthritis progression from incomplete imbalanced longitudinal data	Multimodal	Future
Nelson et al.[81]	2019	A machine learning approach to knee osteoarthritis phenotyping: data from the FNIH Biomarkers Consortium	Multimodal	Future
Landsmeer ett al. [82]	2019	Predicting knee pain and knee osteoarthritis among overweight women	Multimodal	Future
Lazzarini t al.[83]	2017	A machine learning approach for the identification of new biomarkers for knee osteoarthritis development in overweight and obese women	Multimodal	Future
Fernandes et al.[84]	2017	Risk prediction model for knee pain in the Nottingham community: a Bayesian modeling approach	Multimodal	Future
Morales et al. [70]	2021	Uncovering associations between data-driven learned qMRI biomarkers and chronic pain	MRI	Current (Chronic)

4. Conclusions and discussions

Building upon the studies introduced in the previous section, in this section we critically examine the major challenges encountered in ML-based OA pain prediction and propose directions for overcoming these issues in future research.

In general, machine learning models require comprehensive and large datasets to learn patterns and make accurate predictions. When data is missing or sparse, the models may struggle to generalize findings across different patient populations, leading to reduced predictive accuracy.

4.1. Data quality and availability

Lack of longitudinal data:

Many studies on knee OA pain are cross-sectional, limiting the understanding of how pain evolves over time. Although a small number of longitudinal studies exist, there is a need for more research to capture the dynamic nature of pain and its relationship with disease progression [85], [86]. Further studies are necessary to better understand the associations between different risk factors and the occurrence of pain over time. Without longitudinal data, it is challenging to establish causal relationships and predict future pain trajectories.

Temporal alignment:

Variability in disease progression is a known challenge in studying chronic diseases like OA. In large observational studies or registries, participants often enter at different disease stages or under heterogeneous conditions, making it difficult to define a common temporal baseline (Fig. 1b). Even among individuals with similar baseline characteristics, OA progression can vary significantly in speed and pattern (Fig. 1c). Additionally, data collected from diverse sources – such as clinical visits, imaging, and patient-reported outcomes – may not be synchronized, further complicating temporal alignment. These issues are exacerbated by irregular visit intervals, missing data, and time-varying covariates. While such variability is inherent, analytical strategies like time-to-event modeling, alignment based on symptom onset or structural severity, and trajectory clustering can help accommodate differences in disease timelines in ML models.

Fig. 1.

a. Knee osteoarthritis (OA) pain is known to fluctuate, worsen over time, or improve over time.

b. In studies where patients are recruited, even if they experience similar pain trajectories, it is challenging to align their experiences on a timeline, as their starting points may differ.

c. Patients could be recruited with similar baseline characteristics, including previous pain experience, but due to the heterogeneity of disease progression, a single model may not fit each patient, making it difficult to align and stratify patients.

d. Discrete time measurements of pain could flatten the response and may miss important information between visits, which is a limitation of using discrete time intervals for data collection.

Irregular visits and incomplete records:

Irregular patient visits can lead to incomplete records, which pose a significant challenge for developing robust predictive models. Missing data can introduce bias and reduce the statistical power of analyses, making it difficult to draw reliable conclusions about pain trajectories. For example, if patients miss follow-up visits during periods of increased pain, the resulting data may not accurately represent their experiences, leading to potential misclassification of pain patterns.

Furthermore, pain severity may influence patients’ adherence to follow-up appointments, and this could contribute to data skewness. Patients often enter studies during periods of heightened pain as they seek medical care, which can result in regression to the mean—a statistical phenomenon where extreme baseline values (e.g., high pain scores) naturally move toward the average over time, regardless of any actual improvement. This effect may lead to overly optimistic interpretations of symptom changes.

In addition, survivorship bias can occur when individuals with more severe symptoms drop out of longitudinal studies or are lost to follow-up. This can result in datasets that underrepresent patients with worse outcomes, thereby limiting the generalizability and robustness of machine learning models trained on such data.

Discrete observations:

Longitudinal studies typically involve discrete observations at specific time points, which may not adequately capture the continuous and dynamic nature of pain experienced by knee OA patients. Pain is often fluctuating and can change in response to various factors, including activity levels, weather conditions, and psychological states. The inability to capture these fluctuations in real-time can lead to oversimplified models that do not reflect the true complexity of pain trajectories (Fig. 1d). Short-term pain trajectories might differ from long-term patterns, and discrete observations may miss critical variations in pain experiences (Fig. 1).

Small sample sizes and imbalanced data:

Many studies investigating pain trajectories in knee OA have small sample sizes [7], [16], [87], particularly when stratifying by specific pain patterns or demographic groups. This can lead to imbalanced datasets, where certain trajectory groups are underrepresented, making it difficult for machine learning algorithms to learn effectively from the data. There is a clear need for larger and more comprehensive datasets to improve the robustness of predictive models.

Diversity:

Research on knee OA pain often lacks diversity, with many studies focusing on specific demographic groups [88], [89]. This underrepresentation can lead to findings that are not generalizable to broader populations. Furthermore, studies may not adequately address the experiences of minority groups or individuals from diverse socioeconomic backgrounds, which can significantly influence pain perception and treatment outcomes.

Multiple covariates:

Pain in OA is influenced by multiple factors, including biological, psychological, and environmental variables. For example, anxiety and depression can amplify pain perception, creating a cycle that exacerbates both the physical and emotional impact of OA [90]. To make accurate predictions, machine learning models must account for these complex interactions. However, these factors are often insufficiently addressed in research studies, leading to gaps in data and information. Furthermore, the complexity of these interactions can complicate feature selection and model training, increasing the risk of overfitting or underfitting.

4.2. Pain assessment

The subjective nature of pain makes it challenging to measure it accurately. Various pain assessment tools may not fully capture the complexity of pain experiences in knee OA patients. For example, while the WOMAC score is widely used, it may not adequately reflect the multidimensional nature of pain and disability experienced by patients. Additionally, reliance on self-reported pain measures can introduce bias, as individuals interpret and report pain differently based on personal and cultural factors. This variability complicates the interpretation of research findings and the development of effective treatment protocols.

Furthermore, the phenomenon of pain sensitization where the nervous system becomes more responsive to pain stimuli is not well understood in the context of knee OA. Certain subgroups may be at higher risk for developing pain sensitization, yet the underlying mechanisms remain poorly defined. This knowledge gap limits the development of accurate and reliable machine learning models.

4.3. Heterogeneity of osteoarthritis

Knee OA is a heterogeneous condition with diverse phenotypes and pain experiences among patients. This variability complicates the development of generalized predictive models. The existence of multiple subgroups within the knee OA population requires models capable of addressing this heterogeneity, a challenge that is difficult to meet with conventional machine learning approaches, given current data limitations and collection practices.

In conclusion, the current limitations in understanding knee OA pain from a ML research perspective stem largely from the multifactorial nature of pain, variability in patient experiences, inconsistent correlations between structural changes and pain, limited understanding of pain sensitization, challenges in measuring pain, lack of longitudinal data, underrepresentation of diverse populations, and insufficient integration of psychosocial factors. These challenges must be addressed to advance research and develop effective, patient-centered pain management strategies for individuals with knee OA.

Importantly, many of these are not inherent limitations of ML methodologies themselves – which have shown strong performance across various medical domains – but rather arise from underlying issues such as inconsistent data collection, heterogeneous OA phenotypes, diverse and poorly understood pain mechanisms, and missing or sparse psychosocial information. Recognizing and tackling these barriers is essential to enable the successful clinical translation of ML models into meaningful tools that support personalized care.

Future research should focus on integrating advanced imaging techniques, employing machine learning for predictive modeling, exploring the role of biomarkers and inflammation, incorporating patient-centric approaches, and validating findings across diverse populations. Real-time data collection methods, such as mobile health applications or wearable devices, also offer the potential to capture pain experiences more accurately and frequently. These technologies provide continuous monitoring of pain levels and related contextual factors, allowing for a more comprehensive understanding of individual pain trajectories.

Alongside these technological developments, the clinical utility of ML-based pain prediction models depends not only on their predictive performance but also on how well their outputs can be understood and trusted by clinicians. The integration of explainable AI can help visualize and understand the decision-making process of complex models, offering interpretable insights that are more aligned with clinical reasoning. While a comprehensive review of these techniques lies beyond the scope of this work, future research should prioritize model interpretability to ensure clinical relevance and acceptance.

Equally important is how clinicians interpret and act on the outputs of ML models. Visual or decision-based explanations, such as highlighting influential features or regions in imaging data, can increase confidence in model recommendations and facilitate their use in practice. Incorporating these models into user-centered decision support systems will be crucial for their successful deployment in clinical settings, ensuring that predictions are not only accurate but also actionable.

In addition to biological and structural factors, the integration of psychosocial influences into predictive models remains a critical but underexplored area. Pain perception in OA is shaped by a range of psychological and social variables, including mood, stress, and social support. One promising approach is the application of natural language processing (NLP) techniques to extract relevant psychosocial information from unstructured clinical notes or patient surveys. By leveraging free-text data, NLP could help models account for subjective and context-dependent aspects of pain, ultimately contributing to more personalized and holistic predictions.

Taken together, advancing ML-based prediction of OA pain will require attention not only to technical performance, but also to data diversity, interpretability, integration of psychosocial data, and usability in real-world clinical environments. These combined efforts will move the field closer to developing trustworthy, inclusive, and effective tools for improving pain management in knee OA.

Lastly, we would like to emphasize that we do not underestimate the value of publicly available datasets, such as OAI [91] and MOST [92], which have been pivotal in advancing machine learning particularly deep learning applications in OA research. On the contrary, we recognize that conducting a study of similar scale would be extremely challenging for any single research group. We acknowledge the extensive efforts in designing and collecting data in these foundational studies. However, to fully leverage the capabilities of modern machine learning algorithms, it is still essential not only to focus on developing these algorithms, but also to acquire more comprehensive datasets. This challenge would demand significantly greater resources and collaborative efforts from multiple research groups across diverse locations.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.