Du Hyun Ro

Background: Knee osteoarthritis (OA) is a heterogeneous disease with varying progression rates among patients within the same severity grade. This study aimed to identify clinically interpretable radiographic phenotypes among knees with baseline Kellgren and Lawrence (KL) grade 2 osteoarthritis and to compare their 4-year radiographic progression patterns using baseline radiographic features, lifetime exercise patterns, and demographic characteristics. Methods: Data of patients with baseline KL grade 2 from the Osteoarthritis Initiative database were retrieved. Unsupervised clustering was performed using uniform manifold approximation and projection (UMAP) for dimensionality reduction, followed by hierarchical density-based spatial clustering of applications with noise (HDBSCAN) to identify distinct subgroups. The proportion of knees progressing to KL grade ≥ 3 within 4 years was compared between subgroups, followed by a survival analysis until knee replacement or progression to KL grade 4. Results: A total of 1997 knees with a mean follow-up of 103.3 months (maximum, 108 months) were included. A total of seven subgroups were identified and classified into accelerated osteoarthritis progression (AOP) group and slow osteoarthritis progression (SOP) group according to whether their 4-year KL grade progression proportions were above or below the training-set average. In a supportive long-term survival analysis, the AOP group also showed significantly earlier progression to knee replacement or KL grade 4 than the SOP group (training set, P < 0.001; test set, P = 0.049), although the absolute difference in mean survival time was modest (8.3 versus 8.5 years in the training set; 8.4 versus 8.6 years in the test set). The AOP group was characterized mainly by lateral joint space narrowing (JSN) and medial JSN with subchondral sclerosis, whereas the SOP group was characterized by medial JSN without subchondral sclerosis, osteophyte-dominant patterns without JSN, or chondrocalcinosis-dominant patterns; these differences were statistically significant across clusters. Conclusions: Within baseline KL grade 2 knees, clinically interpretable radiographic phenotypes, including lateral JSN, medial JSN with subchondral sclerosis, osteophyte-dominant changes without JSN, and chondrocalcinosis-dominant patterns, were associated with different 4-year KL grade progression rates. These findings support viewing KL grade 2 knee OA as a heterogeneous radiographic stage that may be further stratified into phenotype-based subgroups with different short-term radiographic prognoses.

Musculoskeletal care is being reshaped by population aging, multimorbidity, frailty, osteoporosis, sarcopenia, polypharmacy and increasing vulnerability to postoperative loss of independence. Artificial intelligence (AI) has entered orthopaedic research, but its clinical translation remains constrained. Many models address isolated diagnostic, imaging or short‐term prediction tasks rather than the longitudinal care trajectories that determine outcomes in older musculoskeletal patients. The central limitation is therefore not simply whether older adults are represented in datasets, but whether AI systems are valid across the continuum of geriatric orthopaedic care. This review argues for a transition from task‐specific orthopaedic AI toward continuum‐of‐care AI for the aging musculoskeletal patient. Upstream, radiomics, longitudinal osteoarthritis phenotyping, opportunistic computed tomography (CT)‐based skeletal assessment and AI‐augmented dual‐energy X‐ray absorptiometry (DXA) modelling may improve prediction of disease progression, fracture risk, treatment failure and surgical timing. At the point of intervention, AI‐assisted arthroplasty planning may help move total knee arthroplasty from generic alignment targets toward phenotype‐specific reconstruction strategies that account for constitutional alignment, bone morphology, ligament envelope, neuromuscular control, frailty and functional reserve. Perioperatively, automated frailty and sarcopenia phenotyping, wearable gait analysis and multimodal temporal models integrating imaging, electronic health records, laboratory data, medications, patient‐reported outcomes and wearable signals may support dynamic risk prediction. After discharge, continuous monitoring of mobility, sleep, pain trajectories, wound‐related symptoms and rehabilitation adherence may help close the high‐risk surveillance gap during the 90‐day postoperative episode. However, clinical adoption requires more than high discrimination. Explainable AI should be understood as one component of a broader trust framework that also includes calibration, uncertainty quantification, decision‐curve analysis, external and prospective validation, workflow integration, post‐deployment monitoring and accountability. The future of orthopaedic AI is not simply better diagnosis or more precise surgery, but longitudinal, multimodal support across the full care continuum: predicting progression, selecting treatment, personalizing surgery, optimizing perioperative care, monitoring recovery and preserving mobility and independence. Level of Evidence: Level IV.

Preoperative computed tomography (CT) is routinely acquired for planning robotic-assisted total knee arthroplasty (TKA), yet osteophyte burden is not quantitatively characterized. We developed automated CT-based osteophyte quantification and evaluated it against blinded de novo expert-consensus annotations and an intraoperative accessible resected osteophyte volume reference. A human-in-the-loop 3-dimensional U-Net was developed using 387 knee CT scans. The frozen model was evaluated against blinded de novo expert-consensus annotations in two independent external cohorts (100 knees each). Spatial and volume agreement were assessed using Dice coefficient, surface Dice at 1-mm tolerance, 95th-percentile Hausdorff distance (HD95), intraclass correlation coefficient (ICC), mean absolute error, and Bland–Altman analysis. In a prospective cohort of 66 robotic-assisted TKAs, CT-derived total osteophyte volume was compared with the water-displacement volume of accessible osteophytes that were resected and recovered intraoperatively (accessible resected osteophyte volume reference). In the external cohorts, median total Dice was 0.913 and 0.909, surface Dice was 0.944 and 0.945, HD95 was 1.16 and 1.09 mm, and volume ICC was 0.977 and 0.972, respectively. Automated estimates showed conservative volume bias relative to expert consensus (− 0.499 and − 0.365 cc). In the surgical cohort, automated CT-derived volume showed cohort-level agreement with the accessible resected osteophyte volume reference (ICC, 0.876; mean absolute error, 1.224 cc; mean difference, + 0.401 cc; 95% limits of agreement, − 2.785 to + 3.586 cc). Automated CT-based osteophyte quantification showed consistent spatial and volumetric agreement with blinded de novo expert-consensus annotations across two external cohorts and cohort-level agreement with the accessible resected osteophyte volume reference. This scalable method provides a preoperative measure of osteophyte burden for arthroplasty research.

Background: Accurate alignment correction is essential in medial open-wedge high tibial osteotomy (MOWHTO). However, conventional planning based on standing radiographs may not reflect postoperative intra-articular alignment changes. While differences between standing and supine radiographs have been investigated, the role of valgus-stress whole-leg radiographs in preoperative planning remains largely unexplored. This study evaluated the predictive accuracy of standing, supine, and valgus-stress whole-leg radiographs in MOWHTO planning and determined whether integrating these modalities improves alignment prediction compared with single-modality planning. Methods: This retrospective study included 71 patients who underwent MOWHTO. Preoperative standing, supine, and valgus-stress whole-leg radiographs, along with 1-year postoperative standing radiographs, were analyzed. To minimize the influence of surgical execution error and allow the analysis to focus on planning error, the achieved wedge angle was back-calculated from postoperative bony correction and reapplied to each preoperative modality to simulate the postoperative weight-bearing line ratio (Post-WBLR). Accuracy was assessed by comparing simulated and actual Post-WBLR. Multivariable regression was performed to identify predictors of actual Post-WBLR and to compare a multimodal model with the best-performing single-modality simulation. Results: In the postoperative valgus subgroup ( n = 48), supine-based simulation showed no significant difference from actual Post-WBLR (mean difference, 0.0% ± 6.9%; root mean square error [RMSE], 6.8%). In contrast, standing-based simulation underestimated alignment (−9.3% ± 7.8%; RMSE, 12.1%), whereas valgus-stress simulation overestimated it (15.2% ± 6.7%; RMSE, 16.6%). A multimodal regression model incorporating simulated Post-WBLR from standing and valgus-stress radiographs demonstrated significantly improved predictive accuracy compared with the supine-only simulation ( R ² , 0.606 versus 0.080; RMSE, 4.5% versus 6.8%). Notably, in patients with a preoperative standing-to-supine joint line convergence angle difference (∆JLCA _{standing–supine} ) ≥ 2.7°, the multimodal model significantly reduced prediction errors (absolute error: 3.2% versus 6.8%; RMSE: 4.0% versus 9.0%; all P < 0.015). Conclusions: Supine radiographs generally provided the most accurate single-modality basis for MOWHTO planning. Standing and valgus-stress radiographs offered complementary information, and their integration improved prediction accuracy and reduced planning outliers. While further validation is required, a stepwise planning workflow—using supine radiographs as the default modality and multimodal integration when ∆JLCA _{standing–supine} ≥ 2.7°—may serve as an exploratory framework for reducing the risk of planning-related under- or overcorrection.

Aims: Mechanical varus or valgus alignment measures have been demonstrated to be closely related to knee osteoarthritis (KOA) development, but few attempts have been made to correlate the Coronal Plane Alignment of the Knee (CPAK) classification and disease development. This study hypothesized that certain CPAK types or deviations in CPAK parameters might enhance the risk of KOA development. Methods: KOA was defined as a Kellgren-Lawrence grade of 2 or higher. Data from 2,541 knees without KOA at baseline from the Multicenter Osteoarthritis Study (MOST) over five years of follow-up were examined. Differences in baseline CPAK types and parameters were analyzed between knees that developed KOA and those that did not. The association of CPAK type, arithmetic hip-knee-ankle angle (aHKA), and joint line obliquity (JLO) with the risk of disease development and joint line convergence angle (JLCA) changes were assessed. Results: Knees that developed KOA (480 knees; 18.9%) had significantly lower JLO (p = 0.022) and aHKA (p = 0.008) values than those that did not (2,061 knees; 81.1%). JLO was significantly associated with KOA development (odds ratio 1.048 (95% CI 1.001 to 1.098)). For knees with JLO values under the cut-off value of 171.6°, the risk of KOA development was 1.36 times higher than for those with values over the cut-off value (95% CI 1.053 to 1.770). aHKA showed a significant negative correlation with JLCA change (regression coefficient -0.139 (95% CI -0.241 to -0.036); p = 0.008). However, CPAK type itself showed no significant correlation with KOA development. Conclusion: CPAK parameters, rather than types, were significantly associated with KOA development. Importantly, significant obliquity to the constitutional joint line might be as impactful as lower limb alignment. By potentially adjusting cut-off values, CPAK classification might be applied for predicting KOA development and investigating alignment-related aspects of KOA pathophysiology. Cite this article: Bone Joint J 2026;108-B(7):889–897.

Background: Early diagnosis of knee osteoarthritis (KOA) is often delayed due to reliance on subjective interpretation of radiographs. Recent advances in artificial intelligence (AI)-based automated Kellgren-Lawrence (KL) grading offer the potential to improve diagnostic accuracy and enable earlier nonoperative treatment. However, the cost-effectiveness of such AI tools has not been comprehensively evaluated. This study aimed to assess the economic value of AI-assisted KL grading in comparison with conventional radiographic assessment by health-care professionals. Methods: We developed a model-based cost-effectiveness analysis comparing AI-assisted KL grading versus conventional human-reader assessment. A hybrid decision tree and Markov model simulated disease progression and treatment pathways in a hypothetical cohort of South Korean adults aged ≥50 years. Outcomes included lifetime costs, quality-adjusted life years (QALYs), and incremental cost-effectiveness ratios (ICERs). Analyses were performed from a health-care system perspective. Deterministic and probabilistic sensitivity analyses (DSAs and PSAs) were conducted to evaluate the effects of uncertainty. Results: The AI-assisted strategy yielded a total cost of $137,710 and 22.721 QALYs, compared with $140,834 and 22.461 QALYs for usual care. This resulted in cost savings of $3,125 and a QALY gain of 0.260, leading to an ICER of −$12,031 per QALY gained. One-way sensitivity analysis showed that the model was most sensitive to AI-related costs and the effectiveness of nonoperative treatment. PSA demonstrated a 55.1% probability of the AI-assisted strategy being cost-effective at a willingness-to-pay threshold of $34,642. Conclusions: AI-assisted radiographic grading of KOA demonstrated potential cost savings and a favorable cost-effectiveness profile compared with usual care. However, given the uncertainty reflected in the PSA, its integration into routine imaging workflows should be approached cautiously and supported by further validation in real-world clinical settings. Level of Evidence: Economic and Decision Analysis Level II . See Instructions for Authors for a complete description of levels of evidence.

Objective: To compare alignment parameters in bilateral lower extremities and evaluate the effect of contralateral limb on hip-knee-ankle angle (HKAA) longitudinally. Methods: Native bilateral lower extremities of 3,045 Korean patients with long-leg radiographs were followed for progression of knee osteoarthritis. Concordance rates of Functional Knee Phenotypes (FKPs) and Coronal Plane Alignment of the Knee (CPAK), the proportion of patients with identical bilateral subtypes, were assessed. Varus, neutral, and valgus alignments were defined as baseline HKAA ≤177°, 177°< HKAA P = 0.005) in bilateral varus group, medial proximal tibial angle (MPTA) ( P = 0.046) in ipsilateral varus-contralateral neutral group, MPTA, and mechanical lateral distal femoral angle in bilateral neutral ( P < 0.001; P = 0.012, respectively) and valgus ( P = 0.003; P = 0.004) groups, were associated with follow-up HKAA. However, their standardized β values were smaller than ipsilateral values. Conclusion: Concordance rates of FKPs and CPAK between bilateral limbs were 41.6% and 53.5% at baseline, respectively. Contralateral parameters have limited influence on coronal alignment progression compared with ipsilateral parameters.

Background: Artificial intelligence research in orthopedics has grown rapidly, yet a substantial gap remains between technical development and clinical translation. This narrative review summarizes current applications of artificial intelligence in orthopedic practice and highlights barriers to implementation. Main body: Current work converges on three domains: machine learning for structured perioperative risk prediction, deep learning for standardized musculoskeletal imaging, and large language models for workflow and decision support. Applications such as automated fracture detection, Kellgren–Lawrence grading for osteoarthritis, and transfusion risk modeling are approaching clinical maturity. However, routine adoption is limited by algorithmic opacity, performance degradation in new clinical environments, and poor fit within existing workflows. We argue that progress should shift from increasing model complexity toward rigorous evaluation, including external validation on independent cohorts. In addition, probability calibration and uncertainty estimation are important for trustworthy risk communication. Future directions may include multimodal “digital twin” approaches that integrate electronic medical records, imaging phenotypes, and intraoperative data into patient-specific trajectories. Conclusions: The clinical impact of artificial intelligence in orthopedics will depend on life-cycle governance and demonstrated net benefit, prioritizing reliability and implementation science over retrospective benchmark performance.

Aims: Changes in both native and dynamic joint line obliquity may influence the pathogenesis of osteoarthritis (OA) and subsequent development of coronal limb alignment deformity. This study aimed to characterize joint line alterations relative to OA severity and constitutional alignment. Methods: A retrospective, cross-sectional analysis of 3,320 preoperative weightbearing long-leg radiographs from patients undergoing total knee arthroplasty was performed using deep learning software. OA severity was assessed using the Kellgren-Lawrence (KL) grading system. Constitutional alignment parameters included lateral distal femoral angle (LDFA), medial proximal tibial angle (MPTA), arithmetic hip-knee-ankle (aHKA) angle, and arithmetic joint line obliquity (aJLO). Tibiofemoral joint line alterations were assessed using the mechanical hip-knee-ankle (mHKA) angle, tibial joint line angle (TJLA), femoral joint line angle (FJLA), and joint line congruence angle (JLCA). The primary outcome was the difference in the TJLA relative to the aJLO across KL grades, stratified for the presence of medial or lateral compartment OA. Results: Alterations in TJLA were observed in moderate (KL Grade 3) and severe (Grade 4) OA for medial and lateral compartment OA (both p < 0.001). In medial OA, the TJLA reversed orientation at Grade 3 from medially declined to medially inclined relative to aJLO. The aJLO maintained its constitutional orientation until Grade 4, when non-clinically significant changes (< 1°) were observed (p < 0.001). In lateral OA, the TJLA became more medially inclined in moderate and severe OA, while the aJLO remained stable. Significant changes in aHKA were observed in moderate and severe OA, indicating an apparent change in constitutional limb alignment (p < 0.001 for both medial and lateral OA). Conclusion: The tibiofemoral joint line angle changes dynamically with OA progression, while the constitutional aJLO remains stable until advanced disease occurs. Understanding this distinction between constitutional and dynamic joint line orientation is important when communicating and studying outcomes related to knee alignment. Cite this article: Bone Jt Open 2026;7(2):249–258.