MRI-based patient-specific nomogram for diagnostic risk stratification of patients with early knee OA

Abstract

Objectives

This study is to develop a risk stratification nomogram for early-stage OA based on MRI, especially with potential sequences of MRI called T1rho and T2 mapping.

Methods

Cartilages diagnosed with early-stage OA or normal were collected and allocated into training or validation cohorts after the MRI. Eleven predictors were determined as candidate predictors for OA-anatomical signature-nomogram (OA-ASN). The performance of OA-ASN was evaluated using the concordance index (C-index), the area under the receiver operating characteristic curve (AUC), calibration plots, decision curve analysis (DCA) and clinical impact curve (CIC).

Results

A total of 199 patients were evaluated. Of these, 79 (39.7%) had early OA. Infrapatellar fat pad (IPFP), T1 rho and T2 mapping were independently associated with early-stage OA at multivariable analysis. The nomogram incorporating these variables displayed excellent discrimination (C-index, 0.975; 95% CI: 0.951, 0.999) in the training sample (n = 115) and bootstrap validation (C-index, 0.96), while C-index was 0.904 (95% CI: 0.840, 0.959) in the validation cohort (n = 84). The calibration plots showed favourable consistency between the prediction of the nomogram and actual observations in both the training and validation cohorts. The DCA and CIC showed that the nomogram was clinically useful.

Conclusions

A smaller volume of IPFP, T1ρ value >33 and T2 mapping >35.04 were significantly associated with OA. The OA-ASN demonstrated excellent predictive outcomes with easy-accessible and simple observational screening methods based on physiological MRI, which can provide individual treatment strategies.

Graphical abstract

Rheumatology key messages.

• The nomogram offers a simple, non-invasive tool to accurately predict early-stage OA risk based on physiological MRI.

• Infrapatellar fat pad, T1 rho and T2 mapping were independent risk factors for early-stage OA.

• The nomogram enables risk stratification for early-stage OA, providing precise guidance tailored to individuals of varying risk levels.

Introduction

Early diagnosis and personal monitoring are crucial for OA patients

OA is the most prevalent joint disease with an age-standardized prevalence estimated at 3754.2 per 100 000 in 2017 globally [1]. Moreover, a recent study [2] revealed that prevalent cases of OA increased from 247.51 million in 1990–527.81 million in 2019 worldwide [1, 2]. As the studies stated, the prevalence of knee OA was 2.50% in individuals aged 30–39, while the pooled prevalence of knee OA was 22.9% in individuals aged 40 and over [3]. The dramatically increasing age-dependent prevalence indicates that early diagnosis is crucial to halt OA progression. Furthermore, the final treatment opportunity for OA is joint replacement, whose servicing life is limited, and the postoperative function might be poor, leading to high costs and personal burdens [4]. Strong evidence exists that OA can be prevented through lifestyle changes at the early stage [5]. Therefore, early diagnosis, personal monitoring and preventive treatment significantly prevent or halt development [6].

The goal and status of early OA diagnosis

Patients with OA are currently stratified based on clinical symptoms and traditional imaging (X-ray and morphological MRI). However, early-stage OA symptoms like knee pain and swelling are often subtle and may not lead to diagnosis until years after onset [4]. X-ray, while accessible, cannot reveal early cartilage lesions, while morphological MRI can only identify advanced cartilage damage such as deep fissuring or extensive fibrillation [7–9]. Although morphological MRI can assess other knee structures like the meniscus, bone marrow lesions (BMLs) and synovium [10], current studies lack direct comparisons with histopathology, the gold standard [11, 12]. Thus, X-ray and morphological MRI alone are insufficient for early OA diagnosis [4]. Within the framework of the current standard guidelines [13], the lag in the acquisition of clinical symptoms and traditional imaging has resulted in the treatment paradigm for OA still leaning towards reactive medicine.

With the development and prevalence of predictive, preventive and personalized medicine (PPPM), the focus on the precise stratification of patients based on various biological and clinical characteristics has increased. With more precise stratification of patients, earlier prevention, stratified treatment and further improvements in personalized treatment plans will achieve better prognoses for patients [14]. As pointed out in the EULAR Recommendations (2023), people with knee OA should be offered an individualized, multicomponent management plan that includes the recommended core non-pharmacological approaches [15]. For patients with OA symptoms but without radiological signs, early diagnosis of knee OA (KOA) in asymmetrical stages is essential for the timely management of patients using preventive strategies [16]. Therefore, the need to develop a new tool, which can help clinicians identify patients at early-stage OA and recommend appropriate exercise load and physical therapy, is growing [17].

Risk prediction models based on physiological MRI are effective tools for the prediction and personalized treatment of early OA

Risk prediction models are widely used to identify high-risk individuals with disease. Although a variety of prediction models for OA have already been constructed, they have yet to focus on symptoms, radiology and morphological MRI results [18], which are unable to identify the ‘pre-morphological’ changes of cartilage, such as the disordered arrangement of matrix or loss of cartilage surface lamina. Although some studies have applied machine learning methods to overcome these difficulties, many models have limitations, including the arguable interpretation and limited prevention [19]. Therefore, introducing a new detection method with an established cut-off point, which can effectively assess early-stage cartilage degeneration, is crucial.

Benefiting from the improvements of physiological/functional MRI, such as T1ρ and T2 mapping, is used to assess ‘pre-morphological’ biochemical compositional changes of articular [20]. Micro-pathological features such as disordered cartilage arrangement or loss of surface lamina cannot be fully assessed by morphological MRI. However, these features can be detected by T1ρ and T2 mapping sequences, which can improve treatment outcomes through patient risk stratification [14]. With a significantly smaller magnitude of the magic-angle effect [21], T1ρ images can well assess early-stage knee OA cartilage [22] with a cut-off point [23]. Recently, our research first showed that the cut-off point of the mean T1ρ relaxation time value (>33 milliseconds, ms) could be used to distinguish early-stage OA and moderate-stage OA [24]. However, the utilization of T1ρ/T2 mapping in constructing radiomics prediction models was rarely reported. To the best of our knowledge, the role of T1ρ/T2 mapping in assessing the risk of early OA remains unclear.

Thus, this study takes an advanced diagnostic approach, especially physiological MRI, and constructs an OA-anatomical signature-nomogram (OA-ASN) which can be used to guide treatments tailored to the person in disease prevention, accurate prediction and personalized medicine in the future.

Methods

This study was approved by the Institutional Review Board at Zhujiang Hospital of Southern Medical University (trial register number: 2017-GJGBK-001). All procedures performed in this study involving human participants followed the Declaration of Helsinki, while written informed consent was obtained from all participants. The validation cohort study was registered in the Chinese Clinical Trials Register (ChiCTR2100044698).

Study population

This study enrolled two independent cohorts of patients diagnosed with OA who underwent total knee arthroplasty (TKA) or trauma/tumor who underwent amputation. Patients aged >40 years, with Kellgren–Lawrence grade 3–4 knee OA, who met the criteria of OA according to the American College of Rheumatology were included [25]. Exclusion criteria of the study are given in Supplementary Data S1, available at Rheumatology online.

A total of 115 consecutive patients from the Zhujiang Hospital of Southern Medical University between July 2019 and January 2021 were included in the training cohort, while the validation cohort comprised 84 consecutive patients between June 2021 and March 2022, with the same criteria. The patient recruitment pathway is shown in Fig. 1. Baseline information was recorded for each patient, including age, sex, BMI and body surface area (BSA).

Figure 1.

Flow diagram of inclusion, exclusion, allocation and sample size for analysis. Note: MFDC, medial femoral distal condyle; TKA, total knee arthroplasty

Image acquisition and analysis

MRI was performed before surgery. Details of the protocol are shown in the previous study [24]. Image selection and analysis of T1ρ/T2 mapping were performed according to the method previously described [24]. Briefly, in sagittal images, the regions of interest (ROI) of the medial femoral distal condyle (MFDC) were determined [24]. The T1ρ and T2 mapping relaxation times were calculated. Infrapatellar fat pad (IPFP) maximal areas, hyperintense signal of IPFP and medial tibial plateau cartilage volume (MTCV) were quantified according to the previous study [26]. Cartilage defects and BML of the medial tibial plateau were assessed using previous semiquantitative scoring systems [27] (Supplementary Data S2, available at Rheumatology online).

Pathologic data collection and OA grade ascertainment

MFDCs were obtained and analysed based on the Osteoarthritis Research Society International (OARSI) assessment system [28]. The selection of ROI in MRI was aligned with the slice of MFDC. Cartilage that was diagnosed as OA cartilage pathologically but could not be diagnosed based on morphological MRI was defined as early-stage OA (Supplementary Data S3, available at Rheumatology online).

Development and validation of the nomogram

The least absolute shrinkage and selection operator (LASSO) logistic regression algorithm was used to select the related features [29]. Multivariable logistic regression was applied to construct a competing-risk nomogram (P < 0.05). The performances of the nomogram were assessed using the C-index and the area under the receiver operating characteristic (ROC) curve (AUC). The total risk points for patients were calculated based on the nomogram quantification of risk and stratified by finding the best cut-off score based on the ROC-AUC. Calibration was graphically assessed with the calibration slope and calibration plot. A decision curve analysis (DCA) and clinical impact curve (CIC) were performed. Finally, the clinical usefulness and generalizability of the nomogram were further confirmed in an external validation cohort provided. Detailed methods are presented in Supplementary Data S4, available at Rheumatology online.

Statistical analysis

Appropriately, continuous variables are summarized as means and standard deviations, which were compared by using the Student’s t-test when the distribution was normal and using the Mann–Whitney U-test if the distribution was not normal. Categorical variables are reported as frequencies and percentages, which were compared by using χ² or Fisher's exact tests. Power analysis and sample size (PASS) software (version 12; PASS, NCSS, LLC, Kaysville, Utah, USA; www.ncss.com) was conducted to determine sample size by using a significance level of 0.05 and a power of 0.80, in which the results indicated minimum sample sizes of 47 in the normal group and 31 in the early OA group. All statistical analyses were performed using R software (version 4.0.3; http://www.r-project.org/). A two-sided P < 0.05 was considered significant.

Results

After strict screening, 115 patients (mean age ± standard deviation, 67 years ± 7.23) and 84 patients (68 years ± 4.67) were divided into the training and validation cohorts, respectively (see Table 1).

Table 1.

Patient and clinicopathologic characteristics of the training and validation cohorts^a

	Training cohort (n = 115)			Validation cohort (n = 84)
Characteristic	Overall	Normal	Early OA	Overall	Normal	Early OA
		(n = 70)	(n = 45)		(n = 50)	(n = 34)
Age, years	67 ± 7.23	67 ± 6.62	67 ± 8.07	69 ± 4.67	68 ± 4.58	69 ± 4.87
Sex
Male	41 (35.7)	23 (32.9)	18 (40)	23 (27.4)	14 (28)	9 (26.5)
Female	74 (64.3)	47 (67.1)	27 (60)	61 (72.6)	36 (72)	25 (73.5)
BMI	24.5 ± 3.83	24.9 ± 4.05	25.0 ± 3.40	23.6 ± 3.55	24.0 ± 3.62	22.8 ± 3.39
BSA	1.63 ± 0.14	1.63 ± 0.13	1.66 ± 0.15	1.60 ± 0.14	1.61 ± 0.13	1.58 ± 0.16
IPFP, cm2	3927 ± 833.85	4205 ± 666.72	3493 ± 887.85	3550 ± 892.88	3831 ± 877.70	3269 ± 914.58
IPFP signal intensity, %	37.78 ± 13.23	36.28 ± 13.50	40.12 ± 12.59
T1 rho, ms
≤33	65 (56.5)	61 (87.1)	4 (8.9)	46 (56.5)	39 (78)	7 (20.6)
>33	50 (43.5)	9 (12.9)	41 (91.1)	38 (43.5)	11 (22)	27 (79.4)
T2 mapping, ms
≤35.04	76 (66.1)	65 (92.9)	7 (15.6)	50 (59.5)	43 (86)	7 (20.6)
>35.04	39 (33.9)	5 (7.1)	38 (84.4)	34 (40.5)	7 (14)	27 (79.4)
MTPCV, cm3	294.19 ± 128.65	304.37 ± 130.01	278.36 ± 96.71
MTPCD
≤2	106 (92.2)	66 (94.3)	40 (88.9)
>3	9 (7.8)	4 (5.7)	5 (1.1)
BML
≤1	110 (95.7)	68 (97.1)	42 (93.3)
>2	5 (4.3)	2 (2.9)	3 (6.7)

^aData are presented as mean ± S.D. or number (%) of patients.

BSA, body surface area; IPFP, infrapatellar fat pad; MTPCV, Medial tibial plateau cartilage volume; MTPCD, Medial tibial plateau cartilage defect; BML, bone marrow lesion.

OA diagnosis-related clinical features and radiomic signature selection

We used the LASSO logistic regression algorithm, with the training set to select the nonzero coefficients from among the above features. As shown in Supplementary Fig. S1, available at Rheumatology online, LASSO selected three key features from radiomics signatures and one factor from clinical features, including sex, IPFP, T1ρ and T2 mapping.

Multivariable analysis

The results of multivariable analysis of the above-selected features are summarized in Table 2. In the multivariable analysis, IPFP maximal area (OR, 0.998; 95% CI: 0.995–0.999; P < 0.039), T1ρ relaxation times (OR, 62.42; 95% CI: 3.77–525.36; P < 0.014), T2 mapping relaxation times (OR, 417.26; 95% CI: 19.21–747.74; P = 0.002) were independently associated with the forecast of early OA. The OR of female vs male showed a statistical difference in univariate regression but no statistical difference in multivariate logistic regression. In general, having a smaller volume of IPFP (cm²), T1ρ relaxation times >33 ms and T2 mapping relaxation times >35.04 ms were significantly associated with a greater risk of OA. Moreover, the T2 mapping relaxation times were the strongest predictor of OA (OR = 417.26). Therefore, we constructed a nomogram (a graphic depiction of the model) based on these significant predicting variables (Fig. 2). In the nomogram, all variables are assigned an echoed score on the point scale based on the rank order of the effect estimates. We can draw a straight line down to derive the estimated probability of early OA by adding these points and then evaluating the total score on the ‘Total points’ scale. In addition, examples of clinical use of the nomogram are shown in Fig. 3.

Table 2.

Results of multivariable logistic regression for prediction of advanced OA^a

Variable	Prediction model, OR (95％ CI)	P
Sex
Male	1.00 (reference)	0.146
Female	12.54 (1.99–101.85)
IPFP	0.998 (0.995–0.999)	0.039*
T1 rho, ms
≤33	1.00 (reference)	0.014*
>33	62.42 (3.77–525.36)
T2 mapping, ms
≤35.04	1.00 (reference)	0.002*
>35.04	417.26 (19.21–747.74)

^a P-values with asterisk indicate a statistically significant difference compared with the reference variable (P < 0.05).

IPFP, infrapatellar fat pad.

Figure 2.

Nomogram for predicting early OA, based on MRI-related variables. The T1ρ value, T2 mapping value and IPFP volume are located on the respective axes, then a straight line is drawn from each location upward to the ‘points’ axis, and the sum of points for each variable is calculated. This number is then located on the ‘total points’ axis, and a straight line is drawn down to the ‘risk of complications’ axis. Note: IPFP, infrapatellar fat pad

Figure 3.

Examples of the nomogram in clinical practice. (a) The patient was diagnosed with osteosarcoma and underwent amputation. T1ρ value and T2 mapping value of MFDC ROI are 25.47 and 30.87, respectively. IPFP maximal area is 5406.05 mm². (b) The patient was diagnosed with OA and underwent TKA. Cartilage degeneration was severe in the lateral condyle. T1ρ value and T2 mapping value of MFDC ROI are 33.07 and 37.06, respectively. IPFP maximal area is 3335.17 mm². Note: IPFP, infrapatellar fat pad; ROI, regions of interest

External validation and performance evaluation

First, the nomogram yielded an averaged C-index of 0.975 (95% CI: 0.951–0.999), and the C-index by bootstrap validation (1000 bootstrap samples) was 0.96 in the training cohort. In the external validation cohort, the C-index was still as high as 0.904 (95% CI, 0.840–0.0959), which indicated the favourable discrimination of the nomogram. On the other, the ROC curves of the nomogram to predict early OA are present in Fig. S2A, with a high AUC value of 0.975 in the training set and 0.904 in the validation set (Fig. S2B). These results further confirmed the excellent performance of the nomogram. Furthermore, the calibration curves illustrated an excellent agreement between the nomogram and actual observation, both in the training cohort and the validation set (Supplementary Fig. S2C and D, available at Rheumatology online).

Clinical usefulness

The decision curve and CIC were conducted to evaluate the actual benefit of the nomogram in clinical practice. The results of DCA and CIC are shown in Supplementary Fig. S3, available at Rheumatology online. The DCA exhibited that if the threshold probability in the clinical decision was in the 2–100% range, the nomogram could add more net benefits compared with whether the treat-all-patients scheme or the treat-none scheme in both the training and validation cohorts (Supplementary Fig. S3, available at Rheumatology online). Besides, the CIC visually showed that the nomogram had a superior overall net benefit within almost all ranges of threshold probabilities and impacted patient outcomes in both the training and validation sets, which further demonstrated that the nomogram possesses significant clinical applicability (Supplementary Fig. S3, available at Rheumatology online). The risk probability distribution map (Fig. 4A and B) derived from the model of all patients in both the training and validation cohort indicated that the result was highly consistent with the corresponding event for most of the patients (108/115, 93.9% in the training cohort; 73/84, 86.9% in the validation cohort).

Figure 4.

The risk probability of early-stage OA was calculated using the merged model for each patient. Red bars represent the probability of patients being diagnosed with early-stage OA, while blue bars represent the probability for those who were not diagnosed with early-stage OA. With a cut-off value of 0.5562, bars above the x-axis represent patients who were stratified as having a high risk of early-stage OA by the model, while bars below the x-axis represent patients who were stratified as having a low risk. The red bars above the x-axis or the blue bars below the x-axis indicate the consistency between the predicted results of the model and the actual situation. (a) Patients from the training cohort (108/115, 93.9%). (b) Patients from the validation cohort (73/84, 86.9%). (c) PPPM strategies of Early-stage OA. Note: IPFP, infrapatellar fat pad; PPPM, predictive, preventive, and personalized medicine; SYSADOAs, symptomatic slow-acting drugs for OA

Discussion

Current OA diagnosis guidelines from international organizations rely on typical symptoms (e.g. usage-related pain, short morning stiffness, age >40) and radiography [30]. However, in clinical practice, early-stage OA patients often present with symptoms that do not meet all diagnostic criteria, such as frequent knee pain without X-ray evidence or atypical symptoms occurring less frequently [16, 30, 31]. The onset of biochemical changes leading to irreversible cartilage loss and the corresponding clinical and radiographic signs may lag several years [31]. Thus, traditional treatment for typical OA patients is a delayed approach until years after onset [4]. Early diagnosis of OA and identification of at-risk individuals will facilitate personalized prevention and individual prognosis, which is crucial for early treatment plans.

Conventional radiography, though a low-cost imaging method for confirming OA diagnosis, is limited in detecting early cartilage lesions, often resulting in late-stage diagnosis [4, 32]. Current guidelines recommend exercise programs and physical therapy for early OA, but these interventions are often initiated after traditional imaging, such as X-ray, reveals moderate to severe cartilage wear, leading to a significant compromise in the effectiveness of treatment [13, 33]. Moreover, the long-term efficacy of disease-modifying OA drugs (DMOADs) and symptomatic slow-acting drugs (SYSADs) were undermined by cartilage lesions [13]. The main reason for these dilemmas is that OA is a progressive disease and current guidelines do not provide an effective diagnostic tool for accurately predicting the risk of individual patients, especially when patients have ‘pre-morphological’ biochemical compositional changes of articular, thus leading to the current treatment for OA still being reactive medicine.

Physiological MRI provides reliable predictors for early-stage OA

Effective predictors are crucial for early diagnosis and disease prevention. Researchers are increasingly paying attention to identifying clinical symptoms and image features as predictive factors for diagnosis. However, recent predictive models [28, 34] based on these tools have some limitations. Clinical symptoms often reflect advanced disease stages, while traditional imaging, such as X-ray, could be applied to detect advanced-stage OA [34, 35]. Morphological MRI can reveal moderate OA but struggles to identify early cartilage changes [36]. For atypical presentations, morphological MRI is recommended but remains insufficient for OA early detection [31, 37]. Early OA is subclinical with no visible anatomical changes in the cartilage, despite progress with typical presentations several years later, which makes it challenging to detect using morphological MRI [31]. These models, based on symptoms, X-ray or morphological MRI results, primarily predict moderate to severe OA, failing to meet the need for early diagnosis [6]. Therefore, a tool that detects cartilage micro-morphological changes could help identify at-risk patients, provide early prevention and improve the quality of personal life.

To remedy this, a simple and quantifiable nomogram to identify high-risk persons of early OA with ‘pre-morphological’ biochemical compositional changes is constructed in this study. Physiological MRI, due to its ability to evaluate the arrangement of the extracellular matrix and detect superficial micro-degeneration, has a broader role in detecting early-stage cartilage damage than morphological MRI [20, 24]. Among them, the most recently studied and clinically applied sequences are T1rho and T2 mapping [22, 24, 28]. Physiological MRI, including T1ρ and T2 mapping images, and IPFP maximal area were recently reported to have a predictive potential for early-stage cartilage damage [24, 38]. The reliable non-invasive features identified in this study can effectively stratify patients at risk of early-stage OA into different risk categories. This stratification allows for the implementation of tailored interventions based on individual risk levels. These interventions may include exercise-based activity programs, lifestyle modifications, neuromuscular education, knee bracing (or other supportive measures), intra-articular injections and pharmacological management, to alleviate symptoms and prevent the progression of OA [13, 39].

OA-ASN effectively contributes to initial individualized treatment for early OA

The most important findings of this study were attained through a comprehensive predictive modelling process for predicting early-stage OA based on simple covariates, especially from physiological MRI, with ‘pre-morphological’ changes occurring in cartilage. T1ρ, T2 mapping relaxation times and IPFP total area were significantly associated with outcomes. The OA-ASN is developed and validated to predict the incidence of KOA in non-radiographic osteoarthritic individuals and help at-risk individuals avoid the occurrence of severe OA-related symptoms with early intervention. The effective reduction of false-negative and false-positive results in the diagnosis of early-stage OA would greatly improve the detection rate.

In our study, patients with risk points ≥106.65 or risk probability of ≥55.62% belong to the high-risk group. After OA risk stratification is determined for patients who are troubled by knee pain, preventive approaches and personalized treatment strategies will be carried out. Due to the extensive differences in treatment responses among patients with different risk levels of musculoskeletal pain, personalized treatment is clinically beneficial and cost-effective for patients with knee pain who are primarily diagnosed with early-stage OA [13, 40]. Thus, further stratification of the non-high-risk group was determined. The medium-risk group (80.00 ≤ risk point <106.65 or 20% ≤ risk probability of <55.62%) and the low-risk group (risk point <80.00 or risk probability of <20%) were defined for personalized treatment strategies based on the risk stratification methods [41]. Thus, OA risk stratification was carried out for patients to implement personalized treatment strategies (Fig. 4C).

The initial individualized treatment plan is recommended for patients in the high-risk group, such as arthritis education, neuromuscular education, lifestyle intervention (including diet) for overweight and obesity, rest, low-impact aerobic exercises, activity modification, brace for knee (and other support), biomechanical footwear, physical therapy, the use of NSAIDs, pharmaceutical grade glucosamine sulphf and/or chondroitin sulphf (SYSADOAs) or intra-articular injection [16, 17, 32]. Stepping-up exercise, beginning with a low-resource intensive treatment, is recommended for high-risk patients [31]. If participants were in the high-risk group, high-impact sports should be avoided to prevent torsional loading, while sports that have a limited risk of accelerating joint degeneration should be selected [42]. Intra-articular injection of novel ortho-biologics, such as isolated growth factors, micro-fragmented adipose tissue and platelet-rich plasma, were undoubtedly burgeoning [43]. These customized therapies have some early promising results, which can promote the healing program for targeted patients. It is recommended that high-risk individuals follow-up at 4 weeks and 12 weeks after personalized treatments, which is crucial for adjusting treatment and preventing the progression of OA or further knee dysfunction.

Individuals in the medium-risk group are recommended to exercise. According to 2021 EULAR recommendations [44], patients with OA should be encouraged to exercise as it is particularly beneficial for disease-related outcomes. A reasonable implementation of therapeutic exercise plans is suggested for individuals with different baseline exercise repetition maximums [13, 45]. Meanwhile, core treatment, including physical, psychological and mind-body approaches [46], is recommended. Regular follow-up every 12 weeks is crucial for medium-risk individuals for readjustment of self-management and structured exercise programs.

For individuals in the low-risk group, clinicians should sufficiently consider whether the symptoms are caused by other factors, such as meniscal tears or RA. The patient needs to undergo other sufficient examinations to prevent delaying correct treatment.

Limitations

This study acknowledges several limitations despite the robust performance of the predictive nomogram. First, the cohorts were sourced from a single tertiary clinical institution, potentially introducing selection bias. Although external validation was performed using an independent cohort of 84 patients, multicentre studies are essential to fully assess the model's generalizability across diverse clinical settings. Second, interobserver variability in MRI measurements and the limited accessibility of physiological MRI in developing regions may introduce biases and limit the model's broader application. However, advancements in technology and market competition are expected to enhance MRI accessibility over time. Third, risk stratification may provide a false sense of security. Additionally, the IPFP maximal area differed statistically between the training and validation cohorts in the normal cartilage group, highlighting the need for larger external validation samples. The nomogram also did not account for other potential factors such as theranostics, genetics, lifestyle, mechanical alignment and biomarkers, which could influence knee component damage, malalignment and biochemical indicators, and should be explored further. Furthermore, the overrepresentation of female participants reflects the dataset's composition. While no significant gender differences were found in outcomes, future studies should consider gender and hormonal effects to provide a more comprehensive understanding of OA.

Conclusions

In conclusion, OA-ASN incorporating refined MRI variables accurately predicted the occurrence risk of early OA. This efficient tool demonstrated excellent predictive outcomes with easy-accessible and simple observational screening methods based on MRI. To enhance the model's predictive accuracy and promote its global application, especially amid the growing ageing population, future research should focus on: (1) conducting multicentre validation studies to confirm the model's robustness and generalizability; (2) incorporating additional variables such as patient compliance, intervention types and individual metabolism to boost its predictive power and (3) developing user-friendly online platforms and mobile applications to facilitate easy access to risk stratification results for clinicians. These efforts will improve early knee OA diagnosis, enable timely interventions and ultimately enhance patient outcomes.

Supplementary material

Supplementary material is available at Rheumatology online.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Author contributions

Conceptualization, Zhijian Yang and Chao Xie; Data curation, Zhijian Yang and Chao Xie; Formal analysis, Zhijian Yang, Weiwen Zhu and Chao Xie; Funding acquisition, Zhijian Yang and Chao Xie; Investigation, Huiwen Lu, Zhaowei Lin and Weiwen Zhu; Methodology, Zhijian Yang and Chao Xie; Project administration, Zhijian Yang and Chao Xie; Resources, Chao Xie; Supervision, Chao Xie; Validation, Zhijian Yang, Huiwen Lu and Zhaowei Lin; Visualization, Zhijian Yang and Chao Xie; Writing—original draft, Zhijian Yang, Huiwen Lu, Zhaowei Lin, Weiwen Zhu and Haopeng Guo; Writing—review & editing, Chao Xie. All authors have read and agreed to the published version of the article. NO Large Language Models (LLMs) were used for our article.

Funding

This work was supported by the Presidential Foundation of Zhujiang Hospital (yzjj2023qn03) and Medical Scientific Research Foundation of Guangdong Province of China (Project No. A2023083).

Disclosure statement: The authors have declared no conflicts of interest.