Machine learning-enhanced clinical and ultrasound technology for early detection of difficult-to-treat rheumatoid arthritis

doi:10.21203/rs.3.rs-4891134/v1

Objectives: This study aimed to examine the clinical features, serological indicators, and ultrasound examination results of individuals with difficult-to-treat rheumatoid arthritis (D2T RA) and non-D2T RA, along with the development and validation of two distinct predictive models for the early detection of D2T-RA.

Methods: Enrolled 85 patients with D2T-RA diagnosed as moderate or high disease activity who completed 24 months of follow-up, and randomly matched 83 non-D2T-RA patients with moderate to high disease activity. Baseline clinical information was collected, and US examinations were performed to record the single scores of gray-scale (GS) and power Doppler (PD) for 16 joints and 10 tendons, as well as the EULAR-OMERACT scores. Univariate analysis identified predictive factors, followed by machine learning to create two models: clinical/serological (Model 1) and clinical/serological/US (Model 2). We evaluated the model performance using 5-fold cross-validation, utilizing the F1 score and AUC.

Results: The univariate logistic analysis showed that EULAR-OMERACT＞1 (6 variables) and clinical and serological characteristics (14 variables) were significant predictors of D2T RA. The random forest model performed best on all models, with the AUC and F1 of test set model 1 divided into 0.81 and 0.67, and the AUC and FI of model 2 with ultrasound data increased to 0.83 and 0.69, respectively.

Conclusion: Multi-joint ultrasound score provides important prediction data for early identification of D2T RA, a random forest model improves prediction efficacy, and evaluating limited joints makes this method more feasible in rheumatism clinical practice.

D2T RA

machine learning

ultrasound

prediction

Machine learning models predict D2T-RA effectively; ultrasound improves accuracy.
The Random Forest model, incorporating clinical, serological, and ultrasound data, showed the best predictive efficacy.
Multi-joint ultrasounds with clinical data are key for early D2T-RA detection in practice.

Rheumatoid arthritis (RA) is a systemic, chronic autoimmune disease that is predominantly characterized by erosive and symmetric polyarthritis. Although disease-modifying antirheumatic drugs (DMARDs) have been shown to be widely effective in treating RA, some patients remain unable to achieve low disease activity or symptom relief after treatment with conventional synthetic DMARDs (csDMARDs), biological DMARDs (bDMARDs), and/or targeted synthetic DMARDs (tsDMARDs). These patients may be classified as having Difficult-to-treat Rheumatoid Arthritis (D2T-RA) [1–3]. D2T-RA patients incur higher medical costs, experience significantly reduced physical function, and levels of physical activity[4]. The "window of opportunity" theory emphasizes the critical importance of aggressive early treatment, such as the use of immunosuppressants, in altering the course of inflammatory arthritis[3, 5, 6]. Therefore, the early identification of D2T-RA patients is crucial for effective disease management and reducing healthcare costs.

According to a recent study, the main predictors of D2T-RA include female gender, low body weight, age, disease duration, and serological positivity [7], It can be attributed to the significant complexity and heterogeneity of D2T-RA [8]. In order to accurately apply DMARDs during the early stages of the disease, doctors must have a new and efficient method to help them identify high-risk patients early.

Both gray-scale (GS) and power Doppler (PD) US have been shown to be sensitive to change and predictive of the development of arthritis and radiographic structural damage [9]. Previous studies have shown that models combining clinical, serological, and US characteristics are of significant value in predicting persistent arthritis [10], providing us with a new perspective for research. However, their role in predicting D2T-RA remains unclear [11]. The aim of this study is to analyze the clinical characteristics, serological indicators, and US examination results of patients with D2T-RA and non-D2T-RA, identify the differential factors between the two, and based on this, develop and evaluate predictive models aimed at improving the ability to identify D2T-RA at an early stage.

Study Population

Between January 2020 and January 2024, our hospital's Rheumatology and Immunology Department consecutively collected RA patients who visited our clinic and completed a 24-month follow-up. These patients received continuous treatment with bDMARDs or tsDMARDs, and they were assessed every three months to monitor the potential development of D2T-RA. According to the 2021 ACR/EULAR diagnostic criteria for D2T-RA [3], rheumatologists categorized patients into two groups: those with D2T-RA and those without D2T-RA.

The inclusion criteria consisted of: (1) meeting the 1987 ACR or the 2010 ACR/EULAR classification criteria for RA [12]; (2) having received at least two cs DMARDs, including methotrexate (MTX) or leflunomide (LEF), for at least six months with persistent high disease activity (DAS28CRP ≥ 3.2); (3) having at least one joint with a level 1 PDUS for monitoring purposes. The exclusion criteria consisted of: (1) individuals with missing clinical data; (2) patients who changed the dosage of DMARDs or had received glucocorticoid treatment within four weeks; and (3) patients who refused follow-up. Written informed consent was obtained from each participant. All studies were conducted in accordance with the principles of the Declaration of Helsinki, and the study protocol was approved by the Institutional Review Board. The Ethics Committee number is 2024A-326.

Clinical information

The patients who met the criteria for inclusion filled out a baseline survey that asked about their age, gender, overall health, disease course and onset date, medication usage, comorbidity, and important laboratory markers including anti-CCP, RF, and others. They also filled out disease activity and function scores, such as patient global assessment (PGA), health assessment questionnaire (HAQ), and others.

Sonographic assessment

Patients suspected of having moderate to high disease activity underwent sonographic examination of multiple joints. Baseline US exams were performed on the patients by two experienced musculoskeletal sonographers using a GE LOGIQ E20 US system with an L3-12-D 12MHz linear probe. In both GS and PD modes, they examined the patients' tendons and joints. Prior to the examination, the sonographers underwent standardized comparison practice with static images via an electronic learning platform, and they remained blinded to the patients' clinical and laboratory data.

(1): We scanned multiple bilateral joints and tendons, recording representative images of the following 16 joints and 10 tendons on the severe side: the metacarpophalangeal (MCP) joints 1–5, proximal interphalangeal (PIP) joints 1–5, radiocarpal (RC) joint, midcarpal (MC) joint, ulnocarpal (UC) joint, three main knee joint compartments -lateral (LAT), medial (MED), superior (SUP), wrist extensor compartments (I-VI)-abductor pollicis longus༈APL༉; extensor pollicis brevis༈EPB༉; extensor carpi radialis longus༈ECRL༉; extensor carpi radialis brevis༈ECRB༉; extensor pollicis longus༈EPL༉; EDC: extensor digitorum communis༈EDC༉; extensor indicis propius༈EIP༉; extensor digit minimi༈EDM༉; extensor carpi ulnaris༈ECU༉, and Flexor Digitorum (DF) tendons 2–5 (excluding the Flexor Pollicis Longus tendon) (Fig. 1).

(2) This study employed a binary classification system (0–1 score) and a semi-quantitative(SQ)scoring system (0–3 grades) to assess synovitis [13–15] and tenosynovitis [16]. The presence of synovial hypertrophy༈SH༉or Doppler signal was defined as SH or PD ≥ 1; moderate to severe SH or synovitis was indicated by SH > 1 or PD > 1. The EULAR-OMERACT combined score categorized the degree of Doppler signal from 0 to 3[17], with specific scoring criteria detailed in Table 1. PD was assessed only when the SH score of a joint was ≥ 1.

Table 1

GS, PD, and combined score for grading synovitis and tenosynovitis in RA
Synovitis/Tenosynovitis	Hand and wrist joints		Knee joint		Wrist tendon		Combined score† (SH + PD)
	SH(GS)	PD	SH(GS)	PD	SH(GS)	PD	SH(GS) + PD
Grade 0 (normal)	No SH	No signal	The thickness is less than 2 millimetres.	No signal	No SH	No signal	No SH and no PD
Grade 1 (minimal)	Minimal SH up to the imaginary horizontal line connecting the 2 joints edges	Up to three single signals or one confluent and two single or two confluent	The thickness ranges from 2 to 5 millimetres.	Up to three single signals or one confluent and two single or two confluent	Mild tenosynovial SH	Peritendinous Focal Signal	Grade 1 SH and/or ≤ grade 1 PD
Grade 2 (moderate)	Moderate SH protruding over the joint line along with concave surface	Larger than grade 1, but < 50% of SH area covered by signals	The thickness ranges between 6 and 8 millimetres.	Larger than grade 1, but < 50% of SH area covered by signals	Moderate tenosynovial SH	Peritendinous Multifocal Signal	Grade 2 SH and/or ≤ grade 2 PD; or grade 1 SH and grade 2 PD
Grade 3 (severe)	Severe SH protruding beyond the joint line with convex surface	More than 50% of SH area covered by signals	The thickness > 8 mm	More than 50% of SH area covered by signals	Severe tenosynovial SH	Peritendinous Diffuse Signal	Grade 3 SH and/or ≤ grade 3 PD; or grade 1 or grade 2 SH and grade 3 PD
†EULAR-OMERACT combined score.
EULAR-OMERACT, European League Against Rheumatisms-0utcomes Measures in Rheumatology; GS, Grayscale; SH, synovial hypertrophy; PD, power Doppler.

Statistical analysis

For data following a normal distribution, use the mean ± standard deviation (x ± s). For non-normal data, use the median (Q1, Q3). Categorical data are shown by frequency. Choose the Mann-Whitney U test for non-normal continuous data or the t-test for normal data. Use the chi-square test for categorical data.Baseline clinical variables between groups are compared using either the Mann-Whitney test or Fisher's exact test. In descriptive analysis, P < 0.05 is considered statistically significant.

Consistency analysis is performed using the Kappa test. All statistical analyses were performed using the R 3.3.2 statistical software and Free Statistics software version 1.9. Logistic regression, random forest, and support vector machine (SVM) were implemented using Python (Python 3.8, Scikit-learn 0.24.2).

Reliability analysis

To assess interobserver reliability, two sonographers independently re-evaluated representative US images from 20 patients with RA in a blinded manner and conducted a consistency analysis using the κ statistic.A κ value of 0–0.2 was considered poor, 0.21–0.40 fair, 0.41–0.6 moderate, 0.61–0.8 good, and 0.81–1 excellent.

Logistic Regression, Random Forest, and Support Vector Machine

Initially, univariate analysis will be employed to screen factors associated with D2T-RA. Subsequently, three machine learning techniques—logistic regression, random forest, and support vector machine (SVM)—are utilized to establish two predictive models.

This study employs an 8:2 ratio to partition the data set into training and testing sets. On the training set,the model was systematically trained and evaluated using a five-fold cross-validation method. After five rounds of iteration, The identification of diagnostic predictive factors: We selected the model parameters with the highest average score in the subset to train our final mode.Ultimately, the models were evaluated using an independent testing set based on metrics such as AUC, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and recall rate.This study adopted the F1 score and AUC as the primary indicators for evaluating model performance.

Demographic and disease characteristics

A total of 168 patients were included: 85 with D2T RA and 83 with non-D2T RA. Their baseline characteristics are shown in Table 2. Patients with D2T RA demonstrated a more extended disease duration (3.0 vs. 1.2 years, P = 0.019), symptom onset (3.0 vs. 1.0 year, P = 0.015), time to diagnosis (1.2 vs. 0.6 year, P = 0.015), and initial consultation period (1.5 vs. 1.0 year, P = 0.06). However, the interval from diagnosis to initiation of b/tsDMARD therapy was shorter (0.3 vs. 0.1 year, P < 0.001). Among D2T RA patients, pulmonary disease was prevalent as a comorbid condition, though not statistically significant (49.4% vs. 39.8%, P = 0.208). Nonetheless, the aggregate number of comorbidities was greater in D2T RA patients, reaching statistical significance (P = 0.048).

Table 2

Baseline characteristics of all patients by outcome group: clinical and serological features (n = 168)
Variables	D2T RA (n = 85)	Non-D2T RA (n = 83)	p
Age, years	54.5 ± 11.6	55.5 ± 12.3	0.594
Female, n (%)^*	68 (80)	54 (65.1)	0.03
Disease.duration, years^*	3.0 (0.8, 9.0)	1.2 (0.5, 6.5)	0.019
Onset date.of symptoms,years^*	3.0 (0.8, 9.0)	1.0 (0.5, 6.5)	0.015
Date of diagnosis,years^*	1.2 (0.5, 5.0)	0.6 (0.2, 2.0)	0.001
Date of first visit,years^*	1.5 (0.5, 6.0)	1.0 (0.2, 3.8)	0.006
Diagnosis to treatment initiation with b/tsDMARDs,years^*	0.3 (0.2, 1.0)	0.1 (0.0, 0.2)	< 0.001
Weight (kg)	63.0 ± 24.5	61.6 ± 11.7	0.629
Height, cm	158.4 ± 23.6	163.5 ± 7.5	0.061
BMI	21.9 ± 3.2	22.7 ± 3.7	0.106
Smoking history	7 (8.2)	4 (4.8)	0.371
With osteoporosis^*	36 (42.4)	22 (26.5)	0.031
With anaemia	18 (21.2)	10 (12)	0.112
X-ray hands erosions^*	40 (47.1)	21 (25.3)	0.003
Comorbidity^*	2.0 (1.0, 3.0)	1.0 (0.0, 3.0)	0.018
Hypertension, n (%)	22 (25.9)	25 (30.1)	0.541
Ischemic.heart.disease, n (%)	17 (20)	11 (13.3)	0.241
Stroke, n (%)	2 (2.4)	1 (1.2)	1
Diabetes.mellitus, n (%)	14 (16.5)	9 (10.8)	0.289
Liver, n (%)	3 (3.5)	0 (0)	0.246
Lung, n (%)	42 (49.4)	33 (39.8)	0.208
Depression, n (%)	5 (5.9)	2 (2.4)	0.443
Fracture, n (%)	4 (4.7)	1 (1.2)	0.368
Gastrointestinal.ulcer, n (%)	7 (8.2)	2 (2.4)	0.168
Cancer, n (%)	0 (0)	2 (2.4)	0.243
History of treatment
Number of used csDMARDs	2.0 (2.0, 3.0)	2.0 (2.0, 3.0)	0.411
Number of used b/tsDMARDs	1.0 (1.0, 2.0)	1.0 (0.0, 1.0)	< 0.001
Number of used mode of action	1.3 ± 0.5	0.7 ± 0.5	< 0.001
PSL user	29 (34.1)	37 (44.6)	0.165
Dose of PSL, mg/day	0.0 (0.0, 10.0)	0.0 (0.0, 15.0)	0.22
Anti-CCP, n (%)*			0.004
Negative	23 (27.1)	38 (45.8)
Low positive	25 (29.4)	28 (33.7)
High positive	37 (43.5)	17 (20.5)
RF, n (%)^*			< 0.001
Negative	6 (7.1)	27 (32.5)
Low positive	11 (12.9)	16 (19.3)
High positive	68 (80)	40 (48.2)
ESR, mm/h*	59.0 (28.0, 78.0)	40.0 (18.0, 60.5)	0.002
CRP, mg/l	24.6 (8.3, 63.7)	23.4 (8.4, 56.7)	0.469
TJC (0–28)	8.0 (4.0, 12.0)	6.0 (4.0, 11.0)	0.502
SJC(0–28)	6.0 (2.0, 15.0)	5.0 (3.0, 14.0)	0.48
PGA	1.9 (1.1, 2.6)	1.7 (0.9, 2.5)	0.629
EGA*	7.0 (5.0, 8.0)	6.0 (4.0, 7.0)	0.016
HAQ	1.9 ± 0.5	1.7 ± 1.1	0.199
VAS	5.5 ± 2.2	6.1 ± 2.5	0.097
DAS28 CRP	5.2 ± 1.5	5.0 ± 1.4	0.425
DAS28 ESR	5.7 ± 1.4	5.4 ± 1.4	0.145
BMI, body mass index; csDMARD, conventional synthetic disease-modifying anti-rheumatic drug; b/tsDMARDs, biological/targeted synthetic disease-modifying anti-rheumatic drug; Anti-CCP, anti-cyclic citrullinated peptide antibody; RF, rheumatoid factor; ESR, erythrocyte sedimentation rate; CRP, C-reactive protein; TJC, tender joint count; SJC, swollen joint count; PGA, patient global assessment; EGA, evaluator global assessment; HAQ, health assessment questionnaire; VAS, Visual Analogue Scale DAS28, disease activity score for 28 joints

Regarding the use of biologic or targeted synthetic DMARDs (b/tsDMARDs), the median number administered to D2T RA patients was 1 (IQR 1–2), while the median total number of medications was 3 (IQR 2–4). The utilization of b/tsDMARDs and the overall variety of medications were markedly higher in D2T RA patients compared to non-D2T RA patients (P < 0.001). According to clinical scoring, there were more female patients with D2T RA, as well as more severe osteoporosis and bone erosion in the hands, as well as higher levels of RF and ACPA. The evaluator global assessment (EGA) indicated a statistically significant difference between D2T RA and non-D2T RA patients (7.0 vs. 6.0, P = 0.016), whereas other scores did not differ significantly.

Distribution of synovial US abnormalities

We conducted SQ assessments of GS and PD for joints and tendons(Supplementary Table 1). When GS and PD ≥ 1, the D2T RA group demonstrated statistically significant differences in the positivity rates for the UC, MED, and LAT, as well as for ECRL/ECRB and EDM ( Fig. 2 and Supplementary Table 2). Upon further analysis with GS and PD >1, we observed that the D2T RA group not only sustained high positivity rates in the aforementioned regions but also exhibited pronounced disparities in the small hand joint areas, including MCP1,3,4 PD, and PIP3 PD (Supplementary Table 2).

In accordance with the EULAR-OMERACT composite scoring system (as detailed in Table 3), the D2T RA group had a notably higher prevalence of a score exceeding 1 for MCP1 and MCP4, with 44.7% vs. 26.5% (P = 0.014) and 28.2% vs. 3.6% (P < 0.001), respectively. Moreover, this group showed an elevated proportion in the wrist (UC), knee (MED), and LAT, with UC: 63.5% vs. 43.4% (p = 0.009), MED: 48.2% vs. 25.3% (P = 0.002), and LAT: 50.6% vs. 24.1% (P < 0.001).Furthermore, ECRL/ECRB and DF 4 were also more common in the D2T RA group.

Table 3

Ultrasound variable changes in D2T RA and Non-D2T RA patients with combined score†(SH + PD)>1 ‡with RA
	Non-D2T RA (n = 83)	D2T RA (n = 85)	p
Small joint US variables(SH + PD)>1‡
MCP1,n (%)^*	22 (26.5)	38 (44.7)	0.014
MCP2,n (%)	43 (51.8)	45 (52.9)	0.883
MCP3,n (%)	34 (41)	44 (51.8)	0.16
MCP4,n (%)^*	3 (3.6)	24 (28.2)	< 0.001
MCP5,n (%)	23 (27.7)	31 (36.5)	0.224
PIP1,n (%)	11 (13.3)	19 (22.4)	0.124
PIP2,n (%)	24 (28.9)	32 (37.6)	0.23
PIP3,n (%)	23 (27.7)	35 (41.2)	0.066
PIP4,n (%)	21 (25.3)	14 (16.5)	0.159
PIP5,n (%)	10 (12)	18 (21.2)	0.112
Medium amd large US variables(SH + PD)>1‡
RC ,n (%)	55 (66.3)	54 (63.5)	0.293
MC ,n (%)	51 (61.4)	51 (60)	0.848
UC ,n (%) *	36 (43.4)	54 (63.5)	0.009
SUP,n (%)	35 (42.2)	40 (47.1)	0.524
MED,n (%)*	21 (25.3)	41 (48.2)	0.002
LAT ,n (%)*	20 (24.1)	43 (50.6)	< 0.001
Tendon US variables(SH + PD)>1‡
APL/EPB,n (%)	8 (9.6)	11 (12.9)	0.499
ECRL/ECRB,n (%)^*	3 (3.6)	16 (18.8)	0.002
EPL,n (%)	9 (10.8)	11 (12.9)	0.675
EDC/EIP,n (%)	11 (13.3)	14 (16.5)	0.558
EDM,n (%)	4 (4.8)	8 (9.4)	0.248
ECU,n (%)	16 (19.3)	21 (24.7)	0.396
DF 2,n (%)	10 (12)	13 (15.3)	0.541
DF 3,n (%)	7 (8.4)	11 (12.9)	0.345
DF 4,n (%)^*	1 (1.2)	9 (10.6)	0.018
DF 5 ,n (%)	6 (7.2)	8 (9.4)	0.609
†EULAR-OMERACT combined score.
‡(SH + PD)>1 indicates moderate to severe synovitis.
*Denotes statistical significance at P < 0.05 level.
RC:radiocarpal joint, MC:midcarpal joint, UC:ulnocarpal joint; LAT: lateral compartment, medial, MED: medial compartment, SUP: superior compartment; APL: abductor pollicis longus; EPB: extensor pollicis brevis; ECRL: extensor carpi radialis longus; ECRB: extensor carpi radialis brevis; EPL: extensor pollicis longus; EDC: extensor digitorum communis; EIP: extensor indicis propius; EDM: extensor digit minimi; ECU: extensor carpi ulnaris; DF, digit flexors; GS, Grayscale; SH, synovial hypertrophy; PD, power Doppler.

Univariate logistic regression analyses of clinical, serological, and synovial US variables

Univariate logistic regression analysis identified several predictors, including female, disease duration, onset date of symptoms, date of diagnosis, date of first visit, date of first visit, years, date of first visit, with osteoporosis, X-ray hands erosion, complications, number of used b/tsDMARDs, number of used medications, anti-CCP, RF, and EGA (Table 4).

Table 4

Univariate analyses of clinical, serological and US variables combined score†(SH + PD)>1‡ in the prediction of D2T RA
Variables	Clinical and serological variables
Variables	P	OR score	95%CI
Age, years	0.591	1.01	0.98 ~ 1.03
Female (%)^*	0.032	2.15	1.07 ~ 4.3
Disease.duration, years^*	0.042	0.93	0.86 ~ 1
Onset date of symptoms,years^*	0.034	0.93	0.86 ~ 0.99
Date of diagnosis, years^*	0.007	0.87	0.79 ~ 0.96
Date of first visit, years^*	0.014	0.89	0.81 ~ 0.98
Diagnosis to treatment initiation with b/tsDMARDs,years	0.059	0.64	0.41 ~ 1.02
Weight (kg)	0.628	1	0.98 ~ 1.01
Height, cm	0.091	1.02	1 ~ 1.05
BMI	0.108	1.08	0.98 ~ 1.18
Smoking history	0.366	0.56	0.16 ~ 1.98
With osteoporosis^*	0.032	0.49	0.26 ~ 0.94
With anaemia	0.117	0.51	0.22 ~ 1.18
X-ray hands erosions^*	0.004	0.38	0.2 ~ 0.73
Commodities^*	0.049	0.83	0.2 ~ 0.73
Number of used csDMARDs	0.371	1.2	0.8 ~ 1.81
Number of used b/tsDMARDs^*	<0.001	0.25	0.14 ~ 0.45
Number of used mode of action^*	< 0.001	12.5	4.51 ~ 34.61
PSL user	0.166	1.55	0.83 ~ 2.9
Dose of PSL, mg/day	0.46	0.99	0.96 ~ 1.02
Anti-CCP, n (%)
Negative	Ref
Low positive	0.308	0.68	0.32 ~ 1.43
High positive^*	0.001	0.28	0.13 ~ 0.6
RF, n (%)
Negative	Ref
Low positive	0.059	0.32	0.1 ~ 1.04
High positive^*	< 0.001	0.13	0.05 ~ 0.34
ESR, mm/h^*	0.003	0.98	0.97 ~ 0.99
CRP, mg/l	0.286	1	0.99 ~ 1
TJC (0–28)	0.653	0.99	0.95 ~ 1.04
SJC(0–28)	0.238	0.98	0.94 ~ 1.02
PGA	0.509	0.88 ~ 1.850.88 ~ 1.85	0.62 ~ 1.26
EGA*	0.034	0.85	0.73 ~ 0.99
HAQ	0.203	1.27	0.88 ~ 1.85
VAS	0.098	1.11	0.98 ~ 1.27
DAS28 CRP	0.423	0.92	0.74 ~ 1.14
DAS28 ESR	0.145	0.85	0.69 ~ 1.06
Variables	US variables (SH + PD)>1†
Variables	P	OR score	95%CI
MCP1^*	0.015	0.45	0.23 ~ 0.85
MCP2	0.883	0.96	0.52 ~ 1.75
MCP3	0.41	0.72	0.33 ~ 1.58
MCP4	0.075	0.53	0.27 ~ 1.06
MCP5	0.225	0.67	0.35 ~ 1.28
PIP1	0.127	0.53	0.24 ~ 1.2
PIP2	0.231	0.67	0.35 ~ 1.29
PIP3	0.068	0.55	0.29 ~ 1.05
PIP4	0.161	1.72	0.81 ~ 3.66
PIP5	0.117	0.51	0.22 ~ 1.18
RC	0.71	1.13	0.6 ~ 2.13
MC	0.848	1.06	0.57 ~ 1.97
UC^*	0.009	0.44	0.24 ~ 0.82
SUP	0.524	0.82	0.45 ~ 1.51
MED^*	0.002	0.36	0.19 ~ 0.7
LAT^*	< 0.001	0.31	0.16 ~ 0.6
APL/EPB	0.501	0.72	0.27 ~ 1.88
ECRL/ECRB^*	0.005	0.16	0.05 ~ 0.58
EPL	0.675	0.82	0.32 ~ 2.09
EDC/EIP	0.559	0.77	0.33 ~ 1.82
EDM	0.256	0.49	0.14 ~ 1.69
ECU	0.397	0.73	0.35 ~ 1.52
DF 2	0.541	0.76	0.31 ~ 1.84
DF 3	0.348	0.62	0.23 ~ 1.68
DF 4^*	0.033	0.1	0.01 ~ 0.83
DF 5	0.61	0.75	0.25 ~ 2.26
†EULAR-OMERACT combined score.
‡(SH + PD)>1 indicates moderate to severe synovitis.
^*Denotes statistical significance at P < 0.05 level.
GS, Grayscale; SH, synovial hypertrophy; PD, power Doppler.

Univariate logistic regression analysis of SQ scoring, when SH/PD ≥ 1, revealed UC, MED, and LAT aspects of the knee, and the ECRL/ECR are predictive of refractory RA. These predictors were also observed in cases where the SH/PD>1(Supplementary Table S3).Subsequent analysis focusing on the EULAR-OMERACT combined score >1 through univariate logistic regression identified several predictors. These include MCP1, UC, MED, LAT, ECRL/ECRB, and DF4 (Table 4).

Final logistic regression, linear SVM, and random forest model for the prediction of D2T-RA

Based on significant variables with combined score P<0.05, we established two models: Model 1, which combines clinical and serological factors, and Model 2, which further incorporates US features.

The results indicated that the Random Forest algorithm demonstrated the following performance on the test set: Model 1 achieved an AUC of 0.81 with an F1 score of 0.67; upon the introduction of US features in Model 2, the AUC was enhanced to 0.83, and the F1 score improved to 0.69. However, the test set performance of the Logistic Model 2 was an AUC of 0.72 with an F1 score of 0.44; the SVM Model 2 achieved an AUC of 0.72 with an F1 score of 0.55, which did not show a significant improvement over Model 2(Table 5).The ROC and feature weights (random forest) are shown in Figs. 3 and 4. Supplementary Fig. 1 features the confusion matrix, while Supplementary Fig. 2 details the logistic regression and SVM features.

Table 5

Final logistic regression, linear SVM, and random forest model for the prediction of D2T RA
	Model 1^a						Model 2^b
	AUC	Accuracy	Sensitivity	Specificity	Recall	F1-value	AUC	Accuracy	Sensitivity	Specificity	Recall		F1-value
Training set
Logistic Regression	0.93 (0.88, 0.97)	0.87	0.91	0.85	0.91	0.87	0.96 (0.92, 0.99)	0.89	0.91	0.88		0.91	0.89
SVM	0.93 (0.89, 0.97)	0.88	0.91	0.86	0.91	0.88	0.96 (0.93, 0.99)	0.92	0.94	0.91		0.94	0.92
Random Forest	1.00 (1.00, 1.00)	1	1	1	1	1	1.00 (1.00, 1.00)	1	1	1		1	1
Testing set
Logistic Regression	0.74 (0.56, 0.90)	0.68	0.59	0.71	0.59	0.65	0.72 (0.54, 0.89)	0.56	0.35	0.59		0.35	0.44
SVM	0.73 (0.55, 0.89)	0.62	0.47	0.67	0.47	0.55	0.72 (0.53, 0.88)	0.62	0.47	0.67		0.47	0.55
Random Forest	0.81 (0.66, 0.93)	0.71	0.59	0.77	0.59	0.67	0.83 (0.69, 0.96)	0.74	0.59	0.83		0.59	0.69
a: Clinical + Serology; b: Clinical + Serology + Ultrasonography

Inter-reading agreement

Table 6 presents the Kappa values for each anatomical region's images. Overall, apart from the dorsal extensor tendon compartment of the wrist, which exhibited moderate inter-rater reliability for the SH score binary (κ value of 0.59), the κ values for other regions were good to excellent, ranging from 0.64 to 0.97. The intra-rater reliability for the binary and SQ scores of the MCP PD was the highest (κ values of 0.93 and 0.97, respectively), followed by MCP combined scores (κ values of 0.86 for both) and the MCP SH (κ 0.80 and 0.79, respectively).

Table 6 Inter-readers reliability of scoring synovitis in images according to the joints in joints and tendons in patients with D2T RA and Non-D2T RA

Joints and Tendons	SH* kappa		PD kappa		Combined score† (SH + PD) kappa
Joints and Tendons	(yes/no) +95%CI (min–max)	(0–3) +95%CI (min–max)	(yes/no) +95%CI (min–max)	(0–3) +95%CI (min–max)	(yes/no) +95%CI (min–max)	(0–3) +95%CI (min–max)
MCP	0.80（0.70 - 0.89）	0.79（0.74 - 0.85）	0.93（0.87 - 0.98）	0.97（0.95 - 0.99）	0.86（0.78 - 0.938）	0.86（0.86 - 0.94）
PIP	0.78（0.69 - 0.87）	0.78（0.71 - 0.86）	0.76（0.70 - 0.89）	0.88（0.82 - 0.95）	0.76（0.65 - 0.86）	0.82（0.75 - 0.89）
Wrist	0.76（0.65- 0.87）	0.80（0.74 - 0.87）	0.92（0.86 - 0.98）	0.95（0.92 - 0.99）	0.79（0.69 - 0.90）	0.84（0.78- 0.90）
Keen	0.72（0.62- 0.83）	0.70（0.62- 0.77）	0.93（0.87- 0.99）	0.79（0.72 - 0.85）	0.80（0.70 - 0.89）	0.79（0.72 - 0.85）
Wrist extensor compartments	0.59（0.44 - 0.73）	0.67（0.56 - 0.77）	0.67（0.53 - 0.80）	0.70（0.60- 0.81）	0.64（0.51 - 0.76）	0.68（0.57 - 0.79）
DF	0.71（0.60 - 0.82）	0.65（0.57- 0.73）	0.81（0.70- 0.91）	0.71（0.58 - 0.83）	0.71（0.60 - 0.82）	0.71（0.63 - 0.78）

†EULAR-OMERACT combined score.

DF, digit flexors; SH, synovial hypertrophy; PD, power Doppler.

This study is dedicated to enhancing the accuracy of early diagnosis for D2T RA, for which two machine learning models have been developed. Model 1 is based on clinical and serological data, while Model 2 incorporates clinical, serological, and US data, with the US examination covering the knee, wrist, and hand joints, as well as tendons. Utilizing logistic regression, random forest, and SVM algorithms, we constructed and evaluated these models. The results indicate that the Random Forest model, which includes US data, outperforms the model using only clinical and serological data in terms of AUC, accuracy, specificity, and F1 score, facilitating early diagnosis and timely adjustment of treatment plans.

D2T RA manifests in clinical practice as a complex and heterogeneous condition, often emerging after the exhaustion of all available treatment options [7], highlighting the scarcity of biomarkers for predicting treatment responses and the trial-and-error approach inherent in therapeutic strategies [18]. Our study indicates that pulmonary and cardiovascular diseases are the two primary comorbidities affecting RA treatment. MTX, a pivotal drug in RA therapy, is contraindicated in these comorbidities, potentially leading to initial treatment failure and diminished efficacy of b/tsDMARDs. The presence of multiple comorbidities may be correlated with the progression to D2T RA. Additionally, D2T RA patients had an average of 1.4 ± 0.5 prior b/tsDMARDs, a number higher than that of D2T RA patients. Notably, the use of medications with the number of used modes of action had the greatest weight in the model, significantly increasing the risk of D2T RA. The existence of comorbidities profoundly alters RA's pharmacological treatment strategies, potentially compelling patients to investigate a broader array of b/tsDMARDs and drugs with distinct mechanisms of action in pursuit of optimal therapeutic outcomes [7, 19].

Our research not only elucidates medication use and comorbidities but also highlights the longer disease duration and delayed diagnosis in D2T RA. These findings align with Tan et al.'s study [18]. Additionally, our study identifies osteoporosis and bone erosion as independent risk factors for D2T-RA, associated with disease complexity, steroid use, and persistent inflammation. Aize [20] and Gauri's [21] studies corroborate osteoporosis as an independent predictor of D2T-RA through logistic regression analysis. Bone erosion, a key RA manifestation, is closely linked to disease severity and functional decline, potentially contributing to the development of D2T RA.

Females, high activity serological scores, and some clinical indexes are pivotal in RA. Our systematic analysis of disease-related variables underscores early risk factor identification to preempt treatment post-RA flares. Rigours disease control and targeted treatment are essential for achieving remission and maintaining a state of low disease activity in RA [18]. Nagy G et al. have highlighted [22] that in RA management, the detection of inflammation is critical for guiding both pharmacological and non-pharmacological interventions, thereby minimizing the unnecessary cycles of DMARD therapy. Despite the fact that DAS28, RF, and anti-CCP are helpful for diagnosis and prognosis, they present issues such as complex calculations or insufficient specificity [23].

Musculoskeletal US is a diagnostic tool with high reliability and sensitivity for RA diagnosis and prognostic assessment. [24–26]. Limited target joints are crucial in evaluating the disease activity of RA, and we have designed an US scoring protocol that covers large, medium, and small joints, which are commonly affected in RA, including the hand, wrist, and knee joints. Additionally, some studies suggest that tenosynovitis provides independent predictive data [10, 27], which recommends including tendons as candidate variables. Therefore, we use a multi-joint plus hand tendon semi-quantitative US evaluation scheme.

The current multi-joint US studies have only looked at persistent synovitis and the remission period of synovitis [28–31], but most people with D2T-RA have moderate-to-high disease activity. This is why combined scoring is needed in addition to single scoring to better separate moderate-to-high activity synovitis. Studies indicate that when PD>1, the involvement of the MCP1,3,4, and PIP3 joints is more significant in identifying moderate-to-high activity RA. Compared to GS, the addition of meaningful variables is greater, indicating that GS has lower sensitivity in predicting synovitis activity, while PD assessment is more reliable, consistent with studies. [11, 32]. This underscores the importance of combining GS and PD scores for a more comprehensive inflammatory assessment [33].

In this study, we observed that D2T-RA and non-D2T-RA share similar US characteristics, making it difficult to distinguish between the conventionally involved sites such as the MCP 2, 3, and PIP 2, 3. But MCP 1 may be able to tell you if someone is going to get refractory RA because early damage to it is strongly linked to a loss of hand grip strength [34]. Additionally, UC joint synovitis is the only feature in the wrist joint that helps to differentiate D2T-RA, possibly because the UC joint is less susceptible to involvement in the early stages and does not bear as much functional load as the RC and MC joints [35]. Also, the involvement of the MED and LAT in the knee joint is a strong predictor for D2T-RA. This means that patients may have more disease activity and disability, which shows how important it is to find the disease early [36].

Additionally, involvement of compartment II and the DF4 region may be independent indicators of D2T-RA. Previous studies have mainly focused on early RA and found that dorsal tenosynovitis of the wrist is more common in compartments IV and VI, while the palmar side mainly affects the flexor tendons of the second and third fingers [37]. However, D2T-RA presents a special pattern of involvement in hand joints and tendon sheaths, which may be related to the heavier inflammatory burden associated with high disease activity in patients. Although there is no evidence, this study suggests that in patients with high disease activity and poor treatment response during the active phase, monitoring of the MCP 1, UC joint, medial and lateral regions of the knee joint, as well as the compartments II and the DF4 region, may be considered as early indicators of D2T RA.

After modeling the single-factor variables, we observed the following: Model 1, using the Random Forest algorithm, achieved an AUC of 0.81 and an F1 score of 0.67, indicating the model has a certain diagnostic capability. Model 2 raised its AUC to 0.83 and its F1 score to 0.69, demonstrating that in terms of variable selection, Model 2 is more accurate than Model 1. Although the role of US features in variable selection is limited, they show potential value in identifying patients with a poor response to conventional treatments. Currently, there is a lack of direct model studies on the role of US in D2T RA patients; however, existing evidence suggests that US may play a significant role in distinguishing truly refractory RA[3]. Some studies argue [38] that it does not add extra value in defining the diagnosis of RA-D2T, possibly due to singular US grading, the subjectivity of US without consistency, and only examining synovitis of both hands, all of which have been addressed and corrected in this article. The overall performance of logistic regression and SVM models is relatively lower, which may be constrained by their ability to handle nonlinear data, and the outcomes did not meet expectations. The diagnosis of refractory RA is complicated, suggesting a need for improved models or new biomarkers.

This study compared the consistency between binary scoring and the SQ grading system in joint US assessments. The results indicated that the SH score for the dorsal extensor tendon sheath of the wrist was moderate, while the scores for other regions ranged from satisfactory to excellent. The PD images demonstrated the highest consistency in both binary and SQ grading, which is consistent with the findings of the Maria study [6]. Additionally, the MCP joints showed a high level of consistency between binary and SQ scoring in SH, PD, and SH + PD images (κ values ranged from 0.79 to 0.97). In contrast, the scoring consistency for tenosynovitis was lower. This might be attributed to the fact that the MCP joint, as a standardized joint, is more commonly trained among physicians [11, 17], while tenosynovitis is less common in RA, leading to relatively less experience in exploration among physicians.

The study has limitations: 1) Future research needs stratified analysis for diverse RA patients, considering lifestyle, economy, and demographics. 2) US assessments should cover more joints to improve model accuracy. 3) The model's accuracy might be limited by a small, single-center sample.

In summary, the Random Forest predictive model proposed in this study, based on clinical, serological, and multi-joint US data, has demonstrated good diagnostic efficacy in the early identification of D2T-RA and holds promise as an effective clinical tool for predicting patients with D2T-RA.

D2T-RA difficult-to-treat rheumatoid arthritis

GS gray-scale

PD power Doppler

DMARDs disease-modifying antirheumatic drugs

csDMARDs conventional synthetic DMARDs

bDMARDs biological DMARDs

tsDMARDs targeted synthetic DMARDs

MTX methotrexate

LEF leflunomide

PGA patient global assessment

VAS visual analog scale

HAQ health assessment questionnaire

MCP metacarpophalangeal

PIP proximal interphalangeal

RC radiocarpal joint

MC midcarpal joint

UC ulnocarpal joint

LAT lateral compartment

MED medial compartment

SUP superior compartment

APL abductor pollicis longus

EPB extensor pollicis brevis

ECRL extensor carpi radialis longus

ECRB extensor carpi radialis brevis

EPL extensor pollicis longus

EDC extensor digitorum communis

EIP extensor indicis propius;

EDM extensor digit minimi

ECU extensor carpi ulnaris

DF digit flexors

SQ semi-quantitative

SH synovial hypertrophy

SVM support vector machine

Authors’ contributions

Ting Wang and Zhen Wang conceived and directed this work. Our thanks go to Lele Huang and Yingying Jia for supervising the data analysis, and to Yakun Yu and Rong Zhu for patient recruitment.We acknowledge the crucial role of Yuan Wang, Yan Li, Xuejiao Shen, and Jiaqi Wei in performing ultrasound scans for our study subjects. Special thanks to our corresponding author, Fang Nie, for ongoing guidance.We also appreciate the contributions of Ci Yin and Chunjie Zhang in creating the graphical elements for our presentation.Additionally, we are grateful for the support provided by the Ultrasound Medical Center and the Department of Rheumatology at Lanzhou University Second Hospital for their contribution to this project.

Availability of data and materials

The data that support the findings of this study are available on reasonable request from the corresponding authors.

Declarations

Ethics approval and consent to participate

The institutional review board (Ethics Committee of Lanzhou University Second Hospital, no. 2024A-326) has approved the study protocol, and all patients have provided written informed consent.

Consent for publication

Not applicable.

Competing interests

The authors have declared no conflicts of interest.

Funding

This work was supported by natural science foundation of Gansu province (grant numbers 24JR11RA064).

Footnotes

Ting Wang and Zhen Wang contributed equally to this work.

Acknowledgements

The authors are very grateful to Dr. Xiong Diaohan for his valuable contributions to this research.

Buch MH. Defining refractory rheumatoid arthritis. Ann Rheum Dis. 2018;77(7):966-9.
de Hair MJH, Jacobs JWG, Schoneveld JLM, van Laar JM. Difficult-to-treat rheumatoid arthritis: an area of unmet clinical need. Rheumatology (Oxford). 2018;57(7):1135-44.
Buch MH, Eyre S, McGonagle D. Persistent inflammatory and non-inflammatory mechanisms in refractory rheumatoid arthritis. Nat Rev Rheumatol. 2021;17(1):17-33.
Roodenrijs NMT, Welsing PMJ, van der Goes MC, Tekstra J, Lafeber FPJG, Jacobs JWG, et al. Healthcare utilization and economic burden of difficult-to-treat rheumatoid arthritis: a cost-of-illness study. Rheumatology (Oxford). 2021;60(10):4681-90.
van Nies JAB, Tsonaka R, Gaujoux-Viala C, Fautrel B, van der Helm-van Mil AHM. Evaluating relationships between symptom duration and persistence of rheumatoid arthritis: does a window of opportunity exist? Results on the Leiden early arthritis clinic and ESPOIR cohorts. Ann Rheum Dis. 2015;74(5):806-12.
Contreras-Yáñez I, Pascual-Ramos V. Window of opportunity to achieve major outcomes in early rheumatoid arthritis patients: how persistence with therapy matters. Arthritis Res Ther. 2015;17(1):177.
Takanashi S, Kaneko Y, Takeuchi T. Characteristics of patients with difficult-to-treat rheumatoid arthritis in clinical practice. Rheumatology (Oxford). 2021;60(11):5247-56.
Raza K, Filer A. The therapeutic window of opportunity in rheumatoid arthritis: does it ever close? Ann Rheum Dis. 2015;74(5):793-4.
Nam JL, Hensor EMA, Hunt L, Conaghan PG, Wakefield RJ, Emery P. Ultrasound findings predict progression to inflammatory arthritis in anti-CCP antibody-positive patients without clinical synovitis. Ann Rheum Dis. 2016;75(12):2060-7.
Sahbudin I, Singh R, De Pablo P, Rankin E, Rhodes B, Justice E, et al. The value of ultrasound-defined tenosynovitis and synovitis in the prediction of persistent arthritis. Rheumatology (Oxford). 2023;62(3):1057-68.
D'Agostino M-A, Terslev L, Aegerter P, Backhaus M, Balint P, Bruyn GA, et al. Scoring ultrasound synovitis in rheumatoid arthritis: a EULAR-OMERACT ultrasound taskforce-Part 1: definition and development of a standardised, consensus-based scoring system. RMD Open. 2017;3(1):e000428.
Radner H, Neogi T, Smolen JS, Aletaha D. Performance of the 2010 ACR/EULAR classification criteria for rheumatoid arthritis: a systematic literature review. Ann Rheum Dis. 2014;73(1):114-23.
Szkudlarek M, Court-Payen M, Jacobsen S, Klarlund M, Thomsen HS, Østergaard M. Interobserver agreement in ultrasonography of the finger and toe joints in rheumatoid arthritis. Arthritis Rheum. 2003;48(4):955-62.
Beitinger N, Ehrenstein B, Schreiner B, Fleck M, Grifka J, Lüring C, et al. The value of colour Doppler sonography of the knee joint: a useful tool to discriminate inflammatory from non-inflammatory disease? Rheumatology (Oxford). 2013;52(8):1425-8.
Vreju F, Ciurea M, Roşu A, Muşetescu A, Grecu D, Ciurea P. Power Doppler sonography, a non-invasive method of assessment of the synovial inflammation in patients with early rheumatoid arthritis. Rom J Morphol Embryol. 2011;52(2):637-43.
Naredo E, D'Agostino MA, Wakefield RJ, Möller I, Balint PV, Filippucci E, et al. Reliability of a consensus-based ultrasound score for tenosynovitis in rheumatoid arthritis. Ann Rheum Dis. 2013;72(8):1328-34.
Terslev L, Naredo E, Aegerter P, Wakefield RJ, Backhaus M, Balint P, et al. Scoring ultrasound synovitis in rheumatoid arthritis: a EULAR-OMERACT ultrasound taskforce-Part 2: reliability and application to multiple joints of a standardised consensus-based scoring system. RMD Open. 2017;3(1):e000427.
Tan Y, Buch MH. 'Difficult to treat' rheumatoid arthritis: current position and considerations for next steps. RMD Open. 2022;8(2).
Kearsley-Fleet L, Davies R, De Cock D, Watson KD, Lunt M, Buch MH, et al. Biologic refractory disease in rheumatoid arthritis: results from the British Society for Rheumatology Biologics Register for Rheumatoid Arthritis. Ann Rheum Dis. 2018;77(10):1405-12.
Aizer J, Reed G, Onofrei A, Harrison MJ. Predictors of bone density testing in patients with rheumatoid arthritis. Rheumatol Int. 2009;29(8):897-905.
Wijbrandts CA, Klaasen R, Dijkgraaf MGW, Gerlag DM, van Eck-Smit BLF, Tak PP. Bone mineral density in rheumatoid arthritis patients 1 year after adalimumab therapy: arrest of bone loss. Ann Rheum Dis. 2009;68(3):373-6.
Nagy G, Roodenrijs NMT, Welsing PMJ, Kedves M, Hamar A, van der Goes MC, et al. EULAR points to consider for the management of difficult-to-treat rheumatoid arthritis. Ann Rheum Dis. 2022;81(1):20-33.
Sucher R, Rademacher S, Jahn N, Brunotte M, Wagner T, Alvanos A, et al. Effects of simultaneous pancreas-kidney transplantation and kidney transplantation alone on the outcome of peripheral vascular diseases. BMC Nephrol. 2019;20(1):453.
Andersen M, Ellegaard K, Hebsgaard JB, Christensen R, Torp-Pedersen S, Kvist PH, et al. Ultrasound colour Doppler is associated with synovial pathology in biopsies from hand joints in rheumatoid arthritis patients: a cross-sectional study. Ann Rheum Dis. 2014;73(4):678-83.
Ciurtin C, Jones A, Brown G, Sin FE, Raine C, Manson J, et al. Real benefits of ultrasound evaluation of hand and foot synovitis for better characterisation of the disease activity in rheumatoid arthritis. Eur Radiol. 2019;29(11):6345-54.
Perricone C, Ceccarelli F, Modesti M, Vavala C, Di Franco M, Valesini G, et al. The 6-joint ultrasonographic assessment: a valid, sensitive-to-change and feasible method for evaluating joint inflammation in RA. Rheumatology (Oxford). 2012;51(5):866-73.
Bellis E, Scirè CA, Carrara G, Adinolfi A, Batticciotto A, Bortoluzzi A, et al. Ultrasound-detected tenosynovitis independently associates with patient-reported flare in patients with rheumatoid arthritis in clinical remission: results from the observational study STARTER of the Italian Society for Rheumatology. Rheumatology (Oxford). 2016;55(10):1826-36.
Ramírez J, Celis R, Usategui A, Ruiz-Esquide V, Faré R, Cuervo A, et al. Immunopathologic characterization of ultrasound-defined synovitis in rheumatoid arthritis patients in clinical remission. Arthritis Res Ther. 2016;18:74.
Ramírez J, Ruíz-Esquide V, Pomés I, Celis R, Cuervo A, Hernández MV, et al. Patients with rheumatoid arthritis in clinical remission and ultrasound-defined active synovitis exhibit higher disease activity and increased serum levels of angiogenic biomarkers. Arthritis Res Ther. 2014;16(1):R5.
Baker KF, Thompson B, Lendrem DW, Scadeng A, Pratt AG, Isaacs JD. Lack of association between clinical and ultrasound measures of disease activity in rheumatoid arthritis remission. Ther Adv Musculoskelet Dis. 2020;12:1759720X20915322.
Backhaus TM, Ohrndorf S, Kellner H, Strunk J, Hartung W, Sattler H, et al. The US7 score is sensitive to change in a large cohort of patients with rheumatoid arthritis over 12 months of therapy. Ann Rheum Dis. 2013;72(7):1163-9.
Mandl P, Kurucz R, Niedermayer D, Balint PV, Smolen JS. Contributions of ultrasound beyond clinical data in assessing inflammatory disease activity in rheumatoid arthritis: current insights and future prospects. Rheumatology (Oxford). 2014;53(12):2136-42.
Freeston JE, Wakefield RJ, Conaghan PG, Hensor EMA, Stewart SP, Emery P. A diagnostic algorithm for persistence of very early inflammatory arthritis: the utility of power Doppler ultrasound when added to conventional assessment tools. Ann Rheum Dis. 2010;69(2):417-9.
Rydholm M, Sharma A, Jacobsson L, Turesson C. The relation between synovitis of individual finger joints and grip force over the first 5 years in early rheumatoid arthritis - a cohort study. Arthritis Res Ther. 2023;25(1):231.
Narváez J, Narváez JA, de Albert M, Gómez-Vaquero C, Nolla JM. Can magnetic resonance imaging of the hand and wrist differentiate between rheumatoid arthritis and psoriatic arthritis in the early stages of the disease? Semin Arthritis Rheum. 2012;42(3):234-45.
Sidhu N, Wouters F, Niemantsverdriet E, van der Helm-van Mil AHM. MRI-detected synovitis of the small joints predicts rheumatoid arthritis development in large joint undifferentiated inflammatory arthritis. Rheumatology (Oxford). 2022;61(SI):SI23-SI9.
Sun X, Deng X, Xie W, Wang L, Wang Y, Zhang Z. The agreement between ultrasound-determined joint inflammation and clinical signs in patients with rheumatoid arthritis. Arthritis Res Ther. 2019;21(1):100.
Garcia-Salinas R, Sanchez-Prado E, Mareco J, Ronald P, Ruta S, Gomez R, et al. Difficult to treat rheumatoid arthritis in a comprehensive evaluation program: frequency according to different objective evaluations. Rheumatol Int. 2023;43(10):1821-8.

No competing interests reported.

Supplementarymaterial.docx

Machine learning-enhanced clinical and ultrasound technology for early detection of difficult-to-treat rheumatoid arthritis

Status:

Version 1

Abstract

Figures

Key Message

Introduction

Methods

Statistical analysis

Results

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1