Machine learning methods have the potential to optimise the diagnosis of inflammatory arthropathies and the quantification of damage over time.
Objectives
To summarise the literature investigating the use of machine learning methods to diagnose and detect damage in the hands and feet of patients with Rheumatoid Arthritis (RA) and Psoriatic Arthritis (PsA).
Methods
A literature search was undertaken in PubMed, EMBASE, Web of Science and the Cochrane Central Register for Controlled Trials using a pre-specified search strategy. Full-text articles and conference abstracts published in English between the 1st of January 1995 and the 22nd of June 2022 that utilised machine learning (ML) in an imaging modality (X-Ray [XR], Computed Tomography [CT], Ultrasound [USS] or Magnetic Resonance Imaging [MRI]) to diagnose or assess structural damage in the hands and feet of patients with RA or PsA were included. All abstracts were reviewed by two authors (AA and AB) and all eligible full-text articles were assessed by the same authors, who subsequently independently extracted data according to the review protocol. Reference management and data extraction were conducted using Covidence. Data extracted included patient population and numbers, diagnoses, ML models and validation/testing methods, and algorithm performance.
Results
The literature review identified 851 unique references, of which 123 were eligible for full-text review; 26 references met the inclusion criteria for the review (Figure 1: PRISMA Diagram). There were 21 full-text articles and 5 conference abstracts published between 2000-2022, all of which assessed joints in the hands and/or wrists. Most studies utilised XR (n=23) rather than USS (n=1), MRI (n=1) and CT [1]. The number of patients/radiographs included used ranged from 15-5191 and most studies assessed RA only (n=23), with 2 studies assessing RA and PsA and 1 assessing PsA only. 24 studies were cross-sectional while 2 were longitudinal. ML methods were used to diagnose RA (n=9), diagnose PsA (n=2), estimate joint space score (n=5) or measure joint space width (n=3) or estimate change in joint space width (n=1), detect erosions (n=3) or estimate erosion score (n=7), estimate overall damage score (1), detect subluxation/ankylosis (n=1) and detect periarticular textural change (n=1). Additional tasks performed by ML algorithms were cropping of radiographs, localising joints and segmenting bones. Most studies utilised transfer learning on pre-trained neural networks: VGG15, VGG16, DenseNet, AlexNet, Single Shot Multibox Detector, YOLOv3, YOLOv4, LENet, Network in Network, SqueezeNet, ResNet, ResNet50, DeepTEN, U-net and EfficientNet. Some studies incorporated shape models (e.g. active shape, gaussian and local linear mapping models), multiscale gradient vector flow snakes, Graph Convolutional Networks, support vector machines, support vector regression and ridge regression. Complete model assessment with training, validation and testing cohorts were reported in 12 studies (46.2%), and of these 4 utilised n-fold cross-validation rather than hold-out validation. There was significant heterogeneity in the reporting and the performance of the models used.
Conclusion
The use of machine learning in diagnosing and detecting damage has mostly focused on plain radiography in patients with RA; there are limited data in PsA. Heterogeneity in the reporting of results limits succinct comparison of performance between ML methods.
REFERENCES:
NIL.
Acknowledgements:
NIL.
Disclosure of Interests
Anna Antony Speakers bureau: Lilly, AbbVie, Ann Biju: None declared, Adwaye Rambojun: None declared, William Tillett Speakers bureau: Abbvie, Amgen, Eli Lilly, GSK, Janssen, Novartis, Pfizer, UCB, Consultant of: Abbvie, Amgen, Eli Lilly, GSK, Janssen, Novartis, Ono Pharma, Pfizer, UCB, Grant/research support from: Janssen, UCB, Pfizer, Eli-Lilly.
Machine learning (ML) algorithms could facilitate the standardisation of joint damage assessment in Psoriatic Arthritis (PsA) and improve its accessibility in clinical and research settings. ML algorithms trained on manually annotated hand and wrist radiographs have promising performance characteristics[1]. A large volume of annotated radiographs is needed, and annotation is time consuming and subject to reliability issues given X-Rays (XRs) are 2D representations.
Objectives
To develop a reliable method for the annotation of hand and wrist bones on XRs in order to facilitate the development of supervised ML algorithms for joint damage detection.
Methods
10 bilateral hand and wrist XRs were selected at random from the Bath PsA XR database. 5 XRs were independently annotated by 3 annotators; (AA & WT (rheumatologist) and YHR (radiologist)) using the ASPAX software[2]. Annotations were visually inspected for areas of discordance and consensus annotation guidelines were developed. Annotation was repeated using the annotation guidelines on second set of 5 XRs. With annotator 1 (WT) representing ground truth, the mean error (ME; in pixels) of the annotation (deviation from ground truth) and the mean fractional error (MFE; corrects for the perimeter measurements of the bone), was estimated in pre- and post-training annotations. The ME and MFE within a single annotator (AA) were estimated in 5 radiographs after a 2-month interval.
Results
Visual inspection determined that the areas of discordance in annotation were the 1st interphalangeal joint, the metacarpal bases, the hamate and capitate bones, and the trapezium and trapezoid bones (Figure 1). The MFE between the annotators and ground truth improved for all bones following the development of annotation guidelines, with the largest improvement evident in the annotation of the metacarpal bones (Table 1). The intra-reader and inter-reader MFEs were comparable (Table 1).
Conclusion
Standardised instructions may facilitate reliable hand and wrist bone annotation and enable the acquisition of large volumes of annotated training data for supervised ML algorithms.
References
[1]Adwaye Rambojun, William Tillett, Tony Shardlow, Neill D. F. Campbell; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021, pp. 2043-2052 [2]Machine Learning and Rheumatic Diseases (Website) https://people.bath.ac.uk/amr62/Projects/malard/malard.html
Acknowledgements:
NIL.
Disclosure of Interests
Adwaye Rambojun: None declared, Anna Antony Speakers bureau: Eli Lilly, AbbVie, Ynyr Hughes-Roberts: None declared, William Tillett Speakers bureau: Abbvie, Amgen, Eli Lilly, GSK, Janssen, Novartis, Pfizer, UCB, Consultant of: Abbvie, Amgen, Eli Lilly, GSK, Janssen, Novartis, Ono Pharma, Pfizer, UCB, Grant/research support from: Janssen, UCB, Pfizer, Eli-Lilly.
Background: Telemedicine was widely utilised to complement face-to-face (F2F) care in 2020 during the COVID-19 pandemic, but the impact of this on patient care is poorly understood. Objectives: To investigate the impact of telemedicine during COVID-19 on outpatient rheumatology services. Methods: We retrospectively audited patient electronic medical records from rheumatology outpatient clinics in an urban tertiary rheumatology centre between April-May 2020 (telemedicine cohort) and April-May 2019 (comparator cohort). Differences in age, sex, primary diagnosis, medications, and proportion of new/review appointments were assessed using Mann-Whitney U and Chi-square tests. Univariate analysis was used to estimate associations between telemedicine usage and the ability to assign a diagnosis in patients without a prior rheumatological diagnosis, the frequency of changes to immunosuppression, subsequent F2F review, planned admissions or procedures, follow-up phone calls, and time to next appointment. Results: 3,040 outpatient appointments were audited: 1,443 from 2019 and 1,597 from 2020. There was no statistically significant difference in the age, sex, proportion of new/review appointments, or frequency of immunosuppression use between the cohorts. Inflammatory arthritis (IA) was a more common diagnosis in the 2020 cohort (35.1% vs 31%, p=0.024). 96.7% (n=1,444) of patients seen in the 2020 cohort were reviewed via telemedicine. In patients without an existing rheumatological diagnosis, the odds of making a diagnosis at the appointment were significantly lower in 2020 (28.6% vs 57.4%; OR 0.30 [95% CI 0.16-0.53]; p<0.001). Clinicians were also less likely to change immunosuppressive therapy in 2020 (22.6% vs 27.4%; OR 0.78 [95% CI 0.65-0.92]; p=0.004). This was mostly driven by less de-escalation in therapy (10% vs 12.6%; OR 0.75 [95% CI 0.59-0.95]; p=0.019) as there was no statistically significant difference in the escalation or switching of immunosuppressive therapies. There was no significant difference in frequency of follow-up phone calls, however, patients seen in 2020 required earlier follow-up appointments (p<0.001). There was also no difference in unplanned rheumatological presentations but significantly fewer planned admissions and procedures in 2020 (1% vs 2.6%, p=0.002). Appointment non-attendance reduced in 2020 to 6.5% from 10.9% in 2019 (OR 0.57 [95% CI 0.44-0.74]; p<0.001), however the odds of discharging a patient from care were significantly lower in 2020 (3.9% vs 6%; OR 0.64 [95% CI 0.46-0.89]; p=0.008), although there was no significance when patients who failed to attend were excluded. Amongst patients seen via telemedicine in 2020, a subsequent F2F appointment was required in 9.4%. The predictors of needing a F2F review were being a new patient (OR 6.28 [95% CI 4.10-9.64]; p<0.001), not having a prior rheumatological diagnosis (OR 18.43 [95% CI: 2.35-144.63]; p=0.006), or having a diagnosis of IA (OR 2.85 [95% CI: 1.40-5.80]; p=0.004) or connective tissue disease (OR 3.22 [95% CI: 1.11-9.32]; p=0.031). Conclusion: Most patients in the 2020 cohort were seen via telemedicine. Telemedicine use during the COVID-19 pandemic was associated with reduced clinic non-attendance, but with diagnostic delay, reduced likelihood of changing existing immunosuppressive therapy, earlier requirement for review, and lower likelihood of discharge. While the effects of telemedicine cannot be differentiated from changes in practice related to other aspects of the pandemic, they suggest that telemedicine may have a negative impact on the timeliness of management of rheumatology patients. Disclosure of Interests: None declared.
PCAT attenuation is a non-invasive biomarker of coronary inflammation and is a poor prognostic factor in patients with coronary artery disease.
Objectives
To compare the severity of PCAT attenuation amongst patients with inflammatory rheumatic diseases (IMD) compared to patients with non-inflammatory rheumatic diseases (NIMD) in a retrospective cross-sectional study.
Methods
All patients seen in Monash Health rheumatology clinics over a 12-month period were identified. Patient records were cross-referenced with the hospital computed tomography coronary angiography (CTCA) database to identify patients who had undergone CTCAs. IMD patients who had their CTCA prior to their RMD diagnosis were excluded. Age, sex, co-morbidities, diagnoses, height, weight, and medications were extracted from the patient electronic medical records. CTCA variables were extracted (mA, DLP, contrast volume, kV, CACS, CAD-RADS) or calculated (SIS, SSS, PCAT attenuation).
Results
A total of 117 patients were included (105 with IMD and 12 with NIMD). There were no significant differences in age, sex, co-morbidities or cardiovascular medication use between groups. Patients with IMD had a lower BMI and were less likely to be smokers or have dyslipidaemia, but were more likely to have hypertension and diabetes mellitus, these were not statistically significant. The mean (±sd) PCAT was higher in IMD patients [-84.6 (±12.14) vs. -92.0 (±5.01) p<0.001) compared to non-IMD patients. There were no significant differences in any other CTCA outcomes.
Conclusion
PCAT attenuation is associated with IMD in this retrospective single-center cross-sectional study. Data from larger prospective cohorts with healthy comparator controls are needed to assess the significance of these findings.
Objectives The objective of this article is to validate the Lupus Impact Tracker (LIT), a disease-specific patient-reported outcome (PRO) tool, in systemic lupus erythematosus (SLE) patients in a multi-ethnic Australian cohort. Methods Patients attending the Monash Lupus Clinic were asked to complete the LIT, a 10-item PRO. Psychometric testing assessing criterion validity, construct validity, test-retest reliability (TRT) and internal consistency reliability (ICR) were performed. We compared the LIT scores across patient characteristics, and correlations between LIT scores and SLEDAI-2k, PGA, and SLICC-SDI were examined. Results LIT data were obtained from 73 patients. Patients were 84% female with a median age of 41 years, and 34% were Asian. The cohort had mild-moderate disease activity with a median (IQR) Systemic Lupus Erythematosus Disease Activity Index-2000 (SLEDAI-2k) of 4 (IQR 2–6). The median LIT score was 32.5 (IQR 17.5–50). LIT demonstrated criterion validity against SLEDAI-2k and SDI. Construct validity assessed by confirmatory factor analysis demonstrated an excellent fit (Goodness of fit index 0.95, Comparative Fit Index 1, Root Mean Square Error of Approximation <0.0001). The LIT demonstrated TRT with an overall intraclass correlation coefficient of 0.986 (95% CI 0.968–0.995). ICR was demonstrated with a Cronbach’s alpha of 0.838. Patients with disability, low socioeconomic status, or higher disease activity had significantly worse LIT scores. Conclusion The LIT demonstrated properties consistent with its being valid in this population. Lower socioeconomic status appears to have a significant impact on patient-reported health-related quality of life in SLE.
'/%3e%3c/defs%3e%3cpath fill='%23012169' d='M0 0h10080v5040H0z'/%3e%3cpath stroke='%23fff' stroke-width='.6' d='m0 0 6 3m0-3L0 3' clip-path='url(%23b)' transform='scale(840)'/%3e%3cpath stroke='%23e4002b' stroke-width='.4' d='m0 0 6 3m0-3L0 3' clip-path='url(%23c)' transform='scale(840)'/%3e%3cpath stroke='%23fff' stroke-width='840' d='M2520 0v2520M0 1260h5040'/%3e%3cpath stroke='%23e4002b' stroke-width='504' d='M2520 0v2520M0 1260h5040'/%3e%3cg fill='%23fff'%3e%3cuse xlink:href='%23d' x='2520' y='3780'/%3e%3cuse xlink:href='%23a' x='7560' y='4200'/%3e%3cuse xlink:href='%23a' x='6300' y='2205'/%3e%3cuse xlink:href='%23a' x='7560' y='840'/%3e%3cuse xlink:href='%23a' x='8680' y='1869'/%3e%3cuse xlink:href='%23e' x='8064' y='2730'/%3e%3c/g%3e%3c/svg%3e)
'%3e%3cpath fill='%23012169' d='M0 0v30h60V0z'/%3e%3cpath stroke='%23fff' stroke-width='6' d='m0 0 60 30m0-30L0 30'/%3e%3cpath stroke='%23C8102E' stroke-width='4' d='m0 0 60 30m0-30L0 30' clip-path='url(%23b)'/%3e%3cpath stroke='%23fff' stroke-width='10' d='M30 0v30M0 15h60'/%3e%3cpath stroke='%23C8102E' stroke-width='6' d='M30 0v30M0 15h60'/%3e%3c/g%3e%3c/svg%3e)
