Novel approach uses machine learning for quick and easy rheumatic disease diagnosis

By Hugo Francisco de SouzaReviewed by Sophia CoveneyAug 24 2023

In a recent study published in the journal Frontiers in Medicine, researchers evaluated fluorescence optical imaging (FOI) as a method to accurately and rapidly diagnose rheumatic diseases of the hands.

They used machine learning algorithms to identify the minimum number of FOI features to differentiate between osteoarthritis (OA), rheumatoid arthritis (RA), and connective tissue disease (CTD). Of the 20 features identified as associated with the conditions, results indicate that reduced sets of features between five and 15 in number were sufficient to diagnose each of the diseases under study accurately.

FOI and rheumatic disease diagnosis

‘Rheumatic disease’ is an umbrella term used to describe conditions that affect the joints, tendons, muscles, ligaments, and bones. It is a common condition affecting as many as 2.55% of humans globally, with prevalence higher in older people than in younger adults. Rheumatic diseases may be chronic, with early and accurate diagnoses leading to the best health outcomes for patients.

Conventionally, imagining techniques, including Magnetic Resonance Imaging (MRI) and ultrasonography, have been the gold standard for diagnosing rheumatic diseases. However, these techniques are expensive and require specialized infrastructure not commonly available in diagnostic facilities.

Recently, fluorescence optical imaging (FOI) techniques have been employed for studying these diseases. This non-invasive technique takes advantage of rheumatic inflammation-induced microcirculation impairments to identify and characterize the type of rheumatic disease.

Diagnostic comparisons between MRI, ultrasonography, and FOI have verified the accuracy of the latter. However, until recently, signal enhancement at the joint regions of hands and wrists has been the primary focus of FOI studies, with other image features being largely ignored. Studies exploring the feature selection of specific FOI image features for rheumatic disease diagnosis remain lacking.

About the study

In the present study, researchers aimed to identify specific FOI image features required to accurately and rapidly characterize rheumatic diseases affecting the hands. They employed machine-learning approaches for this identification, given the capabilities of these models and the growing availability of usable data for model training.

Researchers began by collating FOI examination data from 3,690 clinically diagnosed rheumatic disease patients from an online database.

This data was compiled into three cohorts – osteoarthritis (OA), rheumatoid arthritis (RA), and connective tissue diseases (CTD) based on clinical diagnoses. In all cohorts, diagnoses made without knowledge of FOI images (i.e., just using MRI or ultrasonography data) were sufficient for study inclusion.

Since the number of patients in the OA cohort was significantly larger than the others, random choice was used to constrain the dataset and minimize sampling bias.

FOI examinations were conducted using an indocyanine green (ICG) fluorescence contrast agent administered intravenously (0.1 mg/kg body weight) 10 seconds after scanning initiation. The scanning procedure lasted 6 minutes, with images recorded every second for a total of 360 images per patient. Images were compiled into three ‘phases’ using the following methodology:

The end of phase 1 was characterized by the beginning of the apparent backflow of the dye from the nail bed area of fingers II–V. Phase 2 started thereafter and ended with frame 150. Phase 3 comprised all the following images to the end.

Rothe et al. (2023)

Images per phase were summed up and condensed into one image (three total per patient) for downstream analyses. The reader, the researcher responsible for image selection and annotation, perused these images for any of the 20 image features under study.

These image features included five joint-related features, two finger-related features, two nail features, four connective tissue features, and seven other features, five of which have been identified and described in this study for the first time.

The resultant data was pre-processed to address the three-class classification requirement of the study (RA, OA, or CTD). This multinomial requirement was transferred into a binomial classification to improve the identification and classification of disease-specific features via three One-vs-Rest (OvR) and three One-vs-One (OvO) problems.

There were six problems considered: RA-vs.-OA; OA-vs.-CTD; RA-vs.-CTD; RA-vs.-Rest; OA-vs.-Rest, and CTD-vs.-Rest.

Machine learning (ML) models were built, tested, and validated using 90% of the dataset for training and 10% for testing. Feature selection involved identifying and excluding ‘redundant’ features, those that did not provide sufficient diagnostic information, or those highly correlated with other features already providing adequate information.

This resulted in the filtering and ranking features using ML models, achieved using the phi-coefficient, the relief algorithm MultiSURF, Mean Decrease Impurity (MDI), and Mean Decrease Accuracy (MDA).

Finally, sequential forward selection, a feature selection methodology aimed at finding the smallest set of features that achieved near maximum diagnostic performance, was employed.

Study findings

This study included 609 patients in three cohorts (RA = 237, OA = 231, CTD = 141). Most patients were female (80.5%) between the ages of 38 and 74 (mean RA = 61.7, mean OA = 61.6, mean CTD = 51.9).

Collinearity analyses revealed that features were predominantly unique, with no feature directly making another one redundant. Hence, all elements were retained at feature selection initiation.

Ten ML algorithms were trained and tested with cohort data, with the gradient boosting machine (GBM) model identified as the best performing. The GBM model comprises a set of single ML models with stage-wise combinations of these models improving overall performance.

Feature selection analyses revealed that each of the four metrics employed (the phi-coefficient, MultiSURF, MDI, or MDA) depict noticeable differences in feature ranking and total number of essential features.

Area under the curve (AUC) measures were used to identify essential features from GBM output. As expected, AUC scores rose sharply for each additional feature added early in feature ranking, achieved a peak, and then plateaued or oscillated around the maximum on adding the remaining low-ranked features.  

The number of features needed to reach the best performance was five for RA-vs-OA, ten for RA-vs-CTD, 16 for OA-vs-CTD, five for RA-vs-Rest, 11 for OA-vs-Rest, and 15 for CTD-vs-Rest, respectively.

Rothe et al. (2023)

Final AUC scores were below 0.7, which is acceptable for general discrimination. These scores were consistent with computed cross-validated ML performance scores, highlighting their reliability.

Conclusions

In the present study, researchers analyzed FOI data from 609 clinically diagnosed rheumatic disease patients to evaluate using FOI image features as a fast and accurate diagnostic tool for differentiating between OA, RA, and CTD. They trained and tested ML models for feature identification ranking, and to identify the most minimal set of essential features that can accurately and consistently diagnose the specific disease.

Their results identify between five and 16 essential features sufficient to achieve near-maximum diagnostic performance. This highlights how only a reduced subset of specified features suffices to diagnose OA, RA, or CTD.

The most important are features P, M, Y, C, I, and B, representing the proximal interphalangeal joints (P), the metacarpophalangeal joints (M), the muscle-tendon junction of the wrist (Y), the intercarpal joints (C), an inhomogeneous signal in the nail bed (I), and broad, pronounced signals in the area of the dorsal tendons (B). They are relevant to five of the six problems. Moreover, V and F, representing superficial venous structures (V) and punctual sharp signals (F), are found five times. Interestingly, feature D, which is the distal interphalangeal joint, is not used at all in the machine algorithms, even though in the literature, increasing signal intensity at the distal joints is used to diagnose OA.

Rothe et al. (2023)

In summary, FOI feature reading has been elucidated as a fast and accurate tool for identifying RA, OA, and CTD. It shows potential as a helpful tool for health professionals, especially in early arthritis clinics where these diseases are often encountered and sometimes misidentified.

The information gained by the calculations about which features to use in which phase for which problem and the feature-specific improvement of diagnostic performance provides helpful insight for the differential diagnostic process for the presented diseases.

Rothe et al. (2023)