Table 1.
Current applications of deep learning in rheumatology*
| Problem (source ref.) | Data type | Model | Implications |
|---|---|---|---|
| Identifying GCA features from temporal artery biopsy reports (13) | Text | Transformer | Accurate auditing of temporal artery biopsy reports can be performed using deep learning; however, this performance dropped when tested across centers. |
| Classifying HEp‐2 cells based on ANA IIF patterns (29) | Images | CNN | Automated ANA classification based on HEp‐2 cells is approaching expert human performance. |
| OESS from synovial ultrasound (44) | Images | CNN | Deep learning can identify synovitis on ultrasound with a high degree of accuracy. |
| SHS scoring using hand and foot radiographs (51) | Images | CNN | Radiographic scoring for RA is improving but still requires work for clinical implementation. |
| Predicting progression (any increase in K/L score) of knee OA based on baseline knee radiographs plus other clinical features (58) | Images | CNN | Radiographic progression in knee OA can be predicted with a combination of clinical features and baseline radiography using deep learning; however, there are unmeasured factors missing in these models. |
| Identifying halo sign on temporal artery ultrasound images (68) | Images | CNN | Deep learning has significant potential for automated identification of the halo sign; however, ensuring standardized image acquisition is a major barrier to implementation. |
| Predicting future RA disease activity (controlled versus uncontrolled) using clinical data from previous encounters (19) | EHRs | RNN | Deep learning can predict future disease activity from past disease activity and baseline factors; however, performance significantly dropped when the model was tested at a second center, suggesting that there is substantial heterogeneity between centers that must be accounted for in future models. |
*
GCA = giant cell arteritis; ANA = antinuclear antibody; IIF = immunofluorescence; CNN = convolutional neural network; OESS = EULAR Outcome Measures in Rheumatology synovitis scoring; SHS = modified Sharp/van der Heijde; RA = rheumatoid arthritis; K/L = Kellgren/Lawrence; OA = osteoarthritis; EHRs = electronic health records; RNN = recurrent neural network.