近數十年來，風濕病與免疫學相關的研究促成了用於類風溼性關節炎的TNF抑制劑療法。然而，目前市面上有數種TNF抑制劑可供選擇，而接受TNF抑制劑療法的類風溼性關節炎病患可能因各種原因（包含療效不彰或其他因素）而選擇換用別種TNF抑制劑。除妊娠等特殊情況外，目前尚無廣泛認可的算法或標準可用於協助於首次使用時選擇可維持最長使用時間的TNF抑制劑。基於此，我們探討了一個較小但預期有助於解決該問題的研究問題：在接受TNF抑制劑療法的類風濕性關節炎病患中，能否利用常規實驗室檢驗數據預測生物標誌物結果，以作為疾病活動度爆發的替代指標。我們從於三間醫院共獲取了396名使用TNF抑制劑的類風濕性關節炎患者的去識別化數據（不含臨床病歷），並以血清中C反應蛋白（CRP）濃度作為預測目標，比較了不同預測模型的性能。為了滿足臨床上對模型的可解釋性需求，我們在對比中包含了temporal fusion transformer (TFT）。研究結果顯示，在給定實驗室檢驗數據、藥物使用紀錄與族群相關資料的情況下，針對「預測未來四周內血清CRP濃度突然增加」的任務，TFT 在準確率（Accuracy）、召回率（Recall）及F1分數方面均優於線性模型與梯度提升決策樹（GBDT），但會需要更多運算資源。相較於多數深度學習模型，TFT的注意力權重（Attention weights）為特徵的相對重要性提供了一定程度的可解釋性。我們結論為TFT兼具優異的預測性能與可解釋性，在生物醫學領域的時間序列預測中擁有相當的潛力。

Developments in rheumatology and immunology over the past several decades have led to the emergence of TNF inhibitors for the treatment of rheumatoid arthritis. However, there are multiple choices of TNF inhibitor available, and it is possible for a patient using a TNF inhibitor to require a switch to a different one later on in the disease course due to a variety of reasons. Except for specialized cases such as pregnancy, there are no widely accepted algorithms or criteria for selecting the TNF inhibitor that would be most likely to remain in use for the longest period of time. With this in mind, we explored a related but narrower question: the practicality of predicting biomarker outcomes, as a surrogate for disease flares, in patients using TNF inhibitors by using routine lab tests. We obtained the deidentified medical records (excluding clinical notes) of 396 rheumatoid arthritis patients on TNF inhibitors from three sites, and compared the performance of different prediction models, using the serum C-reactive protein (CRP) concentration as a surrogate outcome. To deal with the requirement that the model be interpretable, we included the temporal fusion transformer in our comparisons. We found that, for the task of predicting a spike in serum CRP concentrations within the next four weeks given lab and medication data, the temporal fusion transformer was superior to linear models and gradient-boosted decision trees in terms of accuracy, recall, and F1 score, at the cost of higher computational resource requirements. The attention weights of the trained temporal fusion transformer also provide insights into the relative importance of the features used in prediction. We conclude that the temporal fusion transformer shows much potential in time-series predictions in a biomedical setting, providing good performance and explainability.

[1] D. Aletaha and J. S. Smolen, “Diagnosis and Management of Rheumatoid Arthritis: A Review,” JAMA, vol. 320, no. 13, p. 1360, Oct. 2018, doi: 10.1001/jama.2018.13103.
[2] C. Turesson, W. M. O’Fallon, C. S. Crowson, S. E. Gabriel, and E. L. Matteson, “Extra-articular disease manifestations in rheumatoid arthritis: incidence trends and risk factors over 46 years,” Annals of the Rheumatic Diseases, vol. 62, no. 8, pp. 722–727, Aug. 2003, doi: 10.1136/ard.62.8.722.
[3] M. Cutolo, G. D. Kitas, and P. L. C. M. Van Riel, “Burden of disease in treated rheumatoid arthritis patients: Going beyond the joint,” Seminars in Arthritis and Rheumatism, vol. 43, no. 4, pp. 479–488, Feb. 2014, doi: 10.1016/j.semarthrit.2013.08.004.
[4] Q. Shi et al., “Estimating the response and economic burden of rheumatoid arthritis patients treated with biologic disease-modifying antirheumatic drugs in Taiwan using the National Health Insurance Research Database (NHIRD),” PLoS ONE, vol. 13, no. 4, p. e0193489, Apr. 2018, doi: 10.1371/journal.pone.0193489.
[5] “Cigarette Smoking: A Modifiable Environmental Factor in the Pathogenesis of Rheumatoid Arthritis,” JMA J, vol. 8, no. 4, 2025, doi: 10.31662/jmaj.2025-0029.
[6] R. Zhang et al., “RNA-seq and Network Analysis Reveal Unique Chemokine Activity Signatures in the Synovial Tissue of Patients With Rheumatoid Arthritis,” Front. Med., vol. 9, p. 799440, May 2022, doi: 10.3389/fmed.2022.799440.
[7] K. D. Deane and V. M. Holers, “Rheumatoid Arthritis Pathogenesis, Prediction, and Prevention: An Emerging Paradigm Shift,” Arthritis & Rheumatology, vol. 73, no. 2, pp. 181–193, Feb. 2021, doi: 10.1002/art.41417.
[8] J. S. Smolen et al., “EULAR recommendations for the management of rheumatoid arthritis with synthetic and biological disease-modifying antirheumatic drugs: 2022 update,” Ann Rheum Dis, vol. 82, no. 1, pp. 3–18, Jan. 2023, doi: 10.1136/ard-2022-223356.
[9] S. Banerjee, A. Biehl, M. Gadina, S. Hasni, and D. M. Schwartz, “JAK–STAT Signaling as a Target for Inflammatory and Autoimmune Diseases: Current and Future Prospects,” Drugs, vol. 77, no. 5, pp. 521–546, Apr. 2017, doi: 10.1007/s40265-017-0701-9.
[10] Y.-J. Song, S. W. Nam, C. H. Suh, J. Y. Choe, and D. H. Yoo, “Biosimilars in the treatment of rheumatoid arthritis: a pharmacokinetic overview,” Expert Opinion on Drug Metabolism & Toxicology, vol. 19, no. 11, pp. 751–768, Nov. 2023, doi: 10.1080/17425255.2023.2270407.
[11] C. Salliot et al., “Indirect comparisons of the efficacy of biological antirheumatic agents in rheumatoid arthritis in patients with an inadequate response to conventional disease-modifying antirheumatic drugs or to an anti-tumour necrosis factor agent: a meta-analysis,” Annals of the Rheumatic Diseases, vol. 70, no. 2, pp. 266–271, Feb. 2011, doi: 10.1136/ard.2010.132134.
[12] C. Lin et al., “Automatic Prediction of Rheumatoid Arthritis Disease Activity from the Electronic Medical Records,” PLoS ONE, vol. 8, no. 8, p. e69932, Aug. 2013, doi: 10.1371/journal.pone.0069932.
[13] M. Kalweit et al., “Personalized prediction of disease activity in patients with rheumatoid arthritis using an adaptive deep neural network,” PLoS ONE, vol. 16, no. 6, p. e0252289, June 2021, doi: 10.1371/journal.pone.0252289.
[14] D. Cheng et al., “Inferring rheumatoid arthritis disease activity status from the electronic health records across health systems to enable real-world data studies,” Nov. 17, 2025, Rheumatology. doi: 10.1101/2025.11.13.25340003.
[15] M. Humbert‐Droz et al., “Development of a Natural Language Processing System for Extracting Rheumatoid Arthritis Outcomes From Clinical Notes Using the National Rheumatology Informatics System for Effectiveness Registry,” Arthritis Care & Research, vol. 75, no. 3, pp. 608–615, Mar. 2023, doi: 10.1002/acr.24869.
[16] H. Moradmand and L. Ren, “Multistage deep learning methods for automating radiographic sharp score prediction in rheumatoid arthritis,” Sci Rep, vol. 15, no. 1, p. 3391, Jan. 2025, doi: 10.1038/s41598-025-86073-0.
[17] J. Gandrup, D. A. Selby, S. N. Van Der Veer, J. Mcbeth, and W. G. Dixon, “Using patient-reported data from a smartphone app to capture and characterize real-time patient-reported flares in rheumatoid arthritis,” Rheumatology Advances in Practice, vol. 6, no. 1, p. rkac021, Jan. 2022, doi: 10.1093/rap/rkac021.
[18] L. Stradford et al., “Wearable activity tracker study exploring rheumatoid arthritis patients’ disease activity using patient-reported outcome measures, clinical measures, and biometric sensor data (the wear study),” Contemporary Clinical Trials Communications, vol. 38, p. 101272, Apr. 2024, doi: 10.1016/j.conctc.2024.101272.
[19] on behalf of the RETRO study group et al., “Advanced machine learning for predicting individual risk of flares in rheumatoid arthritis patients tapering biologic drugs,” Arthritis Res Ther, vol. 23, no. 1, p. 67, Dec. 2021, doi: 10.1186/s13075-021-02439-5.
[20] M. S. Van Der Leeuw et al., “Using real-world data to dynamically predict flares during tapering of biological DMARDs in rheumatoid arthritis: development, validation, and potential impact of prediction-aided decisions,” Arthritis Res Ther, vol. 24, no. 1, p. 74, Dec. 2022, doi: 10.1186/s13075-022-02751-8.
[21] B. Lim, S. O. Arik, N. Loeff, and T. Pfister, “Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting,” Sept. 27, 2020, arXiv: arXiv:1912.09363. doi: 10.48550/arXiv.1912.09363.