| 摘要: | 骨質疏鬆是一種由於骨質流失，造成骨密度下降的疾病。在歐盟，骨質疏鬆導致骨折而造成的經濟損失，估計高達每年60億歐元。台灣在2006年的統計，骨質疏鬆於50歲以上女性的盛行率為11.35%，遠低於約同時期歐盟的22.1%，顯示台灣仍有大量潛在的骨質疏鬆症患者。骨質疏鬆的診斷標準乃運用Dual emission x-ray absorption （DEXA） 造影，唯根據健保規範，篩檢性質之檢查均不予給付，導致潛在的骨質疏鬆病患失去早期診斷、早期治療的機會。
儘管骨質疏鬆的臨床診斷由影像為主體，但骨質疏鬆除影像的證據外，由病歷紀錄中也可以發現病人是否有骨質疏鬆的風險因子。因此，我們認為合併病歷資料，應當可以提供更多資訊，以增加骨質疏鬆的診斷率。因此，本研究將同時運X光影像和病歷紀錄，再合併兩者結果，嘗試建立準確的骨質疏鬆預測工具。
本計畫由衛生福利部立雙和醫院收案，收集之資料包含病患之腰部X光影像、DEXA檢查之檢查日期及其結果（T-score），以及由DEXA檢查日期往前半年之病歷資料，包含出生日期、性別、診斷碼、用藥品項及檢驗值。
本計畫使用Google Colab Pro，以Python語言編寫； X光影像使用CheXNet架構訓練，病歷部分則傳統機器學習的羅吉斯迴歸Logistic Regression訓練，最後再合併兩個架構，判斷樣本是否有骨質疏鬆。
研究結果整體而言，（一）二元分類得到的正確率優於三分法；（二）較大的資料集可以增加三分法的正確率，但對二分法沒有幫助；（三）合併模型正確率優於影像模型，正確率最高可高達74%。然而本研究得到之正確率不如過去其他研究，建議應合併採取局部影像作為輸入，或進一步推算各部位骨質密度。此外，使用病歷紀錄預測骨質密度的表現優於原先預期，可見病歷紀錄本身亦為具有潛力的預測工具，極具研究價值，可以進一步發展。合併運用多種資訊是當下深度學習的重要趨勢，本研究在此基礎上得到相當理想的初步成果，希望未來能開發出更準確的篩檢工具，檢查出更多骨質疏鬆病患，早期介入治療，以減少未來骨折發生的可能性與骨折所帶來的醫療與社會經濟負擔。
Osteoporosis is a disease characterized by decreased bone density. The economic burden costed by bone fractures related to osteoporosis was estimated 6 billions euros per year in European Union. In Taiwan, 2006, the prevalence rate of osteoporosis in women over 50 years old was 11.35%, while the rate in Europe was 22.1%. The different suggested that the disease is highly possibly being under diagnosed in Taiwan. The golden standard of diagnosis of osteoporosis is Dual emission x-ray absorption （DXA）, yet the examination for screening purpose is not covered by the national health insurance in Taiwan. As a result, asymptomatic patients might miss the opportunity of early diagnosis and intervention.
Despite osteoporosis is diagnosed by X-ray, risk factors of osteoporosis can also be found in electronic health records （EHR）. Hence, we believe combining EHR to X-ray image analysis, should provide more information, in order to get better prediction. Therefore, we purpose to establish a model combining image and EHR, to predict osteoporosis.
The research was conducted in Taipei Medical University Shuang-Ho Hospital. The data collected including the patient’s lumbar spine （L-spine） X-ray image, examination date and result of DXA examination, and electronic health record （EHR） 6 months prior to the exam. The EHR collected including date of birth, gender, diagnosis code, medication and laboratory examination.
The research was carried out with Google Colab. The CNN models used for images is CheXNet. The models used for EHR is logistic regression. The result from image was combined in the end.
The result of the research showed: 1） Binary classification has better accuracy, comparing to 3-divide classification; 2） larger dataset provides better accuracy in 3-divide classification, but not to binary classification; 3） combine model has better accuracy then individual models, best accuracy is around 74%.
However, the accuracy of this research is not better than previous literatures, indicating that region of interest segmentation is still important in the scenario. Interestingly, the performance of EHR model is better than we expected, suggesting that EHR itself can be a potential tool to predict osteoporosis that might worth further research.
This research combines two different of modality to predict osteoporosis. As multimodal model is popular in recent researches, we consider this work meets the current trend. With current results, we expect to develop a better bone density predicting tool with only EHR and plain L-spine X-ray, in order to provide easy osteoporosis screening, which may provide early intervention thus reducing possible medical and social burden. |
