應用類神經網路建構女性骨質密度預測模式

A Prediction Model for Women’s Bone Mineral Density using Artificial Neural Networks

賴瑞芬(Jui-Fen Lai)

指導教授 ： 劉建財

臺北醫學大學/醫學院/醫學資訊研究所/碩士(2010年)

摘要

台灣65歲以上老年人口比例，自1998年始由8.2％攀升至2008年提升為10.4％，2008年國人兩性之平均壽命分別為男性75.6歲，女性81.9歲，屬於一高齡化社會，中老年人口比例快速增加，中老年族群之「健康促進」及「疾病預防」中骨質疏鬆症為重大健康議題。

世界衛生組織（WHO）於1998年訂10月20日為「世界骨質疏鬆症日」，期望透過早期診斷與預防，進而減少中老年人骨質疏鬆疾病之發生，以控制或降低因骨質疏鬆症所帶來的危害及其他負面影響。骨質流失是隨著年齡的增長日益增加，且為無感性且緩慢發生，鮮少有症狀，因此民眾往往不會發覺骨質流失問題，待發生骨折則為時已晚。伴隨骨折後而引發的相關併發症與後遺症外亦會耗用極大的家族人力及社會資源，無疑帶給醫療照護相當大的社會成本與負擔。

因健保局對於骨質密度篩檢的限制，大部份的民眾仍先自費1200元做第一次篩檢，對於一般大眾而言1200元是一筆不小的開銷，會預算旳限制而失去預防疾病的先機；為此我們期許理想的骨質密度篩檢儀器應具有精確度、敏感度高、操作方法簡便、機動性高等特性。

類神經網路（Artificial Neural Network, ANNs），使用大量簡單的相連人工神經元來模仿生物神經網路透過訓練的方式，讓類神經網路反反覆覆的學習，了解資料的特徵與型態，直到每個輸入能對應到所需的輸出。目前類神經網已應用於醫學重要的領域。

The Osteoporosis Self-Assessment Tool for Asians （簡稱OSTA ），年齡及體重較輕是兩大最主要的危險因素，是以年齡（歲）減去體重（kg）後乘上0.2取其整數，若OSTA 值高或等於-1屬「低風險」、若介於-4至-1之間屬「中度風險」，而小於-4屬「高風險」（即體重較年齡小20以上）。

本研究中使用類神經網路模型（一）卅OSTA模型（二）：年紀小於等於40歲，無論是否已停經；類神經網路模型（二）卅OSTA模型（一）：年紀小於等於50歲，已停經。在類神經網路模型（一）卅OSTA模型（二）中，以類神經網路模型（一） 正確率75.1％、敏感度75.2％、特異度75％、AUC 0.75 ，優於OSTA模型（二）。類神經網路模型（二）卅OSTA模型（一）中以，以類神經網路模型（二）正確率68.82％、敏感度51.28％、特異度81.48％、AUC 0.685，優於OSTA模型（一）。特異度方面是以類神經網路模型（二）為最優。整體而言無論是陽性預測值、陰性預測值，類神經網路模型（一）都是最好的選擇。總結以上各預測模型的效能而言，類神經網路模型較OSTA模型更適用於篩檢骨質密度。

關鍵字

骨質疏鬆 ； 骨質密度類神經網路

並列摘要

Taiwan over 65 years the proportion of elderly population, since the beginning of 1993, up from 7.1％ to 10.2％, and 96 years, the average life expectancy of people sexes, male 75.5 years, women 81.7 years old, belonging to a aging society, the rapid increase in the propor-tion of elderly population , middle-aged population of the "health promotion" and "disease prevention" of osteoporosis as a major health issue.

World Health Organization in 1998 set October 20 as "World Osteoporosis Day", is hoped take to early diagnosis and prevention and reducing osteoporosis in the elderly.Order to con-trol or lower due to osteoporosis’s negative effects. Bone loss is increasing with age, non-
sentimental , happens slowly and rarely have symptoms.So people often do not notice the bone loss until the fracture was already late. Complication and the sequela will also consume enormous manpower of family and social resource after the bone fracture. It will increase so-cial cost and the burden in the medical service.

National Health Insurance Bureau due to bone density screening for the restrictions, the ma-jority of people are still NT$1,200 at their own expense to do before the first screening, 1,200 yuan for the general public in terms of the cost a small fortune, will lose the budget latter's restrictions on disease prevention opportunities; this end, we hopes the ideal instrument should have a bone density screening accuracy, high sensitivity, the operation method is simple, high mobility characteristics.

By employing an artificial neural network （ANN） approach, the present study seeks to train the ANN repetitively to acquaint it with the characteristics and modalities of the data. The result is an output extremely approximating the precise value and thus predictive for BMD.

The Osteoporosis Society of the ROC has recognized and recommended the Osteoporosis Self-Assessment Tool for Asians （OSTA） as a simple self-assessment method in post-menopausal women. OSTA uses weight and age to predict the osteoporosis risk in a decade. More specifically, weight （kg） is subtracted from age （yrs）, the difference is multiplied by 0.2 and only the integer in the product is reserved. If the OSTA value is above or equal to -1, the osteoporosis risk is estimated to be 3％ and is judged to be low; if the value is be-tween -4 and -1, the estimated risk is 15％ and is judged to be intermediate; if the value is below -4 （which means the weight is smaller than the age and the difference is more than 20）, the estimated risk is above 60％ and is judged to be high.

This study used by ANN（1）卅OSTA （2） model: postmenopausal or not females whose ages were 40 or the aboves. ANN（2）卅OSTA （1） model: postmenopausal females whose ages were 50 or the aboves. InANN（1）卅OSTA （2） model, ANN（1） model is 75.1％ accuracy, 75.2％ sensitivity, 75％ specificity, AUC 0.75, better than OSTA （2）. InANN（2）卅OSTA （1） model, ANN（2） model is 68.82 accuracy, 51.28％ sensitivity, 81.48％ specificity, AUC0.685, better than OSTA （1）. So ANN model always is best chances to screening bone mineral density.

並列關鍵字

Osteoporosis ； Bone Mineral Density ； Artificial Neural Network

參考文獻

林欣欣, 陳品玲, 曹麗英, 楊榮森, 蕭美君. 憂鬱症婦女骨密度及相關因素初探. 長庚護理 2008;19:162-71.

衛生署全民健保爭議審議委員會. 骨質疏鬆症:檢查及治療爭議. 台灣醫學 2002; 6:737-43.

李瑋君, 韓岳津, 吳英黛, 曹昭懿. 骨質疏鬆相關骨折住院病患之健保申報分析. 台灣醫學 2007;11:22-7.

李燕鳴. 停經女性骨質疏鬆症的篩檢及預防. 台灣醫學 2008;12:613-9.

洪晧彰, 莊明哲, 歐弘毅, 蕭淑華, 吳達仁. 骨質疏鬆症治療新進展. 內科學誌 2007;18:313-21.