準確地預估子宮內之胎兒大小是當代產科醫學重要的發展，被預估的胎兒體重，也是幫助臨床醫師評估產婦最佳生產方式的重要指標。以高解析度的醫用超音波量測胎兒超音波估測指標數值，再應用類神經網路的非線性模擬分析法可得到最理想的新生兒體重預測值。本研究的目的是發展以統計分組為基礎之倒傳遞類神經網路模式，來改善應用胎兒超音波參數預估新生兒體重的準確性。本研究方法，首先收集2,107位單胞胎新生兒其產前三天內超音波估測指標數值 (頂間最大橫徑、枕額直徑、腹圍、頭圍、股骨長度等) 及出生之真實體重，以上參數經由(1)檢定資料分佈形態；(2)項目分析出最具有鑑別力的參數；(3)逐步迴歸法選取與估測體重最具貢獻度的超音波參數，作為類神經網路模式的輸入參數；(4) 胎兒體重分組之分析。以胎兒腹圍大小(Group I, AC ＜＝29.8cm；Group II, 29.8cm＜AC＜35.6cm；Group III, AC ＞＝ 35.6cm) 作為網路分組之架構，將個案資料隨機分配為二組，其中1,411位作為建立及訓練倒傳遞類神經網路(比例共軛梯度演算法)學習模組，另外696位則作為驗證模組。預測準確度的評估是採用Friedman test 統計分析法。本研究結果以胎兒腹圍分組為基礎之倒傳遞網路輸入參數為AC、BPD、SEX、FL、OFD、GA、FP(胎位)，驗證胎兒體重預估的準確度為 (n = 696，MAPE = 5.52 ± 4.35％，MAE = 160.55 ± 119.06g) 與台灣 (1) Hsieh formula 1B迴歸分析(n = 696，MAPE = 6.12 ± 4.76％，MAE = 177.77 ± 126.56g, p＜0.05)；(2) Hsieh formula 2B迴歸分析(n = 696，MAPE = 6.59 ± 6.80％，MAE = 180.33 ± 128.75g, p＜0.05 )及美國(3) Hadlock 迴歸分析(n = 696，MAPE = 7.68 ± 5.57％，MAE = 230.22 ± 175.98g, p＜0.05)比較後，均有顯著的改善。驗證結果其準確度亦高於無分組訓練的Chuang 類神經網路模擬分析法（n = 362，MAPE = 6.15 ± 4.99％，MAE = 179.91 ± 148.99 g）。
　　本研究的重要性在於應用統計方法將數值範圍大、且變異性高的參數進行合理的分組，以改善各組參數的異質性。故以分組訓練類神經網路來預估胎兒體重的準確度，皆高於上述文獻方法之結果。研究貢獻提供生產方式最安全的選擇可有效的降低嬰兒與產婦的死亡率。

　Accurate estimation of fetal weight is very important in modern Obstetrics. It helps the clinicians to decide which mode is the best for babies to be delivered. With real-time high-resolution ultrasound, a relatively acceptable estimated fetal weight can be obtained by regression methods, yet it remains to be improved. Recent studies showed when artificial neural network (ANN) model is applied, the accuracy of estimation may be much more improved. This research study is in an attempt to develop a group-based ANN model to improve the accuracy of fetal weight estimation through sonographic parameters.
Totally, 2,107 consecutive singleton fetuses were examined by ultrasonography within 3 days before delivery, and the measurements include biparietal diameter (BPD), occipito-frontal diameter (OFD), abdominal circumference (AC), head circumference (HC), femur length (FL), and actual fetal weight. The analyses were initially undertaken as below: (1) test of normality and variation, (2) item analysis, (3) stepwise regression analysis, and (4) fetal weights classification. The data were further stratified into three groups by AC: Group I, AC ＜＝29.8cm; Group II, 29.8 cm＜AC＜35.6 cm; and Group III, AC＞＝35.6 cm. The total cases were further divided into training and validation groups. For training group, 1,411 fetuses were randomly assigned by training with the back-propagation ANN (trainscg) model on three classified groups. For validation group, 696 fetuses were used to validate the ANN model developed by training group. The difference of the mean absolute percent and mean absolute error of estimated weights relative to actual fetal weights between the ANN model and the regression formula was compared by Friedman test(p＜0.05).
After stepwise regression analysis, seven variables (AC, BPD, SEX, FL, OFD, GA, FP (fetal position)) are used to train the group-based ANN model. The accuracy of fetal weight estimation by this ANN modeling in the validation group (n = 696，MAPE = 5.52 ± 4.35％，MAE = 160.55 ± 119.06 g) is significantly better than that by the Taiwanese conventional regression analysis with Hsieh’s formula 1B (n = 696，MAPE = 6.12 ± 4.76％，MAE = 177.77 ± 126.56 g, p < 0.05), or that by Hsieh’s formula 2B (n = 696，MAPE = 6.59 ± 6.80％，MAE = 180.33 ± 128.75 g, p <0.05). Moreover, the accuracy of fetal weight estimation by this ANN modeling in the validation group is also significantly better than that by the American conventional regression analysis with Hadlock’s formula (n = 696，MAPE = 7.68 ± 5.57％，MAE = 230.22 ± 175.98 g, p < 0.05). Finally, the results from the group-based ANN model are also better than that with the unclassified ANN model, previously published by Chuang et al.（n = 362，MAPE = 6.15 ± 4.99％，MAE = 179.91 ± 148.99 g）.
　In conclusion, the importance of this study is to consider and control the heterogeneity among the high variability and broad ranged parameters by statistics, and to choose the best parameter as a reasonable classified group to improve the estimation of fetal weight. This study has proved the accuracy of fetal weight estimation by the group-based ANN model is better than those of previous models. The results of this study may contribute to the best choice of how to deliver a baby safely and thus lower down the maternal-fetal morbidity and mortality.

【1】Jimenez JM, Tyson JE, Reisch JS:Clinical measures of gestational age in normal pregnancies. Obstet Gynecol 1983；61:438
【2】Hadlock FP, Deter RL, Harrist RB, Park SK. Fetal biparietal diameter:a critical re-evaluation of the relation to menstrual age using real-time ultrasound. J Ultrasound Med 1982;1:97-104.
【3】Hadlock FP, Deter RL, Harrist RB, Park SK. Fetal abdominal circumference as a predictor of menstrual age. AJR 1982;139:367-70.
【4】Warda AH, Deter RL, Rossavik IK, Carpenter RJ, Hadlock FP, Fetal femur length: a critical reevaluation of the relationship to menstrual age. Obstet Gynecol 1985;66:69-75.
【5】Hadlock FP, Harrist RB, Deter RL, Park SK. Fetal femur length as a predictor of menstrual age: sonographically measured. AJR 1982;138: 875-8.
【6】Hadlock FP, Deter RL, Harrist RB, Park SK. Estimating fetal age: computer-assisted analysis of multiple fetal growth parameters. Radiology 1984;152:497-501.
【7】Hadlock FP. Sonographic estimation of fetal age and weight. Radiol Clin North Am 1990;28:39-50.
【8】Campbell S.An improved method of fetal cephalometry. J Obstet Gynecol Br Commonw. 1968;75:568-576.
【9】Hadlock FP, Deter RL, Carpenter RJ, et al. Estimating fetal age: Effect of head shape on BPD. AJR 1981;137:83-85.
【10】Campbell S, Thoms A. Ultrasound measurement of the fetal head to abdomen circumference ratio in the assessment of growth retardation. Br J Obstet Gynaecol 1977;84:165-174.
【11】Campbell S, Wilkin D. Ultrasonic measurement of the fetal abdomen circumference in the estimation of fetal weight. Br J Obstet Gynaecol 1975; 82: 689–697.
【12】Warsof SL, Gohari P, Berkowitz RL, Hobbins JC. The estimation of fetal weight by computer assisted analysis. Am J Obstet Gynecol 1977;128: 881–892.
【13】Hadlock FP, Harrist RB, Carpenter RJ, Deter RL, Park SK. Sonographic estimation of fetal weight. The value of femur length in addition to head and abdomen measurements. Radiology 1984; 150: 535–540.
【14】Sabbagha RE, Minogue J, Tamura RK, Hungerford SA. Estimation of birthweight by use of ultrasonographic formulas targeted to large-, appropriate-, small-for-gestational-age fetuses. Am J Obstet Gynecol 1989; 160: 854–860.
【15】Hsieh FJ, Chang FM, Huang HC, Lu CC, Ko TM, Chen HY. Computer-assisted analysis for prediction of fetal weight by ultrasound: comparison of biparietal diameter, abdominal circumference and femur length. J Formosan Med Assoc 1987; 86: 957–964.
【16】Hadlock FP, Harrist RB, Sharman RS, Deter RL, Park SK. Estimation of fetal weight with the use of head, body and femur measurements – a prospective study. Am J Obstet Gynecol 1985; 151: 333–337.
【17】Shepard MJ, Richards VA, Berkowitz RL, Warsof SL, Hobbins JC. An evaluation of two equations for predicting fetal weight by ultrasound. Am J Obstet Gynecol 1982; 142: 47–54.
【18】Farmer RM, Medearis AL, Hirata GI, Platt LD. The use of a neural network for the ultrasonographic estimation of fetal weight in the macrosomic fetus.Am J Obstet Gynecol 1992;166:1467-1472.
【19】Chuang L, Hwang JY, Chang CH, Yu CH , Chang FM. Ultrasound estimation of fetal weight with the use of computerized artificial neural network model Ultrasound in Medicine and Biology 2002;28:991-996.