探讨采用基于"人工智能(AI)的骨龄辅助评价系统(上海初云医疗科技有限公司与四川大学华西第二医院合作开发)"(以下简称为AI系统)对完全性生长激素缺乏症(CGHD)患儿诊断及骨龄评价准确性。

选择2014年7月至2019年11月，于四川大学华西第二医院确诊的66例来自四川地区CGHD患儿为研究对象，纳入研究组。选择同期于病例收集医院儿童保健科进行骨龄测定的67例来自四川地区身高达标儿童作为对照，纳入对照组。对每例受试儿进行左手腕关节正位X射线摄片骨龄测定，由2位医师采用《TW₂骨龄评分法中国未成年人南方标准》(以下简称为TW₂CHN)》与《TW₃骨龄评分法标准》(以下简称为TW₃)，盲法评价受试儿TW₂CHN-桡、尺、掌指骨(RUS)与TW₂CHN-腕骨(carpal)、TW₂CHN-20、TW₃-RUS及TW₃-carpal骨龄(以下简称为5种传统骨龄)，以及以同性别、年龄身高达标儿童5种传统骨龄为标准，计算受试儿5种传统骨龄百分位数。同时，采用AI系统分别对每例受试儿采取TW₂CHN与TW₃法，评价其AI-TW₂CHN-RUS、AI-TW₂CHN-carpal、AI-TW₂CHN-20、AI-TW₃-RUS及AI-TW₃-carpal骨龄(以下简称为5种AI骨龄)及其相应百分位数。以上述5种传统骨龄＋5种AI骨龄(以下简称为10种骨龄)相应的P₅₀、P₂₅、P₁₀、P₃值(统称为P_n值)作为诊断CGHD患儿临界值，分别计算其诊断CGHD患儿的敏感度、特异度、约登(Youden)指数、准确率、阳性似然比、阴性似然比、阳性预测值、阴性预测值。采用Kappa值评价2组受试儿5种传统骨龄百分位数与对应的5种AI骨龄百分位数评价结果的一致性，以及2位医师对2组受试儿TW₂CHN-RUS骨龄百分位数评价结果一致性。绘制上述10种骨龄百分位数诊断CGHD患儿的受试者工作特征(ROC)曲线，并计算曲线下面积(AUC)。采用配对t检验，对2组受试儿TW₂CHN骨龄与TW₃骨龄进行比较。本研究遵循的程序符合2013年新修订的《世界医学协会赫尔辛基宣言》要求。2组受试儿年龄、性别构成比等一般临床资料比较，差异均无统计学意义(P>0.05)。

①采用10种骨龄的P_n值分别作为诊断CGHD临界值，对133例受试儿CGHD诊断结果显示，除了TW₃-RUS骨龄中，以骨龄≤P₁₀作为诊断CGHD患儿临界值时的诊断准确率最高(85.0%)，TW₂CHN-RUS、TW₂CHN-carpal、TW₂CHN-20、TW₃-carpal、AI-TW₂CHN-RUS、AI-TW₂CHN-carpal、AI-TW₂CHN-20、AI-TW₃-carpal、AI-TW₃-RUS骨龄中，均为以骨龄≤P₂₅作为临界值时，对CGHD的诊断准确率最高，分别为81.9%、75.2%、88.0%、78.2%、75.2%、73.6%、81.2%、72.9%、78.9%。②一致性检验结果显示，2组受试儿TW₂CHN-RUS与AI-TW₂CHN-RUS、TW₂CHN-carpal与AI-TW₂CHN-carpal、TW₂CHN-20与AI-TW₂CHN-20、TW₃-RUS与AI-TW₃-RUS、TW₃-carpal与AI-TW₃-carpal骨龄百分位数评价结果均为中等一致性，Kappa值分别为0.445、0.578、0.570、0.446、0.430(均为P<0.001)。③对2位医师对2组受试儿TW₂CHN-RUS骨龄百分位数评价结果进行一致性检验显示，其Kappa值为0.790(P<0.001)，一致性较高。④绘制10种骨龄百分位数评价结果诊断CGHD的ROC曲线分析结果显示，TW₂CHN-RUS、TW₂CHN-carpal、TW₂CHN-20、TW₃-carpal、TW₃-RUS、AI-TW₂CHN-RUS、AI-TW₂CHN-carpal、AI-TW₂CHN-20、AI-TW₃-carpal、AI-TW₃-RUS骨龄百分位数诊断CGHD患儿的AUC分别为0.932、0.859、0.915、0.895、0.844、0.823、0.805、0.866、0.860、0.764(均为P<0.001)。⑤133例受试儿的TW₃-RUS、TW₃-carpal、AI-TW₃-RUS、AI-TW₃-carpal骨龄，均分别显著低于TW₂CHN-RUS、TW₂CHN-carpal、AI-TW₂CHN-RUS、AI-TW₂CHN-carpal骨龄，并且差异均有统计学意义(t＝21.746、25.287、16.498、9.290，P<0.001)。

TW₂CHN法、TW₃法对CGHD患儿骨龄评价及诊断均有临床价值，TW₂CHN-RUS骨龄对于CGDH患儿诊断效能高。四川地区儿童TW₃骨龄较TW₂CHN骨龄低，TW₃法可能不完全适用于四川地区儿童骨龄评价。AI系统对于四川地区CGHD患儿骨龄评价结果与传统骨龄评价结果具有中等一致性，可为临床医师评价受试儿骨龄提供帮助。

To explore the accuracy of " Artificial Intelligence (AI) Assisted Bone Age Evaluation System (co-developed by Shanghai Chuyun Medical Technology Company Limited and West China Second Hospital of Sichuan University)" (hereinafter referred to as AI system) for diagnosis and bone age evaluation of children with complete growth hormone deficiency (CGHD).

A total of 66 children with CGHD from Sichuan area diagnosed in West China Second Hospital of Sichuan University from July 2014 to November 2019 were selected as research subjects and included into study group. And 67 children with normal height from Sichuan area who visited Department of Child Healthcare in cases collected hospital during the same period were selected as control and included into control group. Bone age of each child was examined by anteroposterior X-ray radiography of the left wrist joint. Bone ages of artificial TW₂CHN-radial, ulnar, and short bones (RUS) and TW₂CHN-carpal, TW₂CHN-20, TW₃-RUS and TW₃-carpal (hereinafter referred to as 5 traditional bone ages) and their corresponding percentiles in children with the same gender and age were blindly evaluated by two physicians using TW₂ Bone Age Scoring Method Southern Chinese Standard (hereinafter referred to as TW₂CHN) and TW₃ Bone Age Scoring Method Standard (hereinafter referred to as TW₃). Meanwhile, the AI system was used to evaluate bone age and percentile of AI-TW₂CHN-RUS, AI-TW₂CHN-carpal, AI-TW₂CHN-20, AI-TW₃-RUS and AI-TW₃-carpal (hereinafter referred to as 5 AI bone ages) of each subject by TW₂CHN and TW₃ methods respectively. Sensitivity, specificity, Youden index, accuracy, positive likelihood ratio, negative likelihood ratio, positive predictive value and negative predictive value of diagnosing CGHD children were calculated by using P₅₀, P₂₅, P₁₀, and P₃ values (collectively referred to as P_n values) of 5 traditional bone ages and 5 AI bone ages (hereinafter referred to as 10 bone ages) as critical values for diagnosis of CGHD in children. Kappa coefficient was used to evaluate the consistency of 5 traditional bone age percentile assessment results with the corresponding 5 AI bone age percentile assessment results, as well as the consistency of TW₂CHN-RUS bone age percentile assessment results for 133 subjects by two physicians. The receiver operating characteristic (ROC) curves of 10 bone age percentile assessment results in diagnosis of CGHD were plotted and area under the curve (AUC) was calculated. Bone ages of TW₂CHN and TW₃ were compared by paired t test. The procedures followed in this study were in line with the requirements of newly revised World Medical Association Declaration of Helsinki in 2013. There were no significant differences in age, gender composition and other general clinical data between two groups (P>0.05).

① CGHD diagnosis results of two groups by using P_n values of 10 bone ages as critical values for diagnosis of CGHD children showed that the diagnostic accuracy was the highest (85.0%) when bone age ≤P₁₀ was used as the critical value for diagnosis of CGHD children except for TW₃-RUS bone age. As for TW₂CHN-RUS, TW₂CHN-carpal, TW₂CHN-20, TW₃-carpal, AI-TW₂CHN-RUS, AI-TW₂CHN-carpal, AI-TW₂CHN-20, AI-TW₃-carpal, and AI-TW₃-RUS bone ages, when bone age ≤P₂₅ was taken as the critical value, the diagnostic accuracy of CGHD was the highest, which were 81.9%, 75.2%, 88.0%, 78.2%, 75.2%, 73.6%, 81.2%, 72.9%, 78.9%, respectively. ② Consistency test results showed that bone age percentiles of two groups evaluated by TW₂CHN-RUS and AI-TW₂CHN-RUS, TW₂CHN-carpal and AI-TW₂CHN-carpal, TW₂CHN-20 and AI-TW₂CHN-20, TW₃-RUS and AI-TW₃-RUS, TW₃-carpal and AI-TW₃-carpal were moderately consistent, and the Kappa values were 0.445, 0.578, 0.570, 0.446 and 0.430, respectively ( all P<0.001). ③Consistency test of TW₂CHN-RUS bone age percentile evaluation results of two groups by two physicians showed that Kappa value was 0.790 (P<0.001), indicating a high consistency. ④ ROC curve analysis results of 10 bone age percentile for diagnosis of CGHD in two groups showed that AUC of TW₂CHN-RUS, TW₂CHN-carpal, TW₂CHN-20, TW₃-carpal, TW₃-RUS, AI-TW₂CHN-RUS, AI-TW₂CHN-carpal, AI-TW₂CHN-20, AI-TW₃-carpal, AI-TW₃-RUS bone ages for diagnosis of CGHD children were 0.932, 0.859, 0.915, 0.895, 0.844, 0.823, 0.805, 0.866, 0.860, 0.764 (all P<0.001). ⑤Bone ages of TW₃-RUS, TW₃-carpal, AI-TW₃-RUS and AI-TW₃-carpal in 133 subjects were significantly lower than those of TW₂CHN-RUS, TW₂CHN-carpal, AI-TW₂CHN-RUS and AI-TW₂CHN-carpal, respectively, and all differences were statistically significant (t=21.746, 25.287, 16.498, 9.290; P<0.001).

TW₂CHN and TW₃ method have clinical values in evaluation of bone ages and diagnosis of children with CGHD, and TW₂CHN-RUS bone age has the highest diagnostic efficiency for children with CGDH. Bone ages of children in Sichuan area evaluated by TW₃ method are lower than that of TW₂CHN method, and TW₃ method may not be completely applicable to children in Sichuan area. Bone age of CGHD children in Sichuan area evaluated by AI system has moderate consistency with the results of traditional manual bone age evaluation, but it can be used as an auxiliary tool to help clinicians accurately evaluate the bone age of children.