目的：评估新一代基于人工智能(artificial intelligence,AI)的人工晶状体(intraocular lens,IOL)计算公式的准确性。方法：本研究为回顾性研究，纳入因白内障行晶状体超声乳化联合IOL植入术的262例患者262眼。在术前，通过IOLMaster700获取角膜曲率、角膜白到白、中央角膜厚度、前房深度、晶状体厚度以及眼轴长度。使用第三代公式(SRK/T、Holladay 1和Hoffer Q)、Barrett UniversalⅡ(BUⅡ)、新一代AI公式(Kane、Pearl-DGS、Hill-RBF 3.0、Hoffer QST和Jin-AI)对术后屈光状态进行计算，并与术后实际的屈光状态进行比较。在将预测误差(prediction error,PE)归零后，分析了各公式的标准差(standard deviation,SD)、绝对误差均值(mean absolute error,MAE)、绝对误差中位数(median absolute error,MedAE)以及PE在±0.25、±0.50、±1.00、±2.00 D范围内的百分比。结果：基于AI的IOL屈光力计算公式的SD、MAE和MedAE的范围分别为0.37 D(Kane和Jin-AI)至0.39 D(Hoffer QST)、0.28 D(Hill-RBF 3.0和Jin-AI)至0.31 D(Hoffer QST)以及0.21 D(Hill-RBF3.0和Jin-AI)至0.24 D(HofferQST)；均低于第三代公式(SD：0.43 D~0.45 D；MAE：0.34 D；MedAE：0.25 D~0.28 D)。在所有公式中，Jin-AI公式预测误差在±0.50 D的比例最高，为84.73%，Kane(84.35%)和BUⅡ(83.97%)公式次之。结论：在IOL屈光力预测上，与传统第三代公式相比，新一代基于AI的公式表现出更高的准确性，可以使更多的患者在术后获得预期的屈光状态。

Objective: To evaluate the accuracy of new generation artificial intelligence (AI)-based intraocular lens (IOL)power calculation formulas. Methods: This retrospective study included a total of 262 eyes from 262 patients with cataract who underwent uneventful phacoemulsification combined with IOL implantation. Keratometry, corneal white-to-white, central corneal thickness, anterior chamber depth, lens thickness, and axial length were measured by the IOL Master 700 before surgery. Predicted refractive errors were calculated by the third-generation formulas (SRK/T, Holladay 1, and Hoffer Q), Barrett UniversalⅡ (BUⅡ), and the newer-generation AI formulas (Kane, Pearl-DGS, Hill-RBF 3.0, Hoffer QST, and Jin-AI), and were compared with the actual postoperative refractive value. After adjusting the prediction error (PE) to zero, the standard deviation (SD), mean absolute error (MAE), median absolute error (MedAE), and the percentage of a PE within the range of ±0.25 diopter (D), ±0.50 D, ±1.00 D, and ±2.00 D were analyzed. Results: The SD, MAE, and MedAE of the AI-based formulas ranged from 0.37 D (Kane and Jin-AI) to 0.39 D (Hoffer QST), 0.28 D (Hill-RBF 3.0 and Jin-AI) to 0.31 D (Hoffer QST), and 0.21 D (Hill-RBF 3.0 and Jin-AI) to 0.24 D (Hoffer QST), respectively. These values were all lower than those of the third-generation formula (SD: 0.43 D to 0.45 D; MAE: 0.34 D; MedAE: 0.25 D to 0.28 D). Among all the formulas, the Jin-AI formula had the highest proportion of a PE within ±0.50 D (84.73%), followed by Kane (84.35%) and BUⅡ (83.97%) formulas. Conclusion: The new AI-based IOL formulas show higher accuracy compared with the traditional third-generation ones in predicting IOL power. thereby enabling more patients to achieve the expected refractive outcomes after surgery

目的：比较六种新一代人工晶状体(intraocular lens,IOL)屈光力计算公式[Barrett Universal Ⅱ(BUⅡ)、Emmetropia Verifying Optical(EVO)、Hill-Radial Basis Function (Hill-RBF)、Kane、Ladas Super Formula(LSF)、T2]和传统公式(Haigis、Hoffer Q、Holladay 1、SRK/T)的准确性。方法：纳入2022年1—6月于温州医科大学附属眼视光医院接受白内障手术患者。收集患者的年龄、性别、眼轴(axial length,AL)、平均角膜曲率(mean keratometry,Kmean)、前房深度、IOL常数和屈光力，术后医学验光结果。对上述10种公式进行准确性分析，包括平均预测误差(mean prediction error,ME)及其标准差、平均绝对预测误差(mean absolute prediction error,MAE)、绝对预测误差中位数(median absolute prediction error,MedAE)、绝对预测误差最大值(maximum absolute prediction error,MaxAE)、预测误差落在±0.25、±0.5、±0.75、±1.00 D范围内的百分比(%±0.25 D、%±0.50 D、%±0.75 D、%±1.00 D)。结果：共纳入506例(506眼)。Kane的MAE最低(0.411)。Hill-RBF的%±0.25 D最高(40.91%)，EVO的%±0.50 D或%±0.75 D最高(分别为69.37%、86.17%)，BUⅡ和Hill-RBF的%±1.00 D最高(均为94.07%)。总体上各种公式间，MAE、%±0.50 D、%±0.75 D、%±1.00 D比较差异存在统计学意义(P<0.05)，但两两比较仅发现%±0.75 D中，EVO(86.17%)、Hill-RBF(85.97%)、Kane(85.57%)与HofferQ(81.42%)比较差异存在统计学意义(均P<0.05)。AL亚组中，长AL组的EVO(0.390)、Hill-RBF(0.388)、T2(0.423)、Kane(0.393)四种公式的MAE与Hoffer Q(0.681)、Holladay 1(0.654)比较差异存在统计学意义(均P<0.05)，EVO(74.47%)的%±0.50 D与Hoffer Q(46.81%)比较差异存在统计学意义(P=0.017)。结论：新一代IOL屈光力计算公式在IOL屈光力计算上均具有较好的准确性，但对于不同的眼轴长度与角膜曲率值的眼球，需要选择适合的计算公式，以进一步提高预测准确性。

Objective: This study aimed to compare the accuracy of six new generation intraocular lenses (IOL) refractive power calculation formulas (Barrett Universal Ⅱ [BU Ⅱ ], Emmetropia Verifying Optical [EVO], Hill-Radial Basis Function [Hill-RBF], Kane, Ladas Super Formula [LSF], T2) and traditional formulas (Haigis, Hoffer Q, Holladay 1, SRK/ T). Methods: The patients who received cataract surgery in the Eye Hospital of Wenzhou Medical University from January 2022 to June 2022 were included in this study. Age, gender, axial length (AL), mean keratometry, anterior chamber depth, IOL constant and power, and postoperative refraction results were collected. The prediction accuracy of these ten IOL power calculation formulas was analyzed, including mean prediction error (ME) and its standard deviation, mean absolute prediction error (MAE), median absolute prediction error (MedAE), maximum absolute prediction error (MaxAE), the percentage of eyes of PE within the range of ±0.25 D, ±0.5 D, ±0.75 D, ±1.0 D (%±0.25 D,%±0.50 D, %±0.75 D, %±1.00 D). Results: 506 eyes of 506 patients were included. Kane has the lowest MAE (0.411).%±0.25 D of Hill-RBF was the highest (40.91%), %±0.50 D or %±0.75 D of EVO was the highest (69.37%, 86.17%), and %±1.00 D of BU Ⅱ and Hill-RBF was the highest (94.07%). There are significant differences in MAE, %±0.50 D, %±0.75 D, and %±1.00 D among all formulas (P<0.05). Still, pairwise comparison only found differences between EVO (86.17%), Hill-RBF (85.97%), Kane (85.57%), and Hoffer Q (81.42%) in %±0.75 D (all P<0.05). In AL subgroup, the MAE of EVO (0.390), Hill-RBF (0.388), T2 (0.423) and Kane (0.393) in long AL group was different from that of Hoffer Q (0.681) and Holladay 1 (0.654) (all P<0.05), the difference of %±0.50D of EVO (74.47%) compared with Hoffer Q (46.81%) (P=0.017). Conclusion: The new generation of IOL power calculation formulas have good accuracy in IOL power prediction, but for eyes with different axial lengths and keratometry, it is necessary to optimize the selection of formulas to improve the prediction accuracy further.