从24-2视野预测晚期青光眼眼的Humphrey 10-2视野
本研究用175眼训练集和44眼测试集,从同一眼HFA 24-2最内16个测试点的总偏差值预测HFA 10-2的68个测试点总偏差值。支持向量回归(SVR)预测误差最小(4.0±1.5 dB),约为重测变异性(2.1±1.0 dB)的两倍,误差范围约25%。
要点
- 晚期青光眼(HFA 24-2平均偏差≤-20 dB)可用SVR从24-2预测10-2总偏差值。
- SVR预测误差4.0±1.5 dB,小于双线性插值的5.0±1.7 dB。
- 预测误差范围约25%,约为10-2重测变异性2.1±1.0 dB的两倍。
核心结论
- 对同一晚期青光眼眼的HFA 24-2结果应用SVR,可预测HFA 10-2测试点的总偏差值,误差范围约25%。
研究设计与样本
- 训练集为175眼(175例)开放型晚期青光眼(HFA 24-2平均偏差≤-20 dB),测试集为44眼(44例)同病种眼。
- 用训练集数据,从同一眼HFA 24-2最内16个测试点的总偏差值预测HFA 10-2的68个测试点总偏差值,方法包括图像处理和多种机器学习方法,以双线性插值(IP)为对照标准。
- 将各方法应用于测试集计算绝对预测误差(PredError)。
主要发现
- 测试集HFA 10-2结果的重测变异性均值(SD)为2.1±1.0 dB。
- IP法的PredError为5.0±1.7 dB。
- 支持向量回归(SVR)的PredError最小,为4.0±1.5 dB。
- SVR从同一眼HFA 24-2结果预测晚期青光眼保留「中央岛」的HFA 10-2测试点视网膜敏感度,误差范围约25%。
- SVR预测误差范围约为HFA 10-2重测变异性的两倍。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
AIMS: To predict Humphrey Field Analyzer Central 10-2 Swedish Interactive Threshold Algorithm-Standard test (HFA 10-2) results (Carl Zeiss Meditec, San Leandro, CA) from HFA 24-2 results of the same eyes with advanced glaucoma. METHODS: Training and testing HFA 24-2 and 10-2 data sets, respectively, consisted of 175 eyes (175 patients) and 44 eyes (44 patients) with open advanced glaucoma (mean deviation of HFA 24-2 ≤-20 dB). Using the training data set, the 68 total deviation (TD) values of the HFA 10-2 test points were predicted from those of the innermost 16 HFA 24-2 test points in the same eye, using image processing or various machine learning methods including bilinear interpolation (IP) as a standard for comparison. The absolute prediction error (PredError) was calculated by applying each method to the testing data set. RESULTS: The mean (SD) test-retest variability of the HFA 10-2 results in the testing data set was 2.1±1.0 dB, while the IP method yielded a PredError of 5.0±1.7 dB. Among the methods tested, support vector regression (SVR) provided a smallest PredError (4.0±1.5 dB). SVR predicted retinal sensitivity at HFA 10-2 test points in the preserved 'central isle' of advanced glaucoma from HFA 24-2 results of the same eye within an error range of about 25%, while error range was approximately twice of the test-retest variability. CONCLUSION: Applying SVR to HFA 24-2 results allowed us to predict TD values at HFA 10-2 test points of the same eye with advanced glaucoma with an error range of about 25%.
引用
- Sugisaki K, Asaoka R, Inoue T, et al. Predicting Humphrey 10-2 visual field from 24-2 visual field in eyes with advanced glaucoma. Br J Ophthalmol. 2019. PMID: 31481390. doi:10.1136/bjophthalmol-2019-314170
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。