University of Michigan, Ann Arbor, Michigan.
Wilmer Eye Institute, Johns Hopkins University School of Medicine, Baltimore, Maryland.
Ophthalmology. 2021 Jul;128(7):1016-1026. doi: 10.1016/j.ophtha.2020.12.020. Epub 2020 Dec 25.
Rule-based approaches to determining glaucoma progression from visual fields (VFs) alone are discordant and have tradeoffs. To detect better when glaucoma progression is occurring, we used a longitudinal data set of merged VF and clinical data to assess the performance of a convolutional long short-term memory (LSTM) neural network.
Retrospective analysis of longitudinal clinical and VF data.
From 2 initial datasets of 672 123 VF results from 213 254 eyes and 350 437 samples of clinical data, persons at the intersection of both datasets with 4 or more VF results and corresponding baseline clinical data (cup-to-disc ratio, central corneal thickness, and intraocular pressure) were included. After exclusion criteria-specifically the removal of VFs with high false-positive and false-negative rates and entries with missing data-were applied to ensure reliable data, 11 242 eyes remained.
Three commonly used glaucoma progression algorithms (VF index slope, mean deviation slope, and pointwise linear regression) were used to define eyes as stable or progressing. Two machine learning models, one exclusively trained on VF data and another trained on both VF and clinical data, were tested.
Area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPRC) calculated on a held-out test set and mean accuracies from threefold cross-validation were used to compare the performance of the machine learning models.
The convolutional LSTM network demonstrated 91% to 93% accuracy with respect to the different conventional glaucoma progression algorithms given 4 consecutive VF results for each participant. The model that was trained on both VF and clinical data (AUC, 0.89-0.93) showed better diagnostic ability than a model exclusively trained on VF results (AUC, 0.79-0.82; P < 0.001).
A convolutional LSTM architecture can capture local and global trends in VFs over time. It is well suited to assessing glaucoma progression because of its ability to extract spatiotemporal features that other algorithms cannot. Supplementing VF results with clinical data improves the model's ability to assess glaucoma progression and better reflects the way clinicians manage data when managing glaucoma.
基于规则的方法来确定仅从视野（VF）中青光眼的进展存在分歧并且存在权衡。为了更好地检测青光眼进展，我们使用了合并的 VF 和临床数据的纵向数据集来评估卷积长短期记忆（LSTM）神经网络的性能。
回顾性分析纵向临床和 VF 数据。
从 2 个初始数据集的 213254 只眼中的 672123 个 VF 结果和 350437 个临床数据样本中，选取同时存在 4 个或更多 VF 结果和相应基线临床数据（杯盘比、中央角膜厚度和眼内压）的人。在应用特定的排除标准（即去除具有高假阳性和假阴性率的 VF 和存在缺失数据的条目）以确保可靠的数据后，有 11242 只眼睛被保留。
使用三种常用的青光眼进展算法（VF 指数斜率、平均偏差斜率和逐点线性回归）来定义稳定或进展的眼睛。测试了两种机器学习模型，一种仅在 VF 数据上进行训练，另一种在 VF 和临床数据上进行训练。
在保留的测试集上计算的接收者操作特征曲线下面积（AUC）和精度-召回曲线下面积（AUPRC）以及三次交叉验证的平均准确率，用于比较机器学习模型的性能。
对于每个参与者的 4 个连续 VF 结果，与不同的传统青光眼进展算法相比，卷积 LSTM 网络的准确率为 91%至 93%。同时在 VF 和临床数据上进行训练的模型（AUC，0.89-0.93）比仅在 VF 结果上进行训练的模型（AUC，0.79-0.82；P<0.001）具有更好的诊断能力。
卷积 LSTM 架构可以随时间捕获 VF 中的局部和全局趋势。它非常适合评估青光眼进展，因为它能够提取其他算法无法提取的时空特征。将 VF 结果与临床数据相结合，可以提高模型评估青光眼进展的能力，并更好地反映临床医生在管理青光眼时管理数据的方式。
