Malone Center For Engineering in Healthcare, Johns Hopkins University, Baltimore, Maryland.
Wilmer Eye Institute, Johns Hopkins University, Baltimore, Maryland.
Ophthalmol Glaucoma. 2023 Sep-Oct;6(5):466-473. doi: 10.1016/j.ogla.2023.03.005. Epub 2023 Mar 20.
To assess whether we can forecast future rapid visual field (VF) worsening using deep learning models (DLMs) trained on early VF, OCT, and clinical data.
A retrospective cohort study.
In total, 4536 eyes from 2962 patients. Overall, 263 (5.80%) eyes underwent rapid VF worsening (mean deviation slope less than -1 dB/year across all VFs).
We included eyes that met the following criteria: (1) followed for glaucoma or suspect status; (2) had at least 5 longitudinal reliable VFs (VF, VF, VF, VF, and VF); and (3) had 1 reliable baseline OCT scan (OCT) and 1 set of baseline clinical measurements (clinical) at the time of VF. We designed a DLM to forecast future rapid VF worsening. The input consisted of spatially oriented total deviation values from VF (including or not including VF and VF in some models) and retinal nerve fiber layer thickness values from the baseline OCT. We passed this VF/OCT stack into a vision transformer feature extractor, the output of which was concatenated with baseline clinical data before putting it through a linear classifier to predict the eye's risk of rapid VF worsening across the 5 VFs. We compared the performance of models with differing inputs by computing area under the curve (AUC) in the test set. Specifically, we trained models with the following inputs: (1) model V: VF; (2) VC: VF+ Clinical; (3) VO: VF+ OCT; (4) VOC: VF+ Clinical+ OCT; (5) V: VF + VF; (6) VOC: VF + VF + Clinical + OCT; (7) V: VF + VF + VF; and (8) VOC: VF + VF + VF + Clinical + OCT.
The AUC of DLMs when forecasting rapidly worsening eyes.
Model VOC best forecasted rapid worsening with an AUC (95% confidence interval [CI]) of 0.87 (0.77-0.97). Remaining models in descending order of performance and their respective AUC (95% CI) were as follows: (1) model V (0.84 [0.74-0.95]), (2) model VOC (0.81 [0.70-0.92]), (3) model V (0.81 [0.70-0.82]), (4) model VOC (0.77 [0.65-0.88]), (5) model VO (0.75 [0.64-0.88]), (6) model VC (0.75 [0.63-0.87]), and (7) model V (0.74 [0.62-0.86]).
Deep learning models can forecast future rapid glaucoma worsening with modest to high performance when trained using data from early in the disease course. Including baseline data from multiple modalities and subsequent visits improves performance beyond using VF data alone.
FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
利用基于深度学习模型（DLM）的早期视野（VF）、OCT 和临床数据，评估我们是否可以预测未来的快速 VF 恶化情况。
回顾性队列研究。
共 2962 名患者的 4536 只眼。共有 263（5.80%）只眼发生快速 VF 恶化（所有 VF 中平均偏差斜率小于-1dB/年）。
我们纳入了符合以下标准的眼：（1）接受青光眼或疑似青光眼治疗；（2）至少有 5 次纵向可靠的 VF（VF、VF、VF、VF 和 VF）；（3）在 VF 时具有 1 次可靠的基线 OCT 扫描（OCT）和 1 套基线临床测量（临床）。我们设计了一个 DLM 来预测未来的快速 VF 恶化。输入包括来自 VF 的空间定向总偏差值（某些模型中包括或不包括 VF 和 VF）和基线 OCT 的视网膜神经纤维层厚度值。我们将这个 VF/OCT 堆叠传入一个视觉转换器特征提取器，其输出与基线临床数据串联，然后通过线性分类器将其输入以预测 5 个 VF 中眼睛快速 VF 恶化的风险。我们通过计算测试集中的曲线下面积（AUC）来比较具有不同输入的模型的性能。具体来说，我们使用以下输入训练模型：（1）模型 V：VF；（2）VC：VF+临床；（3）VO：VF+OCT；（4）VOC：VF+临床+OCT；（5）V：VF+VF；（6）VOC：VF+VF+临床+OCT；（7）V：VF+VF+VF；（8）VOC：VF+VF+VF+临床+OCT。
DLM 在预测快速恶化眼中的 AUC。
模型 VOC 以 AUC（95%置信区间[CI]）为 0.87（0.77-0.97）最佳预测快速恶化。按性能降序排列的其余模型及其各自的 AUC（95%CI）如下：（1）模型 V（0.84 [0.74-0.95]），（2）模型 VOC（0.81 [0.70-0.92]），（3）模型 V（0.81 [0.70-0.82]），（4）模型 VOC（0.77 [0.65-0.88]），（5）模型 VO（0.75 [0.64-0.88]），（6）模型 VC（0.75 [0.63-0.87]）和（7）模型 V（0.74 [0.62-0.86]）。
当使用疾病早期的数据进行训练时，深度学习模型可以以中等至较高的性能预测未来的快速青光眼恶化。与仅使用 VF 数据相比，同时使用多种模式和后续就诊的基线数据可以提高性能。
在本文末尾的脚注和披露中可能会发现专有或商业披露信息。
