Purpose

Previously, we introduced a state-based 2-dimensional continuous-time hidden Markov model (2D CT HMM) to model the pattern of detected glaucoma changes using structural and functional information simultaneously. The purpose of this study was to evaluate the detected glaucoma change prediction performance of the model in a real clinical setting using a retrospective longitudinal dataset.

Design

Longitudinal, retrospective study.

Participants

One hundred thirty-four eyes from 134 participants diagnosed with glaucoma or as glaucoma suspects (average follow-up, 4.4±1.2 years; average number of visits, 7.1±1.8).

Methods

A 2D CT HMM model was trained using OCT (Cirrus HD-OCT; Zeiss, Dublin, CA) average circumpapillary retinal nerve fiber layer (cRNFL) thickness and visual field index (VFI) or mean deviation (MD; Humphrey Field Analyzer; Zeiss). The model was trained using a subset of the data (107 of 134 eyes [80%]) including all visits except for the last visit, which was used to test the prediction performance (training set). Additionally, the remaining 27 eyes were used for secondary performance testing as an independent group (validation set). The 2D CT HMM predicts 1 of 4 possible detected state changes based on 1 input state.

Main Outcome Measures

Prediction accuracy was assessed as the percentage of correct prediction against the patient's actual recorded state. In addition, deviations of the predicted long-term detected change paths from the actual detected change paths were measured.

Results

Baseline mean ± standard deviation age was 61.9±11.4 years, VFI was 90.7±17.4, MD was −3.50±6.04 dB, and cRNFL thickness was 74.9±12.2 μm. The accuracy of detected glaucoma change prediction using the training set was comparable with the validation set (57.0% and 68.0%, respectively). Prediction deviation from the actual detected change path showed stability throughout patient follow-up.

Conclusions

The 2D CT HMM demonstrated promising prediction performance in detecting glaucoma change performance in a simulated clinical setting using an independent cohort. The 2D CT HMM allows information from just 1 visit to predict at least 5 subsequent visits with similar performance.

中文翻译：

---

使用二维连续时间隐马尔可夫模型进行结构和功能测量的青光眼进展的临床预测性能

目的

以前，我们引入了基于状态的二维连续时间隐马尔可夫模型（2D CT HMM），以同时使用结构和功能信息对检测到的青光眼变化的模式进行建模。这项研究的目的是使用回顾性纵向数据集在实际临床环境中评估模型检测到的青光眼变化预测性能。

设计

纵向回顾性研究。

参加者

来自134名被诊断为青光眼或怀疑为青光眼的参与者的一百三十四只眼（平均随访时间为4.4±1.2年；平均就诊次数为7.1±1.8）。

方法

使用OCT（Cirrus HD-OCT; Zeiss，Dublin，CA）训练2D CT HMM模型的平均眼眶乳头视网膜神经纤维层（cRNFL）厚度和视野指数（VFI）或平均偏差（MD; Humphrey Field Analyzer; Zeiss） 。使用数据的子集（134眼中的107眼[80％]）对模型进行了训练，其中包括除最后一次拜访外的所有拜访，该次拜访用于测试预测性能（训练集）。此外，其余27只眼睛作为一个独立组（验证集）用于二次性能测试。2D CT HMM根据1种输入状态预测4种可能的检测状态变化中的1种。

主要观察指标

预测准确性被评估为正确预测相对于患者实际记录状态的百分比。另外，测量了预测的长期检测到的变化路径与实际检测到的变化路径的偏差。

结果

基线平均±标准偏差年龄为61.9±11.4岁，VFI为90.7±17.4，MD为−3.50±6.04 dB，cRNFL厚度为74.9±12.2μm。使用训练集检测到的青光眼变化预测的准确性与验证集相当（分别为57.0％和68.0％）。与实际检测到的变化路径的预测偏差在整个患者随访过程中显示出稳定性。

结论

2D CT HMM在使用独立队列的模拟临床环境中，在检测青光眼变化表现中显示出令人鼓舞的预测性能。2D CT HMM允许仅1次访问的信息即可预测至少5次类似性能的后续访问。