Ansari Georg, Schärer Nils, Pfau Kristina, Valmaggia Philippe, Gabrani Chrysoula, Zuche Hanna, Giani Andrea, Esmaeelpour Marieh, Yamaguchi Taffeta Chingning, Feltgen Nicolas, Maloca Peter M, Schmetterer Leopold, Scholl Hendrik P N, Pfau Maximilian
Institute of Molecular and Clinical Ophthalmology Basel (IOB), Basel, Switzerland.
Department of Ophthalmology, University Hospital Basel, Basel, Switzerland.
Invest Ophthalmol Vis Sci. 2025 Aug 1;66(11):34. doi: 10.1167/iovs.66.11.34.
The purpose of this study was to evaluate the effectiveness of different machine-learning models in predicting retinal sensitivity in geographic atrophy (GA) secondary to age-related macular degeneration (AMD) and compare the progression of sensitivity loss using observed versus inferred data over time.
Thirty patients with GA (37 eyes) were recruited for the OMEGA study. Participants underwent fundus-controlled perimetry (microperimetry) and spectral-domain optical coherence tomography (SD-OCT) imaging at baseline and follow-up visits at weeks 12, 24, and 48. Retinal layers were segmented using a custom-written deep-learning algorithm. We used various machine-learning models, including random forest, LASSO regression, and multivariate adaptive regression splines (MARS), to predict retinal sensitivity across three scenarios: (1) unknown patients, (2) known patients at later visits, and (3) interpolation within visits. Predictive accuracy was evaluated using the mean absolute error (MAE), and the models' ability to reduce test variability over time was analyzed using linear mixed models.
The random forest model demonstrated the highest accuracy across all scenarios, with an MAE of 3.67 decibels (dB) for unknown patients, 2.96 dB for known patients at follow-up, and 3.10 dB for within-visit interpolation. The inferred sensitivity data significantly reduced variability compared to the observed data in longitudinal mixed model analysis, with a residual variance of 2.72 dB² versus 8.67 dB², respectively.
Machine-learning models, particularly the random forest model, effectively predict retinal sensitivity in patients with GA, with patient-specific baseline data improving accuracy for subsequent visits. Inferred sensitivity mapping presents a reliable, functional surrogate endpoint for clinical trials, offering high spatial resolution without extensive psychophysical testing.
本研究旨在评估不同机器学习模型在预测年龄相关性黄斑变性（AMD）继发地理萎缩（GA）患者视网膜敏感度方面的有效性，并比较使用观察数据与推断数据随时间变化的敏感度损失进展情况。
招募了30例GA患者（37只眼）参与OMEGA研究。参与者在基线以及第12周、24周和48周的随访时接受了眼底控制视野检查（微视野检查）和光谱域光学相干断层扫描（SD-OCT）成像。使用自定义编写的深度学习算法对视网膜各层进行分割。我们使用了各种机器学习模型，包括随机森林、套索回归和多元自适应回归样条（MARS），来预测三种情况下的视网膜敏感度：（1）未知患者，（2）后期随访时的已知患者，以及（3）随访期间的内插法。使用平均绝对误差（MAE）评估预测准确性，并使用线性混合模型分析模型随时间降低测试变异性的能力。
随机森林模型在所有情况下均表现出最高的准确性，对于未知患者，MAE为3.67分贝（dB）；随访时的已知患者为2.96 dB；随访期间内插法为3.10 dB。在纵向混合模型分析中，与观察数据相比，推断的敏感度数据显著降低了变异性，剩余方差分别为2.72 dB²和8.67 dB²。
机器学习模型，尤其是随机森林模型，能够有效预测GA患者的视网膜敏感度，患者特异性的基线数据可提高后续随访的准确性。推断的敏感度映射为临床试验提供了一个可靠的、功能性替代终点，无需广泛的心理物理学测试即可提供高空间分辨率。
