Dean Arlen, Fu Dun Jack, Zhalechian Mohammad, Van Oyen Mark P, Lavieri Mariel S, Khawaja Anthony P, Stein Joshua D
From the Department of Industrial and Operations Engineering (A.D., M.P.V.O., M.S.L.), University of Michigan College of Engineering, Ann Arbor, Michigan, USA.
National Institute for Health and Care Research Biomedical Research Centre (D.J.F., A.P.K.), Moorfields Eye Hospital National Health Service Foundation Trust and University College London Institute of Ophthalmology, London, UK.
Am J Ophthalmol. 2025 Apr;272:87-97. doi: 10.1016/j.ajo.2025.01.006. Epub 2025 Jan 13.
A previously developed machine-learning approach with Kalman filtering technology accurately predicted the disease trajectory for patients with various glaucoma types and severities using clinical trial data. This study assesses performance of the KF approach with real-world data.
Retrospective cohort study.
We tested the performance of a previously validated KF model (PKF) initially trained using data from the African Descent and Glaucoma Evaluation Study and the Diagnostic Innovations in Glaucoma Study in patients with different types and severities of glaucoma receiving care in the United Kingdom (UK), comparing the predictive accuracy to 2 conventional linear regression (LR) models and a newly developed KF trained on UK patients (UK-KF).
A total of 3116 patients with open-angle glaucoma or suspects were divided into training (n=1584) and testing (n=1532) sets. The predictive accuracy for MD within 2.5 dB of the observed value at 60 months' follow-up for PKF (75.7%) was substantially better than those for the LR models (P < .01 for both) and similar to that for UK-KF (75.2%, P = .70). The proportion of MD predictions in the 95% repeatability intervals at 60 months' follow-up for PKF (67.9%) was higher than those for the LR models (40.2%, 40.9%) and similar to that for UK-KF (71.4%).
This study validates the performance of our previously developed KF model on a real-world, multicenter patient population. Our model substantially outperforms the current clinical standard (LR) and forecasts well for patients with different glaucoma types and severities. This study supports the generalizability of PKF performance and supports prospective study of implementation into clinical practice.
一种先前开发的采用卡尔曼滤波技术的机器学习方法，利用临床试验数据准确预测了不同类型和严重程度青光眼患者的疾病轨迹。本研究使用真实世界数据评估卡尔曼滤波方法的性能。
回顾性队列研究。
我们测试了一个先前经验证的卡尔曼滤波模型（PKF）的性能，该模型最初使用非洲裔和青光眼评估研究以及青光眼诊断创新研究的数据进行训练，研究对象为在英国接受治疗的不同类型和严重程度的青光眼患者，将其预测准确性与2种传统线性回归（LR）模型以及一个基于英国患者新开发的卡尔曼滤波模型（UK-KF）进行比较。
共有3116例开角型青光眼患者或疑似患者被分为训练组（n = 1584）和测试组（n = 1532）。在60个月随访时，PKF对平均偏差（MD）的预测准确性在观察值的2.5 dB范围内（75.7%），显著优于LR模型（两者P <.01），且与UK-KF相似（75.2%，P = 0.70）。在60个月随访时，PKF在95%重复性区间内的MD预测比例（67.9%）高于LR模型（40.2%，40.9%），且与UK-KF相似（71.4%）。
本研究验证了我们先前开发的卡尔曼滤波模型在真实世界多中心患者群体中的性能。我们的模型显著优于当前临床标准（LR），并能对不同类型和严重程度的青光眼患者进行良好预测。本研究支持PKF性能的普遍性，并支持将其应用于临床实践的前瞻性研究。
