Jacquot Robin, Ren Lijuan, Wang Tao, Mellahk Insaf, Duclos Antoine, Kodjikian Laurent, Jamilloux Yvan, Stanescu Dinu, Sève Pascal
Department of Internal Medicine, Hôpital Universitaire de la Croix-Rousse, Hospices Civils de Lyon, University Claude Bernard-Lyon 1, Lyon, France.
Research on Healthcare Performance (RESHAPE), INSERM U1290, University Claude Bernard Lyon 1, Lyon, France.
Eye (Lond). 2025 Apr;39(5):992-1002. doi: 10.1038/s41433-024-03530-2. Epub 2024 Dec 20.
BACKGROUND/OBJECTIVES: The large number and heterogeneity of causes of uveitis make the etiological diagnosis a complex task. The clinician must consider all the information concerning the ophthalmological and extra-ophthalmological features of the patient. Diagnostic machine learning algorithms have been developed and provide a correct diagnosis in one-half to three-quarters of cases. However, they are not integrated into daily clinical practice. The aim is to determine whether machine learning models can predict the etiological diagnosis of uveitis from clinical information.
This cohort study was performed on uveitis patients with unknown etiology at first consultation. One hundred nine variables, including demographic, ophthalmic, and clinical information, associated with complementary exams were analyzed. Twenty-five causes of uveitis were included. A neural network was developed to predict the etiological diagnosis of uveitis. The performance of the model was evaluated and compared to a gold standard: etiological diagnosis established by a consensus of two uveitis experts.
A total of 375 patients were included in this analysis. Findings showed that the neural network type (Multilayer perceptron) (NN-MLP) presented the best prediction of the etiological diagnosis of uveitis. The NN-MLP's most probable diagnosis matched the senior clinician diagnosis in 292 of 375 patients (77.8%, 95% CI: 77.4-78.0). It achieved 93% accuracy (95% CI: 92.8-93.1%) when considering the two most probable diagnoses. The NN-MLP performed well in diagnosing idiopathic uveitis (sensitivity of 81% and specificity of 82%). For more than three-quarters of etiologies, our NN-MLP demonstrated good diagnostic performance (sensitivity > 70% and specificity > 80%).
Study results suggest that developing models for accurately predicting the etiological diagnosis of uveitis with undetermined etiology based on clinical information is feasible. Such NN-MLP could be used for the etiological assessments of uveitis with unknown etiology.
背景/目的：葡萄膜炎病因众多且具有异质性，这使得病因诊断成为一项复杂的任务。临床医生必须综合考虑患者眼科及眼科以外的所有信息。诊断性机器学习算法已经开发出来，在一半至四分之三的病例中能提供正确诊断。然而，它们尚未融入日常临床实践。目的是确定机器学习模型能否根据临床信息预测葡萄膜炎的病因诊断。
本队列研究针对初诊时病因不明的葡萄膜炎患者开展。分析了109个变量，包括人口统计学、眼科及临床信息，以及与之相关的辅助检查结果。纳入了25种葡萄膜炎病因。开发了一个神经网络来预测葡萄膜炎的病因诊断。对模型的性能进行评估，并与金标准进行比较：由两位葡萄膜炎专家达成共识确定的病因诊断。
本分析共纳入375例患者。结果显示，神经网络类型（多层感知器）（NN-MLP）对葡萄膜炎病因诊断的预测效果最佳。在375例患者中，NN-MLP的最可能诊断与资深临床医生的诊断在292例中相符（77.8%，95%置信区间：77.4-78.0）。考虑两个最可能的诊断时，其准确率达到93%（95%置信区间：92.8-93.1%）。NN-MLP在诊断特发性葡萄膜炎方面表现良好（敏感性为81%，特异性为82%）。对于超过四分之三的病因，我们的NN-MLP表现出良好的诊断性能（敏感性>70%，特异性>80%）。
研究结果表明，基于临床信息开发准确预测病因不明的葡萄膜炎病因诊断的模型是可行的。这种NN-MLP可用于病因不明的葡萄膜炎的病因评估。
