【摘要】 目的 基于可视化角膜生物力学分析仪(corneal visualization scheimpflug technology, Corvis ST)的测试数据，探索将包含时间序列的角膜曲率数据用于构建顿挫期圆锥角膜(forme fruste keratoconus, FFKC)诊断模型的可行性。方法 纳入FFKC患者88例(88眼),选择同数量拟行角膜屈光手术患者作为正常对照。利用Corvis ST测试输出的140帧角膜前/后表面的轮廓数据，构建时间序列角膜曲率数据集。分别利用随机森林(random forest, RF)方法和长短期记忆-注意力模型(long short-term memory-attention, LSTM-Attention)构建FFKC分类器，用五折交叉验证对模型进行验证。结果 RF模型的结果显示，将140帧角膜曲率数据作为输入指标构建诊断模型的AUC值仅为0.69;而将特定帧时提取的曲率指标和Corvis ST输出的角膜动力学响应(dynamic corneal response, DCR)参数联合可有效地提升FFKC诊断模型的效率(AUC值最大为0.81)。相较于RF方法，基于时间序列的角膜曲率数据集构建的LSTM-Attention模型的诊断效率可进一步提升(AUC值为0.87)。若将角膜前表面和后表面轮廓的时间序列曲率数据集分别作为模型输入时，模型的AUC值分别为0.82和0.84。结论 利用包含时间序列的角膜曲率数据集构建FFKC诊断模型具有可行性，且LSTM-Attention模型可有效地提升模型的诊断效率。更多还原

【Abstract】 Objective To explore the feasibility of constructing a diagnostic model for forme fruste keratoconus(FFKC) using time-series corneal curvature data obtained from the corneal visualization scheimpflug technology(Corvis ST) test. Methods This study included 88 cases(88 eyes) of FFKC patients, with an equal number of patients undergoing corneal refractive surgery selected as the control group. Utilizing the 140-frame corneal anterior/posterior surface profile data obtained from the Corvis ST test, a time-series corneal curvature dataset was constructed. Subsequently, classifiers for FFKC were constructed using both the random forest(RF) model and the long short-term memory-attention(LSTM-Attention) model. The models were validated using five-fold cross-validation. Results The RF model showed that using the 140-frame corneal curvature data as input features resulted in an AUC value of only 0.69. However, combining specific frame-derived curvature indices with dynamic corneal response(DCR) parameters from Corvis ST significantly improved the model’s diagnostic efficiency, achieving a maximum AUC of 0.81. Compared to the RF methods, the LSTM-Attention model based on the time-series corneal curvature dataset further enhanced diagnostic efficiency, achieving an AUC of 0.87. When using the time-series curvature dataset of the anterior and posterior corneal surfaces as separate model inputs, the AUC values were 0.82 and 0.84,respectively. Conclusions Constructing an FFKC diagnostic model using a time-series corneal curvature dataset is feasible, and the LSTM-Attention model effectively enhances diagnostic efficiency.更多还原