Data_Sheet_3_Machine Learning to Analyze Factors Associated With Ten-Year Graft Survival of Keratoplasty for Cornea Endothelial Disease.PDF

PurposeMachine learning analysis of factors associated with 10-year graft survival of Descemet stripping automated endothelial keratoplasty (DSAEK) and penetrating keratoplasty (PK) in Asian eyes.
MethodsProspective study of donor characteristics, clinical outcomes and complications from consecutive patients (n = 1,335) who underwent DSAEK (946 eyes) or PK (389 eyes) for Fuchs’ endothelial dystrophy (FED) or bullous keratopathy (BK) were analyzed. Random survival forests (RSF) analysis using the highest variable importance (VIMP) factors were determined to develop the optimal Cox proportional hazards regression model. Main outcome measure was 10-year graft survival with RSF analysis of factors associated with graft failure.
ResultsMean age was 68 ± 11 years, 47.6% male, in our predominantly Chinese (76.6%) Asian cohort, with more BK compared to FED (62.2 vs. 37.8%, P < 0.001). Overall 10-year survival for DSAEK was superior to PK (73.6 vs. 50.9%, log-rank P < 0.001). RSF based on VIMP (best Harrell C statistic: 0.701) with multivariable modeling revealed that BK (HR:2.84, 95%CI:1.89–4.26; P < 0.001), PK (HR: 1.64, 95%CI:1.19–2.27; P = 0.002), male recipients (HR:1.75, 95%CI:1.31–2.34; P < 0.001) and poor pre-operative visual acuity (HR: 1.60, 95%CI:1.15–2.22, P = 0.005) were associated with graft failure. Ten-year cumulative incidence of complications such as immune-mediated graft rejection (P < 0.001), epitheliopathy (P < 0.001), and wound dehiscence (P = 0.002) were greater in the PK compared to the DSAEK group.
ConclusionIn our study, RSF combined with Cox regression was superior to traditional regression techniques alone in analyzing a large number of high-dimensional factors associated with 10-year corneal graft survival in Asian eyes with cornea endothelial disease.

研究目的：针对亚洲人群角膜行剥除后弹力层自动内皮角膜移植术（Descemet stripping automated endothelial keratoplasty，DSAEK）与穿透性角膜移植术（penetrating keratoplasty，PK）后10年移植物存活相关影响因素的机器学习分析。
研究方法：本研究为前瞻性研究，纳入1335例因弗uchs内皮营养不良（Fuchs’ endothelial dystrophy，FED）或大疱性角膜病（bullous keratopathy，BK）行DSAEK（946眼）或PK（389眼）的连续入组患者，对供体特征、临床结局及并发症进行分析。采用基于最高变量重要性（variable importance，VIMP）因子的随机生存森林（random survival forests，RSF）分析，构建最优Cox比例风险回归模型。主要结局指标为10年移植物存活率，并通过RSF分析与移植物失败相关的影响因素。
研究结果：本研究纳入的亚洲人群队列以汉族为主（76.6%），患者平均年龄为68±11岁，男性占比47.6%；其中大疱性角膜病患者多于弗uchs内皮营养不良患者（62.2% vs. 37.8%，P<0.001）。DSAEK组的总体10年移植物存活率优于PK组（73.6% vs. 50.9%，对数秩检验P<0.001）。基于最高变量重要性因子构建的随机生存森林分析（最优Harrell C统计量：0.701）结合多变量建模结果显示，大疱性角膜病（风险比HR：2.84，95%置信区间CI：1.89–4.26；P<0.001）、行穿透性角膜移植术（HR：1.64，95%CI：1.19–2.27；P=0.002）、男性受者（HR：1.75，95%CI：1.31–2.34；P<0.001）以及术前视力不佳（HR：1.60，95%CI：1.15–2.22，P=0.005）均与移植物失败相关。PK组的免疫介导移植物排斥反应、上皮病变及伤口裂开等并发症的10年累积发生率均高于DSAEK组（免疫介导移植物排斥反应：P<0.001；上皮病变：P<0.001；伤口裂开：P=0.002）。
研究结论：在本研究中，相较于单一传统回归分析方法，随机生存森林联合Cox回归模型在分析亚洲角膜内皮疾病患者10年角膜移植物存活相关的大量高维影响因素时表现更优。