Accuracy of machine learning classifiers using bilateral data from a Scheimpflug camera for identifying eyes with preclinical signs of keratoconus

详细信息    查看全文

• 作者：Illé ; s Ková ; cs ; MD ; PhD ; ^{kovacs.illes@med.semmelweis-univ.hu" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Kata Mihá ; ltz ; MD ; PhD ; Kinga Krá ; nitz ; MD ; PhD ; É ; va Juhá ; sz ; MD ; Á ; gnes Taká ; cs ; MD ; PhD ; Ló ; rá ; nt Dienes ; MD ; Ró ; bert Gergely ; MD ; Zoltá ; n Z. Nagy ; MD ; DSC
• 刊名：Journal of Cataract & Refractive Surgery
• 出版年：2016
• 出版时间：February 2016
• 年：2016
• 卷：42
• 期：2
• 页码：275-283
• 全文大小：440 K

文摘

To describe the topographic and tomographic characteristics of normal fellow eyes of unilateral keratoconus cases and to evaluate the accuracy of machine learning classifiers in discriminating healthy corneas from the normal fellow corneas.

Setting

Department of Ophthalmology, Semmelweis University, Budapest, Hungary.

Design

Retrospective case-control study.

Methods

Patients with bilateral keratoconus (keratoconus group), clinically and according to the keratoconus indices of the Pentacam HR Scheimpflug camera; normal fellow eyes of patients with unilateral keratoconus (fellow-eye group); and eyes of refractive surgery candidates (control group) were compared. Tomographic data, topographic data, and keratoconus indices were measured in both eyes using the Scheimpflug camera. Receiver operating characteristic (ROC) analysis was used to assess the performance of automated classifiers trained on bilateral data as well as individual parameters to discriminate fellow eyes of patients with keratoconus from control eyes.

Results

Keratometry, elevation, and keratoconus indices values were significantly higher and pachymetry values were significantly lower in keratoconus eyes than in fellow eyes of unilateral keratoconus cases (P < .001). These fellow eyes had significantly higher keratometry, elevation, and keratoconus index values and significantly lower pachymetry values than control eyes (P < .001). Automated classifiers trained on bilateral data of index of height decentration had higher accuracy than the unilateral single parameter in discriminating fellow eyes of patients with keratoconus from control eyes (area under ROC 0.96 versus 0.88).

Conclusion

Automatic classifiers trained on bilateral data were better than single parameters in discriminating fellow eyes of patients with unilateral keratoconus with preclinical signs of keratoconus from normal eyes.

Financial Disclosure

No author has a financial or proprietary interest in any material or method mentioned.