View abstract

This abstract is assigned to session AOB Free Papers 2

Presentation type	Oral presentation
Title	Enhancing Posterior Curvature Analysis using Machine Learning for Early Detection of Keratoconus

Purpose	To develop a machine learning index based only on posterior curvature to differentiate keratoconus, keratoconus suspect and normal eyes.
Methods	520 eyes of 520 patients from the American University of Beirut Medical Center were included in the study, with 264 normal corneas (NL), 197 eyes diagnosed with keratoconus (KC), and 59 keratoconus suspect corneas (KCS). KCS corneas were normal fellow eyes of patients with highly asymmetric keratoconus. Eight parameters were derived from the axial and tangential posterior curvature maps based on the maximal difference in opposing sector (MDOS) at 1.0, 1.5, 2.0, and 2.5 mm of radii, and fed into a machine-learning algorithm to create a posterior curvature index.
Results	In experiment 1 (Normal and KC), a sensitivity of 0.94, a specificity of 0.98 and a positive predictive value (PPV) of 0.98 were achieved by the Random Forest model. The same model in experiment 2 (Normal, KCS and KC) scored a sensitivity of 0.39, a specificity of 0.97 and a PPV of 0.63 in detecting KCS. The MDOS at 1.5 mm radius on axial maps had the most significant contribution in decision making.
Conclusion	A machine-learning index solely based on the corneal posterior curvature was able to differentiate between normal and keratoconus with a decent accuracy, and between KCS and normal with a lower accuracy when compared to available complex models. Posterior curvature incorporation into a multi-model diagnostic system might be beneficial and should be investigated.
Conflict of interest	No
Details of conflicting interests	None

Authors 1

Last name	WEHBI
Initials	Z
Department	American University of Beirut
City	Beirut, Lebanon

Authors 2

Last name	Abou Mrad
Initials	A
Department	American University of Beirut
City	Beirut, Lebanon

Authors 3

Last name	Assaf
Initials	J.F
Department	Casey Eye Institute
City	Portland, Oregon

Authors 4

Last name	Khalil
Initials	J
Department	American University of Beirut
City	Beirut, Lebanon

Authors 5

Last name	Roberts
Initials	C.J
Department	Departments of Ophthalmology and Visual Sciences, Ohio State University
City	Ohio, USA

Authors 6

Last name	Awwad
Initials	S.T
Department	American University of Beirut
City	Beirut, Lebanon