Machine learning techniques in the diagnosis of meibomian glands related alterations from clinical indicators

Abstract

Purpose: There is no "Gold Standard" test that allows the diagnosis and classification of alterations and pathologies related to Meibomian glands (MG). A global evaluation of objective and subjective tests is necessary to determine the final diagnosis. In recent years, Artificial Intelligence (AI) and Machine Learning (ML) techniques have experienced great progress in the field of health sciences, as promising techniques for predicting pathologies from data and images. The main objective of this study has been to train ML classifiers for the classification of three groups of participants with and without MG alterations. The secondary objective was to study the precision, specificity and sensitivity of the ML classifiers.

Methods: A retrospective comparative study was carried out on a total of 135 participants (control, contact lens wearers and MG pathology). Symptomatology and clinical tests were performed to evaluate the ocular surface and adnexa. The numerical data obtained from these tests were used to train ML classifiers and the top 5 were subsequently verified.

Results: Accuracies greater than 76 % were obtained for the training group and greater than 79 % for the verification group, for five classifiers previously described in Matlab. Subspace KNN was the classifier with the highest accuracies, specificities and sensitivities, these being moderate-high (greater than 79 %).

Conclusions: ML algorithms can be useful for classifying groups of participants with various meibomian gland disorders using clinical data. A large number of participants is needed for reliable diagnostic accuracy.