An Innovative Approach Based on Machine Learning to Evaluate the Risk Factors Importance in Diagnosing Keratoconus

Background and objective:
Keratoconus is a non-inflammatory corneal condition affecting both eyes and is present in one out of every 2,000 people worldwide. The cornea deforms into a conical shape and thins, resulting in high-order aberrations and gradual vision loss. Risk factor analysis in the degradation of keratoconus is under-researched.

Methods:
This research work investigates and uses effective machine learning models to gain insight into how much the risk factors of a patient contribute towards the progressive stages of keratoconus, as well as how significant these factors are in the creation of an accurate prediction model. This research demonstrates the value of machine learning approaches on a clinical dataset. This research paper employs several machine learning algorithms to classify the patients' stage of keratoconus using clinical information, such as measurements of the cornea's topography, elevation, and pachymetry taken using pentacam equipment at Sydney's Vision Eye Institute Chatswood.

Results:
Eight different machine learning techniques were investigated over three variations of a dataset and achieved an average accuracy of 68, 80, and 90% for the risk factor, pentacam, and cumulative datasets, respectively. The results show a significant increase in accuracy and a 97% increase in AUC upon addition of risk factor data compared to the models trained on pentacam data alone. The machine learning methods shown in this paper outperform those in current research.

Conclusions:
This research highlights the importance of machine learning methods and risk factor data in the diagnosis of keratoconus and highlights the patient's primary optical aid as the strongest risk factor. The goal of this research is to support the work of the ophthalmologists in diagnosing keratoconus and provide better care for the patient.