A Novel Ai Model For Diagnosing Keratoconus Based On Chaos Theory Compared To A Cnn Model

We developed an AI based model for diagnosing keratoconus using a novel approach based on chaos theory and compared it to a conventional CNN model.

Universal Council of Ophthalmology (UCO)

Many processes in nature (e.g. weather systems) and even in the human body (e.g. blood glucose level) exhibit chaotic behavior, which means they are based on well-known physics principles but are very challenging to predict. The keratoconus process can be considered one such phenomenon. We established a novel AI model based on chaos system principles for diagnosing keratoconus using tomographic data. We also compared overfitting in our model to a conventional CNN model by running simulations using 64 factors with 0, 8, and 16 common sources (redundant data).

Our designed model showed very encouraging results in our preliminary tests (sens. 92%, spec. 96%, N = 64 normal, 44 keratoconus). Analyzing for overfitting showed that adding 8 and 16 common sources to a conventional CNN model caused severe overfitting while adding up to 16 common sources didn’t cause any overfitting in our model.

We developed a new AI model based on a chaotic approach and achieved encouraging results. A current challenge of deep learning models is overfitting which causes severe performance loss in external validation datasets. We showed a high overfitting tendency in CNN, while our model didn’t experience any adding up to 16 CS. This model may be valuable in a broad range of ophthalmologic or other conditions. The outcomes of this model should be confirmed in larger sample sizes.