PurposeTo determine whether combinations of devices with different measuring principles, supported by artificial intelligence (AI), can improve the diagnosis of keratoconus (KC). MethodsScheimpflug tomography, spectral-domain optical coherence tomography (SD-OCT), and air-puff tonometry were performed in all eyes. The most relevant machine-derived parameters to diagnose KC were determined using feature selection. The normal and forme fruste KC (FFKC) eyes were divided into training and validation datasets. The selected features from a single device or different combinations of devices were used to develop models based on random forest (RF) or neural networks (NN) trained to distinguish FFKC from normal eyes. The accuracy was determined using receiver operating characteristic (ROC) curves, area under the curve (AUC), sensitivity, and specificity. Results271 normal eyes, 84 FFKC eyes, 85 early KC eyes, and 159 advanced KC eyes were included. A total of 14 models were built. Air-puff tonometry had the highest AUC for detecting FFKC using a single device (AUC = 0.801). Among all two-device combinations, the highest AUC was accomplished using RF applied to selected features from SD-OCT and air-puff tonometry (AUC = 0.902), followed by the three-device combination with RF (AUC = 0.871) with the best accuracy. ConclusionExisting parameters can precisely diagnose early and advanced KC, but their diagnostic ability for FFKC could be optimized. Applying an AI algorithm to a combination of air-puff tonometry with Scheimpflug tomography or SD-OCT could improve FFKC diagnostic ability. The improvement in diagnostic ability by combining three devices is modest.
Read full abstract