Advancing subclinical keratoconus detection using polarization-sensitive optical coherence tomography and artificial intelligence.

Significance: Early detection of keratoconus (KC), a progressive corneal disorder, remains a major clinical challenge. Polarization-sensitive optical coherence tomography (PS-OCT) is an advanced imaging technique that can quantify corneal birefringence, offering insight into collagen organization and microstructural integrity. Approach: A total of 359 eyes were examined and evaluated. This study explores PS-OCT-derived phase retardation (PR) and corneal sublayer thickness as potential diagnostic biomarkers for healthy, subclinical (SKC), and KC eyes. Further, the performance of AI-based classification models developed from PS-OCT, Pentacam, and MS-39 data were compared, using random forest classifier trained with a leave-one-out methodology and identical hyperparameter settings. Results: All AI models demonstrated comparable accuracy among devices for healthy and KC detection. However, SKC classification differed from Pentacam and MS-39. Here, 39.5% of the SKC eyes were reclassified as healthy by PS-OCT, compared with 27.5% by Pentacam and 30.3% by MS-39. The average diagnostic performance of PS-OCT included an AUC, precision, recall, F1 score, and accuracy of 0.91, 83%, 82%, 0.82, and 82%, respectively. For the Pentacam, the same were 0.95, 87%, 86%, 0.86, and 86%, respectively. For the MS-39, the same were 0.94, 86%, 86%, 0.85, and 86%, respectively. Conclusions: Overall, AI model agreement was strong for healthy and KC groups but varied in SKC. PS-OCT provides complementary diagnostic value and may refine subclinical KC detection for safer refractive surgery decisions.