Related Experiment Video
Updated: Jun 16, 2026
Three Different Protocols of Corneal Collagen Crosslinking in Keratoconus: Conventional, Accelerated and Iontophoresis
Published on: November 12, 2015
A Continuous Severity Index for Keratoconus Diagnosis Based on Kolmogorov-Arnold Networks in a Chinese Population
Ying Qiao1,2,3,4, Lin Wang2,3,4, Lie Ju5,6
1School of Health Science and Engineeering, University of Shanghai for Science and Technology, Shanghai, China.
Objective:
To develop and validate a continuous, interpretable severity index for keratoconus (KC) using a Kolmogorov-Arnold Network (KAN) to improve early detection and disease staging, particularly for forme fruste keratoconus (FFKC) in refractive surgery screening.
Design:
A retrospective case-control study.
Participants:
A total of 384 eyes from 384 participants were included: 101 keratoconic eyes, 132 FFKC eyes (fellow eyes of KC patients), and 151 normal control eyes derived from refractive surgery candidates with at least 2 years of uneventful follow-up.
Methods:
Corneal tomographic parameters were obtained using Pentacam HR, and biomechanical parameters using Corvis ST. A KAN model was trained using categorical diagnostic labels (normal, FFKC, and KC) to generate a continuous, dimensionless Continuous Severity Index (CSI). Continuous Severity Index performance was compared with established indices, including Corvis Biomechanical Index, Corvis Biomechanical Index for Chinese populations, Tomography and Biomechanical Index, Tomography and Biomechanical Index for Chinese populations, and Stress-Strain Index, using receiver operating characteristic analysis. Feature attribution analysis was performed to explore stage-dependent parameter contributions.
Main Outcome Measures:
Area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and optimal cutoff values for differentiating disease stages.
Results:
Continuous Severity Index achieved an AUC of 1.000 for distinguishing KC from normal eyes, and 0.859 for distinguishing FFKC from normal eyes, outperforming all conventional biomechanical and tomographic indices. In combined screening tasks, CSI demonstrated superior performance in identifying any ectatic change (FFKC + KC vs. normal, AUC = 0.920) and advanced disease (KC vs. normal + FFKC, AUC = 0.998). Feature attribution analysis further revealed a stage-dependent shift in dominant contributors, with biomechanical parameters prevailing at lower CSI levels and tomographic asymmetry features increasingly governing advanced disease.
Conclusions:
The KAN-derived Continuous Severity Index enables unified, continuous, and interpretable quantification of KC severity, outperforming existing indices in both early screening and disease staging. By capturing the biologically coherent progression from biomechanical instability to morphological deformation, CSI provides a clinically meaningful tool for refractive surgery screening and KC management.
Financial Disclosures:
The authors have no proprietary or commercial interest in any materials discussed in this article.