Skip to main content

Table 4 Confusion matrix of keratoconus severity prediction by LDA and RF models on validation dataset

From: Development of a classification system based on corneal biomechanical properties using artificial intelligence predicting keratoconus severity

   Reference classification based on TKC  
Final LDA model Healthy Mild KC Moderate KC Advanced KC Overall accuracy
Prediction Healthy 31 7 0 0 71%
Mild KC 2 19 5 0
Moderate KC 0 9 21 4
Advanced K 0 0 9 19
Sn/Sp 82%/97% 73%/81% 62%/83% 68%/95%  
Final RF model (default cut-off) Healthy Mild KC Moderate KC Advanced KC Overall accuracy
Prediction Healthy 33 10 1 0 75%
Mild KC 0 20 6 0
Moderate KC 0 4 22 3
Advanced K 0 1 6 20
Sn/Sp 100%/88% 57%/93% 63%/92% 87%/93%  
Final RF model (optimized cut-off) Healthy Mild KC Moderate KC Advanced KC Overall accuracy
Final prediction (DKI) Healthy 30 5 1 0 78%
Mild KC 3 28 6 0
Moderate KC 0 1 22 5
Advanced KC 0 1 6 18
Sn/Sp 91%/94% 80%/90% 63%/87% 72%/95%  
  1. DKI Dresden keratoconus index based on optimized RF model; KC keratoconus; LDA linear discriminant analysis; RF random forest; Sn sensitivity; Sp = specificity; TBI tomographic and biomechanical index. Bold signifies correct prediction by LDA or RF model