Table 3 Impact of augmented prediction and incorporation of clinical features on vertebral collapse (VC) prediction performance of ViT-PMC-LoRA.

From: Augmented prediction of vertebral collapse after osteoporotic vertebral compression fractures through parameter-efficient fine-tuning of biomedical foundation models

 

AUC

mean (SD)

Specificity

mean (SD)

Sensitivity

mean (SD)

p-value

Internal validation using the development dataset

 ViT-PMC-LoRA + CF

0.8307 (0.0372)

0.8016 (0.1033)

0.7418 (0.1229)

0.179

 ViT-PMC-LoRA

0.8404 (0.0312)

0.8557 (0.0953)

0.7012 (0.1154)

0.442

 ViT-PMC-LoRA + AP + CF

0.8502 (0.0297)

0.8234 (0.0720)

0.7773 (0.0594)

0.775

 ViT-PMC-LoRA + AP

0.8539 (0.0445)

0.8230 (0.0893)

0.7739 (0.0796)

-

External validation using the test dataset

 ViT-PMC-LoRA + CF

0.8103 (0.0169)

0.6741 (0.0969)

0.8611 (0.0878)

< 0.001*

 ViT-PMC-LoRA

0.8113 (0.0519)

0.6963 (0.1155)

0.8111 (0.0915)

0.011*

 ViT-PMC-LoRA + AP + CF

0.8566 (0.0246)

0.7630 (0.0804)

0.8167 (0.0695)

0.172

 ViT-PMC-LoRA + AP

0.8656 (0.0137)

0.8111 (0.1010)

0.7611 (0.1173)

-

  1. Each configuration is compared against the best-performing configuration, ViT-PMC-LoRA + AP.
  2. Bonferroni correction for multiple comparisons across 3 tests was applied for internal and external validation, respectively. Thus, p-values for area under the curve (AUC) are considered statistically significant if less than 0.017. AUC Area under the curve, SD standard deviation, CNN convolutional neural network, ViT vision transformer, LoRA Low-Rank Adaptation, AP, augmented prediction, CF, incorporation of clinical features.
  3. *Indicates a statistically significant difference. The best value for each column is marked in bold.
Back to article page