Table 2.
Glaucoma progression prediction AUCs using earlier AI-based methods as compared with our proposed method.
| Modeling approach | Inputs used | AUC (95% CI) | P value |
|---|---|---|---|
| Traditional machine learning models | VF18,20 | 0.71 | – |
| OCT + Baseline22 | 0.61 | – | |
| VF + OCT25 | 0.58 | – | |
| Deep learning models | VF + Baseline9 | 0.72 | – |
| Single modal deep learning | OCT + Baseline [Current work] | 0.68 (0.62, 0.74) | < 0.001 |
| Single modal deep learning | VF + Baseline [Current work] | 0.72 (0.66, 0.77) | 0.003 |
| Multimodal deep learning | OCT + VF + Baseline at M3 [Current work] | 0.76 (0.71, 0.80) | 0.038 |
| Multimodal deep learning with generative model | OCT + VF + Baseline at M3 + Future Synthetic OCT [Current work] | 0.81 (0.79, 0.84) | Reference |
| Multimodal deep learning with generative model | OCT + VF + Baseline at M6 + Future Synthetic OCT [Current work] | 0.83 (0.80, 0.86) | 0.264 |
Here “Baseline” inputs refer to the demographic and clinical parameters at the first visit as listed in Table 1, i.e. AGE, GENDER, BCVA, REFR, CCT, AXL, RNFL, IOP and VF MD. AUCs and 95% CI are listed for results achieved by our model using different input combinations, where the statistical significance was determined by performing t-test and computing the P values. We considered a P value of less than 0.05 as statistically significant.