Table 2 Associations between risk identification model and participant outcomes
From: A deep learning system for predicting time to progression of diabetic retinopathy
Eyes with DR progression incidence per 1,000 eye-years (number of cases/number of eyes) | ARRa (95% CI) | |||
|---|---|---|---|---|
Integrated hospital–community diabetes management program | IM group (n = 1,076) | DeepDR Plus-low risk (AI-low) | Metadata-low risk (meta-low) | −33.05 (−67.79, 35.76) |
5.11 (16/626) | 7.63 (24/629) | |||
DeepDR Plus-high risk (AI-high) | Metadata-high risk (meta-high) | 14.54 (−28.26, 74.63) | ||
26.67 (60/450) | 23.27 (52/447) | |||
Non-IM (n = 3,294) | DeepDR Plus-low risk (AI-low) | Metadata-low risk (meta-low) | −91.63 (−93.91, −89.06) | |
5.01 (50/1,996) | 11.34 (113/1,993) | |||
DeepDR Plus-high risk (AI-high) | Metadata-high risk (meta-high) | 61.36 (25.36, 109.91) | ||
33.13 (215/1,298) | 23.37 (152/1,301) | |||
Comprehensive interventions: [(AI-high + AI-low) − (meta-high + meta-low)] in IM group − [(AI-high + AI-low) − (meta-high + meta-low)] in non-IM group | 46.80 (12.37, 94.93) | |||
Sankara Nethralaya-Diabetic Retinopathy Epidemiology and Molecular Genetics Studyb | IM group (n = 146) | DeepDR Plus-low risk (AI-low) | Metadata-low risk (meta-low) | −9.39 (−79.77, 287.41) |
4.08 (2/98) | 4.49 (2/89) | |||
DeepDR Plus-high risk (AI-high) | Metadata-high risk (meta-high) | 20.48 (−70.93, 400.0) | ||
25.0 (6/48) | 21.05 (6/57) | |||
Non-IM group (n = 1,798) | DeepDR Plus-low risk (AI-low) | Metadata-low risk (meta-low) | −97.32 (−98.28, −96.32) | |
5.24 (28/1,068) | 13.0 (70/1,077) | |||
DeepDR Plus-high risk (AI-high) | Metadata-high risk (meta-high) | 43.13 (9.1, 87.18) | ||
44.11 (161/730) | 33.01 (119/721) | |||
Comprehensive interventions: [(AI-high + AI-low) − (meta-high + meta-low)] in IM group − [(AI-high + AI-low) − (meta-high + meta-low)] in non-IM group | 88.74 (10.83, 330.25) | |||
