A Deep Learning Prognostic Model for Diabetes Patients Using Bilateral Fundus Imaging

Abstract

Aims:

Diabetes mellitus shortens life expectancy, driven primarily by premature mortality from vascular complications. Mortality models for intensive care units are well established, whereas outpatient mortality prediction remains challenging. To empower proactive, life-extending care in aging societies, we developed a prognostic model that utilizes routine bilateral color fundus photography for identifying high-mortality-risk patients years before irreversible complications, enabling timely interventions that can substantially improve survival.

Methods:

We analyzed 19 029 persons with diabetes who received routine color fundoscopy examinations. We adapted deep learning architectures to align with Cox proportional hazards modeling of mortality events, as opposed to simple classification.

Results:

Our image-trained mortality model stratified the testing cohort into quartiles exhibiting escalating hazard ratios (HRs) of 1.67 (95% confidence interval [CI]: 0.80-3.47), 3.02 (1.55-5.88), and 7.19 (3.91-13.25) relative to the lowest-risk group. The five-year survival rate of the highest-risk quartile was lower than 80%. For comparison, we evaluated a regulatory-approved commercial ophthalmic system, which generated HRs of 1.48, 2.02, and 3.11 across its strata. When evaluated in strata of equivalent sizes to those used by the commercial system, our system demonstrated stronger separation than the commercial system, yielding HRs of 3.64 (95% CI: 3.18-4.16), 4.98 (4.08-6.07), and 6.89 (6.01-7.89). The top 8% patients identified by our model exhibited a five-year survival rate <70%.

Conclusion:

We developed a survival-aware, image-based model that can predict mortality risk directly from routine fundus photography, achieving prognostic discrimination which facilitates diabetes management.