Using an extreme phenotype-derived machine learning framework, the authors identified systemic susceptibility and resilience signatures associated with severe diabetic retinopathy, with renal dysfunction, albuminuria, glycemic burden, and disease duration emerging as key phenotype-defining characteristics.
Key Findings
Results
LightGBM demonstrated the strongest ability to discriminate between susceptible and resilient diabetic retinopathy phenotypes among evaluated algorithms.
Achieved an area under the receiver operating characteristic curve (AUC) of 0.90 (95% CI, 0.84–0.95) in the internal validation cohort.
The extreme phenotype cohort comprised 712 individuals with diabetes mellitus: 437 with proliferative diabetic retinopathy (PDR; susceptible phenotype) and 275 with diabetes duration ≥10 years without retinopathy (resilient phenotype).
LASSO regression identified 21 phenotype-associated features used for model development.
Multiple machine learning algorithms were evaluated before LightGBM was selected as the optimal model.
Results
SHAP analysis identified urinary albumin excretion rate (UAER), diabetes duration, serum creatinine, total protein, age, and hypertension duration as the dominant phenotype-defining features.
SHapley Additive exPlanations (SHAP) was used to interpret the optimal LightGBM model.
UAER was identified as the single most dominant feature in discriminating susceptible from resilient phenotypes.
Serum creatinine and total protein, markers of renal and systemic metabolic status, were among the top contributors.
Age and hypertension duration were also identified as major contributors to phenotype definition.
Results
UAER exhibited a pronounced nonlinear association with susceptibility scores, suggesting a close link between renal microvascular injury and vulnerability to severe diabetic retinopathy.
The nonlinear relationship between UAER and susceptibility scores was revealed through SHAP analysis.
The finding supports the concept that renal and retinal microvascular injury share common pathophysiological pathways.
This nonlinear pattern implies that the relationship between albuminuria and DR susceptibility is not simply proportional but may involve threshold or accelerating effects.
Results
The susceptibility signature showed limited discrimination in a community-based diabetic cohort overall but became progressively enriched among individuals with greater metabolic burden.
An independent community-based diabetic cohort of n=673 was used for external evaluation.
Overall AUC in the community cohort was 0.54, indicating limited discrimination in the general diabetic population.
AUC increased to 0.71 among participants with fasting blood glucose ≥9.0 mmol/L.
This progressive enrichment with metabolic stress supports the concept that severe DR arises through interaction between intrinsic biological susceptibility and cumulative metabolic exposure.
Results
Renal dysfunction, albuminuria, glycemic burden, and disease duration emerged as key phenotype-defining characteristics distinguishing susceptible from resilient individuals.
The susceptible phenotype was defined as proliferative diabetic retinopathy (PDR).
The resilient phenotype was defined as diabetes duration ≥10 years without any retinopathy.
These features collectively reflect cumulative systemic metabolic exposure rather than any single biomarker.
The framework highlights disease heterogeneity, suggesting that not all individuals with equivalent diabetes duration face equivalent retinopathy risk.
Methods
The extreme phenotype sampling strategy was used to maximize biological contrast between susceptibility and resilience to severe diabetic retinopathy.
The cohort was intentionally constructed using extreme phenotypes (PDR vs. long-standing diabetes without retinopathy) to amplify signal for feature identification.
This design choice explains the high internal AUC of 0.90 and the lower AUC of 0.54 when applied to an unselected community cohort.
The framework was described as providing 'new insights into disease heterogeneity' and facilitating 'future precision risk stratification strategies in diabetic eye disease.'
LASSO regression was applied prior to machine learning model development to reduce dimensionality from the full feature set to 21 variables.
What This Means
This research suggests that people with diabetes who develop the most severe form of diabetic eye disease (called proliferative diabetic retinopathy) have distinct biological and clinical characteristics compared to people who have lived with diabetes for a decade or more without any eye complications. The researchers used a machine learning approach that deliberately compared these two extreme groups — those most vulnerable and those most resilient — to identify what sets them apart. They found that kidney-related markers (especially urinary albumin excretion and serum creatinine), how long someone has had diabetes, their age, and how long they have had high blood pressure were the most important factors in distinguishing susceptible from resilient individuals. A notable finding was that the relationship between kidney protein leakage (albuminuria) and retinopathy risk was nonlinear, meaning risk did not simply increase steadily but may jump at certain thresholds, pointing to a shared vulnerability between kidney and eye small blood vessels.
When the model trained on these extreme cases was tested on a broader community population of people with diabetes, it performed modestly overall, but its accuracy improved substantially among people with high blood sugar levels (fasting glucose ≥9.0 mmol/L). This research suggests that the biological signatures of severe diabetic eye disease become most detectable in people who already have a heavy metabolic burden, supporting the idea that severe retinopathy results from a combination of underlying biological susceptibility and accumulated exposure to poor metabolic control over time.
Practically, this research suggests that routine clinical measurements — particularly kidney function tests, urine albumin levels, and blood sugar control — may help identify which people with diabetes are at highest risk of losing vision from retinopathy. The authors propose that this type of machine learning framework could contribute to more personalized risk assessment in diabetic eye disease, potentially allowing for more targeted monitoring or intervention in those with the greatest underlying susceptibility.
Zhang J, Shu Y, Yang F, Cai S, Xie R, Sun R, et al.. (2026). Extreme phenotype-derived machine learning reveals susceptibility and resilience signatures for severe diabetic retinopathy.. Frontiers in endocrinology. https://doi.org/10.3389/fendo.2026.1933878