The authors evaluated whether fluorescence lifetime imaging ophthalmoscopy (FLIO) combined with deep learning can detect retinal metabolic signatures associated with glycemic status and classify type 2 diabetes mellitus (T2DM). The analysis is a cross-sectional assessment of participants from the AI-READI dataset (version 3) who had both FLIO imaging and hemoglobin A1c (HbA1c) measurements. The aim was to determine whether mean fluorescence lifetime maps derived from FLIO could serve as inputs for AI models to distinguish normoglycemic, prediabetic, and diabetic participants.
The study used 1,783 participants from the AI-READI dataset (version 3) with available HbA1c and FLIO scans. In total there were 6,912 FLIO scans across participants. Participants were categorized by glycemic status into three groups: 671 normoglycemic, 726 prediabetic, and 386 diabetic. The authors state that the original AI-READI study was approved by an institutional review board and that participants provided informed consent; this secondary analysis did not require additional IRB approval according to the authors.
Mean fluorescence lifetime maps were created using a center-of-mass approach applied to the FLIO data. These lifetime maps were used as the primary imaging inputs for the AI models. The authors report that group-averaged lifetime maps showed consistent spatial differences across glycemic groups, with a trend toward progressively longer lifetimes from normal to diabetic participants, suggesting that FLIO captures metabolic signals related to glycemic status.
Three types of models were trained: a convolutional neural network (CNN), a ResNet-18 architecture, and XGBoost. Training and evaluation employed nested 5-fold cross-validation with participant-level grouping to avoid leakage of data from the same participant across folds. The modeling strategy included both multi-class and binary classification formulations, and standard performance metrics were calculated.
The authors evaluated two main task formulations:
Primary outcome measures reported were macro-averaged accuracy, F1 score, area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and positive predictive value (PPV). A random classifier baseline for three-class classification was used for comparison (AUROC = 0.50; accuracy = F1 = 0.33).
Group-averaged FLIO lifetime maps demonstrated spatial patterns that differed across glycemic groups. The authors observed progressively longer fluorescence lifetimes when comparing normal to prediabetic to diabetic groups. These observations support that FLIO-derived lifetime metrics contain metabolic information associated with glycemic status.
For the three-class task the best-performing model reported was the CNN. Performance metrics for the CNN were modest: macro-averaged accuracy 0.41 ± 0.03, F1 score 0.39 ± 0.02, and AUROC 0.58 ± 0.02. ResNet-18 and XGBoost produced similar AUROC values in the range of 0.53–0.58. Compared with the random classifier baseline (accuracy and F1 = 0.33; AUROC = 0.50), the trained models showed some improvement but overall limited discrimination among the three classes.
Recasting the problem as a binary classification improved reported performance. In the normal versus diabetic binary task the CNN achieved AUROC 0.63 ± 0.02, while XGBoost reached AUROC 0.67 ± 0.07. The authors report that binary reformulation, particularly normal vs. diabetic, yielded substantially better discrimination than the three-class formulation, although performance remained modest.
Analysis of confusion matrices revealed substantial overlap among classes in the three-class task. A frequent pattern was misclassification toward the prediabetes group, indicating that the models often labeled participants as intermediate even when they belonged to the normal or diabetic groups. This trend underscores difficulty in precisely separating adjacent glycemic categories using current FLIO-derived inputs and modeling approaches.
The authors conclude that FLIO-derived lifetime maps capture metabolic signals associated with glycemic status but that current AI models yield only modest classification performance. The findings indicate both potential for FLIO as a noninvasive metabolic imaging modality and clear challenges before it could be applied as a clinical screening or monitoring tool. Limitations implied by the results include overlap between classes, modest AUROC values even for the best-performing binary models, and model confusion around the prediabetic category. The study is reported as a preprint and has not undergone peer review; authors note that the data are available upon reasonable request.
The authors state that all data produced in the study are available upon reasonable request. The original AI-READI study had prior institutional review board approval and participant informed consent; the present analysis did not require additional IRB approval per the authors. A.Y.L. discloses multiple industry relationships and non-financial support from several companies outside the submitted work; the remaining authors declare no competing interests.