Advancing predictive, preventive, and personalized medicine in eyelid diseases: a concerns-based and expandable screening system through structural dissection

Abstract

Background/aims: Early recognition of eyelid morphological abnormalities was crucial, as untreated conditions could lead to blinding complications. An eyelid screening system that could provide both anatomical and pathological information was essential for formulating personalized treatment strategies. This study aimed to develop a clinically concerns-based framework capable of identifying common eyelid diseases requiring further intervention by evaluating individual anatomical and pathological changes. This approach would enhance individualized and efficient prevention, while supporting targeted treatment strategies.

Methods: The eyelid disorder screening system, Eyetome, was developed based on a morphological atlas and comprised four modules designed to identify 14 common eyelid disorders and pathological changes. A total of 6180 eye patches were analyzed to extract anatomical and pathological features. The performance of Eyetome was evaluated using average accuracy (aACC) and F1 score, with comparisons made against traditional models and ophthalmologists. To assess the system's expandability, an additional test was conducted in a multimorbidity scenario.

Results: Eyetome demonstrated high performance in recognizing single diseases, achieving an aACC of 98.83% and an F1 score of 0.93. The system outperformed classic models, with an aACC of 98.83% compared to 96.72% for Desnet101 and 97.59% for Vit. Additionally, Eyetome's aACC exceeded that of a junior ophthalmologist (JO) (97.11%) and was comparable to a senior ophthalmologist (SO) (98.69%). In the extended multimorbidity dataset, Eyetome maintained robust performance with an accuracy of 97.97%, surpassing JO (95.47%) and closely matching SO (97.81%).

Conclusions: This study developed a clinical concerns-based system for screening and monitoring eyelid disorders, aimed at supporting predictive diagnosis, preventing diseases progression, and facilitating more effective, patient-centered treatment of common eyelid disorders, aligning with the principles of predictive, preventive, and personalized medicine (PPPM/3PM). The system's interpretability, scalability, and user-friendly data acquisition design could further enhance its acceptance among both doctors and patients, facilitating the shift from reactive medicine to proactive precision medicine.

Supplementary information: The online version contains supplementary material available at 10.1007/s13167-025-00401-y.

Keywords: Blepharoptosis; Ophthalmology; Deep learning; Expandability; Eyelid disorders; Eyelid tumors; Eyetome; Interpretability; Monitoring; Predictive preventive personalized medicine (PPPM / 3PM); Screening; Secondary targeted prevention; Thyroid eye disease.

Fig. 1

The architecture of the eyelid morphological atlas, which served as the foundational principle of Eyetome. The atlas was constructed based on the logical relationship between pathological changes and anatomical regions, enabling the precise diagnosis of eyelid diseases. Utilizing the morphological atlas, Eyetome firstly identified the anatomical regions and then discerned the pathological changes in each region to arrive at a final diagnosis

Fig. 2

The main pipeline of the study. Eyetome was developed using the eyelid morphological atlas, which integrated pathological changes and anatomical regions within a clinically logical framework. Comparative tests with image-level models and human specialists were conducted to evaluate Eyetome’s stability and precision. An expanded test in multimorbidity scenarios was performed to further examine the system’s scalability and adaptability

Fig. 3

Receiver operating characteristic (ROC) curves of the Eyetome classifiers, including the inner canthus classifier (a) and the eyelid classifier (b), alongside those of image-level multilabel models, including Desnet101 (c) and ViT (d)

Fig. 4

Comparison of accuracy, sensitivity, specificity, and F1 score between Eyetome, multilabel ViT, multilabel Desnet101, a junior ophthalmologist, and a senior ophthalmologist in identifying common eyelid diseases

Fig. 5

Performance comparison of Eyetome, junior ophthalmologist (JO), and senior ophthalmologist (SO) in diagnosing multimorbidity. The log10 fold-change heatmap (a) illustrates the distribution of correct diagnoses, misdetections, over-detections, and under-detections by Eyetome, JO, and SO across four multimorbidity classes. Log10 fold-change was calculated based on the proportion of correct diagnoses, misdetections, over-detections, and under-detections in each class, as indicated in the heatmap. Eyetome demonstrated higher accuracy in identifying multimorbidity compared to JO and achieved performance comparable to SO (b). *p < 0.05. BPES, blepharophimosis/ptosis/epicanthus inversus syndrome