Developing and validating an explainable clinlabomics-based machine-learning model for screening primary angle-closure glaucoma in the context of PPPM

Abstract

Background

Primary angle-closure glaucoma (PACG) is a common cause of blindness. Early screening is critical to prevent vision loss, yet current methods rely on specialized ophthalmic imaging, which are resource-intensive and reactive, detecting structural damage only after symptom onset. Therefore, we propose a novel clinlabomics-based machine learning prediction model as a screening tool to stratify individuals at high risk for glaucoma, enabling targeted ophthalmic evaluations, preventing progression of optic nerve damage, and facilitating personalized, long-term monitoring in alignment with the principles of predictive, preventive, and personalized medicine (PPPM/3PM).

Methods

This is a multicenter, retrospective study. We retrieved clinical laboratory data from digital medical records between April 2016 and April 2021 in the Eye and ENT Hospital of Fudan University as a discovery set, consisting of 949 normal subjects and 1152 PACG patients. The internal validation was conducted on the dataset of 646 normal subjects and 657 PACG patients from June 2021 to October 2024, also from the Eye and ENT Hospital of Fudan University; the external validation was performed on a dataset of 246 normal subjects and 136 PACG patients from March 2023 to June 2024, from Shanghai Xuhui Central Hospital and Wanbei Coal Electric Group General Hospital. Based on whether there was optic nerve damage, patients were categorized into early PACG patients, namely primary angle closure(PAC) patients, and non-early PACG. Specifically, in the internal validation cohort of 657 PACG patients, 160 were PAC. In the external validation cohort of 136 PACG patients, 41 were PAC. With the inclusion of 50 features, 12 machine learning models were selected and compared to develop the screening model. The feature reduction was performed by SHAP model and Delong test, and the final model was explained by SHAP method. The evaluation parameters of the models include AUC, AUCPR, sensitivity, specificity, and accuracy.

Results

A total of 1841 normal subjects and 1945 PACG patients were included in the study. Among the 12 machine learning models, 4 models, LGBM (AUC = 0.92), XGB (AUC = 0.92), Ada (AUC = 0.91), and GB (AUC = 0.91), performed better than others (P > 0.05). After feature reduction based on feature importance ranking, a final LGBM model of accurate screening PACG ability with six features including TT, PDW, MCV, APTT, TC, and PT was developed, achieving AUC of 0.91, AUCPR of 0.94, sensitivity of 0.89, specificity of 0.79, PPV of 0.84, NPV of 0.85, accuracy of 0.84, and F1 score of 0.86. This final model maintained strong performance in internal validation (AUC = 0.87, accuracy = 0.83, F1 score = 0.85) and external validation (AUC = 0.85, accuracy = 0.89, F1 score = 0.84). The screening efficacy of the final model for PAC was also assessed, where the ROC was 0.85 in the internal validation and 0.84 in the external validation. To enhance its practical application and dissemination, the final model was transformed into an accessible web application.

Conclusion

This study establishes a clinically applicable clinlabomics-based model that implements PPPM principles for glaucoma management through routine blood parameters. Our predictive model enables early identification of high-risk PACG patients, while also facilitating cost-effective population screening and personalized risk assessment through explainable artificial intelligence. The current study demonstrates that routine blood parameters serve as critical indicators for glaucoma risk stratification, predictive diagnosis, and targeted intervention. Consequently, this innovative screening approach provides an essential tool for optimizing clinical outcomes in high-risk populations and improving glaucoma care accessibility, particularly in underserved communities with limited ophthalmic resources.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13167-025-00419-2.

Introduction

The global burden of PACG and the need for PPPM-driven innovation

Glaucoma, the second leading cause of global blindness and the most common irreversible one [1, 2], poses a significant threat to public health. Primary angle-closure glaucoma (PACG), a critical subtype, is characterized by closure of the atrial angle, elevated intraocular pressure, optic disc changes, and visual field damage [3]. With the aging of the population, the number of PACG patients is expected to increase to 32.04 million globally in 2040, with Asia bearing 77% of the burden.Primary angle closure (PAC), characterized by anatomical vulnerability and elevated intraocular pressure (IOP) in the absence of glaucomatous optic neuropathy, is defined as an early stage of PACG, where timely diagnosis is critical to halt progression to sight-threatening glaucoma. Therefore, implementing the principles of predictive, preventive, and personalized medicine (PPPM) in glaucoma care is highly significant [4, 5].

Current screening methods, like slit lamp microscopy [6], intraocular pressure measurement [7], gonioscopy [8], and funduscopic examination [9], are resource-intensive and often detect damage only after symptoms emerge.This reactive approach highlights the urgent need for predictive tools enabling early risk stratification and targeted prevention, especially in resource-limited regions with high PACG prevalence. The integration of artificial intelligence (AI) in ophthalmology shows great promise [10]. AI combined with fundus photography [11, 12] and OCT images [13, 14] has advanced glaucoma screening and diagnosis. Yet, existing AI approaches are limited by reliance on late-stage anatomical data and high-cost infrastructure, restricting their application in areas with scarce specialized ophthalmic resources. Furthermore, these approaches often suffer from poor interpretability [15], referred to as the “black-box” effect [16], which impedes clinical adoption and integration into routine practice. To overcome these limitations and advance the implementation of PPPM in ophthalmology, there is a pressing need for innovative screening tools that are not only accurate and reliable but also cost-effective, accessible, and interpretable. Such tools should enable early detection of high-risk individuals before irreversible vision loss occurs, facilitate personalized risk assessment, and support targeted, preventive interventions.

The rise of clinlabomics and its alignment with PPPM principles

The position paper of EPMA and EFLM emphasized the central role of laboratory medicine in healthcare services [17, 18].Clinlabomics, an emerging field that integrates artificial intelligence with clinical laboratory data, offers a novel approach to optimizing the utilization of routine medical testing data for disease prediction, screening, and diagnosis [19]. By moving beyond one-size-fits-all reference intervals, routine blood parameters can reveal intricate pathophysiological signatures indicative of various disease risks [20]. This approach not only facilitates precision medicine implementation but also enables cost-effective, time-efficient, and accessible dynamic disease monitoring. The occurrence and development of PACG involve a complex interplay of pathogenic processes, including inflammation [21, 22], blood circulation abnormalities [23, 24], metabolic disorders [25–27], and oxidative stress [28, 29]. These systemic alterations leave tell-tale signatures in hematological parameters, biochemical markers, and coagulation profiles [30].The combination of systematic markers such as routine blood indicators with machine learning has become one of the hot topics in the context of PPPM [31].

The EPMA has emphasized the importance of integrating laboratory medicine into healthcare service to advance PPPM, highlighting that laboratory data can serve as a cornerstone for predictive and preventive strategies [32]. In our prior work, routine blood parameter-based models demonstrated utility in continuous monitoring and predicting retinal detachment in high myopia [33]. However, the application of clinlabomics, a novel concept, in ophthalmology still requires further exploration.

Working hypothesis

In this study, we aim to develop and validate an explainable machine learning model based on clinlabomics for screening PACG patients within the context of PPPM. We hypothesize that machine-learning screening models based on clinlabomics can enhance ocular health by enabling early detection of PACG, facilitating the development of personalized, cost-effective preventive and therapeutic interventions. Furthermore, we aim to enhance the interpretability of the model through the SHapley Additive explanation (SHAP) method and transform it into a publicly accessible visualization tool. This approach will enable individualized risk profiling and support cost-effective interventions for prevention. By validating our hypothesis and objectives through a cross-sectional investigation encompassing 1841 normal subjects and 1945 PACG patients, with both internal and external validations performed, we seek to advance the implementation of PPPM in ophthalmology and drive a paradigm shift in glaucoma care.

This study represents a significant step towards catalyzing a paradigm shift in healthcare. By establishing a clinlabomics-based, explainable machine learning model for PACG screening, we provide a powerful example of how routine clinical laboratory data and AI can be integrated to develop innovative screening tools. This approach aligns with the PPPM vision of proactive, personalized healthcare. It empowers clinicians to identify high-risk individuals early, initiate timely interventions, and dynamically monitor disease progression. This innovative screening approach not only addresses existing limitations in PACG detection but also actively drives the healthcare paradigm shift envisioned by PPPM. By democratizing screening in regions lacking specialized ophthalmological infrastructure, our model enables early intervention where it is most urgently needed. This study exemplifies how the integration of AI with routine clinical laboratory data can transform healthcare delivery, moving the entire system closer to the ideal of predictive, preventive, and personalized medicine. Through this work, we aim to contribute to the advancement of PPPM in ophthalmology and pave the way for a future where irreversible blindness is preemptively averted through prediction, prevention, and personalization.

Methods

Study design and population

In this multicenter retrospective study, we compared among 12 machine learning (ML) models to develop and validate a clinlabomics-based machine learning model for screening PACG patients through model evaluation and feature reduction. This work aligns with PPPM principles, aiming to shift from reactive to proactive healthcare. This study was authorized by the Ethics Committee of the Eye and ENT Hospital of Fudan University and followed the principles of the Declaration of Helsinki. Informed consent was obtained from all participants before their enrollment. A comprehensive ophthalmologic examination was conducted for all patients, as previously described [34, 35] and detailed in the Supplementary Materials. The diagnostic, inclusion, and exclusion criteria for PACG were as previously reported [36, 37] and are further explained in the Supplementary Materials.

A total of 949 control subjects and 1152 PACG patients were recruited from the Eye and ENT Hospital of Fudan University, Shanghai, China, from April 2016 to April 2021 as the training set for this study. Meanwhile, the internal validation set was composed of 646 control subjects and 657 PACG patients between June 2021 and October 2024, from the Eye and ENT Hospital of Fudan University. The external validation was from Shanghai Xuhui Central Hospital and Wanbei Coal Electric Group General Hospital with 261 control subjects and 144 PACG patients from March 2023 to June 2024. Patients were classified into two categories according to whether optic nerve damage was present: early PACG patients, also identified as PAC patients, and non-early PACG patients. Specifically, among the 657 PACG patients in the internal validation cohort, 160 were PAC. In the external validation cohort, which consisted of 136 PACG patients, 41 were PAC.

Date acquisition

In this multicenter cohort study, data were retrieved from hospital electronic medical records. The contents of the electronic medical record included demographic data (age and gender) and clinical laboratory data. Such comprehensive data collection aligns with the principles of PPPM, as it allows for a detailed assessment of the patients’ systemic health status. A total of 50 variables were to be collected for each patient, and they included 24 variables of routine blood analysis, including basophil% (BASO%), basophil (BASO), eosinophil% (EO%), eosinophil (EO), hematocrit (HCT), hemoglobin (HGB), lymphocyte% (LYM%), lymphocyte (LYM), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), mean corpuscular volume (MCV), mean platelet volume (MPV), monocyte% (MONO%), monocyte (MONO), neutrophil% (NEU%), neutrophil (NEU), platelet count (PLT), platelet distribution width (PDW), platelet large cell ratio (PLCR), red blood cell distribution width-coefficient of variation (RBCCV), red blood cell distribution width-standard deviation (RBCSD), red blood count (RBC), thrombocytocrit (PCT), and white blood cell count (WBC); 19 variables of biochemistry analysis, including AG, albumin (ALB), alkaline phosphatase (ALP), blood urea nitrogen (BUN), creatine kinase (CK), creatinine (CREA), direct bilirubin (DBIL), gamma-glutamyl transpeptidase (GGT), globulin (GLB), aspartate aminotransferase (AST), alanine aminotransferase (ALT), lactic dehydrogenase (LDH), total bile acid (TBA), total bilirubin (TBIL), total cholesterol (TC), total protein (TP), triglyceride (TG), uric acid (UA), and glucose (GLU); and 7 variables of blood coagulation analysis, including activated partial thromboplastin time (APTT), d-dimer (DD), fibrinogen (FIB), international normalized ratio (INR), prothrombin time (PT), PT%, and thrombin time (TT). The three types of clinical laboratory data were collected as previously described [34].

Model development

The aforementioned 50 features were used as continuous variables. Missing values were filled in using median interpolation. The modeling process mainly includes two processes: model comparison and feature reduction. This rigorous comparison process is designed to identify the model with the best predictive performance for PACG screening, in line with PPPM’s emphasis on selecting the most appropriate tools for preventive healthcare.

Model comparison

A total of 12 machine learning models were employed for screening PACG, including adaptive boosting (AdaBoost), generalized linear model (GLM), decision tree (DT), gradient boosting machine (GBM), K-nearest neighbor (KNN), light gradient boosting machine (LGBM), logistic regression (LR), MLPClassifier (MLP), Naive Bayes (NB), random forest (RF), support vector machine (SVM), and eXtreme gradient boosting (XGB).

In the discovery cohort, fivefold cross-validation was employed to assess each model. The screening efficacy of the models was compared using several metrics: the area under the receiver-operating-characteristic (ROC) curve (AUC), the area under the precision-recall curve (AUCPR), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1 score. Furthermore, pairwise comparisons of the AUC values for the 12 models were performed using the DeLong test to determine statistical significance.

Feature reduction

SHapley Additive explanation (SHAP) is a tool for interpreting the output of machine learning models [38]. It solves the problem of complex models that have difficulty explaining their internal decision-making processes by attributing importance values to features, thereby elucidating the model’s output.

Feature reduction was performed within the cross-validation loop to avoid overfitting.With the function of the SHAP model that can rank the importance of features, the fivefold cross-validated AUC of models with varying numbers of features is compared in combination with the Delong test [39] to determine the final model with the least number of features. This not only streamlines the model for practical clinical application but also aligns with PPPM’s goal of creating efficient and effective predictive tools that can be seamlessly incorporated into healthcare practice.

Model validation

We performed internal and external validation in this study. We developed an internal validation dataset from the Eye and ENT Hospital of Fudan University, which included 646 control subjects and 657 patients with PACG to validate the final model after reduction of features. The efficacy of the model to screen PACG was evaluated by AUC, AUCPR, specificity, sensitivity, NPV, PPV, accuracy, and F1 score in Pycharm software.

The external validation was conducted in a manner analogous to the internal validation process. The external validation cohort consisted of 261 controls and 144 PACG patients recruited from Shanghai Xuhui Central Hospital and Wanbei Coal Electric Group General Hospital. The final model’s performance was assessed using a comprehensive set of metrics, including AUC, AUCPR, specificity, sensitivity, NPV, PPV, accuracy, and F1 score.

To demonstrate the early screening efficacy of the final model in PACG, we conducted internal and external validation specifically for diagnosing those PAC patients. Specifically, in the internal validation cohort of 657 PACG patients, 160 were PAC. In the external validation cohort of 136 PACG patients, 41 were PAC. The evaluation of the final model was the same as that described above.

Model explanation and application

The inherent “black box” effect of machine learning models can limit their clinical application. In this study, the SHAP model provides both global and local interpretation of the model. It can both visualize which features have a greater impact on PACG and help clinicians understand how the model can screen a single patient as a normal person or a PACG patient. By bridging the gap between complex machine learning algorithms and clinical practice, the SHAP model enhances the clinical utility of our approach, matching the PPPM emphasis on clear and practical clinical tools.

To make the model more accessible and advance the equitable healthcare goals of PPPM, the final PACG screening model has been implemented as a web-based application. When visitors enter the values of the relevant features into the application, it calculates and outputs the risk of PACG in real time. This approach simplifies the screening process, enhances user experience, and democratizes access to advanced diagnostic tools. By providing an easy-to-use platform for risk assessment, the web application supports the early detection and prevention of PACG, making predictive healthcare more accessible to diverse populations and advancing the shift toward proactive, patient-centered medicine.

Statistical analysis

Data analysis was performed using Python version 3.6.5 (https://www.python.org), PyCharm 2024.2.4 (https://www.jetbrains.com/pycharm), and IBM SPSS Statistics version 23.0 (https://www.ibm.com/spss). All variables were subjected to descriptive statistical analysis and Shapiro–Wilk normality tests. Continuous variables were analyzed using independent samples t-tests for normally distributed data and Mann–Whitney U tests for non-normally distributed data. Continuous variables were reported as mean ± standard deviation (mean ± SD). Categorical variables were expressed as counts and percentages, with comparisons made using chi-square tests. The screening performance of the model was evaluated using metrics such as AUC, AUCPR, and others. A P value of less than 0.05 was considered statistically significant for all results.

Results

Patient characteristics

The study design is illustrated in Fig. 1. Demographic indicators and characteristics of the discovery and validation sets are shown in Table 1, Table S1, and Table S2. The number of individuals in the control and PACG groups was matched for age and sex (P > 0.05) in all cohorts. In the discovery set, the PACG group consisted of 1152 patients, whose mean age was 67.73 ± 9.1, and approximately 39.5% of the patients were male. Most features were different in the control and PACG groups (P < 0.05), such as TP, TC, PLT, PDW, and LDH. Spearman’s correlation analysis was performed on the 50 variables two by two, and the results are shown in Fig. 2A. The Spearman’s correlation heatmap shows that some of the variables like TP and GLB are positively correlated, while others like AG and GLB are negatively correlated.

Fig. 1.

Study flow chart. This study involves three cohorts: the discovery set, the internal validation set, and the external validation set.The main parts include model development, model validation, model explanation, and application. PACG, primary angle-closure glaucoma; SVM, support vector machine; Ada, adaptive boosting; RF, random forest; LR, logistic regression; GLM, generalized linear model; XGB, eXtreme gradient boosting; NB, Naive Bayes; GB, gradient boosting machine; MLP, MLPClassifier; DT, decision tree; KNN, K-nearest neighbor; LGBM, light gradient boosting machine; ML, machine learning; SHAP, SHapley Additive explanation; DCA, decision curve analysis

Table 1.

The clinical and demographic characteristics of all subjects in the discovery set

Variables	Normal (n = 949)	PACG (n = 1152)	P value
Age,year	67.75 ± 6.1	67.73 ± 9.1	0.971
Sex(male,%)	390 (41.1)	455 (39.5)	0.457
TP,g/L	73.45 ± 4.60	74.23 ± 5.12	< 0.001
TBA, μmol/L	4.59 ± 5.01	5.57 ± 6.02	< 0.001
TBIL, μmol/L	11.20 ± 5.21	12.01 ± 5.80	< 0.001
TC, mmol/L	4.79 ± 0.88	4.97 ± 1.05	< 0.001
Neutrophil, × 109/L	3.71 ± 1.37	3.96 ± 1.42	< 0.001
ALB,g/L	45.57 ± 3.59	45.63 ± 3.41	0.944
neutrophil%	61.27 ± 11.88	61.59 ± 9.11	0.36
AG, mmol/L	1.68 ± 0.33	1.65 ± 0.36	0.13
RBC, × 1012/L	4.42 ± 0.43	4.38 ± 0.44	0.21
PCT,ng/ml	0.21 ± 0.05	0.22 ± 0.05	0.02
PLT, × 109/L	203.38 ± 54.66	220.30 ± 54.97	< 0.001
PDW, fL	13.77 ± 2.72	11.36 ± 1.77	< 0.001
GLU, mmol/L	6.12 ± 1.79	6.23 ± 1.73	0.002
HG,g/L	134.36 ± 13.45	135.62 ± 13.61	0.076
FIB,g/L	3.00 ± 0.66	3.19 ± 0.69	< 0.001
Eosinophil, × 109/L	0.12 ± 0.12	0.13 ± 0.19	0.039
Eosinophil%	1.99 ± 2.00	2.00 ± 2.22	0.789
Basophil, × 109/L	0.02 ± 0.02	0.03 ± 0.02	< 0.001
Basophil%	0.36 ± 0.28	0.44 ± 0.25	< 0.001
LDH,U/L	177.22 ± 31.34	172.84 ± 32.90	< 0.001
GLB,g/L	27.89 ± 4.36	28.60 ± 5.02	0.001
MPV,fL	10.56 ± 1.20	10.01 ± 0.83	< 0.001
PT, s	11.96 ± 1.01	12.83 ± 0.91	< 0.001
TT, s	17.96 ± 2.35	18.65 ± 3.21	< 0.001
UA, mmol/L	0.32 ± 0.08	0.30 ± 0.08	0.003
BUN, mmol/L	5.77 ± 1.53	5.90 ± 1.69	0.15
Lymphocyte	1.63 ± 0.55	1.77 ± 0.59	< 0.001
Lymphocyte%	28.44 ± 8.21	28.69 ± 8.13	0.564
DBIL, μmol/L	4.21 ± 1.82	4.47 ± 2.34	0.009
ALP, U/L	77.99 ± 27.93	79.14 ± 20.85	0.012
CK,U/L	104.64 ± 58.37	90.35 ± 41.89	< 0.001
CREA, μmol/L	70.93 ± 19.59	70.99 ± 17.79	0.608
HCT,%	40.06 ± 3.67	40.58 ± 3.74	0.003
AST,U/L	21.93 ± 10.37	20.90 ± 10.62	< 0.001
ALT,U/L	20.68 ± 13.95	19.52 ± 16.87	< 0.001
GGT, U/L	29.60 ± 63.47	29.19 ± 40.73	0.84
TG, mmol/L	1.67 ± 1.00	1.58 ± 0.98	0.001
Monocyte, × 109/L	0.39 ± 0.14	0.46 ± 0.16	< 0.001
Monocyte%	6.72 ± 1.84	7.28 ± 1.97	< 0.001
PLCR,%	29.34 ± 8.09	25.09 ± 6.58	< 0.001
APTT, s	30.97 ± 5.70	34.93 ± 4.49	< 0.001
INR	0.95 ± 0.06	0.97 ± 0.08	< 0.001
WBC, × 109/L	5.87 ± 1.58	6.34 ± 1.67	< 0.001
RDWSD,fL	42.17 ± 3.12	42.47 ± 3.25	0.099
RDWCV,%	12.73 ± 0.88	12.48 ± 0.76	< 0.001
PT%	110.16 ± 13.02	106.37 ± 11.57	< 0.001
MCV,fL	90.79 ± 4.50	92.90 ± 3.50	< 0.001
MCHC,g/L	335.31 ± 10.45	334.07 ± 10.17	0.002
MCH,pg	30.44 ± 1.76	31.03 ± 1.35	< 0.001
DD, μg/L	0.40 ± 0.64	0.46 ± 0.55	0.002

Fig. 2.

Performance of 12 machine learning models for screening PACG. A Spearman correlation heatmap of 50 features. B The ROC curves of 12 ML models. C Delong test results of the 12 ML models. D–F SHAP summary dot plot for the top 20 features of LGBM (D), XGB (E), Ada (F), and GB (G). The probability of PACG increases with the SHAP value of the feature. In the model, each subject’s SHAP value is visualized as a point on the line for each feature. The color of the point indicates the actual value of the feature, where red denotes higher values and blue denotes lower values

Model development

Model comparison

From a PPPM perspective, the model comparison stage was crucial in identifying the most suitable predictive tool for PACG screening. All 50 features in the discovery set are included to develop 12 machine learning models. fivefold cross-validation was applied to improve the reliability and generalization ability of the models.Complete hyperparameter configurations were documented in Table S6. The AUC values of the 12 ML models (Fig. 2B) were as follows: LGBM (0.92), XGB (0.92), Ada (0.91), GB (0.91), RF (0.90), LR (0.89), GLM (0.89), DT (0.85), NB (0.82), MLP (0.77), SVM (0.73), and KNN (0.62). The ROC and the precision-recall (P-R) curve of the fivefold cross-validation are shown in Supplementary Fig. S1. The 50-featured LGBM model demonstrated prominent performance (AUC = 0.92, sensitivity = 0.89, specificity = 0.80, PPV = 0.84, NPV = 0.86, accuracy = 0.85, F1 score = 0.87), and 12 ML models’ evaluation parameters refer to Supplementary Table S3. Subsequent to the initial evaluation, pairwise comparisons of the AUC values for the 12 machine learning models were carried out using the Delong test. The AUC values of four models, specifically LGBM, XGB, Ada, and GB, were found to be not significantly different from each other (P > 0.05) and were higher than those of the other models. The specific results of the Delong test are depicted in Fig. 2C.

Feature reduction

In the feature reduction phase, guided by PPPM principles, we focused on selecting the most relevant features to ensure the model's efficiency and clinical utility. Feature reduction was performed in the four selected models (LGBM, XGB, Ada and GB) within the cross-validation loop to avoid overfitting. The SHAP model was employed to rank the importance of the top 20 features (Fig. 2D-F), and models incorporating varying numbers of these features were subsequently constructed. The 5-fold cross-validated AUC values of these models with different number of features were then compared by means of the Delong test, as illustrated in Fig. 3A and documented in Supplementary Table S4. Among the four models, the AUC values of the 5-feature models were significantly different from those of the 50-feature models (P 0.05). As shown in Fig. 3B, when the number of features was reduced to 6, the AUC values were 0.91 (P = 0.18) for LGBM, 0.90 (P = 0.07) for XGB, 0.90 (P = 0.15) for Ada, and 0.89 (P 0.05) for GB. Notably, there was no significant difference between the AUC of the 6-feature LGBM model and that of the 50-feature LGBM model (Fig. 3C and Supplementary Table S4). This finding suggests that the 6-feature LGBM model is equivalent to the 50-feature LGBM model in terms of efficacy. The Confusion Matrix Heat Map in Fig. 3D further demonstrates the robust classification performance of the 6-feature LGBM model. Furthermore, as illustrated in Figure 3E and Supplementary Figure S1A, the area under the P-R curve of the 6-feature LGBM model was similar to that of the 50-feature model (AUCPR = 0.94 for both), which indicated that they had comparable and high ability to handle positive instances.

Fig. 3.

Feature reduction. A Delong test results of LGBM, XGB, Ada, and GB with different number of features. LGBM, XGB, and Ada did have no significant difference in performance from their respective 50-feature models when including at least 6 features (p > 0.05), whereas GB required at least 7 features. B AUC for different number of features of LGBM, XGB, Ada, and GB. When the number of features is 6, the AUC of LGBM is 0.91, the AUC of XGB is 0.90, the AUC of Ada is 0.90, and the AUC of GB is 0.89. C The ROC curve of 6-featured LGBM. The AUC of the final model is 0.91. D The confusion matric heatmap of the final model. E The P-R curve of the final model. The AUCPR of the 6-featured LGBM is 0.94. F SHAP summary dot plot of the final model. The six features, in order of importance, are TT, PDW, MCV, APTT, TC, and PT

Finally, the 6-feature LGBM model was selected as the final model based on the combined results of the Delong test and feature count, which were TT, PDW, MCV, APTT, TC, and PT (Fig. 3F). The final model achieved an AUC of 0.91, with sensitivity=0.90, specificity=0.79, PPV=0.84, NPV=0.85, accuracy=0.84, and F1 score=0.86 (Supplementary Table S5). These metrics indicate that the model has excellent predictive performance and aligns with the objective of providing accurate and reliable tools for disease prediction and prevention.

Model validation

LGBM model with 6 features had good performance in both internal and external validation cohorts. Specifically, the AUC was 0.87 in the internal validation (Fig. 4A) and 0.85 in the external validation (Fig. 4E), indicating a high degree of discrimination. The confusion matrix heatmaps presented in Figs. 4B and F further corroborate the model’s satisfactory classification capabilities across both validation sets. Moreover, Fig. 4C and G reveals an AUCPR of 0.81 in the internal validation and 0.83 in the external validation, underscoring the model’s effectiveness in identifying positive cases while maintaining a reasonable balance between precision and recall. In internal validation, the final model had a PPV of 0.76, an NPV of 0.95, an accuracy of 0.83, and an F1 score of 0.85. The PPV of the final model in the internal validation was 0.85, the NPV was 0.90, the accuracy was 0.89, and the F1 score was 0.84, as detailed in Supplementary Table S5. These findings collectively indicate that the final model exhibits stability and generalizability,which is crucial for clinical application.

Fig. 4.

Internal and external validation. A–C The ROC curve, confusion matric heat map, and P-R curve of the final model for screening PACG in the internal validation. D The ROC curve of the final model for screening PAC in the internal validation. E–G The ROC curve, confusion matric heat map, and P-R curve of the final model for screening PACG in the external validation. H The ROC curve of the final model for screening PAC in the external validation

The final model demonstrated robust screening performance for PAC in both the external and internal validation cohorts, highlighting its significance in early glaucoma diagnosis. The internal validation yielded an ROC of 0.85 (Fig. 4 D), accuracy of 0.86, and an F1 score of 0.83, while the external validation achieved an ROC of 0.84 (Fig. 4H), accuracy of 0.86, and an F1 score of 0.83 (Supplementary Table S5). By enabling early screening before irreversible optic nerve damage occurs, this model effectively prevents the progression of PACG, fully aligning with the requirements of PPPM.

Model explanation and application

Models with clinical applications are required to be explainable. The SHAP model is a relatively all-round approach of model explainability, allowing for both global and local explanations. Its interpretability is crucial for gaining the trust of clinicians and patients, making it more likely to be adopted in clinical practice and supporting the shift towards transparent and patient-centered healthcare.

From a global perspective, the SHAP summary dot plot can be viewed as a ranked map of feature importance. As shown in Fig. 3F, the TT features are more important to the model and ranked first. The horizontal dispersion of TT’s SHAP values indicates its strong discriminative power in identifying high-risk PACG individuals. The SHAP dependence plot (Fig. 5A–F) then indicates the relationship between the SHAP value of the feature and the feature value. For instance, patients with PDW ≤ 10.8 or TT ≥ 19.3 had SHAP values exceeding zero, indicating a higher probability of being predicted as PACG. Furthermore, MCV ≥ 94.4, APTT ≥ 38.3, PT ≥ 12.8, and TC ≥ 5.9 also influenced the decision towards the “PACG” class.

Fig. 5.

Model explanation and application. A–F SHAP dependence plot of six features. Each dependency plot shows how individual features affect the output of the screening model, with each point representing a single subject. G–H The waterfall plot of a PACD patient and a normal subject. The Y-axis represents the different feature values, X represents the SHAP value, and E[f(x)] represents the expectation of f(x) for all samples, and the magnitude of the f(x) value is the predicted value for that sample. Red color means the feature is positive gain for the sample; blue color is negative. I–K The DCA curves of the final model in the training set, the internal set, and the external set

In terms of local interpretation, each sample produces a predictive value, with SHAP values assigned to each feature within that sample. For example, from a PACG sample (Fig. 5G), PDW and MCV contributed a higher score of 0.13 and 0.12, respectively, whereas from another control subject (Fig. 5F), it was APTT and PT that had a higher contribution. Local SHAP interpretations decode individual risk profiles, empower clinicians to tailor monitoring intervals and preventive strategies.

While AUC, sensitivity, and specificity can only measure the screening efficacy of a model, decision curve analysis (DCA) can take into account the practical clinical application aspects. Regardless of the training set, internal validation, or external validation, as shown in Fig. 5I–K, the net gain is higher than the treat-all and treat-none line, and the model has practical clinical applications.

In order to enhance the clinical applicability of the model, the 6-feature LGBM model was developed into a web-based application (Fig. 6A). When users input the values of the six features, the application processes this information and outputs the corresponding risk of PACG development. This approach aims to provide a convenient and accessible tool for clinical decision-making.

Fig. 6.

The public Internet calculator for screening PACD by six features.The web server of LGBM model with six features available at https://lzq-pacg.streamlit.app/

The webpage illustrates two examples: a 48-year-old female participant (Fig. 6B) enrolled in Xuhui Central Hospital in 2024 (TT = 20.90, MCV = 98.00, PDW = 13.60, APTT = 36.10, PT = 13.30, TC = 6.51) received a 94.73% risk probability, warranting immediate preventive measures recommendably; a 52-year-old male participant (Fig. 6C) enrolled in the EENT Hospital in 2023 (TT = 16.30, MCV = 103.40, PDW = 16.90, APTT = 36.00, PT = 12.70, and TC = 4.11) had a 1.10% risk, avoiding unnecessary invasive exams. This tool democratizes equitable screening in resource-limited settings and exemplifies PPPM’s paradigm shift. The web application can be accessed online at https://lzq-pacg.streamlit.app/.

Discussion

Clinical significance and advancements in PACG management

This study represents a significant leap in PACG management through the development and validation of a novel clinlabomics-based LGBM model within the PPPM framework. The model’s exceptional screening capability, with high accuracy in identifying high-risk PACG patients, aligns seamlessly with the core tenets of PPPM. Specifically, the model achieved an AUC of 0.91, sensitivity of 0.89, specificity of 0.79, PPV of 0.84, NPV of 0.85, accuracy of 0.84, and F1 score of 0.86, demonstrating its robust performance in distinguishing PACG patients from normal individuals. The model’s performance in internal validation (AUC = 0.87) and external validation (AUC = 0.85) further confirms its stability and generalizability. This stability is crucial for clinical application, as it ensures consistent performance across different populations.

Its reliance on routine blood parameters not only ensures cost-effectiveness and accessibility but also provides profound insights into systemic physiological and pathological conditions. By harnessing machine learning to analyze these parameters, we can now predict PACG risk before symptoms manifest, offering a critical window for timely intervention to prevent disease progression and vision loss. The model’s integration of explainable AI facilitates cost-effective population screening and personalized risk assessment, making glaucoma care more accessible, especially in underserved regions with limited ophthalmic resources. This approach exemplifies the shift from reactive treatment to proactive, personalized healthcare, a cornerstone of PPPM.

Comparison with previous studies and unique contributions

There have also been several previous studies that have developed models for screening, diagnosis, and prediction of glaucoma. Professor Zhang’s team [13] developed a glaucoma onset and progression prediction system based on fundus color photographs (iGlaucoma 4.0), which was able to predict the risk of glaucoma onset and progression 3 to 5 years ahead of time, with AUC of 0.90, 0.88, and 0.87 in the training set and two external validations, respectively. Xu et al. [40]developed a deep learning classifier to analyze the patient’s preoptic node images, which can be used to screen patients with PACG with good performance in both internal and external validations. In one of our previous studies, we built metabolomics-based machine learning models to diagnose PACG and found that serum androstenedione could be used as a potential biomarker for diagnosing PACG and indicating visual field progression [27]. Distancing itself from prior research, this study innovatively merges clinlabomics with the PPPM framework, addressing the limitations of conventional PACG screening methods. Unlike earlier models that depend exclusively on specialized ophthalmic imaging or serum metabolome analysis, our model utilizes routine blood parameters, enhancing accessibility and scalability for widespread use.

Due to the common “black box” effect of machine learning models [38], it is difficult for doctors and patients to understand the decision basis of the models, which may lead to distrust of the models, thus limiting their clinical application. To address this problem, this study uses the SHAP tool to enhance the explainability of the model, which can provide both global and local explanations [41]. In terms of local explanation, it can explain the respective contribution of the included feature values of the patient to the predicting results. This capability allows for personalized risk assessment and aligns with the principles of PPPM by empowering clinicians to make more informed decisions.

Systemic biomarkers in PACG management: a PPPM perspective

The six features included in our final model were TT, PDW, MCV, APTT, TC, and PT, which are features that contribute significantly to PACG screening. These parameters not only reflect individualized systemic physiological and pathological conditions but also align with the predictive, preventive, and personalized medicine (PPPM) approach. Our previous studies have shown that coagulation-related indices are associated with eye diseases such as PACG and high myopia [34, 42]. The onset and progression of glaucoma is thought to involve a variety of pathogenic processes, of which the inflammatory response is considered an important part [43]. Studies have shown that microglia may be activated during the pathologic process of glaucoma, releasing inflammatory factors that lead to damage to retinal ganglion cells [22]. Inflammatory factors such as tumor necrosis factor-α (TNF-α), several interleukins (ILs), and nuclear transcription factor-kappa B (NF-κB) have also been found to be highly expressed in glaucoma patients [44–48]. These inflammatory factors further lead to damage to the structures of atrial outflow, allowing further elevation of IOP [43, 49]. This highlights the important role of inflammatory responses in glaucoma development.

From a PPPM perspective, the model’s ability to identify these systemic biomarkers is a significant step toward early disease prediction and personalized prevention [50]. By integrating these biomarkers with machine learning, routine clinical laboratory data can be transformed into a powerful tool for disease prediction and prevention. This approach supports the holistic view of healthcare advocated by PPPM [51], which emphasizes leveraging systemic health indicators to predict and prevent diseases. The model enables clinicians to develop personalized treatment strategies based on individual risk profiles, enhancing the precision and efficacy of glaucoma management.

Limitations

There are also some limitations in this study. First, although the data in this paper come from three centers, the generalizability to the world is not clear because they are all Chinese populations, and yet further evaluation is needed. Validation in multi-ethnic cohorts is essential to ensure the model’s applicability across diverse healthcare settings—a critical step toward achieving PPPM’s vision of equitable, population-wide prevention. Additionally, the model’s reliance on six blood-based biomarkers, while enhancing accessibility in primary care, excludes established ophthalmic parameters such as anterior chamber depth. This simplification prioritizes scalability and cost-effectiveness, aligning with PPPM’s preventive goals, but may overlook nuanced anatomical risk factors. However, this trade-off reflects a deliberate strategy to democratize screening in regions lacking specialized ophthalmological infrastructure, where early intervention is most urgently needed.

Conclusion and expert recommendation

This study pioneers a PPPM-oriented approach to glaucoma care, marking a transformative step from reactive symptom management to proactive, systemic health preservation. By identifying systemic biomarkers linked to PACG pathology, the model enables early intervention for high-risk individuals, reducing vision loss and healthcare costs. The model’s design and implementation reflect the core principles of PPPM, offering a scalable and cost-effective solution for PACG screening that can be integrated into routine clinical practice.

1. Predictive approach: Central to this innovation is the early detection of systemic biomarkers, which may reflect preclinical pathological processes in PACG. By identifying these subclinical signals years before symptomatic deterioration occurs, the model enables risk prediction, offering a critical window for preventive interventions.

2. Targeted prevention: The model’s reliance on routine blood tests ensures cost-effective risk stratification, optimizing resource allocation in overburdened healthcare systems—particularly in regions with limited access to specialized ophthalmic infrastructure. This preventive utility aligns with PPPM’s emphasis on equitable, population-wide health preservation, prioritizing high-risk populations for targeted interventions while reducing unnecessary referrals. This approach not only democratizes screening in resource-limited regions but also facilitates cost-effective preventive strategies, such as prioritized laser iridotomy for high-risk individuals, potentially reducing acute PACG incidence.

3. Personalized treatments: Clinical laboratory data not only serve as standardized reference intervals but also reflect unique physiological signatures at the individual level. The SHAP method and a web-based tool bridge the gap between complex machine learning algorithms and clinical practice. By translating complex model functioning mechanisms into interpretable risk profiles and scores, clinicians can tailor monitoring intervals or therapeutic plans to individual patients, fostering personalized care and enhancing patient trust.

For the further application of clinlabomics-based model in the context of PPPM in PACG management, we recommend the following:

1. Multi-ethnic cohort validation: To ensure the model’s global applicability and enhance its predictive accuracy, validation across diverse ethnic cohorts is imperative. This will help tailor the model to different populations and solidify its role in equitable, population-wide prevention strategies.

2. Integration with clinical workflows: Future work should focus on seamlessly integrating the model into existing clinical workflows. This includes developing user-friendly interfaces and decision-support systems that can assist healthcare providers in making informed decisions based on the model’s predictions.

3. Biomarker-guided therapies for high-risk subgroups: Biomarker-guided therapies should be explored for high-risk subgroups to translate predictive insights into actionable prevention strategies. This involves developing treatments tailored to individuals at high risk of PACG based on their specific biomarker profiles, thereby enhancing the model’s impact on clinical outcomes and aligning with the personalized treatment goals of PPPM.

Importantly, what is exactly the added value of our study?

The added value of our study lies in its innovative combination of clinlabomics and machine learning within the PPPM framework to address a critical gap in PACG screening. By utilizing routine blood parameters, we offer a cost-effective, accessible, and scalable solution that overcomes the limitations of traditional screening methods. The model’s explainability through the SHAP method further enhances its clinical utility, making it a powerful tool for personalized risk assessment and targeted prevention. This study also provides a blueprint for applying PPPM principles to other disease areas, demonstrating how predictive, preventive, and personalized approaches can transform healthcare delivery and improve patient outcomes.

The paradigm shift from reactive to PPPM/3PM and go beyond the state of the art

Our study exemplifies the paradigm shift from reactive, symptom-driven care to proactive, prediction- and prevention-focused medicine. By leveraging clinlabomics and machine learning, we move beyond the limitations of current practices, which often detect PACG only after significant structural damage has occurred. Instead, we offer a forward-looking approach that identifies individuals at risk before irreversible vision loss happens. This shift is in line with the broader goals of PPPM/3PM, which aim to transform healthcare systems to be more proactive, personalized, and patient-centered. Our work not only advances the technical capabilities of PACG screening but also redefines the approach to disease management, setting a new standard for integrating technology, laboratory medicine, and clinical practice in the pursuit of better health outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviation

PACG

Primary angle-closure glaucoma

AI

Artificial intelligence

ML

Machine learning

SHAP

SHapley Additive explanation

ROC

Receiver-operating-characteristic curve

AUC

Area under the receiver-operating-characteristic curve

PR

Precision versus recall curve

AUCPR

Area under the precision versus recall curve

PPV

Positive predictive value

NPV

Negative predictive value

DCA

Decision curve analysis

SVM

Support vector machine

Ada

Adaptive boosting

RF

Random forest

LR

Logistic regression

GLM

Generalized linear model

XGB

EXtreme gradient boosting;

NB

Naive Bayes

GB

Gradient boosting machine

MLP

MLPClassifier

DT

Decision tree

KNN

K-nearest neighbor

LGBM

Light gradient boosting machine

RBC

Red blood count

PCT

Thrombocytocrit

PLT

Platelet count

PDW

Platelet distribution width

HG

Hemoglobin

MPV

Mean platelet volume

HCT

Hematocrit

PLCR

Platelet large cell ratio

WBC

White blood cell count

RBCSD

Red blood cell distribution width-standard deviation

RBCCV

Red blood cell distribution widthcoefficient of variation

MCV

Mean corpuscular volume

MCHC

Mean corpuscular hemoglobin concentration

MCH

Mean corpuscular hemoglobin

TP

Total protein

TBA

Total bile acid

TBIL

Total bilirubin

TC

Total cholesterol

ALB

Albumin

GLU

Glucose

LDH

Lactic dehydrogenase

GLB

Globulin

UA

Uric acid

BUN

Blood urea nitrogen

DBIL

Direct bilirubin

ALP

Alkaline phosphatase

CK

Creatine kinase

CREA

Creatinine

AST

Glutamic oxaloacetic transaminase

ALT

Glutamic-pyruvic transaminase

GGT

Gamma-glutamyl transpeptidase

TG

Triglyceride

FIB

Fibrinogen

PT

Prothrombin time

TT

Thrombin time

APTT

Activated partial thromboplastin time

INR

International normalized ratio

DD

PT% and d-dimer

TNF-α

Tumor necrosis factor-α

ILs

Interleukins

NF-κB

Nuclear transcription factor-kappa B

Author contribution

Z.L. and J.R. contributed to formal analysis, writing the original draft and reviewing and editing the manuscript. J.W, Y.L, Y.S and M.Z were involved in in data curation, investigation, visualization and reviewing the manuscript. W.C. and S.L. provided resources, supervision, funding acquisition, and manuscript writing and editing.

Funding

This work was supported by the National Natural Science Foundation of China (82302582), Higher Education Industry-Academic-Research Innovation Fund of China (2023JQ006), and Shanghai Municipal Health Commission Project (20224Y0317). The sponsor or funding organization had no role in the design or conduct of this research.

Data availability

Original datasets underlying this investigation can be accessed through direct correspondence with the principal investigator, contingent upon completion of a material transfer agreement and institutional review board approval.

Declarations

Ethics approval and consent to participate

The ethical aspects of this research were reviewed and approved by the Institutional Ethics Committee of Eye and ENT Hospital of Fudan University. All study procedures were performed in strict adherence to the ethical standards set forth in the Declaration of Helsinki. Prior to participation, comprehensive written informed consent was secured from all subjects following a detailed explanation of the study’s purpose, procedures, potential risks, and benefits.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.