Novel Subclassification of Presbyopia Based on Lens Functional and Structural Features

Abstract

Purpose: The purpose of this study was to classify eyes with presbyopia using lens functional and structural features to aid in surgical decision making.

Methods: This cross-sectional observational study included healthy volunteers or those solely affected by presbyopia. Lens function was assessed using dysfunction lens index (DLI), whereas lens density was evaluated with Pentacam (LD-P) and swept-source optical coherence tomography (LD-S). Principal component analysis (PCA) and k-means clustering classified participants into non-lens surgery (NLS) and lens surgery (LS) groups. A scoring system was developed from the optimal cutoff values for simplified diagnosis, with a website created for automatic classification.

Results: Totally, 352 eyes of 176 patients were studied, with a mean age of 46.1 ± 6.2 years and a mean axial length of 24.42 ± 1.62 mm. Among them, 268 (76.1%) eyes were diagnosed with presbyopia. The top 5.2% (14/268) of presbyopia cases with the highest variability were excluded by PCA. Subsequently, 138 eyes (51.5%) were clustered into the NLS group, and 116 eyes (43.3%) into the LS group. The LS group presented lower DLI and higher lens density compared to the NLS group (all P < 0.001). Optimal cutoff values were determined as follows: DLI = 7.75, LD-P = 8.75, and LD-S = 60.05 pixel units. The scoring system demonstrated that combining DLI with LD-P achieved the highest diagnostic efficacy. A website (www.zhu-zhou-lf.com) was developed for automatic classification.

Conclusions: We proposed a novel classification of presbyopia using lens functional and structural features.

Translational Relevance: This classification, integrating DLI and lens density values, offers a diagnostic tool to guide surgical decisions in presbyopia.

Introduction

Presbyopia affects a notable 1.8 billion individuals globally, and the potential productivity loss due to uncorrected or under corrected presbyopia was estimated to be US $25.367 billion, accounting for 0.037% of the global gross domestic product (GDP).^1,2 Therefore, prioritizing the treatment of presbyopia is imperative. Conservative treatment options for presbyopia include topical medications to increase depth of focus or lens softening, as well as spectacle or contact lens correction.^3–5 As the severity of presbyopia increases, the demand for surgical treatment to eliminate the need for glasses and enhance quality of life continues to rise.⁶ 

The selection of surgical plans for patients transitioning from presbyopia to cataracts currently lacks a unified consensus. The concept of dysfunctional lens syndrome (DLS) has been proposed to describe the stages from presbyopia to cataracts.^3,4,7–10 Despite offering a descriptive framework, DLS lacks formal diagnostic classification based on clinical evidence and lens anatomy. The timing and criteria for surgery are not well-defined, leading to varied inclinations among refractive and cataract specialists.⁴ Cornea procedures for monovision design are relatively simple to perform but require a period of adaptation.^3,4,6 Advanced techniques like scleral approaches, extended depth of focus technology, lens-softening procedures, and laser refractive index shaping improve accommodative capacity.^3,6,11 However, they may necessitate periodic retreatment due to age-related progression and anisometropia, diminishing their cost-effectiveness.^3,6,12,13 Lens surgery may provide a permanent solution, avoiding subsequent secondary surgeries, but carries risks such as retinal detachment, intraocular pressure (IOP) elevation, etc. Additionally, some multifocal intraocular lenses (IOLs) have limitations in improving near vision.^13–15 

The key factor in selecting refractive or lens surgery is the comprehensive evaluation of lens function and structure to obtain an objective classification of presbyopia.^4,16 Novel instruments enable accurate measurement of these lens features. The dysfunctional lens index (DLI) from i-Trace provides an objective evaluation of lens function, considering higher-order aberrations (HOAs), pupil size, and contrast sensitivity data.^17,18 Additionally, Pentacam utilizes Scheimpflug photography to generate high-resolution 3D maps and offer precise lens opacity assessment.^19,20 Swept-source optical coherence tomography (SS-OCT) also contributes to the evaluation of lens morphology and density. However, the method for selecting these lens features to establish a classification remains largely unknown. Leveraging artificial intelligence (AI) for comprehensive analysis holds the potential to achieve precise classification of presbyopia.²¹ 

Our study introduces an innovative AI-assisted classification system for presbyopia. This approach integrates features from lens functional and structural characteristics to aid in surgical decision making. 

Methods

Study Design and Participants

The study received approval from the Institutional Review Board of Eye & ENT Hospital, Fudan University, Shanghai, China, and was conducted in accordance with the principles outlined in the Declaration of Helsinki. Written informed consent was obtained from each patient before enrollment. 

This observational cross-sectional study recruited 404 eyes of 202 healthy volunteers aged 35 to 60 years from Eye & ENT Hospital, Fudan University in Shanghai, China, between April 2023 and March 2024. Participants were excluded if they presented with ocular surface diseases, ocular fundus abnormality, glaucoma, ocular hypertension, inflammatory eye diseases, and strabismus (5 eyes, 3 cases). Individuals with a history of ocular trauma or surgeries (4 eyes, 3 cases) and those with significant cataract (Lens Opacity Classification System III, LOCS III grading ≥C3/N3/P2, 18 eyes, 10 cases) were excluded. Patients with systemic diseases with ocular involvement (6 cases) were also excluded. Participants with imaging results of low quality were also excluded from the study (6 eyes, 4 cases). Finally, 352 eyes of 176 cases were included. 

All participants underwent a comprehensive ophthalmic examination, including the measurement of best corrected distant visual acuity (BCDVA), distant corrected near visual acuity (DCNVA) and addition power (ADD) by autorefractor (Tomey RC5000, Phoenix, AZ, USA), IOP, and anterior and fundus photography. Decimal visual acuity was converted to logarithm of the minimum angle of resolution (logMAR). A slit-lamp examination was conducted for the LOCS III grading. Additionally, the IOLMaster 700 (Carl Zeiss AG, Oberkochen, Germany) was used for axial length (AL), anterior chamber depth (ACD), and lens thickness (LT) measurement. Each device was operated by an experienced examiner (Fig. 1). 

Figure 1.

 

Patient enrollment and examination. (A) Patient recruitment. (B) Comprehensive examinations for lens functional and structural parameters, including iTrace for dysfunction lens index (DLI) measurement; Pentacam Nucleus Staging (PNS), average lens density (LD-P) and zones 1 to 3 assessment; swept-source optical coherence tomography (SS-OCT) for lens density (LD-S, calculated by Image J software), lens anterior (Ra) and posterior curvature (Rp); IOLMaster 700 for axial length (AL), anterior chamber depth (ACD), and lens thickness (LT) assessment; and slit lamp for LOCS III grading.

Assessment of Lens Function

The i-Trace Visual Function Analyzer (Tracey Technologies, Houston, TX, USA) calculates DLI values, HOA, and modulation transfer function (MTF) using the i-Trace prime version 7.0.1 software within a 6-mm diameter zone in a dark room.^18,22 DLI values range from 0 to 10, with lower values indicating a more dysfunctional lens and higher values correlating with better lens performance (see Fig. 1). 

Measurement of Lens Structure

Pentacam HR (Oculus Optikgeräte GmbH, Wetzlar, Germany) was used to analyze lens density. The densitometry software conducted analysis in peak, linear, and 3D scan modes based on Scheimpflug images. The grading system provided a Pentacam Nucleus Staging (PNS) ranging from 0 to 4, determined by pixel intensity measurement within the nucleus. Maximum and average lens density values were obtained, and zone values were analyzed in three defined concentric zones centered in the pupil (zone 1 = 2.0 mm, zone 2 = 4.0 mm, and zone 3 = 6.0 mm). 

The anterior segment SS-OCT (TowardPi Yalkaid, Beijing, China) scans were performed in 8 directions at 22.5 degrees intervals and subsequently averaged for further analysis. The software automatically generated cross-sectional images of the lens and calculated the anterior and posterior surface radius of curvature (Ra and Rp). Lens density of SS-OCT images was quantitatively analyzed using ImageJ software. The boundaries of the cortex, adult nucleus, and juvenile nucleus were outlined according to the modified Vogt/Duke-Elder nomenclatures of various zones.^23,24 The adult nucleus zone corresponds to the outer nucleus zone. The juvenile, fetal, and embryonic nucleus zones are collectively designated as the inner nucleus zone. The average and maximum density of the zone areas was automatically measured in pixel intensity units on a scale from 0 to 255 (see Fig. 1). 

Classification by PCA and Clustering Analysis

To establish the classification system, a set of multidimensional parameters was designed, including baseline characteristic (age, sex, BCDVA, DCNVA, and AL), functional parameters (DLI, MTF, and HOA), and lens structural parameters (average lens density by Pentacam, zone 1, zone 2, zone 3, average lens density by SS-OCT, lens density of the cortex, adult nucleus and inner nucleus, Ra, Rp, and LT). Principal component analysis (PCA) was conducted. Based on the ADD results, the population was divided into two clusters: cluster 1 = healthy controls and cluster 2 = presbyopia (ADD≠0). Further analysis was conducted on the presbyopia subgroup. 

To enhance the robustness of the analysis, the top 5.2% (14 cases) of individuals exhibiting the highest variability were excluded to reduce dimensionality and ensure cohort homogeneity. PCA was conducted to uncover underlying patterns and reduce dimensionality by applying a linear and orthogonal transformation. This transformation converted highly correlated parameters into a new set of linearly uncorrelated parameters, retaining enough dimensions to achieve a cumulative variance of 95%. This process facilitated the determination of each individual's contribution to overall variance and informed subsequent clustering by capturing nuanced relationships among variables.^25–27 The correlation matrix was used to select variables with the strongest correlations for two-dimensional projection. Principle dimensions 1 to 3 explained 81.3%, 8.7%, and 7.3% of the total variance, respectively. Cluster analysis was then conducted using the top six parameters with the highest contribution: DLI, average lens density by Pentacam (LD-P), zone 1, zone 2, zone 3 and average lens density by SS-OCT (LD-S; Fig. 2). 

Figure 2.

 

Workflow for artificial intelligence (AI)-based classification for presbyopia treatment. (A) Step 1: Principal component analysis (PCA). (a) Individual selection. Eyes were clustered into healthy control and presbyopia group according to ADD. The presbyopia group was included for further PCA analysis, whereas the top 5.2% of individuals exhibiting the highest variability were excluded. (b) Dimensionality reduction and correlation matrix analyses were performed to select the top six parameters contributing most. (B) Step 2: Presbyopia classification. Three clustering methods—Gaussian Mixture Model (GMM), k-means clustering, and hierarchical clustering—were used to categorize patients with presbyopia into two or three groups. (C) Step 3: Cutoff value for diagnosis. Receiver operating characteristic (ROC) curve analysis was conducted to determine the optimal cutoff value. A scoring system was developed to classify subjects into the lens surgery (LS) and the non-lens surgery (NLS) groups.

Next, unsupervised learning cluster analyses were performed using three clustering methods: K-means clustering, Gaussian mixture model (GMM), and hierarchical clustering for two-categories and three-categories clustering (see Fig. 2). The best model for presbyopia classification was determined based on the clustering pattern and variance explained ratio. K-means clustering with 2 categories was selected as the best model, explaining 90.6% of the total variance, and resulted in the classification of patients into the lens surgery group (LS) and non-lens surgery group (NLS; Fig. 3). 

Figure 3.

 

Presbyopia classification. (A) K-means cluster for two categories. This method provided the best explanation and was deemed optimal subclassification. (B) Gaussian mixture model (GMM) cluster for two categories. (C) Hierarchical cluster for two categories. (D) GMM cluster for three categories. (E) K-means cluster for three categories. (F) Hierarchical cluster for three categories.

Cutoff Value and Scoring System for Diagnosis

Receiver operating curve (ROC) analysis was conducted to determine the optimal cutoff values for lens functional and structural parameters between the NS and NLS groups. The Youden J index was used to identify cutoff points that maximized diagnostic accuracy. The area under the ROC (AUC) and corresponding 95% confidence intervals (CIs) were calculated. 

Based on these established cutoff values, we developed a scoring system for diagnostic efficiency analysis. Each of the six parameters in PCA clustering was assigned a score. A score of 1 was given if the DLI value was less than or equal to the cutoff value, and 0 otherwise. Similarly, a score of 1 was given if any lens density value exceeded the cutoff value, and 0 otherwise. Four different score combinations were calculated: score 1 = score (DLI + LD-P), score 2 = score (DLI + LD-P or LD-S exceeding the cutoff value), score 3 = score (DLI + LD-P /zone 1 /zone 2 /zone 3 /LD-S, with any of the 5 lens density values surpassing the cutoff value), score 4 = score (DLI + LD-S). ROC analysis was performed to determine the diagnostic specificity and sensitivity of different scores. The selection of the optimal scoring combination with the highest diagnostic efficacy was based on AUC and corresponding 95% CIs (see Fig. 2). 

Website Construction With AI-Based Algorithms

We developed a website that uses AI-based algorithms to calculate the total score based on the top six parameters. The calculation takes into account each parameter's contribution to PCA and the proportion of individual examination data at the cutoff value. Users can input their DLI and Pentacam examination results, with the option to include SS-OCT results as supplementary information. The system then automatically provides a precise diagnostic classification. 

Statistical Analyses

Continuous data were presented as means ± standard deviations. Categorical data were expressed as absolute numbers and proportions, with comparisons made using a χ² test. The use of Generalized Estimating Equations (GEEs) helped to account for the correlation between eyes when comparing across groups. During the PCA clustering process, one eye was randomly chosen for correlation analysis and parameter selection for each patient. The six parameters selected for clustering are independent ocular indicators, thereby incorporating data from both eyes in the clustering analysis. The PCA and clustering analyses were conducted using the R Statistical Software (version 4.3.0). Significance was set at P < 0.05. Graphs were generated using Prism 9 (GraphPad, La Jolla, CA, USA) and R 4.3.0 (R Foundation for Statistical Computing, Vienna, Austria). 

Results

Baseline Characteristics

We finally enrolled 176 patients (352 eyes) with a mean age of 46.1 ± 6.2 years (range = 35–60 years), including 72 male patients and 104 female patients. The mean AL was 24.42 ± 1.62 mm (range = 21.37–29.10 mm). Among these eyes, 268 (76.1%) were diagnosed with presbyopia, whereas 84 (23.9%) were healthy controls. Mean DCNVA was 0.22 ± 0.19 logMAR units (range = –0.10 to 0.70 logMAR units). The median ADD was +1.25 diopter (D; range = 0 to +3.75 D). Baseline characteristics for presbyopia and control eyes are summarized in Table 1. 

Table 1.

 

Baseline Characteristics Between Presbyopia and Control Eyes

 

Comparison of Lens Functional and Structural Parameters Between NLS and LS Groups

Using AI-based classification, the top 5.2% (14 cases) of presbyopia cases with the highest variability were excluded by PCA. Subsequently, binary K-means clustering categorized 138 eyes (51.5%) into the NLS group, whereas 116 eyes (43.3%) were classified into the LS group. The mean age of patients in the NLS group (47.48 ± 4.49 years) was significantly younger than that in the LS group (mean = 49.78 ± 4.48 years, GEE, P = 0.004). There was no significant difference in sex (female/ male, 42/27 vs. 33/25, P = 0.545) and AL (24.34 ± 1.70 mm vs. 24.49± 1.78 mm, P = 0.598) between the groups. Eyes in the LS group exhibited higher ADD (+1.45 ± 0.70 D vs. +2.06 ± 0.75 D, GEE, P < 0.001) and worse DCNVA (0.23 ± 0.15 vs, 0.34 ± 0.15 logMAR units, GEE, P < 0.001) compared to the NLS group. 

The comparison of lens functional and structural parameters between the NLS and LS groups is detailed in Table 2, with representative images for each group illustrated in Figure 4. The distribution of DLI, PNS, and LOCS III grades between groups is compared in Figure 5. In terms of lens functional parameters, the LS group presented lower DLI values compared to the NLS group (GEE, OR = 0.47, P < 0.001). The LS group also showed a higher proportion of DLI values below 8 when compared to the NLS group (χ² test, all P < 0.05). Additionally, the LS group had a higher percentage of eyes with LOCS III ≥2 and PNS ≥1 compared to the NLS group (χ² test, all P < 0.05). 

Table 2.

 

Comparison of Lens Structural and Functional Parameters Between Presbyopia Categories

 

Figure 4.

 

Representative images and schematics of lens functional and structural parameters. (A) Representative images and mean values of lens functional and structural examinations in control (C), the non-lens surgery (NLS) and the lens surgery (LS) groups (a) DLI images from iTrace. (b) Pentacam images displaying 3D lens density. (c) Cross-sectional image from SS-OCT, along with mean values for lens density. (d) Lens photography by slip lamp. (B) Schematic representation indicating the patterns of lens structural changes and cutoff values in the LS and NLS groups. Data are presented as mean ± standard deviation. DLI, dysfunction lens index; LD-P, average lens density by Pentacam; LD-S, average lens density by SS-OCT; SS-OCT, swept-source optical coherence tomography.

Figure 5.

 

Associations between the classification and functional and structural parameters in the lens surgery (LS) and the non-lens surgery (NLS) groups. (A) The distribution of dysfunction lens index (DLI) grades. (B) The distribution of Pentacam Nucleus Staging (PNS). (C) The distribution of LOCS III grades.

Notably, the most pronounced difference in lens density was observed in the adult nucleus zone, where the LS group differed significantly from the NLS group (P < 0.001 in both single-factor and multi-factor GEE corrections), in contrast to the differences observed in the cortex or inner nucleus zone areas. 

A Scoring System Based on Optimal CutOff Value for Simplified Diagnosis

ROC analysis was performed to determine the optimal cutoff value between the NLS group and the LS group (Fig. 6). The optimal cutoff values were determined as follows: DLI = 7.75, LD-P = 8.75, zone 1 = 8.95, zone 2 and 3 both = 8.85, and LD-S = 60.05 pixel units. The sensitivity and specificity results for each cutoff value were summarized in Table 3. 

Figure 6.

 

Receiver operating curve (ROC) of the optimal cutoff value for lens structural and functional parameters in the lens surgery (LS) and the non-lens surgery (NLS) groups. (A) ROC for top six lens structural and functional parameters. (B) ROC for four scoring system.

Table 3.

 

Optimal Cutoff Values for Lens Structural and Functional Features to Classification Presbyopia Categories

 

Based on the above cutoff value, diagnostic efficacy of different scoring combinations of DLI with various lens density indicators were evaluated by ROC analysis (see Fig. 6). The results showed that combining DLI with LD-P demonstrated the best diagnostic efficacy (AUC = 0.961, sensitivity = 82.8%, and specificity = 99.3%). Additionally, DLI combined with LD-S can be used as an alternative secondary diagnostic indicator (DLI + LD/P or LD-S, AUC = 0.916, sensitivity = 88.4%, and specificity = 94.8%). Therefore, if DLI ≤ 7.75, and LD-P ≥ 8.75 (or alternatively, LD-S ≥ 60.05 pixel units), the patient can be classified into the LS group, and lens surgery is recommended. Otherwise, the patient is classified into the NLS group. The schematic diagram for diagnosis was presented in Figure 4. 

Website for Automatic Classification

To facilitate precise automatic diagnosis of presbyopia, a website (www.zhu-zhou-lf.com) with AI-based algorithms has been developed. Users can input DLI and Pentacam examination results, including LD-P, zone 1, zone 2, and zone 3. If SS-OCT examination result (LD-S) is available, it can be optionally inputted to enhance the precision of the diagnosis. The website will automatically output the grouping results and provide surgical recommendations. 

Discussion

The transition from presbyopia to cataracts, which occurs due to natural aging-induced changes in the crystalline lens, results in a gradual decline in lens accommodative ability and transparency over a period of more than 10 years.^8,28 This phase presents a therapeutic dilemma as it lacks standardized treatment guidelines for patients, leaving doctors perplexed when choosing between refractive surgery and lens surgery.^3,4 Our study introduces a novel AI-based classification to categorize presbyopic eyes into two groups: those recommending lens surgery and those not, based on lens functional and structural features. We further established a scoring system for simplified diagnosis and developed a website (www.zhu-zhou-lf.com) for AI-based automatic classification, offering a new avenue for presbyopia treatment decision making. 

There is an increasing demand for surgical options to achieve spectacle independence for individuals with presbyopia.^3,7,8 Our AI-based classification system leverages the widely used DLI by i-Trace with Pentacam examination results for rapid diagnosis. Additionally, we have also developed a website that integrates multiple lens functional and structural indicators for a more precise diagnosis. The system's objectivity, efficiency, and precision make it highly suitable for clinical application, enabling the selection of the most cost-effective surgical plan based on the patient's needs. Presbyopic patients in the NLS group with clear crystalline lenses may benefit more from refractive surgeries. This allows them to fully enjoy the advantages of the procedure for an extended period before significant lens changes occur.^11,13,29 Those in the LS group may benefit more from lens surgery, as the crystalline lens quickly becomes more opaquer and functionally compromised after refractive surgery, potentially requiring repeated presbyopia corrections or phakic implantable collamer lens removal. Lens surgery can provide an all-in-one solution, offering greater benefits.^3,28 

According to the latest review in Progress in Retinal and Eye Research, the primary issue driving presbyopia development is rooted in changes to the intrinsic stiffness and elastic modulus of the lens itself. These internal alterations of the crystallin lens hinder the force transmission from the ciliary muscle.^16,23,30,31 Our findings highlight that changes are most pronounced in the adult nucleus zone between the NLS and the LS groups, surpassing those in the cortex or inner nucleus zone. This may be due to a transport barrier that emerges from age 30 years between the cortex and this zone, impeding the movement of water and glutathione, with peak diffusion restriction occurring at ages 50 to 60 years.^23,24 In assessing lens functional changes, DLI by i-Trace, which incorporates HOAs and contrast sensitivity, shows a strong correlation with age, average density, subjective lens grading, and BCDVA, highlighting its value in evaluating presbyopic lens function.^17,28,32 Our research suggests that a DLI cutoff value of 7.75 is crucial for identifying lens dysfunction. In the context of lens density measurement, Pentacam contributes most significantly with the highest diagnostic efficacy. The diagnostic efficacy of SS-OCT is comparatively lower, positioning it as a supplementary tool. The utilization of Pentacam has enabled automatic and quantitative three-dimensional measurement of lens density based on the Scheimpflug tomographic system, which is widely adopted across institutions.^33,34 There is good inter-device agreement in lens densitometry observed between Scheimpflug devices and SS-OCT.^35–37 Considering that SS-OCT may not be accessible in every hospital, its measurements can serve to complement other diagnostic modalities. 

Whereas our cross-sectional design precludes direct validation of surgical outcomes, the classification model identifies distinct lens structural and functional profiles that align with established surgical indicators (e.g. advanced lens density, and reduced DLI). These pathophysiological differentiators, combined with analogous clustering methodologies successfully applied in prior ophthalmic studies, support the biological plausibility of our framework.^38,39 We acknowledge the need for prospective validation of postoperative outcomes in stratified groups – a critical next step to confirm clinical utility. Future studies will evaluate the predictive capacity of this system in guiding personalized surgical decisions. 

In conclusion, our study developed an AI-based presbyopia classification system using lens functional and structural parameters to aid in the decision-making process for refractive or lens surgery. The cutoff value of DLI, in combination with the lens density cutoff values from Pentacam, can enable rapid diagnostic classification. Additionally, our website with AI-based algorithms provides a precise and automated classification tool. 

Acknowledgments

Supported by research grants from Research and Development Program of China (2022YFC2502800), National Natural Science Foundation of China (82271069, 82371040, 82122017, 81870642, 81900839, 81970780, 81470613 and 81670835), Science and Technology Innovation Action Plan of Shanghai Science and Technology Commission (23Y11909800), Outstanding Youth Medical Talents of Shanghai “Rising Stars of Medical Talents” Youth Development Program, Shanghai Municipal Health Commission Project (2024ZZ1025 and 20244Z0015). 

Author Contributions: Conception and design: X.Zhu, X.Zhou, and X.L. Analysis and interpretation: X.Zhou, X.L., W.H., K.C., and J.Q. Data collection: X.Zhu, X.L., W.H., K.C., J.M., J.Q., J.Z., and Y.L. Obtained funding: X.Zhou, Y.L., and X.Zhu Overall responsibility: X.Zhou and X.Zhu. 

Disclosure: X. Zhu, None; X. Liu, None; W. He, None; K. Cheng, None; J. Meng, None; J. Qi, None; J. Zhao, None; Y. Lu, None; X. Zhou, None