The ZEISS AI IOL Calculator outperformed Barrett. In short eyes, the ZEISS AI had more eyes within ±0.50 D of prediction error (76.1%) than did Barrett (71.3%) and Kane (71.9%).
Abstract
Purpose:
To evaluate the ZEISS AI IOL Calculator (ZEISS AI) and compare its accuracy in refractive prediction with the Barrett Universal II (BUII) and Kane formulas.
Setting:
Cullen Eye Institute, Baylor College of Medicine, Houston, Texas.
Design:
Retrospective case series.
Methods:
The ZEISS AI IOL Calculator is an artificial intelligence (AI)–based IOL-optimized formula. The refractive prediction errors (PEs) were calculated in the entire dataset and subgroups of short eyes (axial length [AL] ≤22.5 mm) and long eyes (AL ≥25.0 mm). The SD, root mean square absolute error (RMSAE), mean absolute error (MAE), median absolute error (MedAE), and percentage of eyes within ±0.25 diopter (D), ±0.50 D, ±0.75 D, and ±1.00 D of PEs were calculated. Values with ZEISS AI were compared with those from BUII and Kane. Advanced statistical methods were applied using R.
Results:
A dataset of 10 838 eyes was included. Compared with ZEISS AI, BUII produced significantly greater SDs, RMSAEs, and MAEs in the whole group and short eyes, and the Kane had greater SD, RMSAE, and MAE in short eyes (all adjusted P < .05); the BUII had significantly lower percentages of eyes within ±0.50 D of PEs in the whole group (80.0% vs 81.2%) and in short eyes (71.3% vs 76.1%), and the Kane had lower percentage of eyes within ±0.50 D of PEs in short eyes (71.9% vs 76.1%) (all adjusted P < .05).
Conclusions:
The ZEISS AI IOL Calculator had superior performance compared with the BUII and Kane formulas, especially in short eyes.
In recent years, the application of artificial intelligence (AI) in intraocular lens (IOL) power calculation has garnered significant attention. Of the IOL formulas involving AI algorithms, 2 (Hill RBF and Karmona) are pure data-driven using AI, and the majority are hybrid that combine the AI algorithms and geometric optics or ray tracing in formula development, such as the Ladas Super Formula, Kane, Pearl DGS, EVO, and ZEISS AI IOL Calculator.1–5 The ZEISS AI IOL Calculator is a new IOL constant-free approach of IOL power calculation. It was developed using a combination of paraxial ray tracing and neural networks to perform a physics informed training process.
In a recent study, we compared the prediction accuracy of the ZEISS AI IOL Calculator with other formulas in short eyes from multicenter case series.4 The purpose of this study was to evaluate the ZEISS AI IOL Calculator and compare its accuracy in refractive prediction with the Barrett Universal II (BUII) and Kane formulas in the ZEISS VERACITY Surgery Planner database (Carl Zeiss Meditec AG).
METHODS
Patients
Deidentified patient data from the VERACITY Surgery Planner USA database were used in this study. We reviewed eyes with capsular bag implantation of 3 monofocal nontoric IOLs: SN60WF (Alcon Laboratories, Inc.), ZCB00 (Johnson & Johnson Vision), and MX60E (Bausch & Lomb, Inc.). Inclusion criteria for the testing dataset were eyes that had (1) no prior corneal refractive surgery or additional procedures during the cataract surgery, such as the corneal arcuate incisions for corneal astigmatism correction; (2) available postoperative manifest refraction performed at 3 weeks or later with corrected distance visual acuity of 20/30 or better; and (3) 1 eye from each patient. If both eyes of a patient met the inclusion criteria, 1 eye was randomly selected. Ocular biometric data were obtained from 2 biometers: IOLMaster and Lenstar.
This study was deemed exempt from full review by the Institution Review Board at Baylor College of Medicine because it used an external dataset with deidentified patient data. Consent to use the deidentified patient data was obtained from each site that used the VERACITY Surgery Planner. The deidentification process conformed to the Health Insurance Portability and Accountability Act guidelines.
IOL Calculation Formulas
The ZEISS AI IOL Calculator (ZEISS AI) is a physics-based and AI-based IOL calculation algorithm. Machine learning (supervised learning) was used to train artificial neural networks (multilayer perceptrons) to predict the residual refraction for a given biometry and IOL power. The training was carried on in 2 steps: (1) pretraining on a large amount of simulated paraxial ray-tracing data and (2) fine-tuning using IOL model-specific empirical data obtained from a large patient population (short, medium, and long eyes). The data were collected by the ZEISS VERACITY Surgery Planner software. Therefore, in contrast to purely data-driven AI approaches, the learning process in the ZEISS AI integrated detailed physical knowledge, optical simulations, and fine-tuning for each individual IOL model. This approach was designed to optimize the IOL power and refractive outcome prediction through a machine learning–based optimization for each IOL model instead of optimizing only single-IOL constants.
The BUII was developed based on geometric optics.
The Kane formula used a hybrid approach, based on geometric optics and incorporating both AI and regression components.
Data Analysis
Using a different dataset than was used in the original development of the Zeiss AI, we evaluated the refractive prediction errors (PEs) from the ZEISS AI compared with the PEs calculated from the BUII and Kane formulas after training and optimizing the formulas for the specific IOL models.
To maximize the training and testing data, a cross-validation approach was used to train the ZEISS AI and optimize the BUII and Kane formulas and evaluate their performances. For the cross-validation, 10 folds were used, respectively. The concept of cross-validation has previously been used in publications in the context of IOL calculation, for example about IOL position prediction, posterior cornea curvature prediction, and constant optimization.6–8
The manifest refraction lane length was obtained from each center participating in the VERACITY Surgery Planner. When necessary, the spherical equivalent of manifest refraction was converted to a 6 m distance. The PEs were calculated by subtracting the predicted refraction with each formula from the postoperative manifest refraction. The mean error of all formulas was zeroed by subtracting the corresponding mean error of every formula from each prediction. A negative value indicates a more myopic outcome than the target refraction. We also calculated the SD of refractive PE, root mean square absolute error (RMSAE), mean absolute error (MAE), median absolute error (MedAE), and percentage of eyes within ±0.25 diopter (D), ±0.50 D, ±0.75 D, and ±1.00 D of PEs. Data analyses were performed for the whole group, and subgroups of short eyes (axial length [AL] ≤22.5 mm) and long eyes (AL ≥25.0 mm).
Statistical Analysis
The Shapiro-Wilk test was used to assess the normality of PE distribution. One-sample t test or Wilcoxon signed rank test was used to assess if the mean PEs were significantly different from zero.
The SDs, RMSAEs, MAEs, MedAEs, and percentages of eyes within certain PEs were compared between methods using the advanced statistical approaches outlined in our previous studies.9–12 The heteroscedastic statistical method was used to compare the SD values. A bootstrap-t method was applied to compare the RMSAEs, MAEs, and MedAEs. The McNemar chi-squared test was performed to compare the percentages of eyes within certain PEs. The Holm correction was applied for multiple comparisons. Statistical analysis was performed using the R project for statistical computing and IBM SPSS Statistics (v. 28.0.1.1). A probability of less than 5% (P < .05) was considered statistically significant.
RESULTS
Patient data are summarized in Table 1. A total of 10 838 eyes from 10 838 patients were included in this study.
Table 1.
Summary of clinical data in the entire dataset (n = 10 838)
| Parameters | Mean ± SD | Range |
| Mean average K (D) | 43.95 ± 1.53 | 37.59, 49.54 |
| AL (mm) | 23.95 ± 1.20 | 19.71, 32.11 |
| ACD (mm) | 3.19 ± 0.39 | 2.01, 4.50 |
| LT (mm) | 4.59 ± 0.44 | 3.01, 5.70 |
| IOL power (D) | 20.11 ± 3.40 | 0.00, 34.00 |
| Time between surgery and postop refraction (d) | 43.90 ± 35.65 | 21, 353 |
| Postop SE (D) | −0.33 ± 0.68 | −3.00, +1.00 |
ACD = anterior chamber depth; AL = axial length; K = keratometry; LT = lens thickness; SE = spherical equivalent
Refractive PEs in the Whole Group
Table 2 presents the PEs in the whole group. The MAE values as a function of AL are shown in Figure 1. Compared with the ZEISS AI, significant findings (all adjusted P < .05) were as follows: (1) in SN60WF eyes, the BUII produced greater SD, RMSAE, MAE, and MedAE, and lower percentages of eyes with PE of ±0.25 D, and ±0.50 D; (2) in ZCB00 eyes, the BUII produced greater SD, RMSAE, and MAE, and lower percentage of eyes with PE of ±0.75 D; (3) in MX60E eyes, the BUII produced greater SD, RMSAE, and MAE, and lower percentages of eyes with PE of ±0.50 D and ±0.75 D; the Kane produced greater SD, RMSAE, and MAE, and lower percentage of eyes with PE of ±0.75 D; (4) in all eyes, the BUII produced greater SD, RMSAE, MAE, and MedAE, and lower percentages of eyes with PE of ±0.25 D, ±0.50 D, ±0.75 D, and ±1.00 D.
Table 2.
Refractive prediction errors (D) in the whole group
| Formula | Mean | SD | RMSAE | MAE | MedAE | Range | % ± 0.25 D | % ± 0.50 D | % ± 0.75 D | % ± 1.00 D |
| SN60WF (n = 6041) | ||||||||||
| BUII | 0.00 | 0.391a | 0.391a | 0.301a | 0.245a | −2.04, 1.76 | 50.9%a | 82.2%a | 94.1% | 98.2% |
| Kane | 0.01 | 0.377 | 0.377 | 0.289 | 0.230 | −2.00, 1.79 | 53.7% | 83.3% | 94.7% | 98.5% |
| ZEISS AI | 0.00 | 0.379 | 0.379 | 0.290 | 0.231 | −1.97, 1.78 | 53.2% | 83.3% | 94.6% | 98.4% |
| ZCB00 (n = 2744) | ||||||||||
| BUII | −0.01 | 0.423a | 0.423a | 0.328a | 0.263 | −1.94, 1.40 | 47.8% | 78.4% | 91.7%a | 97.6% |
| Kane | 0.00 | 0.414 | 0.414 | 0.320 | 0.258 | −2.00, 1.41 | 48.8% | 78.9% | 92.5% | 98.1% |
| ZEISS AI | 0.00 | 0.412 | 0.412 | 0.318 | 0.254 | −1.98, 1.56 | 49.3% | 79.0% | 92.7% | 98.1% |
| MX60E (n = 2053) | ||||||||||
| BUII | 0.02 | 0.453a | 0.453a | 0.348a | 0.269 | −1.78, 1.66 | 46.9% | 75.5%a | 89.2%a | 96.7% |
| Kane | 0.01 | 0.444a | 0.444a | 0.339a | 0.268 | −1.97, 1.68 | 47.7% | 76.9% | 90.1%a | 96.8% |
| ZEISS AI | −0.01 | 0.436 | 0.436 | 0.334 | 0.280 | −1.90, 1.66 | 46.5% | 78.0% | 91.3% | 97.1% |
| All (n = 10 838) | ||||||||||
| BUII | 0.00 | 0.411a | 0.411a | 0.316a | 0.253a | −2.04, 1.76 | 49.3%a | 80.0%a | 92.6%a | 97.8%a |
| Kane | 0.00 | 0.400 | 0.400 | 0.306 | 0.242 | −2.00, 1.79 | 51.3% | 81.0% | 93.3% | 98.1% |
| ZEISS AI | 0.01 | 0.399 | 0.399 | 0.306 | 0.244 | −1.98, 1.78 | 50.9% | 81.2% | 93.5% | 98.1% |
BUII = Barrett Universal II; MAE = mean absolute error; MedAE = median absolute error; RMSAE = root mean square absolute error; ZEISS AI = ZEISS AI IOL Calculator
Significant difference compared with ZEISS AI (adjusted P < .05)
Figure 1.
Mean absolute errors vs AL in the entire dataset. To eliminate noise with extremely short and long eyes, this graph only included the average mean absolute error values with more than 200 eyes at each axial length interval. AL = axial length; Av. = average, axial length interval size: 0.7 mm
Refractive PEs in Short Eyes
In SN60WF eyes, the Kane formula produced myopic mean PE (Table 3). In ZCB00 eyes, all 3 formulas had myopic mean PEs. In MX60E eyes, the BUII and Kane had hyperopic mean PEs. In all eyes, Kane had myopic mean PE.
Table 3.
Refractive prediction errors (D) in short eyes with axial length ≤22.5 mm
| Formula | Mean | SD | RMSAE | MAE | MedAE | Range | % ± 0.25 D | % ± 0.50 D | % ± 0.75 D | % ± 1.00 D |
| SN60WF (n = 521) | ||||||||||
| BUII | −0.03 | 0.464a | 0.464a | 0.349a | 0.269 | −2.04, 1.63 | 47.2% | 76.2% | 91.0% | 96.0% |
| Kane | −0.06b | 0.440 | 0.444 | 0.331 | 0.243 | −2.00, 1.35 | 51.4% | 77.0% | 91.7% | 96.5% |
| ZEISS AI | −0.01 | 0.444 | 0.444 | 0.334 | 0.255 | −1.97, 1.36 | 49.3% | 78.3% | 91.2% | 96.7% |
| ZCB00 (n = 224) | ||||||||||
| BUII | −0.23b | 0.474 | 0.525a | 0.418 | 0.358 | −1.63, 0.93 | 37.9% | 64.7% | 83.9% | 95.1% |
| Kane | −0.23b | 0.448 | 0.504a | 0.409 | 0.375 | −1.40, 0.88 | 37.1% | 64.7% | 85.3% | 96.4% |
| ZEISS AI | −0.07b | 0.457 | 0.462 | 0.382 | 0.328 | −1.08, 1.04 | 36.2% | 71.0% | 89.7% | 97.8% |
| MX60E (n = 192 | ||||||||||
| BUII | 0.19b | 0.510a | 0.541a | 0.423a | 0.360 | −1.54, 1.60 | 37.0% | 65.6%a | 84.4%a | 93.8% |
| Kane | 0.16b | 0.520a | 0.543a | 0.420a | 0.356 | −1.47, 1.67 | 39.1% | 66.7%a | 83.9%a | 92.7% |
| ZEISS AI | 0.00 | 0.466 | 0.465 | 0.357 | 0.303 | −1.55, 1.27 | 44.3% | 76.0% | 89.6% | 96.4% |
| All (n = 937) | ||||||||||
| BUII | −0.03 | 0.495a | 0.496a | 0.381a | 0.303 | −2.04, 1.63 | 42.9% | 71.3%a | 87.9%a | 95.3%a |
| Kane | −0.06b | 0.477a | 0.480a | 0.368a | 0.286 | −2.00, 1.67 | 45.5% | 71.9%a | 88.6%a | 95.7%a |
| ZEISS AI | −0.03 | 0.452 | 0.453 | 0.350 | 0.283 | −1.97, 1.36 | 45.1% | 76.1% | 90.5% | 96.9% |
BUII = Barrett Universal II; MAE = mean absolute error; MedAE = median absolute error; RMSAE = root mean square absolute error; ZEISS AI = ZEISS AI IOL Calculator
Significant difference compared with ZEISS AI (adjusted P < .05)
Significantly different from zero (adjusted P < .05)
Compared with the ZEISS AI, significant findings (all adjusted P < .05) were as follows: (1) in SN60WF eyes, the BUII produced greater SD, RMSAE, and MAE; (2) in ZCB00 eyes, the BUII and Kane had greater RMSAEs; (3) in MX60E eyes, the BUII had greater SD, RMSAE, and MAE, and lower percentages of eyes with PE of ±0.50 D and ±0.75 D; the Kane had greater SD, RMSAE, and MAE, and lower percentages of eyes with PE of ±0.50 D and ±0.75 D; (4) in all eyes, the BUII and Kane produced greater SDs, RMSAEs, and MAEs, and lower percentages of eyes with PE of ±0.50 D, ±0.75 D, and ±1.00 D.
Refractive PEs in Long Eyes
In SN60WF eyes, the BUII and Kane produced slight hyperopic mean PEs (Table 4). In ZCB00 eyes, the BUII and Kane had hyperopic mean PEs. In all eyes, the BUII and Kane had hyperopic mean PEs. Compared with the ZEISS AI, significant findings (all adjusted P < .05) were as follows: (1) in SN60WF eyes, the BUII had lower percentage of eyes with PE of ±0.50 D; (2) in ZCB00 eyes, the BUII and Kane had smaller SDs; and (3) in MX60E eyes, the BUII had greater SD, RMSAE, and MAE.
Table 4.
Refractive prediction errors (D) in long eyes with axial length ≥25.0 mm
| Formula | Mean | SD | RMSAE | MAE | MedAE | Range | % ± 0.25 D | % ± 0.50 D | % ± 0.75 D | % ± 1.00 D |
| SN60WF (n = 985) | ||||||||||
| BUII | 0.07b | 0.365 | 0.372 | 0.281 | 0.220 | −1.33, 1.76 | 55.8% | 84.4%a | 94.6% | 98.7% |
| Kane | 0.05b | 0.358 | 0.362 | 0.273 | 0.216 | −1.40, 1.79 | 57.0% | 85.4% | 94.8% | 98.8% |
| ZEISS AI | 0.00 | 0.364 | 0.364 | 0.275 | 0.211 | −1.43, 1.78 | 56.3% | 86.4% | 95.1% | 98.6% |
| ZCB00 (n = 498) | ||||||||||
| BUII | 0.09b | 0.388a | 0.398 | 0.308 | 0.249 | −1.26, 1.40 | 50.8% | 81.5% | 93.2% | 97.6% |
| Kane | 0.08b | 0.387a | 0.395 | 0.302 | 0.240 | −1.25, 1.40 | 52.0% | 80.9% | 94.4% | 98.2% |
| ZEISS AI | −0.03 | 0.404 | 0.404 | 0.308 | 0.237 | −1.42, 1.43 | 52.4% | 80.5% | 92.6% | 97.6% |
| MX60E (n = 330) | ||||||||||
| BUII | 0.02 | 0.455a | 0.455a | 0.355a | 0.273 | −1.21, 1.36 | 46.4% | 74.5% | 89.1% | 96.1% |
| Kane | −0.02 | 0.439 | 0.438 | 0.343 | 0.285 | −1.31, 1.25 | 44.5% | 78.2% | 90.6% | 97.0% |
| ZEISS AI | 0.04 | 0.434 | 0.436 | 0.339 | 0.282 | −1.33, 1.25 | 44.2% | 77.3% | 91.8% | 96.4% |
| All (n = 1813) | ||||||||||
| BUII | 0.07b | 0.390 | 0.395 | 0.302 | 0.238 | −1.33, 1.76 | 52.7% | 81.8% | 93.2% | 97.9% |
| Kane | 0.05b | 0.383 | 0.386 | 0.294 | 0.232 | −1.40, 1.79 | 53.3% | 82.8% | 93.9% | 98.3% |
| ZEISS AI | 0.00 | 0.389 | 0.389 | 0.295 | 0.232 | −1.43, 1.78 | 53.1% | 83.1% | 93.8% | 97.9% |
BUII = Barrett Universal II; MAE = mean absolute error; MedAE = median absolute error; RMSAE = root mean square absolute error; ZEISS AI = ZEISS AI IOL Calculator
Significant difference compared with ZEISS AI (adjusted P < .05)
Significantly different from zero (adjusted P < .05)
DISCUSSION
The usage of AI in IOL power calculation has the potential to improve accuracy in refractive prediction. The ZEISS AI IOL Calculator is an AI-based and physics-based IOL-optimized formula developed with large datasets (minimum of 1000 eyes per IOL model) from the ZEISS VERACITY Surgery Planner. Instead of optimizing only 1 IOL constant for each IOL model, the ZEISS AI IOL calculator was designed to optimize the IOL power and refractive outcome prediction through a machine learning–based optimization for each IOL model. In this study, we compared the accuracy of ZEISS AI in refractive prediction with the BUII and Kane formulas.
Our results showed that the ZEISS AI produced significantly smaller SD, RMSAE, and MAE in the whole group, compared with the BUII. In short eyes, the ZEISS AI had significantly smaller SD, RMSAE, and MAE than did the BUII and Kane. It should be highlighted that for the formula comparison on a subset of data that is not representative of the entire population (eg, short or long eyes), the systematic mean error must be considered. Therefore, the RMSAE is the more informative metric on such populations because it considers the systematic mean error, whereas the SD removes it. As shown in Figure 1, the accuracy improvements with the ZEISS AI are more prominent in short eyes. Although the clinical improvements of the ZEISS AI are small in the whole group, in short eyes, the ZEISS AI produced 4.8% and 4.2% more eyes within ±0.50 D of PEs than did BUII and Kane, respectively.
In 278 short eyes with AL <22 mm, compared with the BUII and Kane, Kenny et al. reported that the ZEISS AI had smaller RMSAE (0.55 D vs 0.60 D for BUII and 0.59 D for Kane) and smaller MAE (0.40 D vs 0.44 D for both BUII and Kane); the percentages of eyes within ±0.50 D of PEs were 73% with ZEISS, and 67% with both BUII and Kane formulas.4 Similar findings were observed in this study with AL ≤22.5 mm. The ZEISS AI produced smaller RMSAE and MAE than did BUII and Kane, and 76.1% of eyes had PEs within ±0.50 D with ZEISS AI, compared with 71.3% with BUII and 71.9% with Kane.
In this study, we compared the ZEISS AI IOL Calculator with 1 geometric optics-based formula (BUII) and 1 hybrid formula based on geometric optics and incorporating both AI and regression components (Kane). These 3 formulas are included in the ZEISS VERACITY Surgery Planner. Accuracy comparisons in refractive prediction between BUII, Kane, and other AI-based formulas have been reported in the literature.1,2,13–15 In general, PEs with these modern formulas are comparable overall.
In this study, we assessed the performance of ZEISS AI IOL Calculator in a dataset from Veracity that included outcomes from multicenters in the United States. Our results are comparable with findings from the study by Melles et al.14 In the study by Melles et al., a dataset from the Kaiser Permanente Northern California was used. In 13 301 eyes implanted with SN60WF, 80.8% of eyes had PEs within ±0.5 D with BUII formula. In our study of 6041 eyes implanted with SN60WF, 82.2% of eyes had PEs within ±0.5 D with BUII and 83.3% with ZEISS AI (Table 2).
With the cross-validation approach, all available data can be used for training and testing, while at the same time, it ensures a clear separation between the training and testing datasets. This approach maximizes the study power and is a useful tool in studies with small sample sizes.
In this study, we also calculated and compared the RMSAE values between formulas. RMSAE is more sensitive to outliers because it gives greater weight to extreme values. Including both RMSAE and MAE in comparisons of PEs between formulas can provide valuable insights.
Our study has limitations: (1) this was a retrospective study with data from multicenters, and personnel who performed the manifest refraction were unknown. For the testing evaluation, we only included cases that had postoperative manifest refraction performed at 3 weeks or later and had corrected distance visual acuity of 20/30 or better; (2) the performance comparison was only assessed for 3 models of monofocal nontoric IOLs; (3) the period between surgery and postoperative refraction is wide (21 to 353 days); and (4) we do not have information about the macular status in these cases. To ensure the accuracy of manifest refraction, we used cutoff value of corrected distance visual acuity of 20/30 or better.
In summary, our results demonstrate that the ZEISS AI produced significantly smaller SD, RMSAE, MAE, and MedAE in the whole group than did the BUII. In short eyes, the ZEISS AI had significantly smaller SD, RMSAE, and MAE, compared with BUII and Kane. More studies are desirable to assess the performance of the ZEISS AI in additional datasets.
WHAT WAS KNOWN
Ever-increasing refractive expectations are being met with the growing predictive accuracy of IOL power calculation formulas, driven by advances in machine learning and a deeper understanding of neural networks.
The ZEISS AI IOL Calculator is a new machine learning–based, IOL-optimized formula that is available on the VERACITY Surgery Planner.
WHAT THIS PAPER ADDS
Comparing the Zeiss AI, Barrett Universal II, and Kane with a dataset of 10 838 eyes, the ZEISS AI outperformed the Barrett Universal II in the overall group, although the differences were small.
In short eyes, compared with the BUII and Kane, the ZEISS AI produced 4.8% and 4.2% more eyes with prediction errors within ±0.50 D of target.
Footnotes
Supported in part by SRB Charitable Corp., Fort Worth, Texas, the Sid W. Richardson Foundation, Fort Worth, Texas, and an unrestricted grant from Research to Prevent Blindness, New York, New York (Wang and Koch).
Disclosures: L. Wang is a consultant for Carl Zeiss Meditec AG, Acufocus, Inc., Cassini Technologies, and Alcon Laboratories, Inc. D.D. Koch is a consultant for Alcon Laboratories, Inc., Carl Zeiss Meditec AG, and Johnson & Johnson Vision. N. Bensaid and H. Burwinkel are affiliated to Carl Zeiss Meditec AG.
First author:
Li Wang, MD, PhD
Department of Ophthalmology, Cullen Eye Institute, Baylor College of Medicine, Houston, Texas
Contributor Information
Li Wang, Email: liw@bcm.edu.
Hendrik Burwinkel, Email: hendrik.burwinkel@zeiss.com.
Nicolas Bensaid, Email: nicolas.bensaid@zeiss.com.
REFERENCES
- 1.Carmona González D, Palomino Bautista C. Accuracy of a new intraocular lens power calculation method based on artificial intelligence. Eye (Lond). 2021;35(2):517–522 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Voytsekhivskyy OV, Hoffer KJ, Tutchenko L, Cooke DL, Savini G. Accuracy of 24 IOL power calculation methods. J Refract Surg. 2023;39(4):249–256 [DOI] [PubMed] [Google Scholar]
- 3.Gökce SE, Zeiter JH, Weikert MP, Koch DD, Hill W, Wang L. Intraocular lens power calculations in short eyes using 7 formulas. J Cataract Refract Surg. 2017;43:892–897 [DOI] [PubMed] [Google Scholar]
- 4.Kenny PI, Kozhaya K, Truong P, Weikert MP, Wang L, Hill WE, Koch DD. Efficacy of segmented axial length and artificial intelligence approaches to intraocular lens power calculation in short eyes. J Cataract Refract Surg. 2023;49(7):697–703 [DOI] [PubMed] [Google Scholar]
- 5.ZEISS AI IOL Calculator. Accessed February 2, 2024. https://www.zeiss.com/meditec/en/products/veracity/ai-iol-calculator.html [Google Scholar]
- 6.Langenbucher A, Szentmáry N, Wendelstein J, Hoffmann P. Artificial intelligence, machine learning and calculation of intraocular lens power. Klin Monbl Augenheilkd. 2020;237(12):1430–1437 [DOI] [PubMed] [Google Scholar]
- 7.Langenbucher A, Cayless A, Szentmáry N, Weisensee J, Wendelstein J, Hoffmann P. Prediction of total corneal power from measured anterior corneal power on the IOLMaster 700 using a feedforward shallow neural network. Acta Ophthalmol. 2022;100(5):e1080–e1087. [DOI] [PubMed] [Google Scholar]
- 8.Langenbucher A, Szentmáry N, Cayless A, Müller M, Eppig T, Schröder S, Fabian E. IOL formula constants: strategies for optimization and defining standards for presenting data. Ophthalmic Res. 2021;64(6):1055–1067 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Wang L, Koch DD, Hill W, Abulafia A. Pursuing perfection in intraocular lens calculations: III. Criteria for analyzing outcomes. J Cataract Refract Surg. 2017;43(8):999–1002 [DOI] [PubMed] [Google Scholar]
- 10.Holladay JT, Wilcox RR, Koch DD, Wang L. Review and recommendations for univariate statistical analysis of spherical equivalent prediction error for IOL power calculations. J Cataract Refract Surg. 2021;47(1):65–77 [DOI] [PubMed] [Google Scholar]
- 11.Wang L, Holladay J, Wilcox RR, Koch DD. Wrap-up R functions for statistical analysis of spherical equivalent and astigmatic prediction errors. J Cataract Refract Surg. 2023;49(8):890–891 [DOI] [PubMed] [Google Scholar]
- 12.Holladay J, Wilcox RR, Koch DD, Wang L. Statistics of prediction error for non-Gaussian dependent and independent datasets. J Cataract Refract Surg. 2023;49(4):440–442 [DOI] [PubMed] [Google Scholar]
- 13.Li T, Stein J, Nallasamy N. Evaluation of the Nallasamy formula: a stacking ensemble machine learning method for refraction prediction in cataract surgery. Br J Ophthalmol. 2023;107(8):1066–1071 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Melles RB, Kane JX, Olsen T, Chang WJ. Update on intraocular lens calculation formulas. Ophthalmology. 2019;126(9):1334–1335 [DOI] [PubMed] [Google Scholar]
- 15.Yoon JH, Whang WJ. Comparison of accuracy of six modern intraocular lens power calculation formulas. Korean J Ophthalmol. 2023;37(5):380–386 [DOI] [PMC free article] [PubMed] [Google Scholar]