Artificial intelligence-based refractive error prediction and EVO-implantable collamer lens power calculation for myopia correction

Yinjie Jiang, Yang Shen, Xun Chen, Lingling Niu, Boliang Li, Mingrui Cheng, Yadi Lei, Yilin Xu, Chongyang Wang, Xingtao Zhou, Xiaoying WangEye 2023 · DOI

TL;DR

This paper introduces novel stacking machine learning (ML) models to predict postoperative refractive error and calculate the power of EVO-implantable collamer lenses (ICLs) for myopia correction. These AI-driven models demonstrate comparable accuracy to existing methods but show particular promise for improving predictions in patients with low-to-moderate myopia.

What Problem Does This Paper Address?

Correcting myopia with implantable collamer lenses (ICLs) offers excellent visual outcomes, but accurately calculating the correct lens power remains a challenge, especially for individuals with low to moderate myopia. Existing methods, like the modified vergence formula (MVF), can lead to refractive surprises (unexpected postoperative refractive errors). This is problematic because patients undergoing ICL surgery often have high expectations for visual outcomes, and even a small percentage of refractive errors can be clinically significant at a population level. The study aims to develop more precise ICL power calculation methods by leveraging machine learning and a comprehensive set of ocular parameters.

Methods

The researchers employed a stacking machine learning approach. This involves training multiple base ML models (Support Vector Regression - SVR, LASSO, Random Forest - RF, and XGBoost) and then using their predictions as input for a meta-model (the "stacker"). This technique aims to combine the strengths of different models to achieve better overall performance.

The stacking process had two levels:

  1. Level 1: Base models were trained on a merged dataset of both non-toric (NT-ICL) and toric (TICL) ICL cases, and then separately on NT-ICL and TICL data to capture specific features of each.
  2. Level 2: The predictions from the Level 1 models were used as new input features to train the final stacking models.

The performance of these ML models was compared against the manufacturer's ICL Power Calculation software, which uses the modified vergence formula (MVF). Evaluation metrics included Mean Absolute Error (MAE), Median Absolute Error (MedAE), standard deviation of prediction error, and the percentage of eyes with prediction errors within ±0.25, ±0.50, and ±0.75 diopters (D). Bland-Altman analyses were used to assess agreement between the ML models and MVF.

Datasets

The study utilized a retrospective dataset of 2767 eyes from 1678 patients who underwent uneventful NT-ICL or TICL implantation between 2014 and 2021 at the Eye and ENT Hospital of Fudan University, Shanghai, China.

  • Training/Validation Set: 2432 eyes from 1343 patients.
  • Testing Set: 335 eyes from 335 patients.

The dataset included a wide range of preoperative ocular dimensional parameters, such as axial length, keratometry, anterior chamber depth, and pupil diameter, along with patient age and refractive error.

Key Insights and Ideas

  • Ocular Dimensions Matter: The study confirms that various ocular dimensional parameters, beyond just refractive error and keratometry, are crucial for accurate ICL power prediction. These parameters likely influence postoperative lens position, which in turn affects refractive outcomes.
  • Stacking ML for Robustness: The stacking approach effectively leverages the diverse predictive capabilities of different ML algorithms, leading to more robust and accurate predictions than individual models alone.
  • Low-to-Moderate Myopia Advantage: The ML models showed a more pronounced improvement over MVF in patients with low-to-moderate myopia. This suggests that existing formulas may struggle more with this specific patient group, where subtle errors can have a greater impact.
  • Feature Importance: Preoperative spherical equivalent (SE) and sphere were consistently the most important features for predicting postoperative outcomes, highlighting their foundational role. For TICL cases, other ocular parameters played a more significant role, indicating the complexity of toric lens calculations.
  • Potential for a Novel Nomogram: The development of these ML models opens the door to creating a novel nomogram (a graphical tool or chart) for more personalized and accurate ICL power calculations.

Key Results

  • Best Performing Models:
    • For NT-ICL, the Random Forest model achieved the lowest MAE (0.339 D) for predicting SE, and the SVR model had the lowest MAE (0.386 D) for predicting sphere.
  • For TICL, the XGBoost model excelled, showing the lowest MAE for both SE (0.325 D) and sphere (0.308 D).
  • Comparison to MVF: The optimal ML models generally showed numerically lower SD, MAE, and MedAE compared to MVF. The percentage of eyes within ±0.25 D, ±0.50 D, and ±0.75 D prediction errors was comparable between ML models and MVF.
  • Subgroup Analysis: The advantage of ML models over MVF was more evident in patients with low-to-moderate myopia (preoperative SE > -6.00 D), where MAEs were numerically lower. This advantage was less pronounced in high myopia.
  • Bland-Altman Agreement: Bland-Altman plots demonstrated strong agreement between the predictive values of the optimal stacking ML models and the MVF, with mean differences close to zero.

Limitations and Future Work

  • Single-Center Study: The data was collected from a single center, and the authors acknowledge the need for validation with external, multicenter datasets to ensure generalizability.
  • No Cylinder Prediction for TICL: The study did not predict postoperative cylinder or calculate TICL cylinder power, as this requires more complex models and vector analysis. The correlation between ML models and MVF for TICL was weaker, and there was a larger disparity in lens sphere calculation.
  • Exclusion of Target Refractive Error: Patients with a myopic target (e.g., for presbyopia or monovision) were excluded from the lens power calculation analysis. Future work should incorporate target refractive error for a more personalized approach.
  • Short Follow-up Period: The study used postoperative data collected at one week and one month. Longer-term follow-up data is needed to assess refractive stability after ICL surgery.
  • Future Directions: The authors suggest exploring more personalized lens power designs for different age and preoperative power groups. They also aim to combine more parameters and explore models for predicting astigmatism with wider applications and higher accuracy. Incorporating longer follow-up data to predict refractive stability is also a key future direction.

Summary updated

Abstract

With various ocular dimensional parameters, ML models demonstrate comparable accuracy than existing MVF models and potential advantages in low-to-moderate myopia, and thus provide a novel nomogram for postoperative refractive error prediction and lens power calculation.
BibTeX
@inproceedings{jiang2023artificial,
  author = {Yinjie Jiang and Yang Shen and Xun Chen and Lingling Niu and Boliang Li and Mingrui Cheng and Yadi Lei and Yilin Xu and Chongyang Wang and Xingtao Zhou and Xiaoying Wang},
  title = {{Artificial intelligence-based refractive error prediction and EVO-implantable collamer lens power calculation for myopia correction}},
  booktitle = {Eye},
  year = {2023},
  doi = {10.1186/s40662-023-00338-1},
}
Figure