Clinical decision support system based on deep learning for evaluating implantable collamer lens size and vault after implantable collamer lens surgery: a retrospective study

Yixuan Yang,Zhengqin Long,Bo Lei,Wei Liu,Jian Ye
DOI: https://doi.org/10.1136/bmjopen-2023-081050
IF: 3.006
2024-02-17
BMJ Open
Abstract:Objectives To aid doctors in selecting the optimal preoperative implantable collamer lens (ICL) size and to enhance the safety and surgical outcomes of ICL procedures, a clinical decision support system (CDSS) is proposed in our study. Design A retrospective study of patients after ICL surgery. Setting China Tertiary Myopia Prevention and Control Center. Participants 2772 eyes belonging to 1512 patients after ICL surgery. Data were collected between 2018 and 2022. Outcome measures A CDSS is constructed and used to predict vault at 1 month postoperatively and preoperative ICL dimensions using various artificial intelligence methods. Accuracy metrics as well as area under curve (AUC) parameters are used to determine the CDSS prediction methods. Results Among the ICL size prediction models, conventional neural networks (CNNs) achieve the best prediction accuracy at 91.37% and exhibit the highest AUC of 0.842. Regarding the prediction model for vault values 1 month after surgery, CNN surpasses the other methods with an accuracy of 85.27%, which has the uppermost AUC of 0.815. Thus, we select CNN as the prediction algorithm for the CDSS. Conclusions This study introduces a CDSS to assist doctors in selecting the optimal ICL size for patients while improving the safety and postoperative outcomes of ICL surgery.
medicine, general & internal
What problem does this paper attempt to address?
This paper attempts to address the problem of how to select the optimal preoperative ICL size in Implantable Collamer Lens (ICL) surgery to ensure the safety and effectiveness of the procedure. Specifically, the authors propose a deep learning-based Clinical Decision Support System (CDSS) aimed at assisting doctors in selecting the most appropriate ICL size by predicting the vault one month postoperatively. ### Main Issues 1. **Selecting the Optimal ICL Size**: Currently, the selection of ICL size mainly relies on parameters such as White-to-White (WTW) and anterior chamber depth provided by manufacturers. However, these indirect measurement methods have errors, and there are only 4 ICL sizes available on the market, which may result in postoperative vault being too high or too low. 2. **Improving Surgical Safety and Effectiveness**: Accurately selecting the ICL size is crucial for maintaining the safety of the postoperative vault and the success rate of the surgery. Existing formula regression methods perform poorly in multi-parameter regression, thus a more precise ICL selection system is needed. ### Solution The authors propose a deep learning-based CDSS, implemented through the following steps: 1. **Data Collection**: A retrospective study collected data from 2772 eyes that underwent ICL surgery at a tertiary myopia prevention center in China from 2018 to 2022. 2. **Feature Parameter Analysis**: Analyzed the impact of 13 preoperative parameters (such as age, pupil size, spherical equivalent, cylindrical equivalent, AOD500, AL, K1, K2, CACD, CT, WTW, STS, etc.) on ICL size and postoperative vault. 3. **Model Construction and Comparison**: Built predictive models using five algorithms (CNN, LSTM, BPNN, RBNN, and RF) and compared the performance of different models through cross-validation and evaluation metrics (such as accuracy and AUC). 4. **System Structure**: The CDSS is divided into two stages: the first stage predicts the recommended ICL size based on preoperative parameters; the second stage adjusts the ICL size based on the predicted postoperative vault value to ensure it is within the normal range (250-750 μm). ### Results - **ICL Size Prediction**: The CNN model showed the highest accuracy (91.37%) in ICL size prediction, with an AUC of 0.842. - **Postoperative Vault Prediction**: The CNN model also showed the highest accuracy (85.27%) in postoperative vault prediction, with an AUC of 0.815. ### Conclusion This study proposes a deep learning-based CDSS that can assist doctors in selecting the optimal ICL size, thereby improving the safety and effectiveness of ICL surgery. Future research can further optimize the algorithm and integrate multi-center data to enhance the system's predictive capability.