Preconditioning of clinical data for intraocular lens formula constant optimisation using Random Forest Quantile...

Area of Science:

• Ophthalmology
• Data Science
• Biostatistics

Background:

• Accurate prediction of refractive outcomes after cataract surgery is crucial for patient satisfaction.
• Intraocular lens (IOL) formula constants require optimization using clinical data to improve prediction accuracy.
• Identifying and removing outliers in clinical datasets is essential for robust formula constant optimization.

Purpose of the Study:

• To implement and assess a fully data-driven strategy for outlier identification in clinical datasets.
• To optimize formula constants for predicting refraction after cataract surgery using an advanced outlier detection method.
• To evaluate the efficacy of random forest quantile regression for outlier detection in ophthalmic datasets.

Main Methods:

• Utilized two clinical datasets (N=888/403) with preoperative biometrics, IOL power, and postoperative spherical equivalent (SEQ).
• Developed a random forest quantile regression algorithm with bootstrap resampling to identify outliers.
• Defined outlier 'fences' based on interquartile range (IQR) of SEQ and predicted refraction (REF) to remove erroneous data points before recalculating formula constants for SRKT, Haigis, and Castrop formulae.

Main Results:

• The random forest quantile regression method successfully identified outliers in both datasets across different formulae (e.g., 4-32 outliers depending on formula and dataset).
• Recalculation of formula constants after outlier removal led to a slight reduction in root mean squared prediction errors (e.g., from 0.4449 dpt to 0.4348 dpt for Haigis in DS1).
• Demonstrated a reduction in formula prediction errors, indicating improved accuracy after outlier exclusion.

Conclusions:

• Random forest quantile regression enables a fully data-driven outlier identification strategy in the response space.
• This method is effective for optimizing IOL formula constants and improving refractive prediction accuracy.
• For real-world applications, combining this response-space method with parameter-space outlier identification is recommended for comprehensive dataset qualification.