New multilayer perceptron method appears highly accurate in predicting IOL power

The retrospective review evaluates the efficacy of a multilayer perceptron (MLP) for IOL power calculation.

Study design

The study comprised 15,728 eyes that underwent uncomplicated cataract surgery by 19 surgeons. Initially, 2 MLP neural networks were developed with the data from 18 of the 19 surgeons. The first MLP divided the original data into 2 groups based on whether the axial length (AL) was longer or shorter than 22 mm. A second MLP did not separate the data based on AL. The neural networks were trained and then used to optimize the predictive MLPs. 

The authors used the predictive MLPs to predict intraocular power needed for the 812 cases performed by the single surgeon whose data had been excluded from the training process.

Outcomes

The MLPs were capable of predicting the IOL power needed to achieve the refractive goal within an error of less than 0.5 diopters in more than 95% of patients. This exceeded the predictive power of the models used at the time of surgery. 

Limitations

The limitations of MLPs include variability between different patient populations and the need to train an MLP to address specific populations. In addition, a prospective, randomized controlled trial would offer a more direct comparison between MLP and currently available methods. 

Clinical significance

Current technology only achieves the desired postoperative refractive goal within 0.5 D about 80% of the time. With further refinement, MLPs might become a new method with greater predictive value than currently available methods for predicting IOL power.