Improving the estimation of the spherical equivalent subjective refraction using objective information on accommodation

Machine learning and deep learning have previously been used to predict the subjective refraction endpoint by objective means with modest success. This study aimed to enhance predictive accuracy by training linear regression models with normal equations using accommodative response and optical quality data. Three models were tested on 176 eyes, with input variables obtained from a Hartmann-Shack aberrometer and an autorefractor. The best model reduced mean absolute error by 40% compared to the objective refraction provided by a commercial autorefractometer and achieved 95% limits of agreement with subjective refraction of ±0.54 D, approaching the subjective refraction inter-examiner variability. Incorporating accommodative response data improved prediction accuracy over objective refraction alone and previous approaches.