Prediction of IOL power
Described are implementations of systems and methods for an improved machine learning-based system that incorporates pre-operative and intraoperative measurements captured during surgery, as well as additional patient-specific data, to provide an individualized, highly accurate post-operative manifest refraction prediction. According to some embodiments, a determination engine generates a predictive feature set of one or more predictors associated with diagnostic measurements of one or more eyes and performs a recursive selection operation using one or more combinations within the predictive feature set and one or more models to produce a most predictive subset, the most predictive subset having a highest prediction accuracy among other predictive subsets for post-operative manifest refraction. The determination engine generates a determination model by refining and retraining the one or more models of the recursive selection operation utilizing the most predictive subset.
1 . A method comprising:
generating, by a determination engine executed by one or more processors, a determination model for predicting zero or near-zero post-operative manifest refraction error by:
iteratively selecting combinations of subsets of a predictive feature set of one or more predictors associated with diagnostic measurements of one or more eyes, and
applying an intermediate determination model on training data comprising diagnostic measurements and measured post-operative outcomes to each of the selected subsets,
wherein a subset having a highest prediction accuracy among the selected combination of subsets is included in a subsequent iteration.
2 . The method of claim 1 , wherein the diagnostic measurements comprise dry data from one or more diagnostic machines, the dry data comprising at least structural anatomy of the one or more eyes or position of an original crystalline lens.
3 . The method of claim 2 , wherein the diagnostic measurements account for post-operative lens settlement absent post-operative lens position calculations.
4 . The method of claim 1 , wherein the intermediate determination model comprises a support-vector machine comprising a radial basis function.
5 . The method of claim 1 , wherein a number of the one or more predictors is equal to or greater than 1000, and a number of predictors of the subset having the highest prediction accuracy in a final iteration of generating the determination model is equal to or less than 50.
6 . The method of claim 1 , wherein the determination engine comprises at least one of a mean absolute error, median absolute error, root means square error algorithm, and proportion of eyes within a diopter range to determine the prediction accuracy for each subset of the selected combination of subsets in each iteration of generating the determination model.
7 . The method of claim 1 , wherein the determination engine acquires a dataset comprising the diagnostic measurements corresponding to a plurality of patients, the diagnostic measurements comprising at least one selected lens attribute for each of the plurality of patients.
8 . The method of claim 7 , wherein the dataset comprises health record information from a first source and the diagnostic measurements are acquired from a second source.
9 . The method of claim 1 , wherein the diagnostic measurements comprise pre-operative intraocular lens biometric data, precision measurement data, three dimensional data, and biometry data derived from the three dimensional data.
10 . The method of claim 9 , wherein the determination model utilizes one or more outputs of one or more algorithms that use the pre-operative data as inputs, based on the pre-operative data being included in the subset having the highest prediction accuracy in a final iteration of generating the determination model.
11 . The method of claim 1 , wherein the subset having the highest prediction accuracy in a final iteration of generating the determination model has a prediction accuracy based on a difference between predicted spherical equivalence and actual spherical equivalence that resulted in zero or near zero.
12 . The method of claim 1 , wherein the subset having the highest prediction accuracy in a final iteration of generating the determination model has a prediction accuracy within about 0.5 D, 0.75 D, or 1.0 D of an absolute error.
13 . The method of claim 1 , further comprising generating a plurality of subsets of the predictive feature set comprising different combinations of the one or more predictors.
14 . The method of claim 1 , wherein iteratively selecting combinations of subsets of the predictive feature set comprises performing a linear regression algorithm on the predictors of the subset having the highest prediction accuracy in a previous iteration.