Methods and devices for increasing the precision of nomogram predictions

Abstract

A computer-implemented method (200) for training an artificial intelligence-based model (M) and a computer-implemented method (600) for supporting surgery on a patient's eye are provided, comprising such a model (M) which serves to calculate and / or provide refractive output data (933) based on input data (931), and in particular to generate nomograms. Prior art nomograms may either have insufficient accuracy or, if the accuracy is improved, generate implausible nomogram progressions. According to the invention, more accurate nomograms based on the output data (933) are enabled without implausible states.This is made possible by a cost function (915) comprising a nonlinear output value (917) of a univariate linear regression function representing a measure of the prediction error and at least one nonlinear penalty term (919). Furthermore, a trained model (M) or a weighting matrix (923) of such a model (M), as well as a computer program product, a data processing device (800), a computer-readable medium, and a laser therapy device are described.

Description

[0001] The present disclosure relates to a computer-implemented method for training an artificial intelligence-based model to predict output data based on input data. The disclosure further relates to a trained artificial intelligence-based model or a weighting matrix of such a trained model, a corresponding computer-implemented method for supporting surgery on a patient's eye comprising such an artificial intelligence-based model, a corresponding computer program product, a data processing device, and a laser therapy device.

[0002] In refractive surgery, nomograms are used to minimize differences, i.e., deviations, between a planned refractive correction (set, for example, on a laser therapy device as the target refractive correction) and the actually achieved refractive correction. In such a nomogram, achieved refractive corrections (for example, in diopters) are plotted against planned refractive corrections. The planned refractive corrections are corresponding to settings on a laser therapy device. If exactly such a planned refractive correction is applied, the actual achieved refractive correction may differ from the planned refractive correction, depending on the laser therapy device used and / or the operator. In such a case, the actual achieved refractive correction does not correspond to the correction requirement (the refractive error of the person being treated).Only when the discrepancies between the planned and the actually achieved refractive correction are taken into account can a refractive correction be achieved that meets the corrective needs. Ideally, after the correction has been successful, the person being treated will no longer require visual aids such as glasses or contact lenses.

[0003] For the treatment of refractive errors, well-known procedures such as PRK, Trans-PRK, LASEK, or LASIK can be used. Preoperative data representing the refractive error and the need for correction can be collected from the patient. This correction requirement is then "adapted" to the corresponding surgical system, i.e., converted into control data representing the location, pulse duration, and / or distance of laser pulses to be applied to the patient's eye. Such control data is readable and executable by laser therapy devices specifically designed for ophthalmic surgery.

[0004] However, the precision of the predictability of existing nomograms is limited, which is partly due to the commonly used methodology of linear regression.

[0005] Therefore, US 8,414,565 or WO 03 075 778 A1 proposes the use of improved algorithms and / or the consideration of additional preoperative or postoperative parameters.

[0006] However, such methods are limited in terms of user-friendliness and the traceability of decisions, as they may deviate from existing clinically integrated solutions. Unlike classical nomograms based on linear regression, such approaches often fail to demonstrate a relationship between a laser setting in a laser therapy device and a predicted refractive correction. This may be because the algorithms used focus solely on minimizing the prediction error, thus neglecting the evolution of a learned function. Particularly when used in or as part of a medical device, the traceability of suggestions from an assistance system to a physician may conflict with a requirement set by the manufacturer or supplier of the medical device and potentially...This is a necessary condition for approval as a medical device.

[0007] It can happen that the (evaluation) function(s) trained by an artificial intelligence-based algorithm delivers results that contradict known medical or physical laws and are therefore implausible.

[0008] As a purely illustrative example, it can be assumed that increasing the refractive correction (the refractive target) set on a laser of the laser therapy device, while keeping all other parameters constant, will also increase the actually achieved refractive correction. Such a plausible relationship may be overlooked by a classical artificial intelligence-based model if the model was trained solely with the aim of minimizing the prediction error.

[0009] The present invention therefore describes methods and devices with which an artificial intelligence-based model, for example a neural network, can be trained and applied, ensuring that invariants (such as, purely by way of example, the above-described plausibility of the relationship between set or planned refractive correction (=refractive target) and achieved refractive correction) are maintained without, however, restricting the model in its flexibility.

[0010] The present invention therefore aims to improve and / or enrich existing solutions from the prior art.

[0011] The aforementioned method solves this problem by comprising the following procedural steps: - Importing training data that includes at least one refractive target data and / or one patient data and / or at least one setting of a laser therapy device data; - Import of ground truth data that includes data representing at least one achieved refractive result.

[0012] The following procedural steps are repeated below: 1) Calculating result data that includes data representing at least one calculated refractive result, using the artificial intelligence-based model based on the training data / or at least on a part of the training data; 2) Determining an overall penalty value of a cost function based on the outcome data and the ground truth data for individual parameter sets of the training data and the outcome data; and 3) Minimizing the calculated penalty value of the cost function by varying at least one weighting of a weighting matrix of the artificial intelligence-based model.

[0013] In this first step of the process, the imported training data is used as a dataset with a large number of data points for the calculation. Preferably, each data point can comprise pairs of values ​​(also called parameter sets, value tuples, data tuples, or tuples) representing a refractive target and / or a patient and / or at least one setting of a laser therapy device and / or an (achieved) refractive result. For example, in the calculation in step 1, only those data points from which a refractive result can be calculated can be considered. Data that represent an achieved refractive result can (still) be disregarded in this step.

[0014] A parameter set or tuple can thus be understood as a list of a multitude of objects describing a parameter or representing a measurement, where a parameter set can be uniquely assigned to a process, i.e., exactly one intervention on exactly one patient.

[0015] However, in an alternative configuration, the training data, which includes at least the data representing a refractive target and / or the data representing a patient and / or the data representing at least one setting of a laser therapy device, may also include the ground truth data, which in turn includes the data representing at least one achieved refractive result.

[0016] In other words, the training data can represent annotated data, where the data representing a refractive outcome can constitute the annotation. This annotated data can be used for supervised learning. The data can comprise a variety of datasets in which a treatment with a corresponding outcome can be assigned to a patient. If more than one treatment has already been administered to the same patient, a second and / or third and / or further treatment can be stored in a separate dataset (or parameter set). The term "patient" is used here to refer to any person of any gender being treated.

[0017] In step 2), the ground-truth data representing an actual refractive outcome can be compared with the outcome data, specifically with the data representing a calculated refractive outcome. An actual refractive outcome is understood to mean a refractive correction or change in a patient's eye resulting from a procedure, as demonstrated, for example, by measurements. This correction or change is verifiable and not merely a (hypothetical) assumption, mathematical calculation, or estimate. Subsequently, in step 3), at least one weight in the weighting matrix of the artificial intelligence-based model is adjusted. In other words, the evaluation of various influencing factors in the mathematical model is modified.

[0018] If, purely as an example and without limitations, an artificial neural network is used, it is formed by interconnected nodes. These nodes can be considered artificial neurons. Such a neuron receives signals from other neurons connected to it in a network-like fashion and processes them using an activation function.

[0019] Such an activation function defines a range of values ​​within which an artificial neuron is activated, i.e., it outputs a signal to the subsequent neuron(s). Activation functions are typically nonlinear, and common examples include the sigmoid function, the tanh function, and the ReLU function (rectified linear unit).

[0020] Furthermore, each connection between neurons has a weight that increases or reduces the influence that the corresponding (output) signal of one neuron has on the activation function of the subsequent neuron.

[0021] Initially, i.e., before the artificial intelligence-based model begins to be trained, all thresholds (of the activation functions) and weights can be random values.

[0022] During training, these thresholds and weights are adjusted to the training data.

[0023] This adjustment is performed by decreasing the calculated penalty values ​​of the cost function until a termination condition is reached. Such a termination condition could, for example, be a maximum number of iterations of the training loop (one iteration of the training loop comprises steps 1), 2), 3) and querying the termination condition(s) and adjusting the weights, and / or falling below a confidence level that represents a sufficiently trained model. "Sufficiently trained" means that the calculated refractive correction (i.e., the refractive correction resulting from the refractive target after applying the deviations of the nomogram function) is within a predefined error interval of the actually achieved refractive correction (the refractive result stored in the corresponding ground-truth dataset).This verification of the confidence value is preferably carried out using validation data, which may represent a subset of the training data that was not used to train the artificial intelligence-based model.

[0024] The artificial intelligence-based model is designed to calculate output data, which includes at least one setting of a laser therapy device and / or at least one refractive target and / or one patient, based on input data that includes at least one predicted refraction.

[0025] A cost function used for training includes: - A nonlinear output value representing a measure of the prediction error of a univariate linear regression function and - At least one nonlinear penalty term that represents a measure of the violation of at least one invariant parameter.

[0026] In this context, the ground truth data can preferably be assigned to or linked with the corresponding (parameter sets of the) learning data.

[0027] A univariate linear regression function (also known as simple linear regression) can serve as the basis of the model. This model can be additively enhanced by additional nonlinear layers. In an artificial intelligence-based model (e.g., a neural network), for example, a nonlinear component can be added to the linear output of the linear regression function.

[0028] This nonlinear component can be used to model the prediction error of linear regression.

[0029] This means that the artificial intelligence-based model can learn to predict the deviations from linear regression - i.e., the residuals.

[0030] The non-linear output value representing the measure of the prediction error can ensure that, during training, the deviation of a prediction from the target values ​​present in the training data is minimized.

[0031] The nonlinear penalty term cannot, in particular, take the value 0 (zero) if the invariants required for the model hold, i.e., if they are satisfied. However, if or to the extent that an invariant is violated, this nonlinear penalty term can take on larger positive values.

[0032] A function value of the cost function can thus provide a performance indicator of a configuration of the artificial intelligence-based model, which can be used for further configuration of the model and, in particular, for its optimization.

[0033] It is a necessary step before any refractive surgery to take preoperative measurements of the patient's eye. These can include established measurements such as OCT, pachymetry, ultrasound, intraocular pressure measurement, fundus examination, or similar tests. Subsequently, based on these measurements (if the measurement does not directly provide the result), a treatment plan is developed, and, for example, in the case of surgery to correct refractive error in one eye, the refractive correction required is calculated.

[0034] This correction requirement can be entered into a nomogram, or a refractive target can be determined from a nomogram based on the correction requirement. For example, to achieve an (actual, postoperative) refractive correction of 1 diopter (1 Dpt; this corresponds to the correction requirement) for an unspecified laser therapy device, it may be necessary to set a refractive target of 1.2 diopters (1.2 Dpt) for the laser therapy device. A nomogram can thus be specific to exactly one laser therapy device in use. Furthermore, for one and the same laser therapy device, its settings can be considered equivalent to the refractive target, i.e., the refractive correction to be achieved (i.e., to be attained), i.e., to be set on the laser therapy device.The data representing at least one setting of the laser therapy device can thus be considered equivalent to the data representing a refractive target. Equivalent does not mean that these data are identical, but rather that they are uniquely convertible into one another. For example, it may be possible that no specific settings for laser energy, pulse duration, pulse repetition rate, etc., are made on the laser therapy device, but instead a specification of the refractive correction to be achieved (i.e., the refractive target) is provided. This specification of the refractive target can then be translated into corresponding settings within the laser therapy device.

[0035] The refractive goal, i.e., the refractive correction to be achieved, can include spherical correction and astigmatic correction.

[0036] If one eye of the person being treated is treated with exactly the refractive target, a complete correction of the person's refractive error cannot be achieved due to the discrepancies between the refractive target (which is set in the laser therapy device) and the predicted refractive correction. Rather, the predicted refractive correction should correspond to the corrective need. For this purpose, the corresponding refractive target can be determined from the predicted refractive correction, or vice versa.

[0037] It is understandable that the goal of both the treating physician and the patient is that the actual refractive correction, i.e., the predicted refractive correction, corresponds to the correction requirement or comes so close to it that any remaining refractive error no longer requires corrective lenses.

[0038] When planning a refractive procedure, the physician generally determines suitable treatment parameters (e.g., type of procedure, location and / or orientation of the tissue to be treated, i.e., removed) and / or laser settings based on patient data (i.e., data representing the patient; e.g., sex, age, etc.), diagnostic data (e.g., manifest refraction, pachymetry, etc.), and data of the desired refraction to be corrected (i.e., data representing the need for correction). The patient data, diagnostic data, and procedure data may be available as numerical values ​​(e.g., measurements), text data, or videos.

[0039] The inventive method describes a method for training a model, which is designed based on input data representing at least one setting of a laser therapy device and / or at least one refractive target and / or one patient, to calculate output data representing at least one expected (also: predicted) refractive correction.

[0040] These calculated output data are determined, among other things, based on anatomical patient variables using machine learning (ML) with a model or an ensemble of nonlinear models. The method can, in particular, consider a variety of preoperative measurements, patient data, and correction needs data, and serves to calculate an achievable refractive correction based on set laser parameters. The model can provide a nomogram function for a given laser therapy device, which describes the relationship between a refractive target and a (realistically) expected predicted refractive correction. The results calculated by the model based on the nomogram function can preferably be displayed in a nomogram itself.

[0041] The procedure can then be the same as with a classic nomogram. For example, by systematically trying out the possible settings of the laser therapy device, precisely those settings can be selected for which the predicted refractive correction best matches the expected result. "Match" in this case means that the predicted refractive correction corresponds to the correction to be corrected, i.e., the correction requirement or the refractive error of the person being treated. Such trial and error is possible because only discrete values ​​with a resolution predetermined by the laser therapy device are supported as possible settings.

[0042] This trial-and-error process is similar to using a classic graphical nomogram. First, the desired correction (corresponding to the correction requirement) is selected on the Y-axis, and then the settings of the laser therapy device are selected on the X-axis for which the curve achieves the desired Y-value. This can be done graphically by finding the intersection point of a line parallel to the X-axis at the level of the desired correction with the nomogram function.

[0043] The artificial intelligence-based model can, in particular, be an artificial neural network comprising a large number of artificial neurons. Such an artificial neuron can receive input values, which are multiplied by so-called weights and thus weighted. The weighted input values ​​can be summed and passed to an activation function, which, depending on its form and / or parameters, provides a (non-linearly calculated) output value. This output value can, in turn, be used as an input value for at least one other neuron. Each neuron in a layer can be connected to all (or a subset of) the neurons of a previous or subsequent layer. "Previous" or "subsequent" refers to the direction of information transfer between the layers.

[0044] When training the artificial intelligence-based model, these weights are varied and connections between neurons of the model are subjected to different strengths, i.e., different relevance.

[0045] The trained artificial intelligence-based model or the weighting matrix of such a trained model described in this disclosure is obtained by an embodiment of the computer-implemented method for training the artificial intelligence-based model described above or below.

[0046] Furthermore, the computer-implemented method described in this disclosure for supporting surgery on a patient's eye comprises an embodiment of an artificial intelligence-based model described above or below. The computer-implemented method, which applies the trained model, comprises the following procedural steps: - Reading in input data that includes at least one setting of a laser therapy device and / or at least one refractive target and / or one patient; - Calculating output data from input data, which includes data representing at least one predicted refraction, using the artificial intelligence-based model; and - Providing the output data.

[0047] A corresponding computer program product described herein comprises instructions which, when executed on a data processing device, cause the data processing device to execute the process steps of an embodiment of a computer-implemented method described in this disclosure (for training and / or supporting an intervention).

[0048] Accordingly, the device presented here is designed for data processing, to implement an embodiment of a computer-implemented method for training and / or a computer-implemented method for supporting an intervention as described herein.

[0049] This disclosure also relates to a computer-readable medium (also: computer-readable non-volatile storage medium) comprising an embodiment of the artificial intelligence-based model described herein, wherein the model is designed to, when executed by a data processing device, induce the device, based on input data representing at least one setting of a laser therapy device and / or at least one refractive target and / or data representing a patient, to calculate output data representing at least one predicted refraction and to provide this output data. The model is stored on the non-volatile storage medium in the form of computer-readable instructions written in a computer-readable programming language.

[0050] The computer program product and / or the trained artificial intelligence-based model and / or the weighting matrix of such a trained model can be provided as a transmittable and computer-readable data signal. Such a data signal can be transmitted centrally or decentrally, for example, from a server or client to another client, a data processing device, or a laser therapy device, or it can be obtained from such a source, i.e., transmitted and received. The data signal may contain additional functional data that does not restrict or definitively enable the interpretation, readability, and / or decryption of the other components of the data signal.

[0051] Finally, a laser therapy device is provided, which includes a data processing device (described above or below) or which is configured to execute an embodiment (described above or below) of a presented computer-implemented method. The laser therapy device can thus enable the determination of a predicted refractive correction from given laser parameters in order to perform a refractive correction that corresponds to the correction requirement.

[0052] The artificial intelligence-based model can be provided in the form of, or as part of, a computer program product. In particular, the artificial intelligence-based model can be designed to be used by a data processing device as part of, or in the form of, a computer program product. The model can, in particular, be a trained model that enables the automated generation of output data based on input data.

[0053] The present computer-implemented method can assist a user in recognizing patterns in constellations of patient data, data on their refractive error, and data on the refractive correction to be achieved (the correction requirement). The output data of the model are always plausible and comprehensible due to adherence to the invariant parameters.

[0054] This can enable operations on a patient's eye with a more precise result, that is, with the smallest possible, ideally no, residual refractive error, which is in the interest of the doctor, but even more so in the interest of the patient.

[0055] Provided a suitable data basis is available for training the presented artificial intelligence-based model, a trained model of the presented invention, based on an input combination of parameters of the laser therapy device, patient data, and data on the patient's refractive error, can determine an achievable refractive correction and optionally display this data in a nomogram. The user is thus provided with information on which refractive correction can be achieved with the input data (such as at least one setting of the laser therapy device).

[0056] Preferably, the data is presented in the form of a nomogram, allowing the calculated information to be compared with a curve of the nomogram. However, this information is not obtained solely through simple linear regression. The application of an artificial intelligence-based model allows for more accurate predictions than simple linear regression, as such a model can calculate deviations from linear regression more precisely and incorporate them into the result. Furthermore, the invention always ensures that no invariants are violated and that the model is not based on implausible behavior.

[0057] A computer program product can be understood as a sequence of instructions or commands provided in a specific programming language. The computer program product enables the execution of certain functions, calculations, or predictions regarding the refractive correction of a patient's eye using a data processing device.

[0058] An artificial intelligence-based model can, in particular, be a machine-learning-generated, i.e., trained, model. In machine learning, an artificial system or model can learn from a multitude of examples and, using algorithms, build a statistical model that evaluates the input data, for example, in a weighting matrix, to determine output data. After training, such a model is preferably tested or validated with test or validation data. An artificial intelligence-based model is not based on memorized examples, but rather recognizes patterns and regularities in the training data and can apply these recognized patterns or regularities to unknown data.

[0059] The previously described computer program, through its artificial intelligence-based model, enables an individualized prediction or calculation of an achievable (i.e., predicted) refractive correction, at least based on input laser parameters. The prediction is individualized because the artificial intelligence-based model performs a calculation based on patient-specific information. This patient-specific information is among the input data for the artificial intelligence-based model.

[0060] The artificial intelligence-based model is not bound to a fixed set of rules, but can draw on knowledge acquired during the learning phase and relationships discovered when determining the achievable refractive correction.

[0061] Based on this acquired knowledge, the artificial intelligence-based model can also react to previously unavailable input data (i.e., input data that was neither entered nor read during training, testing, or validation).

[0062] The artificial intelligence-based model thus allows, taking into account all input data, the modeling of a nomogram that is not merely based on a "simple" linear regression, which would lead to insufficient prediction quality, but rather a nomogram that, in addition to more precise modeling using an artificial intelligence-based model, also takes invariant parameters into account to exclude implausible modeling.

[0063] The following section will explain in detail optional further developments of the aforementioned methods and devices. Features described in the explanation of the method(s) can be applied to the corresponding device. Likewise, explanations of device features can also be mapped to corresponding method steps.

[0064] For example, if the method is described as performing a process step A, it is implied that the corresponding device is designed to be able to perform precisely this process step A. Conversely, a corresponding embodiment of the method includes process step B if a correspondingly designed device has a feature C for performing process step B.

[0065] The training process can optionally include a documentation step. After each iteration, i.e., after completing steps 1 to 3, the variation(s) of the changed weight(s) and the results of the comparison during supervised learning can be stored in a database or memory. This ensures that all learning iterations are traceable. This documentation can be stored, for example, on a standalone server, thus providing auditable documentation of the training of the artificial intelligence-based model. This may be a requirement for the approval of a medical device that incorporates a computer-implemented process.Sufficient documentation of the model or its weighting matrix in the delivery state of a medical device may also be relevant and necessary for its approval, even for the trained artificial intelligence-based model or its weighting matrix.

[0066] The data representing a patient may include at least: information on the biomechanical stability of the eye; information on any pre-existing eye disease; information on any prior eye treatment; information on the shape of the eye to be treated; information on the axial length of the eye; information on the preoperative depth of the anterior chamber; information on the surface curvature and / or corneal contour of the cornea; information on manifest refraction and / or pachymetry; information on the patient's origin and / or ethnicity; and information on the patient's age and / or sex.According to the invention, it can be provided that this list can be supplemented with further information during the use of the artificial intelligence-based model for determining a refractive correction and the necessary laser parameters.

[0067] The data representing a refractive target can include at least the following: information about a target refraction (also: planned refractive correction) that is set or selected before the operation (on the laser therapy device). As described above, the target refraction can be equivalent to laser parameters that need to be set.

[0068] The refractive target can therefore correspond to the setting configured on the laser therapy device. The refractive target can be a spherical refractive target or an astigmatic refractive target.

[0069] This information, particularly the input values, can have different weightings. For example, it is understandable that the need for correction is a far more relevant parameter for evaluating the procedure and calculating the laser parameters than, say, the ethnicity of the person being treated. The various parameters can be grouped into different parameter classes. Nevertheless, it cannot be ruled out that parameters that appear less relevant at first glance may dictate certain regularities in the creation of a nomogram and, in particular, in the function of the nomogram used for calculations. Even if these regularities are not apparent to a human user, they can be discovered and taken into account during the machine learning process of the artificial intelligence-based model.

[0070] The disclosed computer-implemented method, using an artificial intelligence-based model, can support a physician in performing, and especially in planning, refractive surgery by determining an achievable (predicted) refractive correction based on entered laser parameters, which, when applied in an (ophthalmological) laser therapy device during a corresponding procedure, almost completely, ideally completely, corrects the patient's existing refractive error.

[0071] Calculating the necessary laser parameters solely using a nomogram obtained through linear regression can provide an insufficiently accurate prediction. This calculation can lead to a more precise result when an artificial intelligence-based model is used. According to the invention, compliance with the invariants is also ensured.

[0072] The disclosed computer-implemented method can further indicate the quality of the calculated refractive correction and the necessary laser parameters, thus warning a physician if the input data leads to output data with insufficient certainty. The "insufficient" condition can be linked to a predefined threshold for an error or deviation, with exceeding this threshold representing insufficient certainty or excessive uncertainty. An uncertainty factor can be specified for this purpose, representing a deviation of the output data from data input during training, testing, and / or validation.

[0073] Such quality can be provided in addition to or as an alternative to the display of the result data in the nomogram.

[0074] The information or data to be read into, or already read into, the artificial intelligence-based model can be in any format, such as image data, text data, or video data. The information to be read can be requested from a patient record and read into the model. Likewise, the information to be read can be requested from and received from a database or a server.

[0075] If the information to be read—the training data, input data, and / or ground truth data—is presented in a table, for example, then the order of the individual entries, as well as their labels and format, should ideally be consistent for both the training data and the input data. For instance, it is important that age or another parameter is always in the same (predefined) data format and, if necessary, converted to a predefined data format before use. This allows for the use of age or other parameter values ​​with a period (.) as the decimal separator, as well as those with a comma (,). This enables the use of expert knowledge for training the artificial intelligence-based model, regardless of location.

[0076] The information to be read in can be in the form of a table and / or database entries, or in another previously known form. According to the invention, it can be provided that the information to be read in is checked for compliance with a minimum requirement. That is, purely by way of example and without limitation, a minimum number of parameters must be present in order for such a data tuple (i.e., a parameter set) to be read in as a usable entry. Furthermore, it can be checked whether the individual entries of a data tuple contain plausible values ​​(values ​​that lie within an expected interval). For example, entries that show an actual refractive correction of more than or equal to 25 diopters or less than or equal to -25 diopters can generally be marked as erroneous and not used for training the model.

[0077] The expected interval of the entries can be modified by a user to accommodate, for example, future generations of ophthalmic laser therapy devices capable of corrections on the order of, purely exemplified, 0-30 diopters or, purely exemplified, -30 to 0 diopters. In such a case, the expected intervals can be adjusted to the changed limits of the laser therapy device. In other words, the procedure can include the step of removing outliers from the input data and / or training data.

[0078] The data processing device can be a local computing unit. This unit can be controlled wirelessly and / or via a cable, and can receive or output data. The data processing device can also be remotely operated or operate partly locally and partly remotely. Likewise, the data processing device can be a graphics processing unit (GPU) capable of performing the calculations for training and / or executing the artificial intelligence-based model. Such a GPU can be operated in conjunction with and combined with a central processing unit (CPU) and / or memory units and / or input / output interfaces.

[0079] The artificial intelligence-based model can be stored entirely or partially locally, or entirely or partially in the cloud. For example, input data can be transferred from a client-acting data processing device to one or more servers on which the artificial intelligence-based model is hosted. Furthermore, the audit-proof documentation can be stored on a separate data processing device and / or a separate server.

[0080] The artificial intelligence-based model can thus be deployed via cloud computing, also known as Software-as-a-Service (SaaS), for the purpose of inference. In this process, the predicted refractive correction is determined by the data processing device on the server (or servers), based on factors including laser parameters, and output to the client, preferably via the internet. Such a data processing device therefore comprises both the client and the server, as well as a corresponding data connection between them.

[0081] If the artificial intelligence-based model is provided as so-called Software as a Medical Service (SaaMS or SaMS), audit-proof documentation of all learning steps of the model may be necessary.

[0082] By providing the artificial intelligence-based model in such a centralized or decentralized manner, it is conceivable that it can be further trained based on feedback from a large number of users. In this process, the weights in the weighting matrix are continuously modified while adhering to the invariants, in order to enable the model to make more precise predictions.

[0083] The data processing device may have an input interface to receive the learning data and / or the ground truth data and / or the input data and / or test data and / or validation data.

[0084] In particular, it is conceivable that the user can complete previously incomplete data via the input interface. According to the invention, it can therefore be provided that it is checked whether the available data meets a minimum data requirement necessary to determine the predicted refractive correction.

[0085] In particular, the input interface may include an input interface designed to receive a data signal comprising the learning data and / or the ground truth data and / or the input data and / or the test data and / or the validation data.

[0086] Alternatively or additionally, this data can be entered by reading it from at least one database. Therefore, the data can be entered from multiple sources.

[0087] The data processing device can further include an output interface for displaying the predicted refractive correction based, among other things, on the input laser parameters, at which an output signal representing this information can be provided. If the input interface includes an optional display, the output interface can, for example, also use this display to show the predicted refractive correction, at least based on the input laser parameters. These calculated values ​​can also preferably be displayed in a nomogram.

[0088] In the case of the computer program product, this information can be displayed on a screen of the data processing device that executes the computer program product.

[0089] Other known types of display devices can be connected to the output interface to display the predicted refractive correction in a nomogram. Alternatively or additionally, the output interface can be connected to a database and / or transmit the output signal and the predicted refractive correction it represents to the database. The stored values ​​can then be retrieved from the database by the user.

[0090] It is conceivable that the data processing device is designed to provide a warning signal based on the accuracy of the calculated / predicted refractive correction. This warning signal can indicate a condition in which the predicted refractive correction exceeds a predefined deviation from previously entered and / or calculated data in the model. The signal can trigger a warning indicator, such as a warning symbol on a display, and / or activate a warning light and / or an audible warning. The warning signal can thus trigger the display of the warning symbol. This warning is issued during the planning phase, allowing the surgeon to perform a corresponding check and, if necessary, adjust parameters preoperatively.The calculation of the quality and / or the provision of the signal representing a warning can be switched on or off by the user.

[0091] At least one step of the described computer-implemented method can be performed, at least partially, by a data processing device.

[0092] The model is trained using a training dataset in the process steps 1) to 3) in a predetermined number of iterations and / or with at least one predetermined termination condition.

[0093] Subsequent validation using a validation dataset is possible and recommended. The training dataset preferably contains annotated data, i.e., also an actual refractive correction achieved (e.g., postoperatively measured). This actual refractive correction can be provided by the ground-truth data and linked to the corresponding training dataset.

[0094] Similarly, the training data can include data representing the achieved refractive correction associated with each input data, whereby only the input data, and not the data representing the achieved refractive correction, are used to validate the (trained) model. The refractive correction calculated (also: predicted) by the model is then compared with the actual achieved refractive correction stored in the validation dataset.

[0095] If the predicted refractive correction values ​​fall within a predefined training expectation range, the artificial intelligence-based model can be considered "trained." If the training expectation range is not met, further training of the model may be necessary. In this case, a termination condition has not yet been fulfilled.

[0096] The weighting matrix can be understood as an evaluation of individual influencing factors of the artificial intelligence-based model. Depending on the application of the model, the different parameters are assigned correspondingly greater or lesser influences on the result through different weightings within the weighting matrix.

[0097] When training the artificial intelligence-based model, the importance and corresponding weighting of individual parameters are determined using known methods (gradient descent method of the loss function or direct loss minimization), so that parameters with a strong influence on the result are weighted accordingly. According to the invention, the deviation from linear regression is modeled by the nonlinear output value, and compliance with invariant parameters is ensured by the nonlinear penalty term. The sum of these weights after training constitutes the weighting matrix, on the basis of which the evaluation of newly inputted (unknown) data into the model can be performed.

[0098] The trained AI-based model, or its weighting matrix, can be stored locally in a data processing device or in a distributed system, for example, on at least one server. When used to support a procedure on a patient's eye, the trained AI-based model, or its weighting matrix, can thus be accessed locally or remotely. Alternatively, the trained AI-based model, or its weighting matrix, can be requested by the data processing device and transmitted to it as a data signal.

[0099] The trained artificial intelligence-based model or its weighting matrix can therefore be stored in volatile or non-volatile memory, or provided in the form of a data signal.

[0100] Typically, the weights in the weighting matrix are initialized with values ​​from a normal distribution. After training, the weights can take on any (preferably normalized) values.

[0101] The artificial intelligence-based model can have an input and an output layer. Furthermore, the model can have one or more hidden layers. These hidden layers can be fully meshed linear layers, nonlinear layers (e.g., ReLU), or transformer layers. However, the model's layers are not limited to the types listed here.

[0102] The size of the input layer depends on the number of input parameters and their encoding. Input parameter encoding can include, for example, one-hot encoding for categorical variables or normalization. Categorical variables or data refer to variables consisting of label values; for example, a variable "Country of Origin" could have the values ​​"Germany," "Finland," "Australia," etc.

[0103] Some machine learning algorithms can work directly with categorical data, but most require that all input or output variables be a number or numeric value. This means that it is advantageous to map all categorical data to integers.

[0104] One-hot encoding is a possible method for converting data to prepare it for an algorithm and obtain a better prediction.

[0105] Optional normalization of input data has the advantage that the weight values ​​can be within a limited range.

[0106] The output layer can contain a neuron that outputs the predicted refractive correction.

[0107] Furthermore, it is possible to include more than this single neuron in the output layer to output more information. For this purpose, the output layer can be extended with additional neurons.

[0108] The computer-readable (storage) medium according to the invention stores, in computer-readable form, instructions for carrying out the computer-implemented method for training the artificial intelligence-based model or the computer-implemented method for supporting surgery on a patient's eye; that is, the computer-readable medium comprises an embodiment of a previously described computer program product. The computer-readable medium can be any analog or digital data storage device, such as, but not limited to, a storage tape, a floppy disk, an EPROM, a USB flash drive, a hard drive, a CD-ROM, a DVD, an SD card, a holographic storage medium, or an SSD storage medium.

[0109] The computer program product according to the invention can be stored on such a computer-readable (storage) medium in order to be made available to a data processing device. Alternatively, the computer program product can also be provided as a data signal via the Internet or otherwise.

[0110] In summary, this provides an AI-based evaluation of the information input into the artificial intelligence-based model. This evaluation can learn the relationship between input data and the resulting output data through training.

[0111] A nonlinear component of the cost function used in the computer-implemented model training process can be computed, in particular, using a multilayer feedforward network with the training data and the ground truth data. Such a network can also be called a multilayer perceptron (MLP), which is characterized by the fact that information in such an MLP is only passed forward from neuron to neuron. Such an MLP can be trained, for example, using the backpropagation method (also known as backpropagation of error). The nonlinear component of the cost function can include the nonlinear output value and the nonlinear penalty term. Further nonlinear components, which are added to the linear regression used as the basis, are also conceivable.

[0112] In particular, in the procedural step “3),” which is the third procedural step of each iteration of the learning process, the calculated penalty value can be minimized by varying at least one weighting in a direction which reduces both the prediction error and the non-linear penalty term when determining the total penalty value.

[0113] The weighting matrix and its weights are thus optimized for the nonlinear part of the cost function. This ensures that even when the prediction error is reduced through optimization, the invariants are always taken into account and the artificial intelligence-based model is not incorrectly optimized in a direction that violates the invariants.

[0114] It is particularly advantageous if the cost function and the artificial intelligence-based model are differentiable. This allows gradient-based meta-training of the AI-based model. In such training, task-dependent subgroups of weights are formed from the initial weights of the weighting matrix and tracked in sub-learning routines, where the dependence on a task can result in different gradients. Such learning methods are particularly applicable to regression with few data points (i.e., when an arbitrarily large dataset is not available for training).

[0115] In the computer-implemented procedure, a taxation value can be defined for the nonlinear output value and / or the nonlinear penalty term. Such a taxation value can represent the influence of the nonlinear output value and / or the nonlinear penalty term on the linear output value of the univariate linear regression function. If the influence of several parameters is variable, the taxation value can be provided in the form of a taxation vector.

[0116] Thus, it is possible to adjust the influence of the nonlinear output value and / or the nonlinear penalty term relative to the linear term and to change the significance of an output error and / or one or more invariants relative to the linear regression.

[0117] The nonlinear penalty term used in the computer-implemented procedure can represent a measure of the violation of at least one invariant parameter. Possible invariant parameters from the list of parameters include: a monotonicity parameter representing the monotonicity of a regression curve; a smoothness parameter representing the smoothness of a regression curve; a variance parameter representing the variance of a regression curve; and a limit parameter representing a limit dependent on at least one input parameter.

[0118] The invariant parameters listed above can be used in any combination and with different assessment values ​​assigned to the individual parameters in the nonlinear penalty term.

[0119] The monotonicity parameter ensures that a calculated refractive correction always increases with increasing laser power and does not decrease. This prevents the artificial intelligence-based model from calculating implausible states.

[0120] For this monotonicity parameter, the nonlinear penalty term can refer in particular to a partial derivative of the model function with respect to the settings of a laser in the laser therapy device. A partial derivative, especially through its sign, can indicate whether the cost function increases or decreases depending on the laser parameters. If such a derivative yields a negative value, the nonlinear penalty term of the cost function can assume a positive value, which increases with the magnitude of the negative slope. The penalty term, as well as the total penalty value, can thus be understood as criteria that can be used for optimization by the artificial intelligence-based model.

[0121] In other words, a steeper decline in the nomogram function results in a higher total penalty value calculated by the cost function than a smaller negative increase. This portion of the total penalty value due to the nonlinear penalty term only vanishes (i.e., becomes 0) when the slope is 0 or has a positive value.

[0122] Evaluation using partial derivation can be particularly easy to implement, as it may be available in modern deep learning environments (also: deep learning frameworks) in the form of automatic differentiation.

[0123] The smoothness parameter can ensure that the cost function and / or the nonlinear penalty term are always differentiable and have no discontinuities. Here, too, it is conceivable to use a partial derivative of the nomogram function to determine the smoothness parameter. If the nomogram function is differentiable over its entire range of values, it can be assumed that the function can be considered "smooth" and has no discontinuities, since a derivative does not exist at discontinuities.

[0124] The variance parameter ensures that the dispersion around a mean or expected value remains within a predefined range, thus setting limits for the measure of dispersion.

[0125] The limit parameter can represent dependencies on at least one input parameter. For example, a maximum possible correction, which in turn may be defined or limited by a minimum remaining corneal thickness after a procedure, can be incorporated into the limit parameter. If the limits for the remaining corneal thickness are met in such a case, the contribution of the limit parameter to the nonlinear penalty term can be zero. If the limit is exceeded, this can be represented by a contribution of the limit parameter to the nonlinear penalty term greater than zero, and in particular, increasing with the extent of the limit being exceeded. Thus, the penalty term increases the more the limits are exceeded. Each penalty term can depend on the degree of non-compliance with or exceedance of the associated invariant parameter.

[0126] The output value of the univariate linear regression function can represent at least one error measure from the following list: - Root Mean Square Error; - Mean Absolute Error, and - Negative Log Likelihood.

[0127] For example, it is advantageous to formulate the first part of the cost function as a negative log likelihood and the non-linear component as a variance in order to calculate not only the prediction but also a measure of the model's uncertainty. Therefore, it is not possible to directly predict the value of the achieved correction, but rather the mean of a normal distribution.

[0128] The different display modes can be selected by a user and switched between them during calculation by the artificial intelligence-based model.

[0129] To minimize the variance in the prediction (the refractive correction), an ensemble of several artificial intelligence-based models can be used. Optionally, the different models can be trained with different subsets of the training data.

[0130] At least two cases can be distinguished here: a) For an invariant, the fulfillment of the invariant condition by a single artificial intelligence-based model implies fulfillment by the entire ensemble: or b) the condition of a) is not met.

[0131] In the case of an ensemble that satisfies condition a), at least one of the many artificial intelligence-based models can be trained using a configuration of the computer-implemented training procedure as described above or below. If, after training, at least one model satisfies the invariant conditions, then this also applies to the ensemble.

[0132] If condition a) is not fulfilled by the ensemble, the ensemble can be considered a single, overall artificial intelligence-based model, and this overall model can be trained using a modification of the computer-implemented training procedure as described above or below. This also ensures that the overall model fulfills the invariant conditions.

[0133] Since it is desirable for the invariants to apply not only to the data points contained in the training data but to the entire relevant input range, it can be advantageous to supplement the training data with synthetic data. A corresponding design of the computer-implemented method can therefore further include the following procedural steps: - Generating at least one synthetic data tuple for a data tuple already present in the training data; - Evaluating the nonlinear penalty term of the cost function for all synthetic data tuples, and - Varying at least one weighting of the weighting matrix of the artificial intelligence-based model depending on the evaluation of the nonlinear penalty term, whereby the evaluation and variation are repeated until a value of the nonlinear penalty term reaches a predetermined limit ε invariant falls below.

[0134] It can be advantageous if at least one synthetic data tuple satisfies the following conditions: ◯ Elements of the data tuple are functionally related to the invariant parameter under consideration; ◯ The values ​​of the elements of the data tuple lie within a value range predefined for each element,

[0135] These synthetic data tuples therefore do not contain actual values ​​for the target variable (the actual refractive correction achieved and the necessary laser parameters) and thus cannot be used for supervised learning. Instead, these synthetic data tuples are used during model training to evaluate the invariant part of the cost function, i.e., the nonlinear penalty term. These synthetic data tuples have the advantage that, during model training, states in which the invariants are satisfied can be reached with fewer iterations. Furthermore, these states can be more stable.

[0136] Using the monotonicity parameter as an example, a predefined limit value ε can be determined purely by way of example and without limitation. invariantThe value of the partial derivative can be zero. In such an example, a value of the partial derivative can be negated (i.e., multiplied by minus 1) and the result compared to the limit. Other evaluations may also include normalization and / or division.

[0137] It is also conceivable that a large number of synthetic data tuples are generated, the elements of which are essentially - can be evenly distributed over the predefined range of values ​​or - can represent a previously known probability distribution of the value range of the element in question.

[0138] This ensures that sufficiently large value ranges are covered for the input variables referenced in the invariants, so that the invariants are always taken into account for these value ranges.

[0139] In a further embodiment of the computer-implemented training process, the training data and the ground-truth data can be divided into batches comprising a predetermined number of data tuples. For each batch, the process steps 1) to 3) are applied to train the artificial intelligence-based model until at least one termination condition is met. Following such a training loop, the additional process steps of the previously described embodiments of the process, which generate synthetic data tuples, are then performed. A batch is understood to be a subset of the training data, where, depending on the number of entries in the tuples, the subset can comprise only a few data tuples up to several tens, several hundred, several thousand, or even more. Preferably, the training data can be divided into batches of equal size.

[0140] Thus, within a training loop, at the end of each epoch, after the training data has been processed regularly in the form of the aforementioned batches, an additional loop is executed in which only the invariant part, i.e., the nonlinear penalty term of the cost function, is evaluated on all synthetic data tuples. This has the advantage that the model can be optimized until the value of the invariant cost function exceeds a predefined threshold ε. invariant below the threshold, and that from the very beginning of the training it can be prevented that the artificial intelligence-based model remains in impermissible states (in which at least one invariant is violated) for too long.

[0141] The previously described configurations of the computer-implemented procedure for training an artificial intelligence-based model generate or produce a corresponding trained artificial intelligence-based model or a weighting matrix of such a trained model.

[0142] Such a trained, artificial intelligence-based model according to the invention is designed, using its weighting matrix, to calculate output data based on input data, which includes at least one setting of a laser therapy device and / or at least one refractive target and / or patient, and which includes at least one predicted refraction.

[0143] The present invention also relates to a computer-implemented method for supporting surgery on a patient's eye. The method utilizes an embodiment of an artificial intelligence-based model described herein and comprises the following process steps: - Reading in input data that includes at least one setting of a laser therapy device and / or at least one refractive target and / or one patient; - Calculating output data, which includes data representing at least one predicted refraction. The method according to the invention uses the artificial intelligence-based model to calculate the output data and provides the output data.

[0144] The input data, especially the patient data, can include at least one piece of information from the list, comprehensively - information about the axial length of the eye; - information about the preoperative depth of the anterior chamber; - information about the surface curvature and / or corneal contour of the cornea of ​​the eye); - information about manifest refraction and / or pachymetry; - information about a target refraction; - information about the biomechanical stability of the eye; - information about a pre-existing eye condition; - information about pretreatment of the eye; - information about the shape of the eye to be treated; and - include information about the patient's age and / or gender and / or ethnicity.

[0145] Possible patient-representing data were also mentioned before the list above. According to the invention, it may be possible that some parameters, for example, and not limited to, a patient's ethnicity, do not primarily influence the output data (i.e., that patients of different ethnicities with the same refraction to be corrected may, after calculation by the model, have comparable values ​​for the refraction to be achieved and the associated laser parameters to be set in the laser therapy device, but different remaining deviations may occur).

[0146] Such influences can be taken into account by the artificial intelligence-based model according to the invention and applied to support an intervention on a patient's eye.

[0147] In general, it is preferred that the adherence to invariants is always checked when using the artificial intelligence-based model, and that the model is corrected accordingly if necessary.

[0148] To ensure compliance with invariants when using the artificial intelligence-based model, in particular the trained artificial intelligence-based model, even for data that differ from the training data, an embodiment of the computer-implemented procedure to support an operation can optionally perform the additional procedural steps of an embodiment of the computer-implemented procedure for training an artificial intelligence-based model that generates at least one synthetic data tuple.

[0149] Furthermore, the procedure can be designed to include the following procedural steps: • Checking the similarity of input data with the training data used to train the artificial intelligence-based model; and ◯ If a similarity factor representing similarity falls below a predefined similarity threshold: Perform the additional procedural steps of a design of the computer-implemented procedure for training an artificial intelligence-based model that generates at least one synthetic data tuple.

[0150] Such a similarity check may require the definition of a similarity measure. With multivariate data, it can be advantageous to always perform the invariant compliance check, rather than depending on the similarity of the input data. This is because it is not readily possible to define a similarity measure for multivariate data that guarantees invariant compliance for input data that is similar to the training data.

[0151] The similarity factor can be based on a range of values ​​for the parameters or elements of the input data and the training data. The similarity factor can be defined in binary terms with specific possible values ​​of 0 and 1, or continuously. For example, if the input data falls within a range of values ​​for at least one parameter used during training, a similarity factor of 1 can represent high similarity. Conversely, input data that lies outside the parameter ranges used during training can result in a similarity factor of 0, which can represent low similarity.

[0152] For example, if a parameter lies outside a specified range of values ​​during training, synthetic data tuples are generated. These tuples preferentially extend the previously considered range of the parameter(s) such that the input parameter is included within this extended range. Subsequently, the nonlinear penalty term is checked against the synthetic data tuples for compliance with the invariants and / or trained accordingly, ensuring that the invariants are also satisfied for the input dataset (which triggered this re-verification using the synthetic data tuples).

[0153] Such an adaptation to the invariants requires only a few optimization steps, so that the retraining time or the overall inference time for calculating the output data is not significantly increased. Nevertheless, it is ensured that the model can continue to be trained during use, particularly with regard to satisfying the invariants.

[0154] In some embodiments, the computer program product according to the invention also includes commands which, when executed on a data processing device, cause it to additionally generate synthetic data tuples when training or applying the artificial intelligence-based model and to use these for verification or further training of the model in order to ensure compliance with the invariants.

[0155] The following section will explain the aspects of the present invention in more detail with reference to the accompanying drawings. The drawings merely illustrate possible embodiments of the present invention, whereby the described features can be combined or omitted as desired, provided that the technical effect achieved by the omitted feature is not essential. Identical features or features with the same function are further indicated by the same reference numerals. Repetitive descriptions of features are avoided, so that explanations of features described in preceding figures can also be applied to other figures, unless differences are explicitly indicated.

[0156] They show: Fig. 1 a schematic representation of an invariant parameter; Fig. 2 a schematic flowchart of a computer-implemented procedure according to revelation for training an artificial intelligence-based model; Fig. 3. A schematic flowchart of part of the procedure. Fig. 2; Fig. 4 a schematic representation of a further embodiment of the procedure for training an artificial intelligence-based model; Fig. 5 a schematic representation of a further embodiment of the procedure for training an artificial intelligence-based model; Fig. 6 a schematic representation of a design of a computer-implemented procedure to support an operation on a patient's eye; Fig. 7 a schematic representation of a further development of the procedure Fig. 6; and Fig. 8 A schematic representation of a device for data processing.

[0157] In the Fig. Figure 1 shows two schematic diagrams illustrating an exemplary invariant parameter 101. In both diagrams, an achieved refractive correction AC (achieved correction) is plotted against a laser therapy device setting LS (laser setting). A nomogram curve 103 describes this relationship.

[0158] In a case where the invariant parameter 101 (also: L) invariant ) is not fulfilled, the nomogram curve 103 describes a range 109 within which a physically inexplicable behavior is considered.

[0159] In the example shown, the nomogram curve 103 shows a (sign) change in the slope (dAC / dLS) from point P1 (with the values ​​(LS1, AC1)) to point P2 (with the values ​​(LS2, AC2)), which may be due to the calculation, i.e. the underlying functional relationship of the nomogram curve 103, but does not describe a real relationship between the achieved refractive correction AC and the settings of the laser therapy device LS used for this purpose.

[0160] Such a nomogram without considering the invariants 105 is shown on the left in Fig. Figure 1 is shown. Here, a monotonicity parameter 101a is shown as an exemplary representative of an invariant parameter 101.

[0161] A nomogram taking into account the invariants 107 is shown on the right. Fig. Figure 1 shows the nomogram curve 103', which displays a positive (or at least not negative) increase over the entire course of the settings of the laser therapy device LS. The monotony parameter 101a is therefore fulfilled.

[0162] The shape of nomogram curve 103 in nomogram 105 is not plausible, as it contradicts known medical or physical laws. Rather, it is to be expected that (purely by way of example and not limiting) an increase in the refractive correction set in the laser therapy device (i.e., an increase in the settings of the laser therapy device LS), with all other input parameters remaining identical, would also result in an increase (but at least not a decrease) in the achieved refractive correction AC.

[0163] In this revelation, an improvement of nomogram curves 103 from a course 105 to a course 107, i.e. taking into account the invariant parameters 101, is presented.

[0164] In Fig. Figure 2 shows a schematic diagram of the computer-implemented method for training an artificial intelligence-based model 200. The method 200 comprises a first step 205 in which a training data set 901 is acquired. This acquisition 201 of the data set 901 can be performed by reading it from an external source and / or by reading it from an internal source (related to the device that executes the method). Preferably, the training data set 901 can be stored in and read from a local or remote memory location.

[0165] In the embodiment shown, process step 205 comprises the reading 201 of training data 903 and the reading 203 of ground truth data 905. In other embodiments, the training data 903 can include the ground truth data 905.

[0166] The training data 903 can be referred to as annotated data 907 due to the presence of the ground truth data 905. The ground truth data 905 can be understood as the desired outcome of a (yet to be performed) computation of the artificial intelligence-based model. These desired outcomes can be used, in particular, during the training of the artificial intelligence-based model, to determine and optimize its individual weights and thus the entire weighting matrix of the artificial intelligence-based model.

[0167] In procedure 200, after data acquisition 205, a learning loop 207 takes place. At the beginning of this learning loop 207, individual weightings 921 of the weighting matrix 923 are generally determined with random numbers or according to a previously defined distribution in step 217.

[0168] In step 209, individual value pairs 909 from the training dataset 901 are entered into the artificial intelligence-based model M. In the configuration shown, this input 209 includes the training data 903 and / or the ground truth data 905. In another configuration, the training data 903 can also include the ground truth data 905.

[0169] However, only the training data 903 are used to calculate the result data 911. More precisely, depending on the task, a subset of data from the imported data can be used to calculate the result data 911.

[0170] If the inference model is to reflect the approach of a classical nomogram, then at least one setting of the laser therapy device is used as an input parameter, and the refractive correction achievable with these settings (taking into account other input data, such as patient-specific data), i.e., the predicted refractive correction, is defined as the dependent variable, i.e., as the output value. This approach can effectively illustrate the causality between a setting of the laser therapy device and the achievable refractive correction.

[0171] As mentioned above, in a further embodiment using the same model, this approach can also be deviated from by defining the achievable refractive correction as the input parameter and the setting or settings of the laser therapy device as the output parameter.

[0172] The calculation of the result data 911 takes place in step 209. The calculation therefore includes the prior data input.

[0173] Subsequently, in step 211, a total penalty value 913 is determined. For this purpose, a cost function 915 is used, which uses a univariate linear regression as its basis, wherein the nonlinear part of the cost function 915 is composed of a nonlinear output value AW representing a measure of a prediction error, or 917 of a univariate linear regression function, and at least one nonlinear penalty term NLS representing a measure of the violation of at least one invariant parameter 101.

[0174] For this purpose, for example, the result data 911 calculated using the artificial intelligence-based model M can be compared with the ground truth data 905 and a calculated deviation can be used for further processing.

[0175] In step 213, it is checked whether at least one termination condition 927 is met. Possible termination conditions 927 can be falling below a threshold value for the deviation between the result data 911 and the ground truth data 905, or a maximum number of iterations to be performed in the learning loop 207.

[0176] If multiple termination conditions 927 exist, they can be combined arbitrarily using the operations *and*, *or*, or *exclusive or*. For the sake of simplicity and without restriction, it is assumed below that exactly one termination condition 927 is satisfied or not.

[0177] If termination condition 927 is met, the learning loop 207 is terminated. If termination condition 927 is not met, at least one weighting 921 of the weighting matrix 923 is varied in step 215.

[0178] The weighting matrix 923 initially set in step 217 is thus modified in step 215, and the learning loop 207 is executed again with this modified weighting matrix 923. In particular, the learning loop 207 can be executed until the termination condition 927 is met (or the termination conditions 927 are met).

[0179] In the Fig. 3 is a specific form of procedure step 215 of the Fig. 2 shown in detail.

[0180] Since a nonlinear portion of the cost function 915 consists of the nonlinear output value AW, 917, and the nonlinear penalty term NLS, 919, the weights 921 of the weighting matrix 923 can be optimized for both the output value 917 and the nonlinear penalty term 919. In the illustrated configuration of process step 215, this is done additionally or alternatively to process steps 209 and 211 before the learning loop 207 is repeated with a modified weighting matrix 923.

[0181] Step 315A of varying at least one weighting 921 is identical to step 215 of the in Fig. 2. In the embodiment of procedure 200 shown, only the non-linear output value 917 is calculated in procedure step 315B, and the non-linear penalty term 919 is calculated in the next step 315C. In steps 315D and 315E, it is checked whether the newly calculated non-linear output value 917 or the newly calculated non-linear penalty term 919 is smaller than the previously calculated non-linear output value 917 or the previously calculated non-linear penalty term 919, respectively.

[0182] In the Fig. In the embodiment shown, the process steps 315D and 315E are carried out according to the respective calculations 315B and 315C, however, it is also conceivable that the steps are carried out in the sequence 315B, 315D, 315C and 315E.

[0183] If queries 315D and 315E are both negative, this information is passed to an "AND gate" 316, which, even with a single "No" response, continues the procedure with process step 315A in a loop until both queries 315D and 315E return a "Yes".

[0184] If this is the case, a new weighting matrix 923 (or its weights 921) is defined in process step 315F and, if necessary, combined with step 209 (see Fig. 2) continued.

[0185] In the Fig. 4 is the one in Fig. 2. Procedures 200 are shown schematically in the form of a flowchart.

[0186] The training dataset 901 comprises individual data tuples 401, which can also be referred to as the value pair 909 of the training dataset 901. In the configuration presented here, a data tuple 401 comprises training data 903 and ground truth data 905. The training data 903 can, in particular and purely by way of example, include data 403 representing a refractive target, patient data 405 (i.e., data representing a patient), and data 407 representing the settings of a laser therapy device (also: LTG data).

[0187] A large number of data tuples (401) can be grouped into a batch (952), and multiple batches (952) can exist. A certain number of these batches (952) can be processed simultaneously in parallel data processing.

[0188] The ground truth data 905 can, purely as an example, include data 409 representing a refractively achieved result. This data 409 thus represents an actual refractive correction achieved, the value of which was determined in subjective and / or objective measurements.

[0189] As previously described, in a different (not shown) embodiment of the procedure, the training data 903 can also include the ground truth data 905 and thus the data representing a refractive result.

[0190] Such a data tuple 401 is then fed into the artificial intelligence-based model M.

[0191] The model M shown here is a neural network 411, which has an input layer 413, two hidden layers 415, and an output layer 417. The number of hidden layers 415 and / or the number of individual neurons 419 (for clarity, only one neuron 419 in each layer is labeled with a reference symbol) and / or whether a neuron 419 is connected to all preceding or subsequent neurons 419 may differ in other configurations.

[0192] Due to the number of parameters of the learning data 903 and the number of possible results, three neurons 419 in the input layer 413 and two neurons 419 in the output layer 417 are shown here.

[0193] Each connecting line between neurons of adjacent layers (419) is assigned a weight (921). For clarity, only three weights (921) are shown. The totality of all weights (921) of the artificial intelligence-based model M is represented by the weighting matrix (923).

[0194] An input I into the model M leads, according to the internal calculations based on the individual weightings 921, to an output O. Using this output O and the ground truth data 905, the nonlinear output value AW, 917 and the nonlinear penalty term NLS, 919 of the cost function 915 can be calculated.

[0195] The calculated nonlinear output value 917 and the calculated nonlinear penalty term 919 can then be used for optimization. See here. Fig. 2 and in particular learning loop 207 or Fig. 3.

[0196] In Fig. 5 is a further development of the procedure of Fig. 4 shown schematically.

[0197] Since the invariant parameters 101 must apply not only to the training data 903 of the training dataset 901, but to an entire relevant input range, that is, a range of possible values ​​of the input data, it is conceivable to supplement the training data 903 with synthetic data 501. This can be done by a synthetic data tuple generator 503, which generates synthetic data tuple 505.

[0198] These synthetic data tuples 505 are additional hypothetically generated data points that are created for each existing data point, in particular for each existing data tuple 401, by inserting values ​​from a permissible range of values ​​for all input data that are relevant for the invariant parameters 101 and by retaining the values ​​for all other variables.

[0199] Such a synthetic data tuple 505 can include synthetic refractive target data 507 and synthetic laser therapy device data 509. Since, in the example shown, the patient data 405 generally have no influence on the output data (which is at least a predicted refractive correction), they can be adopted unchanged. This is indicated by an asterisk (*).

[0200] Using the extended data encompassed or provided by the synthetic data tuple 505, the model M can be optimized with regard to compliance with the invariant parameter 101. For this purpose, a query 511 can be performed at output O to check whether the calculated result was generated by synthetic data tuple 505 or not. If this is not the case, the ground truth data 905 is used, and the nonlinear output value 917 and the nonlinear penalty term 917 are calculated, and the optimization described above is performed.

[0201] If query 511 reveals that the calculation was performed with a synthetic data tuple 505, only the nonlinear penalty term 919 is used for optimization and the model M is trained so that the invariant parameters 101 are also satisfied for the value ranges of the input parameters extended with the synthetic data tuples 505.

[0202] In the Fig. Figure 6 is a computer-implemented procedure for supporting surgery on a patient's eye, schematically represented. This support is provided by applying (inference to) the artificial intelligence-based model M.

[0203] The artificial intelligence-based model M can be used to solve various tasks. These tasks differ in that the model M can calculate different output data, while the type and number of different data points that the model M processes (in total) remain unchanged. Two possible approaches have been described beforehand: either a predicted refractive correction is calculated based on at least one setting of the laser therapy device, or at least one setting of the laser therapy device is calculated based on a predicted refractive correction (this can be chosen so that the predicted refractive correction corresponds to the correction requirement).

[0204] To apply model M, 601 input data 931 are read in a first step. When using model M, an intervention on a patient's eye is to be supported based on these input data 931. This is done by creating a nomogram or a nomogram function in which invariant parameters 101 are also taken into account. The input data 931 can, in particular, include patient-representing data, or in short: patient data 405. Furthermore, the input data 931 can include data 937 representing at least one property of the laser therapy device. The input data 931 can also include data 403 representing a refractive target. If the data 937 representing at least one property of the laser therapy device are defined as input data, the procedure according to the invention can be carried out similarly to reading a classical nomogram.

[0205] A refractive target is defined as the value for the refractive correction to be performed on a patient's eye, which is set on the laser therapy device. The refractive target can be a spherical refractive target or an astigmatic refractive target.

[0206] The refractive target can be translated into a predicted refractive correction, taking the nomogram into account. Ideally, this correction, after surgery performed with the appropriate settings, corresponds to the corrective requirement, thus eliminating the need for corrective lenses such as glasses or contact lenses after the operation. The refractive target, along with information about the patient, is used to determine and provide output data 933 with the aid of the artificial intelligence-based model M. This occurs in process step 603.

[0207] In this case, the output data 933 can, in particular, include data 935 (hereinafter referred to as "prog. Refr. data") representing the refractive correction achievable with the input parameters, i.e., the predicted refractive correction. Specifically, this output data 933 can be obtained based on data 937 (hereinafter referred to as "LTG data") representing at least one setting of the laser therapy device. This means that, given the patient data 405 and the data 937 representing at least one property of the laser therapy device, an achievable refractive correction can be determined by the model M. This achievable refractive correction is output in the form of the prog. Refr. data 935.

[0208] In the further procedure, the input data 931 can be varied until the data 935 representing a predicted refractive correction correspond to a correction that is identical to the correction requirement.

[0209] In procedure 600, the input data 931 is applied as input I to the neurons 419 of the input layer 413. The artificial intelligence-based model M calculates the output data 933 based on the previously trained weighting matrix 923, which is provided at an output O.

[0210] The provided output data 933, here representing the achievable refractive correction, data 935 can be plotted in a nomogram representation 605 for further use as a purely exemplary example. This is in the Fig. 6 is sketched purely schematically.

[0211] Shown is a nomogram curve 103, which shows an actually achieved refractive correction 609 over a refractive correction to be chosen 607.

[0212] This means that the nomogram represents the relationship between a selected refractive correction 607 and a resulting actual (expected) refractive correction 609. The selected refractive correction 607 can be equivalent to the refractive target and / or equivalent to at least one property of the laser therapy device, such that the data representing the selected refractive correction can be transformed into those data 937 that represent at least one property of the laser therapy device.

[0213] The nomogram curve 103 is only schematically exaggerated for illustrative purposes and is not drawn as a straight line. The calculated output data 933 can be plotted as data points 611 in such a nomogram representation 605.

[0214] From the nomogram representation 605, an achievable refractive correction 613 can be read, which is represented by the data 935 representing the achievable refractive correction.

[0215] A refractive correction 607a to be selected for this purpose can also be read out, whereby the necessary settings of the laser therapy device 615 (represented by the LTG data 935) can be derived from this refractive correction 607a. The equivalence or convertibility of the refractive correction 607a to be selected (i.e., the refractive target) and the settings of the laser therapy device 615 are described in advance.

[0216] According to the invention, model M can determine different predicted refractive corrections 613 for different settings of the laser therapy device 615. This can be carried out until the predicted refractive correction 613 corresponds to the refractive correction to be performed, i.e., the correction requirement.

[0217] This inference 600 method can be used in a Fig. The design of the procedure shown in section 7 is improved to support an operation 700.

[0218] Essentially, in procedure step 701, it is checked whether a similarity factor SF is less than or equal to a predefined delta Δ.

[0219] A calculation of the similarity factor SF can take into account the mathematical distances between the individual parameters of the provided input data 931 and a range of values ​​for the corresponding parameters already known to the artificial intelligence-based model M. If the values ​​of the input data 931 are within a range known to the model M for this parameter (for example, through training data 903, which already includes values ​​of the newly entered input data 931), then the similarity factor SF is smaller than the predefined Δ, and the input data 931 is entered into the artificial intelligence-based model M as shown in procedure 600.

[0220] Provided that the individual parameters of the input data 931 lie outside a range known to the model M, at least one synthetic data tuple 505 (preferably several or a multitude of synthetic data tuples 505) can be generated in a step 703.

[0221] This ensures that the invariant parameters 101 apply to the entire relevant input range (i.e., also to the newly entered parameters of the input data 931) and not only to the values ​​contained in the training data 903.

[0222] As demonstrated in procedure 500 during training, in procedure 700 shown, values ​​from the permissible range for each parameter can be used for those parameters (for example, the refractive correction to be achieved 613 and the associated data representing a refractive target 935) that are relevant for the invariant parameter 101. Other variables, such as those represented by the patient data, can remain constant. Subsequently, the nonlinear penalty term NLS, 919, can be calculated in step 705 using this synthetic data tuple 505.

[0223] Provided that the nonlinear penalty term NLS in query step 707 is smaller than a predefined limit ε inv (also: ε invariant) means that a synthetic data tuple 505 generated based on the newly entered input data 931 already satisfies the invariant parameters 101 and the input data 931 is passed on to the artificial intelligence-based model M to calculate the output data 933.

[0224] If the invariant parameters 101 are not satisfied by the newly entered input data 931 of the generated synthetic data tuple 505, then in process step 709 at least one weighting 921 of the weighting matrix 923 of the model M is varied.

[0225] Thus, a new adjustment of the weighting matrix 923 takes place with regard to the fulfillment of the invariant parameter 101 (by evaluating the nonlinear penalty term NLS).

[0226] The calculation in step 705, the query in step 707 and the variation in step 709 are iterated 711 until the condition(s) of query 707 is / are satisfied.

[0227] This new optimization of model M using the additional data from the synthetically generated data tuple 505 (in step 703) thus makes it possible to further train and optimize the weighting matrix 923 of the artificial intelligence-based model M, particularly with regard to compliance with the invariant parameters 101, even during operation. Therefore, due to this adaptation, model M is not in impermissible states during inference.

[0228] In the Fig. Figure 8 is a schematic representation of a data processing device 800. This data processing device 800 can be configured as a self-contained computing unit 800A, such as a personal computer 800B, a client 800C, or a server 800D. The data processing device 800 can comprise the following units in any number and / or combination: - a reading unit 801, which is designed to read in and, if necessary, classify any form of data, such as the training data 903, the ground truth data 905, or the input data 931. The reading unit 801 is therefore designed at least to perform the process steps 205, 601, and 701; - a weighting matrix setting unit 803, which is configured to modify individual weights or weightings 921 of the weighting matrix 923 of the artificial intelligence-based model M. This weighting matrix setting unit 803 is thus specifically configured to perform the process steps 215, 315 F and 709; - a cost function calculation unit 805, which is configured to calculate and provide a total penalty value 913 and / or a non-linear output value AW, 917 and / or a value of the non-linear penalty term NLS, 919. The cost function calculation unit 805 is thus specifically configured to execute the procedural steps 211, 315 B, 315 C, 915 and 705. The cost function calculation unit 805 can provide different output values ​​depending on the input data. The output values ​​to be provided can be determined by the type of input data and / or the type of invocation of the unit 805 and / or by additional input information, which is optionally passed to the cost function calculation unit 805 at startup and which represents the result to be calculated. - a loop unit 807, which is configured to control all steps to be repeated and the necessary parameters to be changed or retained. The loop unit 807 can thus be used in particular to execute the learning loop 207, the optimization loops of the Fig. 2, Fig. 4 and Fig. 5 and step 711; - a storage unit 809, which is designed to store any type of data and / or information in a designated memory 813 and to document information about the storage location in the memory 813; - a CPU 815, which acts as a central control unit, controlling all units 801 to 809 connected to the CPU 815 via a bus 823 and their data exchange via the bus 823 and / or their timing; optional - a GPU 817 (graphics processing unit), which can perform the individual epochs of a learning loop 207 in a highly parallelized manner, especially in the process of training or learning 200, 400.

[0229] Ultimately, the data processing device 801 can comprise an input interface 821 connected to bus 823 and an output interface 819 also connected to bus 823. The data processing device 800 can thus be controlled by a user and information can be provided to the user via the input interface 821 and the output interface 819. 101 L invariant Invariant parameters 101a Monotonicity parameter 103 Nomogram curve 103' Nomogram curve of 107 105 Nomogram without considering the invariants 107 Nomogram taking into account the invariants 109 area 200 computer-implemented methods for training an artificial intelligence-based model 201 Record learning data set 203 Reading Ground Truth Data 207 Learning loop 209 Entering value pairs into the model / Calculating result data 211 Determine the total penalty value 213 Check whether the termination condition is met 215 Vary at least one weighting of the weighting matrix 217 Initializing the weighting matrix 315a Vary at least one weighting 315B Calculation of non-linear output value 315C Calculation of nonlinear penalty term 315D Check if nonlinear output value is less than previously calculated nonlinear output value 315E Check if the nonlinear penalty term is smaller than the previously calculated nonlinear penalty term 316 AND gates 315F Setting the weights of the weighting matrix 401 Data Tuple 403 Data representing a refractive target 405 patient-representing data 407 Data representing the settings of a laser therapy device (LTG data) 409 data representing a refractive result 411 neural network 413 Input layers 415 hidden layers 417 output layers 419 (artificial) neuron 501 synthetic data 503 Synthetic Data Tuple Generator 505 synthetic data tuple 507 synthetic refractive target data 509 synthetic laser therapy device data Query 511: Result generated by synthetic data tuple 600 computer-implemented procedures to support surgery on a patient's eye 601 Read input data 605 Nomogram Representation 609 actual achieved refractive correction 607 refractive correction to be selected 611 data points 613 refractive correction to be achieved 607a for 613 refractive correction to be selected 700 procedures to support surgery Query 701: Similarity factor less than Δ 703 Generating at least one synthetic data tuple 705 Non-linear calculation of penalty term with synthetic data tuple 707 Comparison of NLS with limit value 709 Vary at least one weighting of the weighting matrix 711 Iterating steps 705, 707 and 709 800 Data processing device 800A self-contained computing unit 800B Personal Computer 800C Client 800D Server 801 Reading unit 803 Weighting matrix setting unit 805 Cost function calculation unit 807 Loop unit 809 storage unit 813 memory 815 CPU 817 GPU 819 Output interface 821 Input interface 823 Bus 901 Learning data set 903 learning data 905 Ground Truth Data 907 annotated data 909 pairs of values 911 results data 913 Total penalty value 915 Cost function 917, AW issue value 919, NLS nonlinear penalty term 921 weighting 923 Weighting matrix 927 Termination condition 931 input data 933 output data 935 data representing a refractive target (in short: prog. Refr. data) 937 settings of the laser therapy device representing data (abbreviated: LTG data) 952 Batch AC achieved refractive correction LS setting of a laser therapy device M artificial intelligence-based model I Input O Output SF Similarity Factor ε inv (also: ε invariant ) X Input vector QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 8,414,565

[0005] WO 03 075 778 A1

[0005]

Claims

[1] Computer-implemented method (200) for training an artificial intelligence-based model (M) to compute and / or provide refractive output data (933) based on input data (931), comprising - Reading in training data (903) that includes at least one refractive target data (403) and / or one patient data (405) and / or at least one laser therapy device setting data (407); - Reading in ground truth data (905) that includes data (409) representing at least one achieved refractive result; - Repeat the procedural steps: 1) Calculating result data (911) comprising at least one calculated refractive result data (935) using the artificial intelligence-based model (M) based on the training data (903); 2) Determining an overall penalty value (913) of a cost function (915) based on the outcome data (911) and the ground truth data (905) for individual parameter sets (909) of the training data (903) and the outcome data (911); and 3) Minimizing the calculated penalty value (913) of the cost function (915) by varying at least one weight (921) of a weighting matrix (923) of the artificial intelligence-based model (M) until a termination condition (927) is reached, wherein the artificial intelligence-based model (M) is designed based on input data (931) representing a setting of a laser therapy device (937) and / or at least one refractive target (403) and / or one patient (405), to calculate output data (933) representing at least one predicted refraction (935), wherein the cost function (915) used for training: - A nonlinear output value (917) representing a measure of the prediction error of a univariate linear regression function and - Includes at least one nonlinear penalty term (919) which represents a measure of the violation of at least one invariant parameter (101). [2] Computer-implemented method (200) according to claim 1, wherein a non-linear part of the cost function (915) is calculated using a multi-layered feed-forward network with the training data (903) and the ground-truth data (905). [3] Computer-implemented method (200) according to claim 1 or 2, wherein the method step 3) comprises minimizing the calculated penalty value (913): - A variation (315A) of at least one weighting (921) in a direction which, when determining the total penalty value (913), reduces both the prediction error (917) and the non-linear penalty term (919). [4] Computer-implemented method (200) according to one of claims 1 to 3, in which the cost function (915) and the artificial intelligence-based model (M) are differentiable and gradient-based meta-training of the artificial intelligence-based model (M) is performed. [5] Computer-implemented method (200) according to one of claims 1 to 4, wherein a taxation value can be determined for the non-linear penalty term (919) which represents an influence of the non-linear penalty term (919) and / or the non-linear output value (917) of the univariate linear regression function. [6] Computer-implemented method (200) according to any one of claims 1 to 5, wherein the non-linear penalty term (919) comprises a measure of violation of at least one invariant parameter (101) from the list of parameters (101): - A monotonicity parameter (101a) representing the monotonicity of a regression curve, - A smoothness parameter representing the smoothness of a regression curve, - A variance parameter representing the variance of a regression curve, and - Represents a limit parameter that depends on at least one input parameter. [7] Computer-implemented method (200) according to any one of claims 1 to 6, wherein the non-linear output value (917) of the univariate linear regression function comprises at least one error measure from the list: - Root Mean Square Error; - Mean Absolute Error, and - Negative Log Likelihood represented. [8] Computer-implemented method (200) for training an ensemble of a large number of artificial intelligence-based models (M), wherein in the case a) that for an invariant (101) the satisfaction of the invariant condition by a single artificial intelligence-based model (M) implies the satisfaction by the entire ensemble: - at least one of the plurality of artificial intelligence-based models (M) is trained according to a method (200) according to one of claims 1 to 7, and wherein in the case b) that the condition of a) is not met: - The ensemble is considered as a single overall model based on artificial intelligence and this overall model is trained according to a method (200) according to any one of claims 1 to 7. [9] Computer-implemented method (200, 500) according to any one of claims 1 to 8, further comprising the method steps: - Generating at least one synthetic data tuple (505) for a data tuple (909) present in the training data (903), wherein the at least one synthetic data tuple (505) satisfies the following conditions: ◯ Elements of the data tuple (505) are functionally related to the invariant parameter under consideration (101); ◯ The values ​​of the elements of the data tuple (505) lie for each element within a value range predefined for that element, - Evaluating the nonlinear penalty term (919) of the cost function (915) for all synthetic data tuples (505), and - Varying at least one weight (921) of the weighting matrix (923) of the artificial intelligence-based model (M) depending on the evaluation of the nonlinear penalty term (919), wherein the evaluation and variation is repeated until a value of the nonlinear penalty term meets a predetermined limit (ε invariant ) falls below. [10] Computer-implemented method (200, 500) according to claim 9, wherein a plurality of synthetic data tuples (505) is generated, and wherein the elements of the generated synthetic data tuples (505) are substantially - are evenly distributed over the predefined range of values ​​or - represent a previously known probability distribution of the value range of the element in question. [11] Computer-implemented method (200) according to any one of claims 1 to 8, wherein the training data (903) and the ground truth data (905) are divided into batches (952) comprising a predetermined number of data tuples (909) and the method steps 1) to 3) are applied to each of the batches (952) to train the artificial intelligence-based model (M) until at least one termination condition (927) is reached, wherein following such a training loop (951) the additional method steps of any one of claims 9 or 10 are carried out. [12] A trained artificial intelligence-based model (M) or weighting matrix (923) of such a trained model (M), wherein the trained model (M) or its weighting matrix (923) was obtained by a computer-implemented method (200, 500) according to any one of claims 1 to 11. [13] Computer-implemented procedure (600) to support an operation on a patient's eye, comprising an artificial intelligence-based model (M) and comprising the procedure steps: - Reading (601) input data (931) that includes at least one setting of a laser therapy device representing data (937) and / or data representing a refractive target (403) and / or data representing a patient (405); - Calculating output data (933) which includes at least one predicted refraction representing data (935), using the artificial intelligence-based model (M); and - Provide (603) the output data (933). [14] Method according to claim 13, wherein the input data (931) includes at least one piece of information from the list, comprising - information about the axial length of the eye; - information about the preoperative depth of the anterior chamber; - information about the surface curvature and / or corneal contour of the cornea of ​​the eye); - information about manifest refraction and / or pachymetry; - information about a target refraction; - information about the biomechanical stability of the eye; - information about a pre-existing eye condition; - information about pretreatment of the eye; - information about the shape of the eye to be treated; and - includes information about the patient's age and / or gender and / or ethnicity. [15] Method according to claim 13 or 14, wherein after reading (601) input data (931) and before calculating output data (933) the following method steps are performed: ◯ Checking the similarity of input data (931) read in with the training data (903) used to train the artificial intelligence-based model (M); and ◯ If a similarity factor (SF) representing the similarity falls below a predefined similarity threshold: Perform the additional process steps of one of claims 9 or 10. [16] Computer program product comprising instructions which, when executed on a data processing device (800), cause the data processing device (800) to execute the process steps of a method according to any one of claims 1 to 15 [17] Device for data processing (800) which is configured to carry out a method according to any one of claims 1 to 15. [18] A computer-readable medium comprising an artificial intelligence-based model (M) according to claim 12, wherein the model (M) is configured, when the model (M) is executed by a data processing device (800), to cause the data processing device to calculate output data (933) based on input data (931) representing at least one setting of a laser therapy device (937) and / or at least one refractive target (403) and / or patient (405), and to provide these output data (933). [19] Laser therapy device comprising or configured to perform a method according to any one of claims 1 to 15, including a data processing device (800) according to claim 17.

Images