Super-personalized treatment based on coronary artery motion field and big data

Abstract

A system (SYS) and related methods for predicting patient treatment options. The system may include an input interface (IN) for receiving input data including biodynamic measurements regarding a patient. A predictor module (PM) is configured to process the biodynamic measurements to obtain output data comprising an indication of a treatment option for the patient.

Description

Technical Field

[0001] The invention relates to a system for predicting treatment options for a patient, a training system for training such a system, related methods, a computer program element and a computer-readable medium. Background Art

[0002] Personalized medicine (or "precision medicine") is an approach to medicine that some define as "the tailoring of medical treatments to each patient's individual characteristics. This does not literally mean creating a drug or medical device that is unique to the patient, but rather the ability to classify individuals into subgroups that differ in their susceptibility to a particular disease or response to a particular treatment. Preventive or therapeutic interventions can then be focused on those who will benefit, sparing costs and side effects for those who will not." See a US government report cited by J Watkins et al., Therapeutic Delivery, 1(5), Nov 2010, pp. 651-65.

[0003] Modern clinical management guidelines are often based on cohort studies that consider a wide range of patient characteristics.

[0004] However, due to the static nature of guidelines and associated decision mechanisms, there are limits to the number of parameters and decision points that can be considered in such guidelines or current efforts at personalized medicine. Summary of the invention

[0005] Therefore, there may be a need for a technical solution for improving medical practice.

[0006] The objects of the invention are achieved by the subject-matter of the independent claims, wherein further embodiments are incorporated in the dependent claims.It should be noted that the following described aspects of the invention apply equally to the associated method, training system, computer program element and computer readable medium.

[0007] According to a first aspect of the present invention, there is provided a system for predicting patient treatment options, comprising:

[0008] an input interface for receiving input data including biokinetic measurements about a patient (to be treated);

[0009] A predictor module is configured to process the biokinetic measurements to obtain output data including an indication of a treatment option for the patient.

[0010] In an embodiment, the biokinetic measurements are time series.

[0011] In an embodiment, the predictor module is preferably based on a trained machine learning model previously trained on patient data from a patient cohort.

[0012] In an embodiment, the machine learning model is implemented as or based on any of the following: i) a clustering algorithm, ii) one or more artificial neural networks, iii) a support vector machine.

[0013] An implementation based on a clustering algorithm is preferred. The training data may be clustered into clusters.

[0014] In an embodiment, different treatment options correspond to different clusters, and wherein said indication comprises an indication of one or more of said clusters.

[0015] In an embodiment, the clustering algorithm is based on a similarity metric, based on which similarity metric the indication is provided. The similarity metric may relate to the similarity between the measurement data of two given patients. Similarity may be used in the training phase to divide the training data into different clusters based on patient groups. After training, in deployment or testing, the treatment option prediction for a given patient is calculated based on the similarity metric or a related metric. Specifically, the similarity metric is used to establish the similarity between a given patient and a predefined cluster. Conceptually, in some embodiments, the similarity metric may also be understood as a distance metric. A given patient may fall into multiple clusters. Therefore, the use of a clustering algorithm avoids "hard" classification, and better takes into account that a given patient may benefit from a variety of treatment options or a mixture of these options. Although it has been found that clustering type algorithms are particularly beneficial in this article, other machine learning model methods, such as ii) and iii) listed above or not excluded in this article, can still be used in combination with clustering or in place of clustering. In some embodiments, a fuzzy classification model and / or a multi-class classification model may be used.

[0016] In an embodiment, the system comprises a graphical display generator configured to generate a graphical display for display on a display device, the graphical display for visualizing the indication.

[0017] In an embodiment, the graphical display includes a visualization of the clusters and a graphical indicator for the patient indicating similarity or proximity to the clusters. Visualizing predicted treatment option indicators for predefined clusters based on training data allows a user to quickly determine an appropriate treatment option, particularly for patients with mixed needs.

[0018] In an embodiment, the biodynamic measurement results include any of the following: coronary vasomotor data, perfusion data, ECG (electrocardiogram) data, EEG (electroencephalogram) data, oxygenation data.

[0019] It has been found that coronary motion data is a particularly robust (bio)marker / indicator for treatment options, in particular but not limited to cardiac related treatments, such as stenosis treatment types etc., or for other cardiac treatments, as well as extra-cardiac treatments. The coronary motion data may be represented by a 4D (3D+t) or 2D+1 motion field, e.g. based on segmented cardiac images. The motion field may be based on a registration over time frames. The coronary motion data may be prepared and used as a descriptive vector or matrix representing such a motion field.

[0020] In an embodiment, the biodynamic measurements comprise image data in the projection domain or in the image domain.

[0021] In an embodiment, the coronary vessel data is based on back-projected vessel segmentation in multi-directional projection images acquired from the patient.

[0022] In an embodiment, the coronary vessel data is based on registration of a series of 2D projection images.

[0023] In an embodiment, the output data comprises outcome data of the treatment options.

[0024] In another aspect, a training system is provided for training a machine learning model of the system of any one of the preceding claims based on training data.

[0025] In an embodiment, the training system is implemented based on a clustering algorithm and a similarity measure, wherein the training system defines clusters in the training data based on the similarity measure.

[0026] In another aspect, a computer-implemented method for predicting patient treatment options is provided, comprising:

[0027] receiving input data including biokinetic measurements about a patient; and

[0028] The biokinetic measurements are processed to obtain output data including an indication of a treatment option for the patient.

[0029] In another aspect, a computer-implemented method for training a machine learning model of a system according to any one of the preceding claims 1-13 based on training data is provided.

[0030] In another aspect, at least one computer program element is provided which, when executed by at least one processing unit or system, is adapted to cause the processing unit to perform a method according to any one of the above embodiments.

[0031] On the other hand, at least one computer-readable medium is provided, on which the program unit is stored or on which the machine learning model used in the system described in any one of the above embodiments is stored.

[0032] In another aspect, use of a trained machine learning model for predicting treatment options for a given patient is provided.

[0033] The proposed system or method allows personalized medicine to be implemented by taking into account a large number of characteristics of the current patient when necessary. The system provides a technical solution for assisting medical decision-making. Costs can be reduced, patient comfort can be increased, and wear and tear on expensive medical equipment due to unnecessary use can be reduced. For example, in an X-ray instrument, the anode disk is subjected to very high temperature gradients. If the X-ray instrument is used in the wrong way, such as unnecessary use, it may cause premature degradation due to the patient's selection of the wrong / unnecessary treatment option. And, more worryingly, the patient is deprived of the treatment he or she does need. The proposed system allows a hyper-personalized approach because the patient can benefit from a customized "transformation" of the group data to their specific characteristics. The group data can include a large amount of data in many dimensions. The biodynamic measurement results can be represented as a potential high-dimensional descriptive vector or matrix / tensor, thus increasing the expressiveness. The descriptive vector can be used as a search keyword for searching for treatment options for a given patient, and in some embodiments, this database search operation is contemplated herein.

[0034] In some embodiments, the predictor module may perform a lookup operation in a patient database to perform a more personalized analysis of the available data while taking into account a potentially large number of patient characteristic parameters. This high degree of personalization may lead to better customized outcome predictions, thereby helping, for example, to formulate an optimal treatment plan. In some preferred embodiments, machine learning ("ML") methods are used to train an ML model on a large amount of patient data stored in a database, and the predictor module uses this trained model to predict (i.e., calculate) treatment options. The predictor module may preferably operate in real time.

[0035] In some embodiments, the biodynamic measurements can be expressed as descriptive vectors describing the movement of the patient's anatomical features, such as the movement of the coronary arteries in particular. Patients with coronary artery disease have lesions in their coronary arteries, usually stenosis, which can lead to impaired heart function. The function of the coronary arteries is to provide oxygenated blood to the myocardium, and this function is severely impaired in such patients. The stenosis can be a single focal lesion or it can be severe, with multiple stenoses in multiple blood vessels. The degree of stenosis may vary, it may last for many years or be acute. All of these factors and others affect the severity of myocardial damage, resulting in different movement patterns that have been discovered herein. In addition, this movement pattern can express aspects that cannot be assessed when simply measuring the degree of stenosis, such as the duration of the disease and the viability and vitality of the muscles. It seems that these factors do contribute to possible treatment outcomes and treatment options.

[0036] In some embodiments presented herein, coronary artery motion patterns are determined based on a preferably large pool of prior data that may exist in current (large) medical databases of patients with coronary artery disease. These patterns can be clustered into descriptive clusters. Correlations between motion patterns and disease progression, optimal treatment decisions, optimal follow-up regimens, risk factors, etc. can be determined. Since motion patterns have been found to represent disease severity, location, and diversity, correlations with the above factors can be expected. Clustering is performed by comparing descriptive motion vectors using a suitable similarity metric.

[0037] Then, once clustering is complete, and a (new) specific patient is to be diagnosed or treated, the motion pattern of that patient can be established. Thus, the cluster (and sub-clusters) that the patient falls into can be established by comparing the patient's motion pattern with the motion patterns in the cluster according to a similarity measure. Since the clusters express patient outcomes and optimal treatment options, this information can be used to determine the best treatment strategy, perform device selection, develop follow-up plans, and risk monitoring plans.

[0038] However, the motion pattern of the cardiovascular vessels is one example of a biodynamic measurement, and other examples are contemplated herein, such as perfusion, EEG, ECG, etc. Thus, a biodynamic measurement is not necessarily related to the physical motion of one or more anatomical features, but can be related to other bioquantities and their evolution over time. Clustering methods are preferred herein because clustering can capture the concept of treatment options well with the right level of fuzziness.

[0039] "User" refers to a person who operates an imaging device or at least partially supervises or manages a medical procedure involving a patient, such as a medical staff or other person. In other words, the user is usually not the patient, however, it is not excluded in this document that in some applications, such as for information purposes, the user is a patient who can use the system as proposed in this document.

[0040] "(One or more) biodynamic measurement results" may not necessarily relate to raw data, but may relate to such raw data or derived data, possibly prepared to facilitate efficient processing, such as scaling or other types of mapping. (One or more) biodynamic measurement results may be represented as one or more time series of data measurement values ​​or (one or more) sequences of data measurement values, and may be represented in the singular or plural in this document. (One or more) biodynamic measurement results should be broadly interpreted to include any time-varying measurement results acquired in a single measurement period or over multiple such periods (which may be interrupted by one or more time gaps). During such one or more periods, suitable sensors / transducers, preferably for medical purposes, are applied to the patient to acquire (one or more) biodynamic measurement results. The biodynamic measurement results may vary in space and time, such as a sequence of medical image data. The data items may be scalars, or may be higher dimensional, such as 2D, or 3D or 4D or higher dimensional static data. The biodynamic measurement results may be acquired by suitable sensors of the patient and may represent the time evolution of one or more medical, anatomical features or biological / biomedical quantities that vary over time. The biodynamic measurement(s) may relate to movement, deformation, etc., or other changes with respect to one or more organs, one or more tissues, or one or more anatomical features. The biodynamic measurement(s) may be associated with treatment options and, optionally, with outcome data.

[0041] In general, the term "machine learning" includes a computerized device (also referred to herein as a "module") that implements a machine learning ("ML") algorithm. Some such ML algorithm operations are used to adjust a machine learning model that is configured to perform ("learn") a task. This adjustment or update of the model is called "training". In some applications (implicit modeling), the training data is the model or part of the model. In such an implicit setting, the adjustment of the model can broadly include structuring the training data, such as structuring the training data in clustering. Typically, the task performance of an ML module may improve significantly with the accumulation of training experience. The training experience may include suitable training data and the model's exposure to these data. The better the data represents the task to be learned, the better the task performance may be. "If the training data well represents the distribution of examples that measure the performance of the final system, the training experience helps improve performance". Performance can be measured by objective testing based on the output generated by the module in response to feeding test data to the module. Performance can be defined based on a specific error rate to be achieved for given test data. See, e.g., T. Mitchell, “Machine Learning,” p. 2, section 1.1, p. 6, section 1.2.1, McGraw-Hill, 1997. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Exemplary embodiments of the present invention will now be described with reference to the following drawings, which are not drawn to scale unless otherwise specified, and in which:

[0043] Figure 1 A computer-implemented medical apparatus for calculating treatment options for a given patient is shown;

[0044] Figure 2 In the embodiment shown in FIG. Figure 1 Graphical display generated by the system;

[0045] Figure 3 shows a representation of medical image data as an example of biodynamic measurements that may be used by the system herein in some embodiments;

[0046] Figure 4 shows a machine learning model that can be used in embodiments;

[0047] Figure 5 A block diagram of a training system for training a machine learning model that may be used in embodiments herein is shown;

[0048] Figure 6 A computer-implemented method of calculating treatment options for a given patient based on biokinetic measurements is shown; and

[0049] Figure 7A flowchart of a method for training a machine learning model based on training data is shown. DETAILED DESCRIPTION

[0050] Reference now Figure 1 , in which a medical apparatus AR is shown. As will be explained in more detail below, the apparatus AR is capable of generating (one or more) treatment options T for a patient based on medical measurement data ("medical measurement results") of a given patient PAT, optionally with associated result data o.

[0051] The apparatus AR includes in some embodiments, but not necessarily in all embodiments, a medical measurement device MD. The medical measurement device MD may use one or more sensors S to acquire medical measurements of a given patient PAT. The measurements m may be provided to a measurement processing system SYS, which is configured to generate (e.g., calculate) a treatment option T for a given patient PAT based on the provided measurements. That is, as said, the result data o associated with the treatment option T is optional, and the treatment option T will be simply referred to as an output provided by the system SYS hereinafter, it being understood that the output may or may not include the associated result data o.

[0052] As used herein, medical measurements m refer to biokinetic measurements (data) m. In an online embodiment, the system SYS may receive biokinetic measurements m via a communication channel from a measurement device MD, which is operable to acquire such measurements about a given patient's PAT. Alternatively or additionally, in an offline embodiment, the measurements m may be received from a data storage device MD, such as a medical database or other database storing previous measurements of a given patient. Thus, the biokinetic measurements may be first stored in the storage device MD after acquisition and may then be retrieved by the system SYS for processing as needed (e.g., upon a query by a medical user).

[0053] In either case, one or more biokinetic measurements m are received at a data input port IN of the system SYS and processed by a predictor module PM of the system SYS. The predictor module PM generates output data based on the biokinetic measurements m thus received. The output data may include an indication of an associated treatment option T based on the one or more biokinetic measurements m. The output data including the calculated / predicted treatment options may be stored in a data storage device, such as in the medical database MD, or may be stored elsewhere. The calculated treatment option T may be stored as an entry in the patient's health record. Any one or more of a variety of uses of the output data are contemplated herein. For example, the output data may be displayed on a display device DD, or may be processed in other ways, such as by a medical data analysis system, or may be used to control a medical device D to achieve treatment of a patient's PAT, etc.

[0054] The output data including the treatment options T may be provided in a suitable form, such as text data (natural language or encoded), digital data, graphical data in the form of a graphical display, or as a combination of two or more of the above. Specifically, the graphic display generator GDG is operable to generate a graphic display GD. For example, the graphic display GD may be displayed on a display device DD. Figure 2 An example of such a graphic display GD including an indication of (one or more) treatment options T is shown in FIG, which will be explained in more detail below. As an alternative or in addition to such a display, treatment option data T may be provided to a medical device D via a communication or control interface CI. The medical device D may be controlled based on the treatment data T, for example, to treat a patient, or to initiate one or more steps related to a future treatment of the patient. The treatment device may include a radiation treatment device such as a linear accelerator, a contrast agent pump, a surgical robot, etc.

[0055] The system SYS for predicting treatment recommendations T can be implemented by one or more computing devices PU. For example, in one embodiment, a cloud-based architecture is used, in which the system SYS is implemented by one or more servers that preferably work in conjunction. The system SYS can provide calculated treatment options to terminal (user) devices via a suitable communication channel. In other embodiments, the system SYS is implemented entirely or partially on a terminal device or other edge device of a communication network. The terminal device can be a desktop computer, a laptop computer, a tablet computer, a smart phone, a smart watch, or other computing device of a clinical user or the patient himself. Preferably, but not necessarily, the computing device PU on which the system is implemented is powered by one or more processors capable of parallel processing. For example, a processor of a multi-core design can be used to parallelize the calculations performed by the predictor module PM.

[0056] In an embodiment, the output treatment option T may be a specific treatment regimen, such as a radiation treatment plan or a drug regimen with specific parameters. However, more preferably, in other embodiments, the treatment option data T calculated by the predictor module PM may instead define a specific treatment template or treatment option type, which typically does not provide detailed parameters, but rather provides a set of ranges or types of treatment parameters that may result in a favorable outcome for the patient given the biodynamic measurement result m. ​​For example, in radiation therapy, such as external or internal radiation beam therapy, which is contemplated as a treatment option herein, the treatment option T may involve a specific type of radiation distribution with respect to the organ of interest and / or organ at risk, or involve a specific type of flux map, etc.

[0057] In some embodiments, the predictor module PM may be implemented by explicit analytical modeling techniques, where the relationship between the measurement results m and the treatment options T is modeled by (one or more) explicit functions. For example, in some embodiments, the predictor module PM may be implemented by a lookup table that lists the types of measurements and the associated treatment options.

[0058] However, preferably, the predictor module is implemented in machine learning ("ML") by using one or more trained machine learning models. Machine learning allows for the collection of historical medical knowledge that exists in previous medical data collected from a cohort of patients in past historical examinations or cases. The previous medical data may be found in an existing medical database MD', for example a data set comprising medical records of previous patients and / or indeed previous records of a given patient's PAT.

[0059] Figure 1 The lower left portion of MD' illustrates a medical dataset as a plurality of medical records j stored in a database MD'. A medical record j of a patient group (index j refers to a different patient) may include biodynamic measurements m acquired in the past as part of diagnosis / medical fact finding. j , the associated treatment option T for patient j j and possible result data O j For different patients j, j', there may be different treatments T j 、T j’ , producing different results O j , O j’ Such different treatment options may be based in part on different medical aspects, such as different patient medical histories, different patient physiology, different genetic makeup of the patients, etc. Each of the aforementioned medical aspects may explain why different patients respond differently to different treatment options.

[0060] It can be assumed that it was the treating doctors at the time who used their medical knowledge to decide the corresponding treatment options that were considered suitable for the patients in the group recorded in the historical data record j of the medical database MD'. Therefore, the collection of historical medical records MD can be considered as an encoding of such medical knowledge. In particular, on the one hand, the medical measurement results m collected from the patients in the past j The disposition option associated with j and optionally their associated results Oj may have a potential relationship between It may be difficult or impossible to model this underlying relationship analytically in terms of analytical functions of set types. This is referred to herein as the measurement versus treatment (“MvT”) relationship. Machine learning offers a practical way forward, allowing for a more flexible modeling approach that is not constrained from the outset by a specific analytical function expression of this relationship. ML can collectively use prior medical data to implicitly model the underlying MvT relationship.

[0061] In a preferred embodiment, it is proposed herein to use any of a variety of different machine learning models to learn this relationship from the prior medical data MD' Based on appropriately large and representative prior data from different cohorts of patients After sufficient training, the predictor module PM can be constructed in a manner that allows for the computation of reliable treatment option recommendations / predictions for a given patient based on currently or previously acquired biokinetic measurements of the given patient's PAT. The system SYS allows for the reduction of overall healthcare costs by computing suitable treatment options personalized to the respective patient in a computationally efficient manner. It is assumed here that the different potential parameters that determine whether a treatment option is suitable for a given patient (i.e., has a positive or favorable outcome) are implicitly encoded in the overall biokinetic measurements available for the given patient.

[0062] Biodynamic measurements represent an aspect of a patient's constitution that changes over time and can be viewed as a tell-tale of his or her response to treatment options. Biodynamic measurements as used herein may be stored or provided as time series data or spatially varying data, depending on the type of sensor used to acquire them. Some measurements may vary over time and space. For example, a biodynamic measurement may include a measurement m=m over time t acquired by (one or more) sensors S of a medical measurement device MD (e.g., an ECG device, an EEG device, a laboratory device that establishes blood parameters over time). t, or a transmission / emission image of a patient. Some medical devices MD work by having a sensor(s) S pick up a signal of interest, such as electrodes attached to the patient's skin, for acquiring EEG or ECG data.

[0063] In other embodiments, the medical measurement device MD comprises a signal source SS that generates an interrogator signal to which the patient is exposed. The interrogator signal interacts with the patient tissue and generates a response signal in response thereto, which is then picked up by the sensor S. This is the case, for example, with medical imaging data acquired over time, which is an example of temporally and spatially varying data, and in turn an example of biodynamic measurements contemplated herein. For example, in a CT or C-arm setting, during an imaging session, an X-ray source as the signal source SS is rotated around an examination region in which the patient is located to acquire projection images from different directions and preferably over a period of time. The projection images are detected by an X-ray sensitive detector S. The X-ray sensitive detector S may be rotated relative to the X-ray source SS around the examination region, although such co-rotation is not necessarily required, for example in CT scanners of the fourth generation or higher. The signal source SS, for example an X-ray source such as an X-ray tube, is activated so that an X-ray beam is emitted from a focal spot in the tube during the rotation. The X-ray beam passes through the examination region and the patient tissue therein and interacts therewith, thereby generating modified radiation due to, for example, differential attenuation and scattering effects. The modified radiation is detected by an X-ray sensitive detector S as, for example, intensity. The X-ray sensitive detector S is coupled to an acquisition circuit such as a DAQ, capturing the projection image in digital form as a digital image. The same principle applies to (planar) radiography, except that the source SS is not rotated during imaging. In such a radiographic environment, the projection image can then be examined by a radiologist. In a tomography / rotational setup, the multi-directional projection images are first processed by a reconstruction algorithm that transforms the projection images from the projection domain into cross-sectional images in the image domain. The image domain is located in the examination region. A C-arm imager can be used to acquire projection images of a region of interest (such as the human heart) over time. Dual / multi-plane C-arm scanners allow multi-directional imaging without rotation during imaging. Fluoroscopic (particularly angiography) imaging is envisaged herein in a preferred embodiment, which will be discussed in more detail below with reference to some examples.

[0064] The image thus acquired is an example of a biodynamic measurement. However, such an image may not necessarily be from X-ray imaging. In contrast to the previously mentioned transmission imaging modalities, other imaging modalities are also contemplated herein, such as emission imaging, such as SPECT or PET, etc. Furthermore, in some embodiments, magnetic resonance imaging (MRI) is also contemplated herein, as is optical coherence tomography, and so on.

[0065] In an MRI embodiment, the signal source SS is formed by a radio frequency coil, which may also serve as a detector sensor S, configured to receive, in a receive mode, a radio frequency response signal emitted by a patient residing in a magnetic field. Such a response signal is generated in response to a previous RF signal emitted by the coil in a transmit mode. However, in some embodiments, there may be dedicated transmit and receive coils, rather than using the same coil in the different modes.

[0066] In emission imaging, a source SS is present in the patient's body in the form of a previously administered radioactive tracer that emits radioactive radiation that interacts with the patient's tissue. This interaction results in a gamma signal that is detected by a suitably configured sensor S (in this case a gamma camera), which is preferably arranged in a ring around the examination region where the patient is located during imaging.

[0067] As an alternative or in addition to the above mentioned modalities, ultrasound (US) is also envisaged, wherein the signal source SS and the sensor S are suitable acoustic US transducers.

[0068] As previously mentioned, some imaging devices can provide biodynamic measurements that may vary in both space and time. For example, time-varying images (such as projection angiography, computed tomography angiography ("CTA"), etc.) represent spatially and temporally varying data, and in some examples contemplated herein, such as medical images that capture the motion of cardiac vessels during one or more cardiac cycles.

[0069] In some embodiments, it may be preferred herein to combine biodynamic measurements from acquisition devices of different sensor types into a "super" biodynamic measurement to obtain better results. Thus, for example, ECG data and medical images may be provided by a predictor model PM for joint processing. The underlying ML model M of such a module PM may be jointly trained on such data.

[0070] In an embodiment, machine learning is typically performed in two / three phases: a training phase, an optional testing phase, and a deployment / inference phase.

[0071] In the training phase, the model M is adjusted based on the previous training data. The training phase can be a one-time operation, or can be repeated as more training data becomes available. After the training phase, there can be an optional testing phase, and after the testing phase, there is also a deployment / inference phase. Thus, the trained model can be used for testing or deployment.

[0072] In embodiments, model adjustments in the training phase may include implicit or explicit modeling. For example, a neural network type model may be used as an example of explicit ML modeling, while a data clustering method (also contemplated herein) represents an example of implicit ML modeling.

[0073] In explicit modeling, in a previous training phase, the parameters of the model M are adjusted based on training data and according to predefined criteria, for example measured by one or more objective functions.

[0074] Once the model is sufficiently trained, which can be established by a target criterion (such as the behavior of an objective function), the model can be used for testing or deployment in daily medical practice. Thus, in explicit modeling, a (preferably large) collection of parameterized functions can be used to model the relationship Perform modeling. Adjust parameters based on training data. Such modeling can include neural network type modeling, support vector machine, etc.

[0075] In other types of ML, such as the implicit modeling, the training data set itself represents the model, optionally including a set of descriptors, as a way to construct data for deployment. One such type of implicit modeling includes the clustering type machine learning algorithm specifically contemplated in the embodiments herein. Clustering type algorithms are preferred herein because the "fuzziness" inherent in treatment options sometimes can be better graphically represented as clusters shown for the patient's current measurement results m. It has been found that treatment options are not always well suited for hard classification because hard classification tends to set treatment options too low. Clustering better captures the fluidity or trend between and throughout such hard classifications. It can be more easily and intuitively determined by the user which category or cluster the patient tends to fall into. Each cluster is considered to be associated with a specific type or template of the treatment option type (optionally with associated result data), rather than with a hard set of predefined parameters. Therefore, medical users can more easily determine in which "treatment selection area" the patient is more likely to obtain a positive treatment result without being forced to make a hard (difficult) choice at the parameter level.

[0076] Figure 2 Examples of output data that may be generated by a predictor module PM are shown, in particular when implemented based on a clustering type algorithm. Figure 2 Also illustrated is an embodiment of a graphical display GD that may be generated by a graphical display generator GDG processing output data received by a predictor module PM. For example, in the described clustering type algorithm, after training, the predictor module PM generates a graph for a given predefined cluster C into which a patient may fall. i One or more of T provide an indication, such as a centroid or other descriptor. Each cluster represents one (or more related) (different) treatment options T iFor example, Figure 2 As shown, cluster C i Can be generated as a point cloud in vector space.

[0077] Each point corresponds to a set of parameters represented in an appropriate high-dimensional vector space. Appropriate encodings may include one-hot encoding, embedding or other encodings. The measurement result preprocessor PP processes the raw measurement results into a suitable representation. The measurement result preprocessor PP may perform mapping, scaling or normalization operations as required. The preprocessor PP itself may use ML to obtain a suitable data representation for fast and efficient processing by the module PM. In more detail, the preprocessor PP may obtain a suitable encoding by, for example, simply mapping to lists or symbols or by ML techniques such as mapping to the latent space of an autoencoder (AE) network or variational AE, etc. Although the vector space may be a high-dimensional space with a dimension higher than 3, it may still be possible to obtain a 3D or 2D (such as) dimensionality reduction technique such as PCA or other. Figure 2 The parameters spanning the vector space may include appropriately encoded vectors for the measurements, and may optionally be enriched with patient context information κ, such as biometrics (weight, age, sex, etc.), patient medical history, etc. In some preferred embodiments, the coronary artery motion pattern is used as an example of a biodynamic measurement. The context information κ parameter may also include one or more of the following: baseline clinical demographics, clinical characteristics of the patient, laboratory values, medications, surgeries, and follow-up data coronary artery motion pattern.

[0078] The nature of the preprocessor PP may depend on the clinical domain and the nature of the measurement results. Figure 3 A detailed example based on cardiovascular image data is described in , where the preprocessor PP performs appropriate data preparation operations to capture the vascular motion pattern as an example of the biodynamic measurement result m.

[0079] Continue to refer Figure 2 , point cloud C i Some or each of them may be provided by the corresponding cluster C i An indication of the characteristics of the treatment option represented is associated with a graphical or textual description.

[0080] When processing the medical measurements of the current patient, the predictor module calculates the similarity or proximity to a given cluster. Figure 2 A graphical widget or indicator, shown as an example cross ('X') in FIG. 1 , may be superimposed on the point cloud representation of the clusters so that a user can discern which treatment type category (or indeed, multiple categories) a given patient tends to fall into. Figure 2In the embodiment of FIG. 5 , the current patient X may fall into the upper left cluster and may benefit from the associated treatment options of the associated treatment results. It is likely that the patient PAT falls into multiple clusters C i , in this case, the indicator X computed by the module PM is plotted between a set of clusters.

[0081] Figure 3 An example of a biodynamic measurement that can be used in the embodiments contemplated herein is shown. The biodynamic measurement m may include an X-ray image or other type of medical modality image acquired of a cardiac vessel (such as a coronary vein or artery of a patient). Preferably, a contrast agent is administered prior to acquisition to improve contrast. In an embodiment, at least two X-ray projection images π are acquired along different projection directions α, β. α , Π β The back-projection operation BP is used to α , Π β Obtain the measurement results of the region of interest (such as the heart and blood vessels) α ,Π β A 3D spatial representation of the recorded intensity distribution (image domain). A cross-sectional (tomographic) image I* (e.g. Figure 3 The reconstruction I* may be refined by using a greater number of projections. However, in some embodiments, it may be sufficient to use only two projections as shown.

[0082] In more detail, according to some embodiments, the measurement result m may include a motion pattern description vector of a 4D vector field representing the motion of the coronary arteries during one cardiac cycle. ML may be used for a preprocessor PP, for example based on a (preferably convolutional) neural network trained to obtain from at least two angiographic X-ray sequences II α t ,Π β t The coronary arteries are segmented in , where the index t indicates the acquisition time. Alternatively, the preprocessor PP may use non-ML techniques, or use ML methods other than NN. The segmentation may be back-projected into 3D space by the preprocessor PP, as described above and as Figure 3 As shown. CTA data with time-varying segmented coronary artery structures can be used as the basic truth. Optionally, but preferably, the angiographic image II α t , Π β t The ECG data recorded together can be used to form a super measurement result, which can be jointly processed by the predictor PM to obtain better robustness, because the ECG data indicates the cardiac phase that changes over time and can be combined with the contrast-enhanced angiography image Πα t , Π β t Used together to select frames in the angiographic sequence that represent the associated cardiac phase (for projection α, β, or both). This selection can be done automatically (e.g., based on minimal contrast fill or vessel overlap), or manually, or in any other clinically meaningful manner. Two (or more) angiographic sequences II α t , Π β t It can be preferably obtained from different angles (e.g., at least 30 degrees apart). α t , Π β t Imaging parameters, such as C-arm angiography system parameters (including, for example, rotation and angulation angles, field of view, source image distance, etc.), can be used to perform a (spatial perspective) backprojection m=I* over time t .

[0083] The above process can be performed for multiple phases (e.g., 15 phases) in the cardiac cycle. Then, each back projection m=I* t This forms a 4D sequence (3 spatial dimensions that vary with time). t can be used directly as the above-mentioned descriptive vector. Alternatively, and optionally, a spatial registration step between the various stages can be used (eg, using an iterative closest point algorithm or other algorithm) to build a 4D vector field, which will then be used as the descriptive vector for preparing the measurement result m.

[0084] As an alternative to the above process, the measurement result m can be prepared by a preprocessor PP as a 2D+t motion field, preferably within at least one cardiac cycle in at least two projection directions α, β. The angiogram Π at time t can be obtained by image registration of a plurality of (e.g., all) frames in the sequence relative to a preselected reference frame (e.g., the frame with the most contrast agent filling). α t , Π β t 2D+t motion fields are computed for some or each of the . The descriptive vector m then comprises a table of these 2D+t motion fields, optionally together with the system parameters of these motion fields (rotation and angulation angles, field of view, source image distance, etc.). Still optionally, the 2D+t motion fields are constrained to a single cardiac cycle. The pre-processor stage PP can be used during training to prepare the training data appropriately, and / or during deployment to prepare incoming new measurements encountered in deployment or testing.

[0085] The above examples of CTA-based cardiac vascular motion are not limiting in this context. For example, other vessels, other regions of interest not actually related to the heart, may be imaged instead. Additionally or alternatively, tomographic imaging, while preferred, is not necessarily required in this context. Pure projection-based imaging over time is also contemplated. Imaging modalities other than X-ray-based, such as MRI, SPECT / PET, US, etc., may be used instead.

[0086] Reference now Figure 4 , which shows an artificial neural network ("NN") type machine learning model M according to one embodiment of explicit modeling. Preferably, deep learning is used, where the neural network model includes multiple hidden layers Lj between the input layer IL and the output layer OL.

[0087] The model M can be obtained by Figure 5 The training is performed by a computerized training system TS described more fully in . During training, the training system TS adjusts an initial set of (model) parameters θ of the model M. In the context of a neural network model, these parameters are sometimes referred to herein as weights.

[0088] The machine learning model M may be stored in one (or more) computer memories MEM'. The pre-trained model M may be deployed as a machine learning component that may be run on a computing device PU (such as a desktop computer, workstation, laptop computer, etc.) or multiple such devices in a distributed computing architecture. Preferably, in order to achieve good throughput, the computing device PU includes one or more processors (CPUs) that support parallel computing, such as a processor of a multi-core design. In one embodiment, a GPU (graphics processing unit) is used, particularly (but not only) when the model is of an artificial neural network type, which embodiment will now be explained in more detail.

[0089] Specifically, now refer to Figure 4 , which shows a convolutional neural network M in a feed-forward architecture as an example of a NN. The network M comprises a plurality of computational nodes arranged in layers IL, Lj, OL in a cascaded manner, wherein the data flow proceeds from left to right, and therefore from one layer to another. Recurrent networks are not excluded herein, in particular when processing temporal data such as biokinetic measurements. Convolutional networks have been found to produce good results when processing image data. In some preferred embodiments, the network may be a convolutional network and be of recurrent type.

[0090] In the deployment, input data including measurement results m are applied to the input layer IL, optionally supplemented with context data κ. The input data m is then propagated through the sequence of hidden layers L1-L N(Only two hidden layers are shown, but there may be only one hidden layer or more than two hidden layers) then appears at the output layer OL as an estimated output M(m)=T, which represents the associated treatment option T.

[0091] To improve computational and memory allocation efficiency, the layers of the network, and indeed the inputs m and / or outputs T, and the inputs and outputs between hidden layers (referred to herein as feature maps), can be represented as vectors or two-dimensional or higher-dimensional matrices ("tensors").

[0092] Preferably, some or all of these layers are convolutional layers, i.e., they include one or more convolutional filters CV that process the input feature maps from earlier layers into intermediate outputs, sometimes called logits. For example, an optional bias term can be applied by addition. The activation layer processes the logits into the next generation of feature maps in a nonlinear manner, which are then output and passed as input to the next layer, and so on. The activation layer can be implemented as a rectified linear unit RELU as shown, or as a Softmax function, a sigmoid function, a hyperbolic tangent (tanh) function, or any other suitable nonlinear function. Optionally, there may be other functional layers, such as a pooling layer P or a dropout layer to facilitate more robust learning. The pooling layer reduces the dimension of the output, while the dropout layer cuts off the connection between nodes in different layers.

[0093] In an embodiment, downstream of the sequence of convolutional layers there is an output layer OL configured for a classification or regression task. In particular, a classification result may be sought, such as classifying a treatment option T from a set of predefined treatment templates. For example, in RT therapy, the treatment option T may represent a specific flux map, or a metric characterizing such a flux map, or more generally a treatment plan template. Downstream or upstream of the run of convolutional layers there may be one or more fully connected layers. When configured for a classification task, the output layer OL may be implemented as, for example, a Softmax layer.

[0094] It should be understood that Figure 4 The above-mentioned model M is only according to one embodiment and is not intended to limit the present disclosure. Other neural network architectures are also contemplated herein, which have more or less or different functions than those described herein, such as pooling layers or dropout layers, etc. More importantly, the model M contemplated herein is not necessarily of the neural network type at all. In alternative embodiments, other statistical regression methods based on sampling from training data are also contemplated herein. Other techniques may include Bayesian networks or random fields, such as Markov type random fields, etc.

[0095] Figure 5An embodiment of a training system TS is shown that can be used to train a machine learning model M (such as in explicit or implicit modeling). The system TS can include an updater UP that updates / adjusts model parameters for the training data set structure.

[0096] Training is the process of adjusting the parameters of the model M based on the training data. An explicit model is not necessarily required, as in some examples, the training data itself constitutes the model, such as in clustering techniques or the k-nearest neighbor algorithm, as described above.

[0097] In explicit modeling, such as in NN-based methods and many other methods, the model can include a system of model functions / computation nodes whose inputs and / or outputs are at least partially interconnected. The model function or node is associated with a parameter θ that is adjusted during training. The model function can include convolution operators and / or nonlinear units (as described above in Figure 4 In the NN case, the parameters θ may include the weights of the convolution kernel of the operator CV and, optionally, the weights of the nonlinear unit. The parameterized model can be formally written as M θ . Parameter adjustment can be achieved through a numerical optimization process. The optimization can be iterative. The objective function F can be used to guide or control the optimization process. Adjust or update the parameters to improve the objective function. Input training data x = m' k is applied to the model. k is the index of the pairwise training data traversal in the supervised setting. The model M responds to the training data to produce output M θ (m' k )=T' k The objective function is a mapping from parameter space to a set of numbers. The objective function F can measure the output M of the training data. θ (m' k ) and the corresponding target y = T' k The parameters are iteratively adjusted to reduce the combined deviation until the stop condition preset by the user or designer is met. The objective function can use a distance / similarity metric ||.|| to quantify the deviation.

[0098] In some embodiments, but not all embodiments, the combined bias may be implemented as a function of the training data instance / pair (m' k ,T' k ) k The sum of some or all residuals of , and in terms of the objective function, the optimization problem can be formulated as:

[0099] argmin θ F=∑ k ||M θ (m′ k ),T′ k|| (1)

[0100] In setting (1), the optimization is formulated as the minimization of a cost function F, but this is not limiting in this context, as a dual formulation of maximizing a utility function may be used instead. The sum is over the training data instances k.

[0101] The cost function F may be pixel / voxel based, such as L1 or smoothed L1 norm, L2 norm, Hubert function or Soft Margin cost function. For example, in a least squares method or similar method, the (squared) Euclidean distance type cost function F in (1) may be used for the above regression task to regress to the treatment option T. When the model M is configured as a classifier that classifies from a set of such treatment options to the treatment option, the sum in (1) may be alternatively formulated as one of the cross entropy or negative log-likelihood (NLL) divergence or similar values.

[0102] The exact functional composition of the updater UP depends on the optimization procedure implemented. For example, an optimization scheme such as back / forward propagation or other gradient-based methods can then be used to adjust the parameters θ of the model M in order to reduce all or a subset of the training pairs (m') from the full training dataset. k ,T' k ). Such subsets are sometimes called batches, and optimization can be performed in batches until all of the training dataset has been exhausted, or until a predetermined number of training data instances have been processed.

[0103] Reference is now made to equation (2) below, which represents an example of the implicit type modeling approach described previously, which can be used to train a system TS, such as in clustering or similar ML methods:

[0104] argmin θ＝{Ck} F(C,m') (2)

[0105] In such a model, the training data itself represents the model, which is optionally enhanced by descriptors δ (such as centroids, etc.) of the training data τ or other structured mechanisms. The descriptors can be calculated based on the training data and a suitable similarity measure ∑ defined for the elements of the training data set. For example, k-means clustering or a similar algorithm can be used. Equation (2) or (1) can be implemented by one or more iterations. The objective function F can be formulated based on a distance function Σ that quantifies the distance between the training data measurement m' and the existing descriptor δ of the cluster calculated in the previous iteration cycle.

[0106] In clustering type algorithms, the objective function F is iteratively improved for some or each m' (a sample from the training data), and the optimization is performed on a subset C of the training data. kperformed on. F typically corresponds to a potential similarity ∑ or a distance metric. For each m’ in the current training dataset, either a new class m’ ∈ C is opened k , or m’ is considered a member of a class C that was opened in an earlier iteration k’ < k k’ . Thus, in such an implicit scheme, adapting the model involves structuring the dataset, e.g., by finding a subset that defines the clusters in the training dataset. Once the clustering of the current dataset (learning phase) is complete, newly received measurements in testing or deployment can be assigned to one (or more) classes according to F or ∑, as appropriate.

[0107] Clustering or similar implicit ML modeling methods are typically unsupervised and do not pair the training data into training inputs with the desired targets as in some of the explicit ML modeling methods described above. The similarity metric ∑ can be based on, e.g., the Euclidean squared distance, or it can be based on other metrics. In this or other embodiments, it can be assumed that the biokinematic measurements are mapped as elements of a suitable vector space so that the similarity metric can be clearly defined to compute the descriptors of the clusters in the learning step. An example of such a mapping is the descriptive vector as described above in Figure 3 .

[0108] At the end of training, the training data is clustered into clusters C i , optionally with an associated descriptor δ i . In the testing or deployment phase, a new measurement m is received. The new measurement m is applied to the descriptors to find the most suitable one or more descriptors (according to F), thereby finding the cluster C i into which the measurement m’ falls. Then, the corresponding disposition Ti associated with the one or more clusters so identified can be used to inform further medical steps for a given patient.

[0109] Whether using implicit or explicit modeling, training can be a one-time operation or can be repeated once new training data becomes available.

[0110] More specifically and referring to the example in Figure 3 , when the clustering is driven by a similarity metric as described above, grouping of similar motion patterns is preferably achieved, e.g., for the cardiac vascular motion described above. A suitable similarity metric between two given motion vector fields can be configured by first spatially scaling the 4D motion field so that the bounding boxes of the regions of interest (e.g., the heart) are the same, and then normalizing the vector field based on the vector magnitude by averaging. The dot product integral of the normalized vectors can be used to represent the similarity between the vector fields.

[0111] This clustering may be performed by the training system TS for a plurality of (eg all) patient data items in the training data set.

[0112] As an alternative to being similarity metric driven, clustering by System TS can be performed by first clustering patient data on one or more selected or predefined context parameters κ (e.g., mortality, complications, appropriately scaled outcomes, etc.) and then training an ML model (e.g., a convolutional network) to match the descriptive vectors m to the predefined clusters.

[0113] After training, during testing or deployment, a (new) coronary artery motion description vector m is established by the data result preprocessor PP for a given patient currently being treated. The predictor module PM then performs a lookup in the pre-clustered data in order to determine which cluster the current patient falls into. The medical information of this cluster can be displayed on a display device DD for clinical staff. These data can be used to determine the (optimal) treatment strategy, make device selection, establish a follow-up plan and / or a risk monitoring plan, etc.

[0114] Optionally, segmented structures (such as the mentioned 4D coronary arteries) that may be extracted as described above or similarly may be used for overlay and roadmap purposes.

[0115] Instead of using the motion descriptors of the coronary arteries as the biodynamic measurement m in this article, other descriptive features may be used as biodynamic measurements instead or in combination with the above. This may be in the same clinical area or in other clinical areas. Examples contemplated herein include stroke, where the descriptive vector m includes cerebral perfusion or oxygen consumption, or for peripheral arterial disease, where the descriptive vector m includes blood perfusion, etc.

[0116] Reference now Figure 6 , which shows a flow chart of a method for predicting treatment options and / or associated outcome data based on a patient's biodynamic measurement m. It should be understood that the steps described below are not necessarily associated with the above-described architecture, but can be understood as a teaching. The optional outcome data, for example, can also be suitably prepared by the processor PP in vector form, and can relate to a time period indicating how long the patient is symptom-free after treatment. Any other medically recognized outcome measure may alternatively be used.

[0117] At step S610 , one or more biokinetic measurements m of a given patient are received.

[0118] In step S620, treatment option data T is calculated based on the measurement results m. This calculation step may be based on a trained machine learning model as described above and applied to the measurement results m. Thus, the method may be used during testing or deployment after training.

[0119] Optionally, the output data includes outcome data o associated with the calculated treatment option T.

[0120] The output data T may contain an indication of a cluster in a predefined space of clusters, which clusters were previously derived from the training data set. Thus, the machine learning model may be one of: a clustering type model such as k-means clustering or other models such as more generally a mixture model or EM type algorithm or PCA etc. The indication indicates one or more clusters into which the patient may fall given the measurement result m. ​​Thus, the patient is considered to benefit from the treatment option pre-associated with the corresponding cluster. In other ML embodiments, a (kernel) support vector machine type algorithm may be used, or a K nearest neighbor algorithm, a decision tree or any other ML classification scheme.

[0121] In ML methods other than clustering, the output data calculated at step S620 may be provided as a vector. The entries of the vector may encode the corresponding treatment option type. A matrix may alternatively be used for representation.

[0122] At step S630, the treatment data and optionally associated outcome data are provided for display, storage or other processing.

[0123] In step S640, the output data T, o may be visualized in a graphical display on a display device. The graphical display may include, for example, Figure 3 A graphical representation of the clusters shown, where each cluster represents a particular treatment option. The result data may be data associated with each such cluster, such as a pop-up text field or otherwise. Details of the treatment options may also be presented textually and / or graphically when the user interacts through a suitable user interface (such as a GUI or other). When the user specifies the indicated cluster, a text or graphical widget may appear (see Figure 2 Specifies that the action can be performed by a pointing tool, such as a computer mouse, a stylus, or via a touch screen.

[0124] Thus, the outcome data predicted herein in step S640 may include an indication of treatment options that may be appropriate for the patient at the point in time when the prediction is made or at some point in time in the feature. Alternatively, it has been determined that the patient requires treatment for certain diseases or conditions, and the cluster represents treatment options for that particular disease or condition. Thus, in the first case, the predicted (one or more) treatment options may be related to a disease condition that was unknown prior to the prediction. Thus, the proposed method and system may be used in a centralized sense to find treatment options for a given disease, or in a more global, exploratory sense to find treatments for a disease / condition that the patient may benefit from at some point in the future, based on the measurement results m for a given patient.

[0125] Reference now Figure 7 A flowchart illustrating a method of explicitly or implicitly training a machine learning model that can be used to Figure 6 method.

[0126] At step S710, prior medical data, including biodynamic measurements m' of the current or other patients, is accessed, acquired or otherwise obtained. The data so accessed, acquired or obtained represents training data. The training data may be accessed, acquired, procured or otherwise obtained by querying historical data sets, such as health data records from prior patients and / or from prior records of a given patient. Preferably, it is ensured that data from a suitably large and diverse group of patients is queried to ensure appropriate representativeness of such training data sets.

[0127] At step S720, the machine learning model is adjusted based on the training data. The adjustment may also be referred to herein as training of the ML model. The adjustment may be implicit, such as in clustering, or may be explicit, such as in neural network type modeling or others. The adjustment may be performed in an iterative manner. The adjustment may be driven by an objective function F() such as a cost function to improve the objective function. The parameters of the model may be adjusted and / or the training data set may be structured.

[0128] In step S730 after training S720, the appropriately trained model is then provided for testing or deployment. In testing, the trained model is provided with some training data that was not used during training but was set aside in advance, so that these data are provided as input to the training model to evaluate the performance of the model.

[0129] In deployment, clinical staff can use the model in daily clinical practice by applying new data ((one or more) biodynamic measurement results) m, rather than extracting from the training / testing data. In an offline embodiment, the new data m can be retrieved from memory, or can be currently acquired by the measurement device M as a given patient's measurement result m in an online embodiment. As described above, the new data m is applied to the trained model to calculate the associated treatment options and / or outcome data for the given patient.

[0130] The components of the system SYS may be implemented as one or more software modules, running on one or more general purpose processing units PU, such as a workstation associated with an imager IA, or on a server computer associated with a group of imagers.

[0131] Alternatively, some or all components of the system SYS may be arranged in hardware, such as a suitably programmed microcontroller or microprocessor, such as an FPGA (field programmable gate array) or a hardwired IC chip, an application specific integrated circuit (ASIC). In yet another embodiment, the system SYS may be implemented partly in software and partly in hardware.

[0132] The different components of the system SYS may be implemented on a single data processing unit PU. Alternatively, some or more components are implemented on different processing units PU, possibly remotely arranged in a distributed architecture and connectable in a suitable communication network, such as in a cloud setting or a client-server setting, etc. For example, the trained module PM may run on a computing system, while the training system TS runs on another. The preprocessor PP may be implemented on yet another computing system. Conversely, any two or all three of the module PM, the preprocessor PP and the training system TS may be implemented on a single computing system PU.

[0133] One or more features described herein may be configured or implemented as circuits encoded in a computer-readable medium and / or a combination thereof. The circuits may include discrete and / or integrated circuits, systems on a chip (SOCs) and combinations thereof, machines, computer systems, processors and memories, computer programs.

[0134] In a further exemplary embodiment of the present invention, a computer program or a computer program element is provided, which is characterized by being adapted to execute the method steps of the method according to one of the preceding embodiments on a suitable system.

[0135] Therefore, the computer program element can be stored on a computer unit, which can also be part of an embodiment of the present invention. The computing unit can be adapted to perform or induce the steps of the above method. In addition, it can be adapted to operate the components of the above device. The computing unit can be adapted to automatically operate and / or execute the user's commands. The computer program can be loaded into the working memory of a data processor. Therefore, the data processor can be equipped to perform the method of the present invention.

[0136] This exemplary embodiment of the invention covers both a computer program that right from the beginning uses the invention and a computer program that by means of an update turns an existing program into a program that uses the invention.

[0137] Furthermore, the computer program element may be able to provide all necessary steps to complete the procedures of an exemplary embodiment of the method as described above.

[0138] According to another exemplary embodiment of the present invention, a computer readable medium, such as a CD-ROM, is proposed, wherein a computer program element is stored on the computer readable medium, the computer program element being described in the previous section.

[0139] The computer program may be stored and / or distributed on a suitable medium, in particular but not necessarily a non-transitory medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems.

[0140] However, the computer program may also be presented over a network like the World Wide Web and may be downloaded from such a network into a working memory of a data processor. According to a further exemplary embodiment of the invention, a medium for making a computer program element available for downloading is provided, the computer program element being arranged to perform a method according to one of the previously described embodiments of the invention.

[0141] It must be noted that embodiments of the present invention are described with reference to different subject matters. Specifically, some embodiments are described with reference to method type claims, while other embodiments are described with reference to device type claims. However, those skilled in the art will appreciate from the above and following descriptions that, unless otherwise stated, any combination of features related to different subject matters, in addition to any combination of features belonging to one type of subject matter, is also considered to be disclosed with the present application. However, all features can be combined to provide synergistic effects, rather than just a simple addition of features.

[0142] Although the present invention has been shown and described in detail in the drawings and the foregoing description, such illustration and description are to be considered illustrative or exemplary rather than restrictive. The present invention is not limited to the disclosed embodiments. Other variations of the disclosed embodiments may be understood and effected by those skilled in the art in practicing the claimed invention by studying the drawings, the disclosure, and the dependent claims.

[0143] In the claims, the word "comprising" does not exclude other elements or steps and the attributive "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items repeatedly cited in the claims. The fact that certain measures are repeatedly cited in mutually different dependent claims does not mean that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope. Such reference signs may include numbers, letters or any alphanumeric combination.

Claims

1. A system (SYS) for predicting patient treatment options, comprising: an input interface (IN) for receiving input data including biokinetic measurements about the patient; a predictor module (PM) configured to process the biokinetic measurements to obtain output data including an indication of a treatment option for the patient, wherein the predictor module (PM) is based on a trained machine learning model (M) previously trained on patient data from a patient cohort, Wherein, the machine learning model is implemented based on a clustering algorithm.

2. The system according to claim 1, wherein: The biokinetic measurements are time series.

3. The system according to claim 1 or 2, wherein: Different treatment options correspond to different clusters (Cj), and wherein the indication comprises an indication of one or more of the clusters.

4. A system according to any one of the preceding claims, wherein: The system comprises a graphic display generator (GDG) configured to generate a graphic display (GD) for display on a display device (DD), the graphic display for visualizing the indication.

5. The system according to claim 4, when claim 4 is dependent on claim 2 or 3, wherein the graphical display (GD) comprises a visualization of the clusters and a graphical indicator of the patient indicating proximity or similarity to the clusters.

6. A system according to any one of the preceding claims, wherein: The biodynamic measurement results include any one or more of the following: coronary vascular movement data, perfusion data, electrocardiogram data, electroencephalogram data, oxygenation data.

7. A system according to any one of the preceding claims, wherein: The biodynamic measurements include image data.

8. A system according to any one of the preceding claims, wherein: The output data includes outcome data for the treatment option.

9. A training system (TS) for training a machine learning model (M) of a system according to any of the preceding claims based on the training data.

10. A computer-implemented method for predicting patient treatment options, comprising: receiving ( S610 ) input data including biokinetic measurements regarding a patient; and processing (S620) the biokinetic measurements to obtain output data including an indication of treatment options for the patient, Therein, the processing is based on a clustering algorithm.

11. A computer-implemented method for training a machine learning model of a system according to any one of the preceding claims 1-8 based on training data.

12. At least one computer program element which, when run by at least one processing unit, is adapted to cause the processing unit to perform the method according to any one of claims 10 or 11.

13. At least one computer-readable medium having stored thereon a program element according to claim 12, or having stored thereon a machine learning model of a system according to any one of claims 1-8.

Images