Tree-based data exploration and data driving scheme

Abstract

The invention relates to tree-based data exploration and data-driven protocols, and discloses a method of providing treatment suggestions to a physician to treat a patient. The method includes determining, from a processor in communication with a patient data repository, a first treatment suggestion based on a combination of selected patient demographic information from the patient data repository suitable for the patient and operating parameters of a plurality of ventricular assist devices (VADs) suitable for treating the patient, the first treatment suggestion has a first survival rate and includes the use of a first VAD. The method then obtains a first signal by using the first VAD on the patient. The method then determines a second treatment suggestion based on the first signal and the first treatment suggestion, the second treatment suggestion having a second survival rate. Then, if the second survival rate is higher than the first survival rate, the method provides a second treatment suggestion to the physician.

Description

[0001] This application is a divisional application of Chinese patent application 201980065045.8, whose application date is October 4, 2019 and whose name is “Tree-based Data Exploration and Data-Driven Solutions”.

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] This application claims the benefit of priority under 35 U.S.C. §119 to U.S. Provisional Patent Application Serial No. 62 / 741,985, filed on October 5, 2018, the contents of which are incorporated herein by reference in their entirety. Background Art

[0004] Acute and chronic cardiovascular conditions reduce quality of life and life expectancy. A variety of treatment modalities have been developed for treating the heart in such conditions, ranging from medications to mechanical devices and transplants. Ventricular assist devices (VADs) such as heart pump systems and catheter systems are commonly used to treat the heart to provide hemodynamic support and facilitate recovery. Some heart pump systems are inserted percutaneously into the heart and can operate in parallel with the native heart to supplement cardiac output. Such heart pump systems include the VaD ... Equipment series.

[0005] Currently, the physician is provided with a selection of treatment plans for the use of a VAD for a patient with a cardiovascular condition by a VAD controller, and the selection is largely based on statistics of previous success using the VAD to treat patients with similar conditions. Conventional data analysis is able to analyze the survival rates associated with the use of a VAD based on a single factor, such as the patient's gender or age. As the type of myocardial condition to which a patient is susceptible develops, a treatment plan that does not properly account for other factors that affect the patient's condition may worsen the patient's condition by treating the patient with a suboptimal VAD. Summary of the invention

[0006] The methods and systems described herein use a tree-based predictive model to provide a physician with a treatment recommendation optimized for a patient. The method begins by determining, using a processor, a first treatment recommendation based on a combination of selected patient demographic information obtained from a patient data repository applicable to the patient and operating parameters of multiple ventricular assist devices (VADs) suitable for treating the patient, the first treatment recommendation having a first survival rate and including using a first VAD. The method then proceeds to obtaining a first signal from the first VAD by treating the patient using the VAD. The first signal is obtained from a controller that communicates with the first VAD. The processor then determines a second treatment recommendation based on the first signal and the first treatment recommendation, the second treatment recommendation having a second survival rate. The processor then determines whether the second survival rate is higher than the first survival rate, and if so, provides the second treatment recommendation to the physician.

[0007] In some embodiments, the method further includes notifying the physician to continue using the first VAD to treat the patient if the second survival rate is not higher than the first survival rate. Therefore, if the processor determines that the second survival rate is equal to or less than the first survival rate, the patient continues to be treated using the first VAD. In certain embodiments, each VAD includes at least one sensor for providing a first signal to the controller. The sensor can be any input transducer configured to convert patient data into an electrical signal. In certain embodiments, the first signal includes information related to the patient's vitals, such as at least one of the following: for example, mean arterial pressure (MAP), left ventricular pressure (LVP), left ventricular end-diastolic pressure (LVEDP), pulmonary artery wedge pressure (PAWP), pulmonary capillary wedge pressure (PCWP), pulmonary artery occlusion pressure (PAOP).

[0008] In some embodiments, if the second survival rate is determined to be higher than the first survival rate, the method further comprises treating the patient using a second treatment recommendation. The second treatment recommendation may include treating the patient using at least one of the following: the first VAD, the second VAD, and no VAD. In some embodiments, the second treatment recommendation may be based on the first treatment recommendation, wherein an additional VAD may be recommended for use with the first VAD. In some embodiments, the VAD includes at least one of the following: pumps, extracorporeal membrane oxygenation (ECMO) pumps, balloon pumps, and Swan-Ganz catheters. Pumps include any of the following: Impella Pump, Impella Pump, Impella Pump, Impella Pump and Impella The aforementioned method is used to treat patients with cardiogenic shock.

[0009] In some embodiments, the first treatment recommendation is determined by a predictive model executed by a processor. In certain embodiments, the predictive model is based on a machine learning algorithm comprising any of: bagging and random forest algorithms, logistic regression algorithms, classification decision tree algorithms, deep learning algorithms, naive Bayes algorithms, and support vector machine algorithms. In some embodiments, the predictive model uses patient demographic information applicable to the patient in its calculations. Such demographic information includes gender, age, region, duration of support, indication for use, and insertion site. By using patient-specific demographic information, the final treatment recommendation provided to the patient will be more appropriate for each patient, thereby improving treatment outcomes.

[0010] In certain embodiments, the processor may display the survival rate for each available VAD; and find the VAD with the highest survival rate. Additionally, the processor may display the combination of selected patient demographic information used to determine the survival rate using a branching tree representation. Such a feature combination tree provides the physician with a visualization of the features that have an impact on the recommended VAD (and associated survival rate) for treating the patient. With respect to the present disclosure, the survival rate includes the probability of survival of a patient belonging to the combination of selected patient demographic information when treated with the VAD.

[0011] In some embodiments, the patient data repository comprises an Acute Myocardial Infarction Cardiogenic Shock (AMICS) database or a High-Risk Percutaneous Coronary Intervention (High-Risk PCI) database.

[0012] According to a second embodiment of the present disclosure, the method and system obtain data from a patient repository, wherein the data is stored in the repository according to patient demographic information. The method obtains data from the patient repository using a processor. The processor then determines at least one ventricular assist device (VAD) suitable for treating the patient. The processor then uses a predictive model to determine the survival rate of each suitable VAD based on the data from the patient data repository for a combination of selected patient demographic information applicable to the patient. The processor then provides a controller with a recommended first VAD associated with the highest survival rate. The physician then uses the recommended first VAD to treat the patient.

[0013] In some embodiments, the method further provides the physician with survival rates for each suitable VAD for all combinations of selected patient demographic information applicable to the patient. In certain embodiments, the processor also provides the physician with survival rates without the use of the VAD for each combination of patient demographic information applicable to the patient. In certain embodiments, the first VAD may include at least one of: pumps, extracorporeal membrane oxygenation (ECMO) pumps, balloon pumps, and Swan-Ganz catheters. Pumps can include any of the following: Impella Pump, Impella Pump, Impella Pump, Impella Pump and Impella Pump. Patient demographic information may include: age, sex, region, year of implantation, support device, duration of support, insertion site, and ejection fraction.

[0014] In some embodiments, the prediction model uses a machine learning algorithm to determine the survival rate. The machine learning algorithm may include any of the following: bagging and random forest algorithms, logistic regression algorithms, classification decision tree algorithms, deep learning algorithms, naive Bayes algorithms, and support vector machine algorithms. In certain embodiments, the combination of selected patient demographic information follows a tree model. The tree model may have any of the following orders: two, three, four, five, and six.

[0015] In some embodiments, the method may further include displaying the survival rate for each available VAD and identifying the VAD with the highest survival rate. In certain embodiments, the method may further include displaying the combination of selected patient demographic information used to determine the survival rate using a branching tree representation. The survival rate may include the probability of survival of patients belonging to the combination of selected patient demographic information when treated with the VAD. In some embodiments, the patient data repository may include an Acute Myocardial Infarction Cardiogenic Shock (AMICS) database or a high-risk percutaneous coronary intervention (high-risk PCI) database.

[0016] According to a third embodiment of the present disclosure, a system for providing treatment recommendations to physicians to treat patients is provided. The system includes at least one ventricular assist device (VAD) including a sensor. The system further includes a processor in communication with the VAD, the processor being configured to communicate with an acute myocardial infarction cardiogenic shock (AMICS) repository or a high-risk percutaneous coronary intervention (high-risk PCI) repository. In addition, the system includes a controller in communication with the VAD and the processor, the controller being configured to execute the method described in any one of the aforementioned embodiments.

[0017] According to a fourth embodiment of the present disclosure, a system for providing treatment suggestions to a physician to treat a patient is provided. The system includes a processor and a controller, and the controller is configured to execute the method according to any one of the aforementioned embodiments.

[0018] According to a fifth embodiment of the present disclosure, there is provided a computer program comprising computer executable instructions, which, when executed by a computing device comprising a processor and a controller, causes the computing device to perform the method according to any one of the preceding embodiments.

[0019] According to a sixth embodiment of the present disclosure, a non-transitory computer-readable storage medium having computer-readable code stored thereon is provided. When the computer-readable code is executed by a computing device including a processor and a controller, the computing device performs the method according to any one of the aforementioned embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The foregoing and other objects and advantages will become apparent upon consideration of the following detailed description taken in conjunction with the accompanying drawings, in which like reference numerals refer to like parts throughout, and in which:

[0021] Figure 1 An illustrative system according to an embodiment of the present disclosure is shown;

[0022] Figure 2 An illustrative flow chart of a method for providing treatment recommendations to a physician to treat a patient according to an embodiment of the present disclosure is shown;

[0023] Figure 3A-3C An illustrative dendrogram with patient demographic factors contributing to survival with VAD use in patients is shown;

[0024] Figure 4A The accuracy of the prediction model based on the order / depth of the dendrogram of the method disclosed in the present invention is shown;

[0025] Figure 4B shows the output of the prediction model with an optimized tree depth; and

[0026] Figure 5 An illustrative flow chart of a method for providing a second treatment recommendation to a physician to treat a patient according to another embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0027] In order to provide an overall understanding of the methods and systems described herein, certain illustrative embodiments will be described. Although the embodiments and features described herein are described specifically for use in conjunction with survivability of ventricular assist devices, it should be understood that all of the components and other features outlined below may be combined with one another in any suitable manner and may be adapted and applied to other types of medical therapies that have survivability associated therewith.

[0028] The systems and methods described herein use predictive modeling to determine the best treatment recommendations for patients with cardiogenic shock. The treatment recommendations may include the use of a single ventricular assist device (VAD) or multiple VADs in combination with each other. The predictive model extracts data and statistical information from archived ventricular assist routines performed in the past. Such patient data can be stored in a patient data repository, such as, for example, an Acute Myocardial Infarction Cardiogenic Shock (AMICS) database or a high-risk percutaneous coronary intervention (high-risk PCI) database. The system and method use a machine learning algorithm to predict survival when a patient is treated with a VAD. In order to further customize the predictive model, patient demographic information and operating parameters of the VAD are also included in the model using a selected tree-based feature combination. Here, the physician is able to combine any number of patient demographic information and / or device features to obtain treatment recommendations with realistic survival rates.

[0029] Additionally, real-time patient data from patients using VADs can also be fed into the predictive model to further optimize treatment recommendations. Such real-time patient data can include, but is not limited to, mean arterial pressure (MAP), left ventricular pressure (LVP), left ventricular end-diastolic pressure (LVEDP), pulmonary artery wedge pressure (PAWP), pulmonary capillary wedge pressure (PCWP), pulmonary artery occlusion pressure (PAOP). VADs can include, but are not limited to: pumps, extracorporeal membrane oxygenation (ECMO) pumps, balloon pumps, and Swan-Ganz catheters. Pumps can include: Impella Pump, Impella Pump, Impella Pump and Impella Pumps, both manufactured by Abiomed, Inc. of Danvers, MA.

[0030] Figure 1 A block diagram of a system 100 for providing a first treatment recommendation 110 to a physician to treat a patient 120 is shown. The first treatment recommendation 110 includes instructions to the physician to use the most appropriate VAD for the condition of the patient 120. The VAD may include, but is not limited to: pumps, extracorporeal membrane oxygenation (ECMO) pumps, balloon pumps, and Swan-Ganz catheters. Pumps can include Impella Pump, Impella Pump, Impella Pump, Impella Pump and Impella pumps, both of which are manufactured by Abiomed, Inc. of Danvers, MA. Using the predictive model, a first treatment recommendation 110 is determined based on selected demographic information applicable to the patient 120 and the operating parameters of all VADs appropriate for the patient's condition. Here, survival rates for all VADs are then determined based on the selected demographic information for the patient 120 and the operating parameters of the VADs. The VAD associated with the highest survival rate is then recommended to the physician. This indication may be made via a display of the computing unit 130. Figure 1 , the first recommendation 110 includes instructing the physician to treat the patient 120 using a first VAD having a first survival rate (SR1).

[0031] The system 100 also includes a computing device 130, such as a laptop computer, in communication with a patient data repository 140. Figure 1 Only the processor 135 of the computing device 130 is shown. However, it will be understood that the computing device 130 also includes other components commonly associated with a computing device, such as, for example, volatile memory (e.g., random access memory RAM), non-volatile memory (e.g., read-only memory ROM), a display, and a connection bus that enables communication between these components, all of which are included in the present disclosure.

[0032] The computing device 130 includes a processor 135 capable of performing operations on data using a predictive model. The computing device 130 communicates with a patient data repository 140, which includes patient data obtained from various medical institutions. According to certain embodiments of the present disclosure, the patient data repository 140 may include an acute myocardial infarction cardiogenic shock (AMICS) database or a high-risk PCI database compiled and maintained by a CRM (such as Salesforce.com). The patient data repository 140 stores data from, for example, acute myocardial infarction (AMI) patients, high-risk PCI patients, and cardiogenic shock patients. The patient data includes patient demographic information, such as, for example, gender, age, and region. The patient data also includes data from previous treatments, such as, for example, duration of support, indications for use, insertion sites, treatment devices, and ejection fractions. Table 1 shows exemplary AMICS data. In some cases, the patient data repository 140 also includes a database 145 of available VADs and their associated operating parameters.

[0033] Processor 135 of processing unit 130 is also in communication with controller 150, which controls the operation of any VAD used to treat patient 120. Each VAD includes a sensor that collects data from the patient while the VAD is in use and transmits the data as a signal (such as Figure 1The first VAD of the first treatment recommendation 110 is connected to the controller 150 and transmits the first signal (SIG1) to the controller 150. Such data may include, but is not limited to, mean arterial pressure (MAP), left ventricular pressure (LVP), left ventricular end-diastolic pressure (LVEDP), pulmonary artery wedge pressure (PAWP), pulmonary capillary wedge pressure (PCWP), and pulmonary artery occlusion pressure (PAOP). The first VAD of the first treatment recommendation 110 is connected to the controller 150 and transmits the first signal (SIG1) to the controller 150. In some embodiments, the controller 150 includes an Automatic Controller (Abiomed, Inc. of Danvers, MA) Controller 150 may be housed in a servicing hub 160, which may include other components to ensure that the respective VADs connected thereto are in operational order. In certain embodiments, processor 130, processing unit 135, controller 150, and servicing hub 160 may be housed in a workstation 170. Workstation 170 may further include a display (not shown) to indicate treatment recommendations to a physician.

[0034]

[0035]

[0036]

[0037] Table 1 Exemplary AMICS data

[0038] According to certain embodiments of the present disclosure, after the physician uses the first VAD to treat the patient 120, the controller 150 may additionally provide the physician with an indication of a second treatment recommendation. Here, the controller 150 receives a first signal SIG1 transmitted from the first VAD. The controller 150 then determines a second treatment recommendation, which includes using at least a second VAD, operating parameters of the first VAD, and operating parameters of all suitable VADs in the database 145 using a comparison model based on SIG1. The survival rates of all suitable VADs are also determined, and the VAD with the highest survival rate is selected as the second VAD. The controller 150 also determines a second survival rate (SR2) of the second VAD from the AMICS database 140.

[0039] The controller 150 then compares the first survival rate SR1 with the second survival rate SR2. If the second survival rate SR2 is higher than the first survival rate SR1, i.e., SR2>SR1, a second suggestion is provided to the physician via a display on the workstation 170. The second suggestion may include instructing the physician to use a second VAD instead of the first VAD to treat the patient 120. The second suggestion may also include instructing the physician to use a combination of VADs selected from the database 145. The VAD used in the combination may include a VAD other than the first VAD, or at least one second VAD in addition to the first VAD. In addition, the second suggestion may include not using a VAD at all, i.e., the second suggestion may instruct the physician to stop treating the patient with any VAD. If the second survival rate SR2 is not higher than the first survival rate SR1, i.e., SR2≤SR1, the controller 150 indicates to the physician that the first treatment suggestion should not be changed.

[0040] Figure 2 1 is a flowchart of a method 200 for providing a physician with a first treatment recommendation for treating a patient with cardiogenic shock according to an embodiment of the present disclosure. The method 200 is based on the previously described Figure 1 The method begins at step 210, where the processor 135 of the computing unit 130 accesses a patient data repository 140. In some embodiments, the repository 140 may include an AMICS or high-risk PCI database, which includes a VAD database 145. Then, in step 220, the processor 135 determines a VAD that is suitable for treating the patient. Such suitability may be based on the clinical indication of the patient and the operating parameters of the VAD obtained from the patient data repository 140. Therefore, in step 220, the processor 135 determines a candidate list of suitable VADs from the patient data repository 140 for treating the patient.

[0041] In step 230 of the method, the processor 135 uses a prediction model to determine the survival rate of each VAD in the candidate list from step 220. The prediction model is based on a machine learning algorithm, which in turn includes but is not limited to bagging and random forest algorithms, logistic regression algorithms, classification decision tree algorithms, deep learning algorithms, naive Bayes algorithms, and support vector machine algorithms, the details of which are omitted in this disclosure for brevity.

[0042] For example, the logistic regression algorithm is based on an equation that represents a predictive model using coefficients learned from training data. The representation of the model can be stored in the memory of the computing unit 130 as a series of coefficients, each coefficient corresponding to a weight indicating the relative importance of a particular feature (e.g., a particular patient demographic information), and can be used to calculate a probability, which is then converted into a patient's survival rate. For any number of features (Feature_α, Feature_β, Feature_γ) and associated coefficients (α, β, γ), the probability can be calculated as (1+exp(-x)) -1 , where x is equal to α*Feature_α+β*Feature_β+γ*Feature_γ+….

[0043] In another example, a decision tree algorithm uses a decision tree as a predictive model to draw conclusions about the target value of an item from observations of the item. Tree depth can be a hyperparameter in decision tree learning. A hyperparameter is a value that cannot be estimated from the data used in the model. Hyperparameters are often used to help estimate model parameters and can be adjusted for a given predictive modeling problem. Accuracy can be used as a performance metric for a predictive model. By adjusting hyperparameters (such as tree depth) to determine the maximum accuracy of a decision tree, the system can provide an optimized machine learning model and therefore better provide predictions (such as survival rates). Receiver operating characteristics (ROC) and area under the curve (AUC) can also be used as indicators for comparing prediction algorithms.

[0044] In step 240 of the method, the processor 135 compares the survival rates obtained for each VAD and determines the VAD with the highest survival rate. In step 250, the processor 135 provides a first treatment recommendation to the physician via an indication on a display connected to the processor to treat the patient using the VAD with the highest survival rate.

[0045] Using the tree-based data-driven approach described above, Figure 2 The first treatment recommendation is customized to the needs of the patient by providing a first treatment recommendation in the method 200. By finely tailoring the treatment to the patient's demographic information, the most effective treatment plan with the best survival rate can be provided to the patient (e.g., based on historical analysis of the data in the data repository 140), thereby ensuring effective treatment of the cardiogenic shock patient.

[0046] Figure 3AA patient demographic information / feature combination tree according to an embodiment of the present disclosure is shown. Tree 300 shows how various types of features and treatments can be combined to determine the treatment with the highest predicted survival rate. Features are determined based on the attributes of data 310 stored in the data repository 140. As can be seen from Table 1, the data stored in the repository 140 can have various attributes, such as gender, age, ejection fraction, the type of catheter used (e.g., Swan-Ganz), and the type of pump used (e.g., ECMO). The attributes of the data related to the patient's demographic information are used as state-level features in the combination tree 300, such as features 322-325. State-level features are features that are patient-related and cannot be changed, that is, state-level features do not have adjustable values. Examples of state-level features include, but are not limited to, gender, age, and ejection fraction. The repository 140 also stores data related to the types of available VADs. The features of such VADs are used as protocol-level features in the combination tree 300, such as features 326-327. Protocol-level features are physician-selectable VAD options that contribute to the controlled operation of the VAD. Regimen layer features have associated values ​​that may be specified by the physician.Examples of regimen layer features include, but are not limited to, rotor speed and flow rate.

[0047] The selected state-level features and the selected regimen-level features extract patient data 310 from the data repository 140 so that the extracted data can be used in a prediction model to determine predicted survival rates 328-329 for each corresponding combination of the selected features 322-327 and provide the predicted survival rates 328-329 to the physician. The predicted survival rates 328-329 are provided in the prediction layer in the tree 300. In some embodiments, the prediction model can identify the combination of features 322-327 that gives the highest predicted survival rate. The combination of features 322-327 associated with the highest predicted survival rate is provided to the physician as a treatment recommendation. The treatment recommendation can be provided to the physician on a monitor. The predicted survival rates 328-329 can be provided to the physician on a monitor such as a Figure 3A The feature combination tree in tree 310 presents the physician with combinations of features that make up a treatment recommendation.

[0048] It will be appreciated that cardiac assist technology evolves over time. Additionally, increased exposure to and use of VADs by physicians may improve their impact on patient conditions (e.g., once a physician is better trained in the use of a VAD, the impact of the use of the VAD on patients with a particular condition will be seen in the patient data - e.g., the ejection fraction of a particular VAD may increase). In accordance with an embodiment of the present disclosure, to cater for such factors, the predictive model may specify a date range when extracting patient data 310 from the data repository 140. In such a case, the predictive model effectively weights the data and only uses patient data 310 from the repository 140 that falls within the specified date range. Although such data weighting is illustrated above by date range, other factors may be considered when weighting the extracted data 310.

[0049] Figure 3B An exemplary feature combination tree 360 ​​of various types of patient demographic data used in a predictive model according to an embodiment of the present disclosure is shown. As previously described, patient demographic data is stored in the patient data repository 140 and may include, for example, gender, age, and region. The tree 360 ​​shows combinations of gender 362 and age 363, 364 of a patient. These combinations are used for specific types of VADs (e.g., 2.5 pumps). Based on the data extracted from the patient data repository 140, feature 362 has the values ​​"male" and "female", and features 363, 364 have the values ​​"50-59" and "60-69". Figure 3B The tree-based data combination shown, using data from the patient data repository 140 for selected patient demographics and the machine learning algorithm (described above), the survival rates 365-368 for each combination of the tree 360 ​​are shown in FIG. Figure 3B shown. Figure 3B Shows that a specific VAD has been implanted (e.g. 2.5 pumps) male patients aged 60-69 years had the highest survival rate of 84.86%. Figure 3B The data on survival rates in are based on 1,071+1,697+1,307+2,317=6,392 records in patient data repository 140 having features 362-364. Figure 3B The example in illustrates the effect of a combination of features of patient data determined by a predictive model on survival.

[0050] exist Figure 3CAnother example of a tree-based data-driven approach of the present disclosure is shown in the feature combination tree 370 shown. Tree 370 shows a combination of three types of patient data: gender 372, ejection fraction 374, and feasibility of hemodynamic monitoring 376. The feasibility of hemodynamic monitoring depends on the type of VAD used. For example, it is known that Swann-Ganz catheters have pressure sensors that can be used to facilitate such hemodynamic monitoring. Figure 3C The selected patient demographics for the combination were: male sex, ejection fraction less than 30%, and hemodynamic monitoring. Figure 3C The tree-based data combination shown, using the data in the patient data repository 140 for the selected patient demographics and the prediction model (described above), the survival rate 378-379 for this particular combination of patient demographic information is as shown in FIG. Figure 3C As shown. For male patients with an ejection fraction of less than 30%, the highest survival rate using a VAD capable of hemodynamic monitoring is 58.43%. Therefore, the treatment recommendation to the physician is to use a VAD capable of hemodynamic monitoring (such as a Swan-Ganz catheter) for male patients with an ejection fraction of less than 30%. For completeness, branch 378 in tree 370 indicates the survival rate when a VAD without hemodynamic monitoring is used for male patients with an ejection fraction of less than 30%. Figure 3C The data of survival rates in is based on 584+421=1,005 records having features 372, 374, and 376 in patient data repository 140. Graphical representations of feature combination trees 360 and 370 may also be displayed to a physician, for example, on a monitor along with the first treatment recommendation and / or the second treatment recommendation.

[0051] Figure 4A A data graph 400 is shown for optimizing the accuracy of a prediction model using a decision tree algorithm. As previously described, the decision tree algorithm can use a hyperparameter such as tree depth. By adjusting a hyperparameter (such as tree depth) to determine the maximum accuracy of the decision tree algorithm, the system can provide an optimized machine learning model and thus better provide predictions (such as survival rates). Figure 4A As shown in Figure 2, when we change the hyperparameter (tree depth) to a larger value, the decision tree algorithm is able to capture all the noise in the training data. Figure 4A As can be seen in the line graph 410 in , this results in a high training score. However, at such a large tree depth, the model overfits the data and does not generalize well enough. As a result, as shown by Figure 4A As can be seen in the line graph 420 in FIG. 4 , the cross validation score deteriorates. For a decision tree algorithm, if the tree is too shallow, for example, if the tree depth is 2 or 3, the prediction model is too simple and cannot make any correct predictions, such as Figure 4AAs shown in , both the training score and the cross-validation score are low. Therefore, the optimal tree depth for the decision tree algorithm is the depth that maximizes the cross-validation score and the training score. Figure 4A As shown, this occurs at the peak of the cross validation curve (point 430) at a tree depth of six.

[0052] In machine learning, a false positive is the result of a machine indication that incorrectly indicates the presence of a condition or attribute. Similarly, a false negative is the result of a machine indication that incorrectly identifies the absence of a condition or attribute. Ideally, a predictive model should predict survival with VAD that matches the actual survival when the VAD is used on the patient. However, as with most things in machine learning, there are no perfect models in real life, so there is a trade-off between false positives and false negatives, and vice versa, when evaluating machine learning models. Figure 4B The ROC curve for a tree depth of 6 is shown in Figure 4. ROC curve 450 scans from 0 to 100% false positive rate and checks what the true positive rate given by the model is. Line graph 460 is ideal, with an area under the curve AUC of 100%, while line graph 470 is based on a random model with an AUC of 50%. Any reasonable prediction model should stay between line graphs 460 and 470. Figure 4A For the decision tree algorithm discussed, the true positive rate is shown by line graph 480, and the AUC is 87.4%, indicating good predictive power of the prediction model when the decision tree algorithm is used. For clarity, the decision tree algorithm is one of many machine learning algorithms that can be used as a prediction model according to the present disclosure. Figure 3B and Figure 3C The feature combination tree in is separate from the predictive model and provides the physician with a graphical representation of how the selected features and / or patient demographic information are combined in the decision model.

[0053] Figure 5 A flowchart of a method 500 for providing a second treatment recommendation to a physician to treat a cardiogenic shock patient according to another embodiment of the present disclosure is shown. The method 500 is based on the previously described Figure 1 The features of the described system 100 are as follows. The method 500 works off Figure 2 Step 250 of method 200 in which a first treatment recommendation is provided to a physician, the first treatment recommendation comprising using a first VAD having a first survival rate (SR1). Method 500 begins at step 510, in which the first treatment recommendation is determined by processor 135. As previously described with respect to Figure 2As described, the processor 135 accesses the patient data repository 140 and determines a VAD that is suitable for treating the patient. For example, such suitability can be based on the patient's clinical indications and the operating parameters of the VAD obtained from the patient data repository 140. The processor 135 determines a first candidate list of suitable VADs from the patient data repository 140 for treating the patient, and uses a prediction model to determine the survival rate of each VAD in the first candidate list. As previously described, the prediction model is based on a machine learning algorithm, which in turn includes but is not limited to bagging and random forest algorithms, logistic regression algorithms, classification tree algorithms, deep learning algorithms, naive Bayes algorithms, and support vector machine algorithms. The VAD with the highest survival rate is used for the first treatment recommendation, and the VAD is referred to as the first VAD with a first survival rate (SR1).

[0054] The physician is informed of the first treatment recommendation via a display unit connected to the processor 135, and the physician treats the patient using the first VAD (step 520). A feature combination tree (e.g., trees 300 and 350) may also be displayed. Each VAD includes a sensor that collects data from the patient when the VAD is used (step 530) and transmits the data as a signal (SIG1) to the controller 150 and the processor 135. In some embodiments of the present disclosure, during the period of treatment of the patient with the first VAD, a collection of patient data or patient vitals, referred to as a patient vitals check, is performed at predetermined time intervals. As previously described, such data may include, but is not limited to, mean arterial pressure (MAP), left ventricular pressure (LVP), left ventricular end-diastolic pressure (LVEDP), pulmonary artery wedge pressure (PAWP), pulmonary capillary wedge pressure (PCWP), and pulmonary artery occlusion pressure (PAOP).

[0055] In step 540, the processor 135 determines a second candidate list of VADs suitable for treating the patient from the patient data repository 140 based on the patient data contained in SIG1. It will be understood that as the patient is treated with the first VAD, the patient's vitality may change, and therefore the patient data in SIG1 may be different from the patient's clinical indication used when determining the first treatment recommendation. Therefore, the VADs in the second candidate list may be different from the VADs in the first candidate list. As with the VADs in the first candidate list, the processor 135 uses a predictive model to determine the survival rate of each VAD in the second candidate list. The predictive model is based on a machine learning algorithm, which may include but is not limited to bagging and random forest algorithms, logistic regression algorithms, classification and regression tree algorithms, deep learning algorithms, decision tree algorithms, naive Bayes algorithms, support vector machine algorithms, and vector quantization algorithms.

[0056] The VAD with the highest survival rate, referred to as the second VAD, is determined (step 540), and in step 550, the processor 135 compares its survival rate (SR2) with the survival rate of the first VAD (SR1). If SR2>SR1, the second VAD is used in a second treatment recommendation to the physician (step 560). The second treatment recommendation is communicated to the physician via a display unit connected to the processor 135. If SR2≤SR1, the second treatment recommendation is not provided to the physician, but the physician is instructed (via the monitor) to continue using the first treatment recommendation and continue to treat the patient with the first VAD as described in step 520 until the next patient vitals check. After providing the second treatment recommendation to the physician, the method 500 may continue to perform patient vitals checks to further refine the treatment process, such as Figure 5 shown.

[0057] Using the tree-based data-driven approach described above, Figure 5 The second treatment recommendation provided in the method 500 further customizes the first treatment recommendation to the progress of the patient being treated. This further customization ensures that the treatment recommendation with the highest survival rate is provided to the physician for further treatment of the patient. By fine-tuning the treatment to the progress of the patient, the treatment of the cardiogenic shock patient will be more effective.

[0058] In some embodiments, feature combination trees are also displayed on a monitor attached to the processor running the decision model. These feature combination trees are similar to Figure 3B and Figure 3C Such a feature combination tree provides the physician with a visualization of the features that have an impact on the suggested VAD for treating the patient.

[0059] In some embodiments, the second treatment recommendation may include the use of a single second VAD or multiple second VADs in combination with one another. For example, the first treatment recommendation may include the use of a balloon pump, while the second treatment recommendation may include the use of an ECMO pump in combination with a Swann-Ganz catheter. As another example, the first treatment recommendation may include the use of an Impella pump, and the second treatment recommendation may include continued use of Impella in combination with a balloon pump pump. In this case, the second candidate list will generate the two VADs with the highest survival rates. In certain embodiments of the present disclosure, the processor may be configured to determine the top n VADs with higher survival rates, where n ≥ 1. Then, Figure 5In step 550 of method 500, the survival rates of the n VADs are compared with the survival rate of the first VAD. According to some embodiments of the present disclosure, the second treatment recommendation may be to stop using any VAD, such as in the "No" option 391 in FIG. 3, whereby the "No VAD" option of the first VAD gives the highest survival rate.

[0060] In connection with the present disclosure, computer-readable media may include computer-readable storage media, which may be any tangible media or component that can contain or store instructions for use by or in connection with an instruction execution system, apparatus, or device (such as a computer as defined above) for performing any of the methods described herein.

[0061] According to various embodiments of the present disclosure, the computer program may be implemented in a computer program product including a tangible computer-readable medium carrying a computer program code embodied therein, which may be used together with a processor to implement the above functions or methods.

[0062] References to "computer-readable storage medium", "computer program product", "tangibly embodied computer program", etc., or "processor" or "processing circuit", etc. should be understood to include not only computers with different architectures (such as single processor architecture / multi-processor architecture and sequencer / parallel architecture), but also special-purpose circuits (such as field programmable gate arrays FPGA, special-purpose circuits ASIC, signal processing devices and other devices). References to computer programs, instructions, codes, etc. should be understood to mean software for programmable processor firmware, such as the programmable content of a hardware device, such as instructions for a processor, or configured settings or configuration settings for a fixed function device, gate array, programmable logic device, etc.

[0063] As an example and not limitation, such "computer-readable storage medium" may represent a non-transitory computer-readable storage medium, which may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage devices, flash memory, or any other medium that can be used to store the desired program code in the form of instructions or data structures and can be accessed by a computer. Moreover, any connection is appropriately referred to as a "computer-readable medium". For example, if a coaxial cable, optical fiber cable, twisted pair, digital subscriber line (DSL) or wireless technology (such as infrared, radio and microwave) is used to transmit instructions from a website, server or other remote source, the definition of the medium includes coaxial cable, optical fiber cable, twisted pair, DSL or wireless technology (such as infrared, radio and microwave). However, it should be understood that "computer-readable storage medium" and data storage medium do not include connections, carriers, signals or other transient media, but are directed to non-transient tangible storage media. Disks and optical disks used herein include compact disks (CDs), laser disks, optical disks, digital versatile disks (DVDs), floppy disks and blue-ray disks, where disks usually reproduce data magnetically, while optical disks reproduce data optically with lasers. Combinations of the above should also be included within the scope of “computer-readable media”.

[0064] Instructions may be executed by one or more processors, such as one or more digital signal processors (DSPs), general purpose microprocessors, application specific integrated circuits (ASICs), field programmable logic arrays (FPGAs), or other equivalent integrated or discrete logic circuits. Thus, the term "processor" as used herein may refer to any of the foregoing structures or any other structure suitable for implementing the techniques described herein. Additionally, in some aspects, the functionality described herein may be provided within dedicated hardware and / or software modules. Moreover, these techniques may be fully implemented in one or more circuits or logic elements.

[0065] If desired, the different steps discussed herein may be performed in different orders and / or simultaneously with each other. In addition, if desired, one or more of the above steps may be optional or may be combined.

[0066] The foregoing is merely illustrative of the principles of the present disclosure, and the apparatus may be practiced in other ways than the described embodiments presented for purposes of illustration and not limitation. It should be understood that the methods disclosed herein, although shown for use in an automated ventricular assist system, may be applied to systems to be used in other automated medical systems.

[0067] After reviewing this disclosure, variations and modifications will occur to those skilled in the art. The disclosed features may be implemented in any combination and sub-combination (including multiple dependent combinations and sub-combinations) with one or more other features described herein. The various features described or illustrated above (including any components thereof) may be combined or integrated in other systems. Moreover, certain features may be omitted or not implemented.

[0068] Examples of changes, substitutions, and alterations are ascertainable by those skilled in the art, and changes may be made without departing from the scope of the information disclosed herein. All references cited herein are incorporated herein by reference in their entirety and become a part of this application.

Claims

1. A method for providing treatment recommendations to a physician for treating a patient, the method comprising: determining, from a processor in communication with a patient data repository, a first treatment recommendation based on a combination of selected patient demographic information from the patient data repository applicable to the patient and operating parameters of a plurality of ventricular assist devices (VADs) appropriate for treating the patient, the first treatment recommendation having a first survival rate and comprising use of a first VAD; treating the patient using the first VAD; obtaining a first signal from a controller; determining, by the processor, a second treatment recommendation based on the first signal and the first treatment recommendation, the second treatment recommendation having a second survival rate; as well as The processor provides the second treatment recommendation to the physician if the second survival rate is higher than the first survival rate.

2. The method according to claim 1, further comprising: If the second survival rate is not higher than the first survival rate, the physician is notified to continue using the first VAD to treat the patient.

3. The method according to any of the preceding claims, wherein each VAD comprises at least one sensor for providing the first signal to the controller.

4. The method according to any of the preceding claims, wherein the first signal comprises information related to the patient's vitality.

5. The method according to any one of the preceding claims, wherein the first signal comprises at least one of the following: mean arterial pressure MAP, left ventricular pressure LVP, left ventricular end-diastolic pressure LVEDP, pulmonary artery wedge pressure PAWP, pulmonary capillary wedge pressure PCWP, pulmonary artery occlusion pressure PAOP.

6. The method according to any one of the preceding claims, further comprising: The patient is treated with the second treatment recommendation.

7. A method according to any one of the preceding claims, wherein: The VAD includes at least one of the following: Pump, extracorporeal membrane oxygenation pump, ECMO pump, balloon pump and Swan-Ganz catheter.

8. The method according to claim 7, wherein the Pumps include any of the following: Impella Pump, Impella Pump, Impella Pump, Impella Pump and Impella Pump.

9. The method of claim 1, wherein the second treatment recommendation comprises: The first VAD from the first treatment recommendation continues to be used in addition to the second VAD.

10. The method of any one of claims 1 to 8, wherein the second treatment recommendation comprises treating the patient using at least one of: the first VAD, a second VAD, and no VAD.

Images