阅读说明：本技术 使用机器学习模型的左心室容积和心脏输出估计 (Left ventricular volume and cardiac output estimation using machine learning models ) 是由 A·埃尔卡特吉 Q·谭 E·克罗克 R·王 于 2020-01-15 设计创作，主要内容包括：本发明公开了用于创建和使用神经网络模型来估计患者的心脏参数,以及使用所估计的参数来提供血泵支持以改善患者心脏性能和心脏健康的方法和系统。特定适配包括调整血泵参数,以及确定是否以及如何增加或减少支持,或者使患者完全脱离血泵。该模型基于对来自第一患者集的数据的神经网络处理来创建,并且与原位测量的心脏参数相比包括所测量的血流动力学参数和泵参数,例如通过米勒导管(在动物中)或印加导管(在人类中)测量的左心室容积。在基于第一患者集开发模型之后,将模型应用于第二集中的患者以估计心脏参数而不使用附加导管或直接测量。(Methods and systems for creating and using neural network models to estimate cardiac parameters of a patient, and using the estimated parameters to provide blood pump support to improve patient cardiac performance and cardiac health are disclosed. The specific adaptation includes adjusting blood pump parameters and determining whether and how to increase or decrease support or to completely disengage the patient from the blood pump. The model is created based on neural network processing of data from the first patient set and comprises measured hemodynamic and pump parameters, such as left ventricular volume measured by a miller catheter (in animals) or a printed catheter (in humans), compared to in situ measured cardiac parameters. After the model is developed based on the first set of patients, the model is applied to the patients in the second set to estimate cardiac parameters without using additional catheters or direct measurements.)

1. A method of estimating a cardiac parameter of a patient, the method comprising:

operating a blood pump within each patient of a first set of patients, the blood pump having at least one measurable pump parameter;

measuring at least one hemodynamic parameter and the at least one measurable pump parameter for each patient in the first set of patients to obtain first hemodynamic parameter measurements and first pump parameter measurements,

constructing a model of a cardiac parameter based on a relationship between the at least one first hemodynamic parameter and the at least one measurable pump parameter for the first set of patients,

operating a second blood pump in a second patient of a second patient set; and

applying the model to the second patient by:

measuring the at least one measurable pump parameter in the second patient to obtain a second pump parameter measurement;

measuring the at least one first hemodynamic parameter in the second patient to obtain a second hemodynamic parameter measurement; and

estimating a cardiac parameter of the second patient, wherein the cardiac parameter of the second patient is output by the model based on the second pump parameter measurement and the second hemodynamic parameter measurement.

2. The method of claim 1, wherein measuring at least one hemodynamic parameter comprises measuring the aortic pressure.

3. The method of claim 1 or 2, further comprising determining the aortic pressure at a pressure sensor located on the blood pump.

4. The method of any of claims 1-3, wherein measuring the at least one measurable pump parameter comprises measuring pump flow.

5. The method according to any of claims 1-4, further comprising determining the estimated cardiac parameter based on the at least one hemodynamic parameter and at least one measurable pump parameter at least one point in time.

6. The method of any of claims 1-5, further comprising inserting a sensing catheter separate from the blood pump into each patient within the first set of patients.

7. The method of claim 6, further comprising measuring the measured cardiac parameter at the sensing catheter.

8. The method of claim 7, further comprising comparing the estimated cardiac parameter to the measured cardiac parameter.

9. The method of claim 6, wherein the sensing catheter is an imprinted catheter.

10. The method of any one of claims 1-9, wherein the cardiac parameter is left ventricular volume.

11. The method of any one of claims 1-9, wherein the cardiac parameter is one of cardiac output, cardiac power output, stroke volume, or compliance.

12. The method of any one of claims 1-11, further comprising associating the model with patient information describing the first set of patients.

13. The method of claim 12, wherein the patient information comprises a diagnosis or demographic profile of each patient in the first set of patients.

14. The method of claim 13, wherein the diagnosis is one of cardiogenic shock or myocardial infarction.

15. The method of claim 13, wherein the demographics are one or more of gender, risk factors, outcome, or age.

16. The method of any one of claims 1-15, further comprising determining whether the model is applicable to the second patient based on the patient information associated with the model.

17. The method of any one of claims 1-16, further comprising displaying the second pump parameter measurement and the second hemodynamic parameter measurement for the second patient on a display.

18. The method of any one of claims 1-17, further comprising displaying the estimated cardiac parameter of the second patient on the display.

19. The method of any of claims 1-18, further comprising calculating a suggested change in pump speed based on the estimated cardiac parameter in the second patient.

20. The method of claim 19, further comprising implementing the proposed change in the pump speed.

21. The method of claim 19, further comprising displaying the suggested change in the pump speed on a display.

22. The method of any one of claims 1-21, wherein constructing a model of cardiac parameters comprises using a neural network to extract a model from the at least one first hemodynamic parameter and the at least one measurable pump parameter for the first set of patients.

23. The method of claim 22, wherein the neural network comprises a plurality of cells.

24. The method of claim 23, wherein a first unit of the plurality of units comprising the neural network accepts as inputs the at least one first hemodynamic parameter and the at least one measurable pump parameter for the first set of patients at a first point in time.

25. The method of claim 24, wherein the first unit transforms the at least one first hemodynamic parameter and the at least one measurable pump parameter based on one or more model fits prior to transmitting the transformed hemodynamic parameter and transformed pump parameter to a second unit of the plurality of units.

26. The method of claim 25, wherein the first unit updates hidden and unit states at the first point in time.

27. The method of claim 26, wherein the first unit receives the at least one first hemodynamic parameter and the at least one measurable pump parameter at a second point in time and updates the hidden state and the unit state at the second point in time.

28. The method of any one of claims 22-27, wherein the neural network is a recurrent bidirectional neural network.

29. The method of any one of claims 1-28, wherein the first set of patients comprises one patient.

30. A method of estimating cardiac parameters of a patient based on a model, the method comprising:

operating a blood pump in a patient;

measuring at least one measurable pump parameter of the blood pump in the patient to obtain a pump parameter measurement;

measuring at least one hemodynamic parameter in the patient to obtain a hemodynamic parameter measurement;

accessing from a database a model of a relationship between the at least one measurable pump parameter, the at least one hemodynamic parameter, and a cardiac parameter; and

estimating a cardiac parameter estimate for the patient, wherein the cardiac parameter estimate for the patient is output by the model based on the pump parameter measurements and the hemodynamic parameter measurements.

31. The method of claim 30, wherein measuring at least one hemodynamic parameter comprises measuring the aortic pressure.

32. The method of claim 30 or 31, further comprising determining the aortic pressure at a pressure sensor located on the blood pump.

33. The method of any of claims 30-32, wherein measuring the at least one measurable pump parameter comprises measuring pump flow.

34. The method of any one of claims 30-33, wherein the cardiac parameter is left ventricular volume.

35. The method of any one of claims 30-33, wherein the cardiac parameter is one of cardiac output, cardiac power output, stroke volume, or compliance.

36. The method of any one of claims 30-35, further comprising displaying the pump parameter measurement and the hemodynamic parameter measurement of the patient on a display.

37. The method of any one of claims 30-36, further comprising displaying the cardiac parameter estimate for the patient on the display.

38. The method of any of claims 30-38, further comprising calculating a suggested change in pump speed based on the cardiac parameter estimated in the patient.

39. The method of claim 38, further comprising implementing the proposed change in the pump speed.

40. The method of claim 38, further comprising displaying the suggested change in the pump speed on a display.

41. The method of any of claims 30-40, wherein accessing a model comprises determining a selected model from a plurality of models.

42. The method of claim 41, wherein determining the selected model from a plurality of models comprises selecting a model based on information associated with the patient.

43. The method of any one of claims 30-42, wherein accessing a model comprises choosing a model formed by a neural network.

44. The method of claim 43, wherein the neural network is a recurrent bidirectional neural network.

45. The method of any of claims 30-44, further comprising determining a recommended change in the operation of the blood pump based on the estimated cardiac parameter.

46. A method for developing an estimate of a cardiac parameter in a patient, the method comprising:

measuring one or more parameters derived from operation of the medical device and measuring cardiac parameters in a first patient population;

developing a model of the cardiac parameter based on the one or more parameters derived from operation of the medical device and the cardiac parameter in the first patient population;

applying the model to patients in a second patient population to estimate the cardiac parameter of the patients.

47. The method of claim 46, further comprising:

labeling the model according to common characteristics of one or more patients in the first patient population.

48. The method of claim 46 or 47, further comprising:

determining whether the model is applicable to the patients in the second patient population by comparing characteristics of the patients in the second patient population to characteristics of the one or more patients in the first patient population based on the indicia of the model.

49. The method of any one of claims 46-48, wherein developing the model further comprises:

utilizing a machine learning algorithm to develop a model of the cardiac parameter based on the one or more parameters derived from operation of the medical device and the measured cardiac parameter in the first population of patients.

50. The method of any one of claims 46-49, wherein applying the model to the patients in the second patient population further comprises:

operating the medical device in the patients in the second patient population;

measuring the one or more parameters derived from operation of the medical device in the patients in the second patient population;

inputting the measured one or more parameters derived from the operation of the medical device into the model of the cardiac parameter; and

estimating the estimated cardiac parameters of the patients in the second patient population based on the model.

51. A system for estimating cardiac parameters of a patient based on a predetermined model, the system comprising:

a blood pump, the blood pump comprising:

a drivable rotor configured to be driven at one or more pump speeds; and

a sensor configured to measure a hemodynamic parameter; and

a controller, the controller comprising:

a memory configured to receive hemodynamic parameter measurements from the sensor and record the hemodynamic parameter measurements, the memory further storing a predetermined model of a cardiac parameter based on the hemodynamic parameter and a pump speed of the one or more pump speeds;

a driver configured to drive the rotor, the driver configured to transmit a pump speed of the driven blood pump rotor to the memory for recording;

a display configured to display one or more parameters recorded in the memory;

wherein the memory is configured to:

determining an associated cardiac parameter from the predetermined model based on the hemodynamic parameter measurement and the pump speed, an

Transmitting the determined cardiac parameter to the display.

52. The system of claim 51, wherein the memory is configured to store a plurality of predetermined models of the cardiac parameter based on the hemodynamic parameter and the pump speed.

53. The system of claim 52, wherein the controller is configured to select one predetermined model from the stored plurality of predetermined models based on at least one of the hemodynamic parameter and the pump speed.

54. The system of claim 52, wherein the controller is configured to select one predetermined model from a plurality of stored predetermined models based on input to the display.

55. The system of any of claims 52-54, wherein the plurality of predetermined models are formed by a neural network comprising a plurality of cells.

56. The system of claim 55, wherein the neural network is a recurrent bidirectional neural network.

57. The system of any one of claims 51-56, wherein the memory is configured to wirelessly connect to a database containing a plurality of predetermined models of the cardiac parameter based on the hemodynamic parameter and the pump speed.

58. The system of claim 57, wherein the controller is configured to select one of the predetermined models from the database and retrieve the selected one of the predetermined models for storage in the memory.

59. The system of claim 57 or 58, wherein the plurality of predetermined models are formed by a neural network comprising a plurality of cells.

60. The system of claim 59, wherein the neural network is a recurrent bidirectional neural network.

61. The system according to any of claims 51-60, wherein said controller is configured to determine a recommended change to said pump speed based on said determined cardiac parameter.

62. The system of claim 61, in which the controller is further configured to generate the recommended change to the pump speed for display on the display.

63. The system of claim 61 or 62, wherein the controller is configured to implement the recommended change to the pump speed.

64. The system of any of claims 51-63, wherein the sensor is configured to measure at least one of aortic pressure, left ventricular end diastolic pressure, and capillary wedge pressure.

65. The system according to any one of claims 51-64, wherein the cardiac parameter is left ventricular volume.

66. The system according to any of claims 51-64, wherein said cardiac parameter is cardiac power output.

67. A method of estimating cardiac parameters of a patient using a database, the method comprising:

operating a blood pump in a first patient;

measuring at least one measurable pump parameter of the blood pump in the first patient to obtain a pump parameter measurement;

measuring at least one hemodynamic parameter in the first patient to obtain a hemodynamic parameter measurement;

accessing a database comprising patient data of patients other than the first patient, wherein the patient data comprises at least one of measurable pump parameters, hemodynamic parameters, and cardiac parameters; and

estimating a cardiac parameter of the first patient based on the pump parameter measurements in the first patient, hemodynamic parameter measurements in the first patient, and stored patient data from the database.

68. The method of claim 67, wherein the cardiac parameter is cardiac power output.

69. The method of claim 67 or 68, wherein the database is a global database storing data from patients with different characteristics and different medical conditions.

70. The method of claim 69, wherein a characteristic comprises age, weight, gender, or BMI.

71. The method of any of claims 67-70, wherein the database is periodically updated with new data.

72. A pump system having a controller configured to implement the method of any of claims 1-50 and 67-71.

73. A memory configured to perform the method of any of claims 1-50 and 67-71.

74. The method of any of claims 1-50 and 67-71, wherein a neural network is used to derive the model to be applied to input data.

75. The method of claim 74, wherein the neural network comprises a plurality of units in communication with each other and wherein the units:

accepting as input one or more measured parameters,

transforming the one or more measured parameters based on model fitting, an

The transformed parameters are transmitted to neighboring cells having one or more of a hidden state and a cell state.

Background

Cardiovascular disease is a leading cause of morbidity and mortality and places a burden on health care around the world. Various treatment modalities have been developed for cardiovascular diseases, ranging from drugs to mechanical devices and finally transplants. Temporary heart support devices, such as ventricular assist devices, provide hemodynamic support and facilitate cardiac recovery. Some ventricular assist devices are inserted percutaneously into the heart and can be run in parallel with the native (native) heart to supplement cardiac output, such asA series of devices (Abiomed, inc., Danvers MA).

The amount of support (e.g., by volumetric flow measurement of blood delivered by the pumping device) or the duration of support required for each patient may vary. It is difficult for a clinician to directly and quantitatively determine how much support the device should deliver or when to terminate use of the cardiac assist device, especially for patients recovering from an intervention or other cardiac care. Thus, clinicians often rely on the judgment and indirect estimation of heart function, such as the use of fluid-filled catheters to measure intracardiac or intravascular pressure.

While fluid-filled catheters may provide important measurements of cardiac parameters that enable healthcare professionals to make decisions about the cardiac care and health of a patient, the presence of diagnostic equipment in a blood vessel may be at risk to the patient and may be less accurate than desired; in some cases, the equipment may interfere with the function of the pumping device.

Disclosure of Invention

The methods, systems, and devices described herein enable creation and use of a model relating blood pump parameters to cardiac parameters based on a first patient population, which can then be applied to a second patient population to estimate cardiac parameters without the use of additional measurement catheters or other diagnostic devices. In particular, the method and system enable the use of machine learning to develop a model representing the relationship between measured parameters of a blood pump of a first set of patients and cardiac parameters (such as left ventricular volume or cardiac output). Machine learning algorithms construct models of measured cardiac parameters with respect to one or more measurable parameters of the blood pump based on data from a large number of patients having various characteristics, such as gender, weight, disease state, cardiac outcome, diagnosis, or other characteristics. After developing a model that predicts cardiac parameters measured by a diagnostic device (e.g., a fluid-filled catheter), the model can then be accessed and applied to patients in a second patient set to estimate cardiac parameters (such as cardiac output) based on the pump parameters without the use of additional catheters or other diagnostic devices.

Specifically, the model is created by: tracking blood pump performance parameters such as pump speed, current, flow, and pressure in the vessel in which the pump is located (such as aortic pressure measured by on-board optical or other pressure sensors on the pump itself); and measuring one or more hemodynamic parameters, such as left ventricular volume, left ventricular pressure, pulmonary artery pressure, or other cardiac parameters (such as through a pressure sensing catheter) over a period of time in a plurality of patients comprising a training set of models. The data is collected, stored and then analyzed using machine learning algorithms to extract a curve fit for a set or specific subset of patients. For example, a model indicative of cardiac output may be extracted based on pump performance parameters and measured hemodynamic parameters from a population of patients in the patient set. The model may be applicable to all patients in the patient set, or to one or more patients in the patient set, or a model may be extracted that is applicable to a subset of patients in the set that have particular characteristics. For example, in some embodiments, different models may be determined for all patients diagnosed with cardiogenic shock, myocardial infarction, or based on patient demographics (such as gender, weight, or risk factors). In another example, the model is applicable to all types of patients regardless of their diagnosis or various demographics.

The model is created by using a neural network to fit a large amount of stored data to the model. At each point in time of measured pressure and flow data in a particular patient in a patient population, the neural network may calculate a cardiac parameter (such as left ventricular pressure) using pressure and flow data (or pump speed or other parameters) extracted from the blood pump and compare the calculated cardiac parameter to a true measurement of the parameter determined by the catheter. The neural network may include a plurality of units in communication with each other to develop a model based on relationships between pump parameters (e.g., pump speed, pressure, and flow data) and cardiac parameters. The unit receives as inputs pump performance data (e.g., pump speed, pressure, and flow) and hemodynamic parameters at a first point in time and transforms the inputs based on a model fit. The inputs to the model may be hemodynamic parameters and pump parameters that may be correlated to the measured cardiac parameters. The neural network may be a stacked neural network, such as a stacked bidirectional recurrent neural network, that communicates over time in a hidden state and develops the model based on a plurality of activation functions to iteratively develop the model. For example, a cell of the neural network may transform the input based on model fitting and then transmit the transformed input to the next cell in the stack along with the updated hidden state and cell state. The final model output from the neural network can accurately represent the cardiac output or left ventricular volume (or other cardiac function) based on the pump parameters without the use of a catheter.

The model may then be applied to patients outside the training set. In the case of a model that applies to patients regardless of demographics or diagnosis, the model may apply to all patients in a second group that is not part of the training set of models. In another embodiment, the health care provider may enter various demographics of the patient and select an appropriate model based on the patient demographics. The model is then applied to the blood pump parameters measured for the patient and the estimated cardiac parameters are extracted. For example, blood pump speed and aortic pressure measured in a patient may be used with a model to extract estimated left ventricular pressure or cardiac output. The estimated left ventricular pressure shows the patient's heart health over time.

When the pumping device is in a patient, the model may be used to provide a healthcare professional with a continuous or nearly continuous estimate of the cardiac parameters, thereby enabling the healthcare professional to make real-time decisions on the care of the patient. For example, the estimated cardiac parameters provided may be used by a healthcare professional for decisions related to cardiac health, thereby removing or adding support to the patient from the pumping apparatus. The cardiac parameter may be left ventricular volume, cardiac output, cardiac power output, compliance, native (native) flow, stroke volume, volume in diastole or systole, or other relevant cardiac parameter, or any combination of the preceding. Other hemodynamic or cardiac parameters may be determined using the estimated cardiac parameter and also provided to the healthcare professional.

In one aspect, a method of estimating a cardiac parameter of a patient includes: operating a blood pump within each patient of the first set of patients, the blood pump having at least one measurable pump parameter; measuring at least one hemodynamic parameter and at least one measurable pump parameter for each patient in the first set of patients to obtain a first hemodynamic parameter measurement and a first pump parameter measurement; and constructing a model of the one or more cardiac parameters based on a relationship between the at least one first hemodynamic parameter and the at least one measurable pump parameter for the first set of patients. The model may include a neural network having inputs of hemodynamic parameters and pump parameters from a plurality of patients within the first set. The method further comprises the following steps: operating a second blood pump in a second patient of a second patient set; and applying the model to a second patient by: measuring at least one measurable pump parameter in a second patient to obtain a second pump parameter measurement; measuring at least one first hemodynamic parameter in a second patient to obtain a second hemodynamic parameter measurement; and estimating a cardiac parameter of the second patient, wherein the cardiac parameter of the second patient is output by the model based on the second pump parameter measurement and the second hemodynamic parameter measurement. In some embodiments, the method further comprises determining the estimated cardiac parameter based on the at least one hemodynamic parameter and the at least one measurable pump parameter for the at least one point in time. In some embodiments, the method includes inserting a sensing catheter separate from the blood pump into each patient within the first set of patients (e.g., placing the catheter in the left ventricle or pulmonary artery), and measuring a hemodynamic parameter (such as left ventricular end diastolic pressure or pulmonary capillary wedge pressure) at the sensing catheter. The measured hemodynamic parameters may be used to calculate cardiac output or other cardiac parameters as the measured parameters. In some embodiments, the method further comprises comparing the estimated cardiac parameter based on the output from the model with a measured cardiac parameter based on input provided from readings of the sensing catheter. Finally, pump operation may be established and adjusted based on the estimated cardiac parameters from the model, for example by using the estimated cardiac parameters from the model as input to a pump controller configured to receive these parameters and adjust pump output.

In some embodiments, the method comprises: displaying the second pump parameter measurement and the second hemodynamic parameter measurement for the second patient on a display, displaying the estimated cardiac parameter for the second patient on the display, and/or calculating a suggested change in pump speed based on the estimated cardiac parameter in the second patient. In some embodiments, the method further comprises implementing the proposed change in pump speed.

In some embodiments, constructing the model of the cardiac parameter includes using a neural network to extract the model from the at least one first hemodynamic parameter and the at least one measurable pump parameter for the first set of patients. The model may be extracted from a plurality of parameters derived from one or more patients, including a plurality of hemodynamic parameters and a plurality of pump parameters. The model is stored in memory and may be onboard or otherwise accessible to the pump controller over a network. The neural network may comprise a plurality of cells. In some embodiments, multiple units communicate with each other, and the units accept as input one or more parameters (measured parameters, such as pump parameters and hemodynamic parameters, or a combination of pump parameters and hemodynamic parameters), and transform the one or more parameters based on model fitting. One or more units may transmit the transformed parameters to neighboring units, such as units having a hidden state or a unit state. In some embodiments, a first unit in the neural network receives as inputs one or more hemodynamic parameters and one or more measurable pump parameters for a first set of patients at a first point in time. A first unit in the neural network may receive a plurality of parameters or a combination of parameters, such as a plurality of hemodynamic parameters and a plurality of pump parameters. In some embodiments, the first unit transforms the at least one first hemodynamic parameter and the at least one measurable pump parameter based on one or more model fits prior to transmitting the transformed hemodynamic parameter and the measurable pump parameter to the second unit in the neural network. In some embodiments, the first unit updates the hidden state and the unit state at the first point in time. In some embodiments, the first unit receives the at least one first hemodynamic parameter and the at least one measurable parameter at the second point in time and updates the hidden state and the unit state at the second point in time. In some embodiments, the first set of patients is formed from a single patient.

In one aspect, a method of estimating cardiac parameters of a patient based on a model includes: operating a blood pump in a patient; measuring at least one measurable pump parameter of a blood pump in a patient to obtain a pump parameter measurement; measuring at least one hemodynamic parameter in a patient to obtain a hemodynamic parameter measurement; and accessing from the database a model of a relationship between the at least one measurable pump parameter, the at least one hemodynamic parameter, and the cardiac parameter. The method also includes estimating a cardiac parameter estimate for the patient, wherein the cardiac parameter estimate for the patient is output by the model based on the pump parameter measurement and the hemodynamic parameter measurement.

In some embodiments, the method and system access the model by determining a selected model from a plurality of available models. In some embodiments, the selected model is determined based on information associated with the patient. In some embodiments, the method includes choosing a model formed by a neural network comprising a plurality of cells. In some embodiments, the neural network is a recurrent bidirectional neural network. In some embodiments, the neural network comprises a plurality of cells. In some embodiments, a plurality of units communicate with each other and the units accept one or more measured parameters as input, transform the one or more measured parameters based on model fitting, and transmit the transformed parameters to a neighboring unit having a hidden state or a unit state. In some embodiments, the method includes determining a recommended change in operation of the blood pump based on the estimated cardiac parameter.

In one aspect, a method for developing estimates of cardiac parameters in a patient includes: measuring one or more parameters derived from the operation of the medical device and measuring cardiac parameters in a first patient population, developing a model of the cardiac parameters based on the one or more parameters derived from the operation of the medical device and the cardiac parameters in the first patient population; and applying the model to patients in the second patient population to estimate cardiac parameters of the patients.

In some embodiments, the method further comprises: the model is labeled according to common characteristics of one or more patients in the first patient population, and/or based on the labeling of the model, it is determined whether the model is applicable to patients in the second patient population by comparing characteristics of patients in the second patient population with characteristics of one or more patients in the first patient population. In some embodiments, the method further comprises utilizing a machine learning algorithm to develop a model of the cardiac parameter based on the one or more parameters derived from the operation of the medical device and the measured cardiac parameter in the first patient population. In some embodiments, the model is developed using a neural network. In some embodiments, the neural network comprises a plurality of cells. In some embodiments, a plurality of units communicate with each other and the units accept one or more measured parameters as input, transform the one or more measured parameters based on model fitting, and transmit the transformed parameters to a neighboring unit having a hidden state or a unit state.

In some embodiments, applying the model to the patients in the second patient population comprises: operating the medical device in a patient in a second patient population; measuring one or more parameters derived from operation of the medical device in patients in a second patient population; inputting the measured one or more parameters derived from the operation of the medical device into a model of cardiac parameters; and estimating the estimated cardiac parameters of the patients in the second patient population based on the model.

In one aspect, a system for estimating cardiac parameters of a patient based on a predetermined model (such as a model formed by any of the techniques disclosed herein) includes a blood pump and a controller. The blood pump comprises a drivable rotor designed to be driven at one or more pump speeds, and a sensor capable of measuring a hemodynamic parameter. The controller includes a memory that receives hemodynamic parameter measurements from the sensor and records the hemodynamic parameter measurements, the memory further storing (or accessed from a network) a predetermined model of a cardiac parameter based on the hemodynamic parameter and a pump speed (or current, flow, or other pump parameter) of the one or more pump speeds. The controller also includes a driver designed to drive the rotor and to transmit the pump speed (or one or more other pump parameters) of the driven blood pump rotor to the memory for recording, and a display to display the one or more parameters recorded in the memory. The memory determines an associated cardiac parameter using the predetermined model and the hemodynamic parameter measurements and the pump parameter (e.g., pump speed), and transmits the determined cardiac parameter to the display.

In some embodiments, the memory stores a plurality of predetermined models of cardiac parameters based on hemodynamic parameters and pump parameters (e.g., pump speed, motor current). In some embodiments, the controller selects one predetermined model from a plurality of stored predetermined models based on one of a hemodynamic parameter or a pump parameter (e.g., pump speed, motor current). In some embodiments, the controller selects one predetermined model from a plurality of stored predetermined models based on input to the display. In some embodiments, the plurality of predetermined models is formed by a neural network comprising a plurality of cells. In some embodiments, the neural network is a recurrent bidirectional neural network. In some embodiments, the neural network comprises a plurality of cells. In some embodiments, a plurality of units communicate with each other and the units accept one or more measured parameters as input, transform the one or more measured parameters based on model fitting, and transmit the transformed parameters to a neighboring unit having a hidden state or a unit state.

In some embodiments, the memory is wirelessly connected to a database containing a plurality of predetermined models of cardiac parameters based on hemodynamic parameters and pump speed. In some embodiments, the controller selects one of the predetermined models from the database and retrieves the selected one of the predetermined models for storage in the memory. In some embodiments, the plurality of predetermined models is formed by a neural network comprising a plurality of cells. In some embodiments, the neural network is a recurrent bidirectional neural network.

In some embodiments, the controller determines a recommended change to the pump speed based on the determined cardiac parameter. In some embodiments, the controller generates a recommended change to the pump speed for display on the display. In some embodiments, the controller implements the recommended change to the pump speed for display on the display. In some embodiments, the sensor measures aortic pressure. In some embodiments, the cardiac parameter is left ventricular volume. In some embodiments, the cardiac parameter is cardiac power, cardiac power output, or another cardiac parameter.

In one aspect, a method of estimating cardiac parameters of a patient using a database includes: operating a blood pump in a first patient; measuring at least one measurable pump parameter of a blood pump in a first patient to obtain a pump parameter measurement; measuring at least one hemodynamic parameter in a first patient to obtain a hemodynamic parameter measurement; and accessing a database comprising patient data of patients other than the first patient, wherein the patient data comprises at least one of a measurable pump parameter, a hemodynamic parameter, and a cardiac parameter. The method also includes estimating a cardiac parameter of the first patient using the pump parameter measurement in the first patient, the hemodynamic parameter measurement in the first patient, and the stored patient data from the database.

In some embodiments, the blood pump is operated in a first patient, and measurable input from the first patient is used in conjunction with a database including patient data from patients other than the first patient to estimate a cardiac parameter of the first patient. For example, the database may include a series of patients' cardiac power outputs, as well as other measured data. The database includes data from a series of patients having different characteristics (e.g., age, gender, weight, height, etc.). In one example, the database includes data from a series of patients with different medical conditions. The database may be updated periodically to include new data. In some embodiments, the database includes a model of the relationship between hemodynamic parameters, pump parameters, and cardiac parameters. In some embodiments, the model is derived by using a neural network on the patient data. In some embodiments, the neural network from which the model is derived comprises a plurality of units. In some embodiments, a plurality of units communicate with each other and the units accept one or more measured parameters as input, transform the one or more measured parameters based on model fitting, and transmit the transformed parameters to a neighboring unit having a hidden state or a unit state.

Drawings

The above 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 characters refer to like parts throughout, and in which:

fig. 1 shows a block diagram of a system for estimating cardiac parameters of a patient based on a predetermined model;

FIG. 2 shows a block diagram of a stacked bidirectional recurrent neural network;

FIG. 3 shows a block diagram of a long-short term memory unit of the stacked bidirectional recurrent neural network of FIG. 2;

FIG. 4 illustrates a method of developing and using a model for estimating cardiac parameters of a patient;

FIG. 5 illustrates a method of estimating cardiac parameters of a patient using a model;

FIG. 6 illustrates a method for developing estimates of cardiac parameters in a patient;

FIG. 7A illustrates an exemplary graph of measured left ventricular volume and predicted left ventricular volume based on an exemplary relationship between aortic pressure and pump flow;

FIG. 7B illustrates an exemplary graph of measured cardiac output and predicted cardiac output based on an exemplary relationship between aortic pressure and pump flow;

FIG. 7C illustrates example measured aortic pressures used to predict the left ventricular volume of FIG. 7A and cardiac output of FIG. 7B;

FIG. 7D illustrates an example measured pump flow rate for predicting the left ventricular volume of FIG. 7A and the cardiac output of FIG. 7B;

FIG. 8A illustrates an exemplary graph of measured left ventricular volume and predicted left ventricular volume based on an exemplary relationship between aortic pressure and pump flow and at a pump power level of 2;

FIG. 8B shows an exemplary graph of measured heart beat volume and predicted heart beat volume based on an exemplary relationship between aortic pressure and pump flow and at a pump power level of 2;

FIG. 8C illustrates an exemplary graph of measured left ventricular volume and predicted left ventricular volume based on an exemplary relationship between aortic pressure and pump flow and at a pump power level of 3;

FIG. 8D shows an exemplary graph of measured heart beat volume and predicted heart beat volume based on an exemplary relationship between aortic pressure and pump flow and at a pump power level of 3;

FIG. 9A illustrates an exemplary graph of measured left ventricular volume and predicted left ventricular volume based on an exemplary relationship between aortic pressure and pump flow for an irregular waveform;

FIG. 9B shows an exemplary graph of measured heart beat volume and predicted heart beat volume for an irregular waveform based on an example relationship between aortic pressure and pump flow;

FIG. 9C illustrates an exemplary graph of measured left ventricular volume and predicted left ventricular volume based on an exemplary relationship between aortic pressure and pump flow for irregular waveforms;

FIG. 9D illustrates an exemplary graph of measured heart beat volume and predicted heart beat volume for an irregular waveform based on an exemplary relationship between aortic pressure and pump flow;

Detailed Description

Certain illustrative embodiments will be described in order to provide a general understanding of the methods and systems described herein. Although the embodiments and features described herein are specifically described in connection with a blood pump apparatus, it will be understood that all of the components and other features outlined below may be combined with each other in any suitable manner and may be adapted and applied to other types of hearts and medical treatments.

In some embodiments, the blood pump is operated in a first patient, and measurable input from the first patient is used in conjunction with a database including patient data from patients other than the first patient to estimate a cardiac parameter of the first patient. For example, the database may include a series of patients' cardiac power outputs, as well as other measured data. The database includes data from a series of patients having different characteristics (e.g., age, gender, weight, height, etc.). In one example, the database includes data from a series of patients with different medical conditions. The database may be updated periodically to include new data.

Fig. 1 shows a block diagram of a system 100 for estimating cardiac parameters of a patient based on a predetermined model. The system 100 includes a controller 102 and a blood pump 104. The controller includes a memory 106 with a predetermined model 118, a driver 108 and a display 110. The blood pump 104 includes a rotor 114 and a sensor 116. The controller 102 is communicatively coupled to the blood pump 104 via a line 112, which may be an electrical wire and/or a mechanical drive shaft. A driver 108 within the controller 102 controls the blood pump 104, including the speed of operation of the rotor 114. Driver 108 is communicatively coupled to memory 106 through channel 107 and is also communicatively coupled to display 110 through channel 109. The sensor 116 of the blood pump 104 may be coupled to the controller 102 via line 112, or may be wirelessly coupled to the controller 102.

The blood pump 104 operates in the patient's vascular system to provide cardiovascular support by pumping blood in the patient's heart or vascular system. The rotational speed of the rotor 114 controls the flow rate of blood through the blood pump 104. The sensor 116 is located on the blood pump 104 such that when the blood pump 104 is in place within the patient's vascular system, the sensor 116 can measure a hemodynamic parameter of the patient. The sensor 116 transmits the measured hemodynamic parameters to the controller 102 wirelessly or via the line 112. In some embodiments, the sensor 116 is an on-board optical sensor or pressure sensor located on the blood pump 104. In some embodiments, sensor 116 measures aortic pressure. In some embodiments, sensor 116 measures other hemodynamic parameters.

The controller 102 controls the speed of the rotor 114 by varying the power supplied to the blood pump 104. Driver 108 also measures the load on rotor 114 by measuring the current supplied to rotor 114 to maintain a particular rotor speed. The driver 108 stores the measured pump parameters in the memory 106. The driver 108 receives the measured hemodynamic parameters from the sensor 116 and also stores these parameters in the memory 106. The driver 108 may also include processing hardware or software (not shown) to enable processing of hemodynamic parameters and pump parameters in the controller 102, such as averaging or for calculating other cardiac parameters. The controller 102 tracks blood pump parameters (such as pump speed, current, flow, and pressure in the blood vessel) based on the performance of the blood pump and hemodynamic parameters measured by the sensors 116. The driver 108 transmits hemodynamic parameters, pump parameters, or other measured or calculated parameters to the display 110.

The memory 106 includes a predetermined model 118 that relates pump parameters to one or more hemodynamic parameters. The creation of such a model is described below. The memory 106 and/or driver 108 uses the measured pump parameters and hemodynamic parameters and the stored predetermined model 118 to estimate specific cardiac parameters based on the measured pump parameters. The cardiac parameter may be left ventricular volume, cardiac output, cardiac power output, compliance, native flow, stroke volume, volume in diastole or systole, or other relevant cardiac parameter, or any combination of the preceding. No additional catheters or diagnostic devices may be needed to measure the cardiac parameters, as the model provides estimated cardiac parameters based on a model constructed from other patient data from the first set of patients. In some implementations, the memory 106 includes more than one predetermined model 118, and a particular predetermined model 118 is selected based on one or more of the measured pump parameters and hemodynamic parameters. In some embodiments, the particular predetermined model 118 is selected from a plurality of stored models by input from a healthcare professional. In some embodiments, the memory 107 stores or is linked to a database from which the predetermined model is selected.

In some embodiments, driver 108 displays the estimated cardiac parameter on display 110. In some implementations, the controller 102 uses the estimated cardiac parameters to determine a recommended course of action with respect to the support of the increase or decrease of the blood pump 104. For example, the controller 102 may display on the display 110 recommended changes to the operation of the blood pump 104 based on the measured hemodynamic and pump parameters and the estimated cardiac parameter. In particular, the controller 102 may determine the recommended course of action based on a comparison of the estimated cardiac parameter of the patient with a previously estimated cardiac parameter. In some embodiments, the controller 102 may make changes to the support provided by the blood pump 104 based on the proposed course of action. In some embodiments, the controller 102 presents options to the health professional via the display 110 and allows the health professional to select options to control or change the operation of the blood pump 104.

In some embodiments, hemodynamic parameters and pump parameters or other data stored in memory 106 may be retrieved from memory 106 for use with data from other patients for creating an algorithm that correlates blood pump parameters with one or more cardiac parameters. The extracted data may be combined with other health data (such as gender, weight, disease status, cardiac outcome, diagnosis, or other characteristics) and used to create algorithms based on machine learning or neural networks. In some embodiments, the controller 102 is coupled to a database that stores data from which the predetermined model is derived, and the controller 102 uploads the data to update the database.

Fig. 2 illustrates a block diagram of an exemplary stacked bidirectional recurrent neural network 200 that may be used to create a model, such as the predetermined model 118, that may be used in the blood pump system 100 of fig. 1 to interpret and estimate cardiac parameters from pump parameters measured in a patient. The neural network 200 is used to fit a large amount of data from a training data set comprising measured parameters from a first set of patients. The exemplary neural network may be implemented in creating a model that relates blood pump parameters to cardiac parameters, as described above in fig. 1. The exemplary neural network 200 is a stacked bidirectional recurrent neural network, although other neural network models suitable for creating the models described herein are also available. The neural network 200 communicates over time in the hidden state and develops the model in an iterative development model based on a plurality of activation functions, as will be described in more detail below. The model created using the neural network may then be stored in a controller memory (e.g., in memory 106 of fig. 1) and used to estimate cardiac parameters of a patient in which the blood pump is being operated. In this exemplary neural network 200, the processing units 220a-220j (labeled "LSTM" for long and short term memory) are organized in a grid having rows 222 and 224 and columns 226 and 234. Processing units 220a-220j communicate with each other along rows 222 and 224 and columns 226 and 234. There are multiple stages or rows 222 and 224 stacked between the input and output, and there are multiple columns 226 and 234.

The lowest row 224 is the input row with inputs 236a-236e for aortic pressure (AOP) and pump flow (flow). The highest row 222 is the output row, which outputs an estimated output parameter 238, such as the Left Ventricular Volume (LVV). The number of lines between the input lines 224 and the output lines 226 indicates the model depth or complexity (sophistication). For example, the models may be bi-directionally stacked, as shown by the neural network 200 in fig. 2. Alternatively, the model may have three, four, five, or more levels of cells stacked between the input row 224 and the output row 222. Each estimate of the neural network 200 is based on the number of states at different sampling times, represented by the number of columns 226-234 in the exemplary neural network 200. For neural network 200, at time t, neural network 200 receives AOP and input 236e of traffic and uses information from neural network 200 for a plurality of previous states (e.g., 75 states shown as t-74 for input 236a in column 226 to t-1 for input 236d in column 232). The neural network 200 calculates an estimated cardiac parameter based on a set of at least 25 previous sampling instances (output 238). In some embodiments, the neural network calculates the estimated cardiac parameter based on a set of at least 50 or at least 75 or more previous sampling instances (output 238). At each point in time of measured aortic pressure and flow data in a particular patient in a patient population, the neural network may calculate a cardiac parameter (such as left ventricular pressure) using pressure and flow data extracted from the blood pump and compare the estimated cardiac parameter to a true measurement of the parameter determined by the catheter. In some embodiments, the cardiac parameter is left ventricular volume, cardiac output, cardiac power output, compliance, native flow, stroke volume, volume in diastole or systole, or other relevant cardiac parameter, or any combination of the foregoing.

Specifically, within each cell of the neural network 200, the neural network 200 generates so-called hidden states and shares these hidden states between different cells. By utilizing a stacked neural network system, complex relationships between the input data 236a-236e can be extracted to produce accurate estimates of the output parameters 238.

The neural network 200 may be used in a machine learning algorithm that constructs a model of a measured cardiac parameter (e.g., aortic pressure) with respect to one or more measurable parameters of a blood pump (such as the blood pump 104 in fig. 1), such as pump speed or flow, based on data from a large number of patients having various characteristics, such as gender, weight, disease state, cardiac outcome, diagnosis, or other characteristics. Patient data is input into a machine learning algorithm to develop a model based on the relationships between the various input data determined by the algorithm. The final model can represent an accurate left ventricular volume or cardiac output (or other cardiac function) curve based on the pump parameters without the use of catheters, and as described above, can be a global model equipped to handle all physiological conditions of any situation in a patient population. After developing a model that predicts cardiac parameters measured by a diagnostic device (e.g., a fluid-filled catheter or other internal sensor), the model may then be applied to patients in a second patient set other than the training set to estimate cardiac parameters based on the pump parameters without using additional catheters or other diagnostic devices.

FIG. 3 shows a block diagram of a long short term memory unit of the stacked bidirectional neural network of FIG. 2. For example, the cells 220a-220j of the neural network 200 of FIG. 2 may be long and short term memory cells. Alternatively, the cells 220a-220j of the neural network 200 may be other types of cells. Similarly, the neural network 200 itself may be a neural network such as shown in the example of FIG. 2, orAnother type of neural network, such as a fully recursive, Ellman, Hopfield, echo state, hierarchical, etc. In the example of fig. 2, the neural network element is a long-short term memory element. Fig. 3 shows a single long-short term memory cell 300. As shown in FIG. 3, the long-short term memory unit 300 has four activation functions and their associated functions, represented by four blocks, each including a first function "f_t340, second function342. Third function "i_t"344 and fourth function" o_t"346. First function "f_t"340 is an inverse function (sigmoidal) function that generates a gating variable, the second function342 is a hyperbolic tangent function that produces candidate states for the memory cell, a third function "i_t"344 is an inverse function that produces a gating variable, and a fourth function" o_t"346 is an inverse function that produces a gating variable. Although the first function "f_t340, second function342. Third function "i_t"344 and fourth function" o_t"346 is an example, and other functions may be used to process information in unit 300, but an exemplary first function" f is defined below_t340, second function342. Third function "i_t"344 and fourth function" o_t”346：

f_t＝σ(W_f[h_t-1，x_t]+b_t)

i_t＝σ(W_i[h_t-1，x_t]+b_i)

o_t＝σ(W_o[h_t-1，x_t]+b_o)

The cell 300 receives the cell state 348a ("c") from the previous cell_t-1") and the unit state 348a is processed by: the first function ("f) indicating which elements the unit 300 should no longer consider_t") 340, a second function that indicates which information the unit 300 should extract342. Third function "i" indicating which information the cell should update_t"" 344, and a fourth function "o" providing an output for updating the candidate cell_t"346 or summary gate (summary gate). The updated cell state 348b is communicated to the neighboring cells in the neural network. In this example, the cell state is defined by:

cell 300 receives hidden state 349a ("h") from a previous cell_t-1") and process the hidden state 349 a. Hidden state 349a serves as an input to: the first function ("f) indicating which elements the unit 300 should no longer consider_t") 340, a second function that indicates which information the unit 300 should extract342. Third function "i" indicating which information the cell should update_t"" 344, and a fourth function "o" providing an output for updating the candidate cell_t"346, or a summary gate. The updated hidden state 349b is passed to the neighboring cells in the neural network. As shown, the updated hidden state 349b is passed to cells adjacent to cell 300 in the same row or column. In this example, the hidden state is defined by:

the activation function or gate may correspond to a series of functions including an sigmoid function, a hyperbolic tangent function, an sigmoid function, or any combination of these or other functions. The processing of inputs by the various functions of unit 300 enables a neural network comprising many such units to access complex relationships between data inputs to produce algorithms that can be applied to other data to predict results.

Fig. 4 illustrates a method 400 of developing and using a model for estimating cardiac parameters of a patient based on blood pump parameters (e.g., the blood pump 104 in fig. 1). The method 400 includes a step 402 in which a blood pump is operated within a first set of patients. In some embodiments, another intravascular medical device (such as a balloon pump, a centrifugal pump such as an ECMO, pulsatile pump, roller pump, or other ventricular assist device) may be used in a similar manner, rather than a blood pump. At step 404, hemodynamic parameters and pump parameters are measured for each patient in the first set of patients. More than one hemodynamic parameter and/or pump parameter may be measured for each patient in the first set of patients. In some embodiments, the hemodynamic parameters and the pump parameters measured for each patient are one or more of pump speed, current, flow, and pressure in a blood vessel, and the measurements are based on performance of the blood pump. In some embodiments, aortic pressure is measured as a hemodynamic parameter. The hemodynamic parameter is measured by a measurement catheter (such as a fluid-filled catheter, an imprints (inca) catheter, a miller (millar) catheter (for animals)) or another diagnostic device.

In some embodiments, one or more of pump speed, flow rate, pump pressure are measured as pump parameters. The pump parameters are measured by the blood pump controller based on the current supplied to the pump, the load on the pump, or other characteristics of the blood pump operation. At step 406, cardiac parameters are measured for each patient in the first set of patients. In some embodiments, the cardiac parameter is left ventricular volume, cardiac output, cardiac power output, compliance, native flow, stroke volume, volume in diastole or systole, or other relevant cardiac parameter, or any combination of the foregoing. For each patient in the first set of patients (which is the model training set), cardiac parameters, hemodynamic parameters, and pump parameters may be measured over a period of time.

At step 408, a model of the cardiac parameter is constructed using the hemodynamic parameter and the pump parameter based on a relationship between the hemodynamic parameter and the pump parameter. Data from each patient in the first patient set is collected and stored, and then the data is analyzed using a machine learning algorithm to extract a curve fit for the entire patient set or a particular subset of patients. For example, a model may be extracted that is applicable to one or more patients in the set of patients, or a model may be extracted that is applicable to a subset of patients in the set that have a particular characteristic. For example, in some embodiments, different models may be determined for all patients diagnosed with cardiogenic shock, myocardial infarction, or based on patient demographics (such as gender, weight, or risk factors). In another example, the model is applicable to all types of patients regardless of their diagnosis or various demographics.

The model may be constructed using machine learning or neural networks (such as described above in fig. 2 and 3) or any other available machine learning arrangement. Neural networks can be used to fit large amounts of stored data to a model. Once constructed, the model may be stored in a controller of the blood pump (e.g., in memory 106 of fig. 1) or may be hosted in a server or processor coupled to the blood pump controller or another processor that receives the measured parameters from the blood pump controller.

At step 410, the blood pump is operated in a patient of the second patient set to provide cardiac support. At step 412, the model generated in step 406 is applied to the patients in the second patient set by: measuring pump parameters and hemodynamic parameters in the patient, and estimating cardiac parameters of the patient based on the model and the measured pump parameters and hemodynamic parameters in the patient in the second set. In this way, the estimated cardiac parameters may be determined for the patients in the second patient set based on the model without using additional catheters or diagnostic tools.

In the case of a model that applies to patients regardless of demographics or diagnosis, the model may apply to all patients in a second group that is not part of the training set of models. In another embodiment, the health care provider may enter various demographics of the patient and pick the appropriate model based on the patient demographics. The model is then applied to the blood pump parameters measured for the patient and the estimated cardiac parameters are extracted. For example, the blood pump speed and aortic pressure measured in the patient may be used with the model to extract estimated cardiac parameters, such as left ventricular volume, cardiac output, cardiac power output, compliance, native flow, stroke volume, volume in diastole or systole, or other relevant cardiac parameters, or any combination of the foregoing.

Fig. 5 illustrates a method 500 of estimating cardiac parameters of patients in a second patient set using a model constructed from data of a first patient set. At step 502, a blood pump is operated within the vascular system of a patient in a second patient set. At step 504, at least one measurable pump parameter of the blood pump is measured in the patient to obtain a pump parameter measurement. In some embodiments, the pump parameter may be pump speed, flow rate through the pump, or pressure within the pump, and may be measured based on current supplied to the pump, load on the pump, or other characteristics of blood pump operation. The pump parameters may be measured at a controller of the blood pump (e.g., controller 102 in fig. 1) or at the blood pump itself. At step 506, at least one hemodynamic parameter is measured in the patient to obtain a hemodynamic parameter measurement. In some embodiments, the hemodynamic parameter is aortic pressure. The hemodynamic parameters may be measured by sensors placed on the blood pump or sensors placed on a catheter coupled to the blood pump.

At step 508, a model of a relationship between the at least one measurable pump parameter, the at least one hemodynamic parameter, and the cardiac parameter is accessed. The model may be generated by a machine learning or neural network algorithm (e.g., by the neural networks described in fig. 2 and 3, or by any available machine learning process) to estimate cardiac parameters from measured hemodynamic parameters and pump parameters. The model may be stored in a controller of the blood pump (e.g., in memory 106 of fig. 1) or may be hosted in a server or processor coupled to the blood pump controller or another processor that receives the measured parameters from the blood pump controller. At step 510, the model is used to estimate cardiac parameters of patients in the second patient set based on the pump parameter measurements and the hemodynamic parameter measurements in the patients. The cardiac parameter may be left ventricular volume, cardiac output, cardiac power output, compliance, native flow, stroke volume, volume in diastole or systole, or other relevant cardiac parameter, or any combination of the preceding. The patient's cardiac parameters are not measured otherwise, so that no additional catheters or diagnostic devices need to be inserted into the patient's vascular system. The estimated cardiac parameter may be used to inform health decisions made by the health care professional and may be displayed to the health care professional and/or used to recommend changes to the health care professional in the support provided by the blood pump.

Fig. 6 shows a method 600 for developing an estimate of a cardiac parameter of a patient. At step 602, one or more parameters derived from operation of a medical device and cardiac parameters are measured in a first patient population. At step 604, a model of cardiac parameters is developed based on the measured cardiac parameters and one or more parameters derived from operation of the medical device in the first patient population. The model may be developed using machine learning or neural networks (such as those described in fig. 2 and 3) or by any other available machine learning process. At step 606, the model is applied to patients in the second patient population to estimate cardiac parameters in the patients. No additional determination of the cardiac parameters is required in the patient.

At step 608, the estimated cardiac parameters of the patient are displayed, for example, on a display associated with a medical device (such as a blood pump). The health care professional can use the displayed estimated cardiac parameters to make health care decisions related to the treatment and use of the medical device.

When a medical device (such as a blood pump) is in a patient, the model may be used to provide a healthcare professional with a continuous or nearly continuous estimate of cardiac parameters, thereby enabling the healthcare professional to make real-time decisions on the care of the patient. For example, in the case of a blood pump being used in a patient, the estimated cardiac parameters provided may be used by a healthcare professional for decisions related to cardiac health, thereby taking the patient out of or adding support to the pumping apparatus. The cardiac parameter may be left ventricular volume, cardiac output, cardiac power output, compliance, native flow, stroke volume, volume in diastole or systole, or other relevant cardiac parameter, or any combination of the preceding. Other hemodynamic parameters or cardiac parameters may be extracted from the estimated cardiac parameters and also provided to the healthcare professional.

In some embodiments, the controller of the blood pumping device may use the estimated cardiac parameter to determine a recommended course of action regarding increased or decreased support of the blood pumping device. In particular, the controller may determine the recommended course of action based on a comparison of the estimated cardiac parameter of the patient with a previously estimated cardiac parameter. In some embodiments, the controller may make changes to the support provided by the blood pumping apparatus based on the proposed course of action.

7-9 show exemplary curve graphs of parameters over time comparing model predicted traces with "true" traces obtained by direct measurement. As described above with respect to fig. 4-6, a model of cardiac parameters may be developed for a second patient population using hemodynamic and pump or medical device data of the first patient population. The model of the cardiac parameters enables estimation of the cardiac parameters of the patient without the need for additional diagnostic or sensing catheters in the vascular system of the patient, which is safer and more efficient. Because well-developed algorithms can also take into account additional patient data such as gender, weight, disease status and outcome, the estimated cardiac parameters can be highly accurate. Additionally, additional data considered in the development of the algorithm may be used to suggest a treatment regimen or change to the use or operation of the blood pump or other medical device in order to improve the patient's heart health based on the measured parameters and application of the developed model. Fig. 7-9 illustrate the accuracy of an example model for predicting cardiac parameters based on measured pump parameters and hemodynamic parameters compared to real measured cardiac parameters.

Fig. 7A-7D show exemplary graphical plots of various parameters over time during use of a particular blood pump operating at a pump power level ("pwevel") of 4. Fig. 7A includes a graph 701 showing the measured and estimated left ventricular volumes at a particular pump power level (P-level 4). The curve graph 701 includes: an x-axis 702 showing time in seconds, a y-axis 704 showing volume in milliliters, a measured ("true") trace 706 of the measured left ventricular volume, and an estimated ("predicted") trace 708 of the left ventricular volume predicted by the example model based on the aortic pressure and pump flow as shown in fig. 7C and 7D.

Fig. 7B shows a graph 711 showing the measured and estimated stroke volume of the pump. The curve graph 711 includes: an x-axis 712 showing time in seconds, a y-axis 714 showing heart beat volume, a measured ("true") trace 716 of the measured heart beat volume, and an estimated ("predicted") trace 718 of heart beat volume predicted by the example model based on aortic pressure and pump flow as shown in fig. 7C and 7D. Stroke volume is related to cardiac output.

Fig. 7C shows a graph 721 showing the trace of measured aortic pressure over time for a pump as used to predict left ventricular pressure in fig. 7A and stroke volume in fig. 7B. Curve graph 721 includes an x-axis 722 that shows time in seconds, a y-axis 724 that shows aortic pressure in mmHg, and traces of measured aortic pressure ("AOP").

Fig. 7D shows a graph 731 showing a trace of pump flow rate over time for a pump as used to predict left ventricular pressure in fig. 7A and heart beat in fig. 7B. The graph 731 includes an x-axis 732 showing time in seconds, a y-axis 734 showing flow rate in ml/s, and a trace of the measured flow rate. The x-axis of the four curve patterns in fig. 7A-7D are the same.

7A-7D illustrate metrics of an example model to accurately predict left ventricular volume and cardiac output based on input of aortic pressure and flow rate of a pump operating at a constant pump power level.

Fig. 8A-8D show exemplary graphical plots of both left ventricular volume and stroke volume at different pump power levels. Fig. 8A includes a graph 801 illustrating the measured and estimated left ventricular volumes at a particular pump power level of 2(P level 2). The curve graph 801 includes: an x-axis 802 showing time in seconds, a y-axis 804 showing volume in milliliters, a measured ("true") trace 806 of the measured left ventricular volume, and an estimated ("predicted") trace 808 of the left ventricular volume predicted by the example model for a pump operating at a pump power level of 2.

Fig. 8B shows a graph 811 showing the measured and estimated stroke volume for the pump at P-level 2. The curve graph 811 includes: an x-axis 812 showing time in seconds, a y-axis 814 showing heart beat volume, a measured ("true") trace 816 of the measured heart beat volume, and an estimated ("predicted") trace 818 of the heart beat volume predicted by the example model for a pump operating at a pump power level of 2.

Fig. 8C shows a curve graph 821 showing the measured and estimated left ventricular volumes at a pump power level of 3(P level 3). The curve graph 821 includes: an x-axis 822 showing time in seconds, a y-axis 824 showing volume in milliliters, a measured ("true") trace 826 of the measured left ventricular volume, and an estimated ("predicted") trace 828 of the left ventricular volume predicted by the example model for a pump operating at a pump power level of 3.

Fig. 8D shows a curve graph 831 showing the measured heart beat volume and the estimated heart beat volume for a pump at a power level of 3. The curve pattern 831 includes: an x-axis 832 showing time in seconds, a y-axis 834 showing heart beat volume, a measured ("true") trace 836 of the measured heart beat volume, and an estimated ("predicted") trace 838 of the heart beat volume predicted by the example model for a pump operating at a pump power level of 3. The x-axis of the four graph plots in fig. 8A-8D are identical and all four graph plots are based on aortic pressure and pump flow as inputs to the model. 8A-8D show that the example model is accurate in predicting cardiac parameters even for blood pumps operating at various power levels.

Fig. 9A-9D show exemplary curve graphs for both left ventricular volume and stroke volume for irregular waveforms. Fig. 9A includes a graph 901 showing measured and estimated left ventricular volumes during the occurrence of an irregular waveform in the heart. The graph 901 shows the measured left ventricular volume and the estimated left ventricular volume. The curve graph 901 includes: an x-axis 902 showing time in seconds, a y-axis 904 showing volume in milliliters, a measured ("true") trace 906 of the measured left ventricular volume, and an estimated ("predicted") trace 908 of the left ventricular volume predicted by the example model based on the irregular waveform.

Fig. 9B shows a graph 911 showing the measured and estimated stroke volumes of a pump during the occurrence of irregular waveforms in the heart. The curve graph 911 includes: an x-axis 912 showing time in seconds, a y-axis 914 showing heart beat volume, a measured ("true") trace 916 of the measured heart beat volume, and an estimated ("predicted") trace 918 of the heart beat volume predicted by the example model based on irregular waveforms.

Fig. 9C shows a third curve graph 921 showing the measured and estimated left ventricular volumes during the occurrence of an irregular waveform in the heart. Fig. 9C shows predicted and true traces of left ventricular volume during operation of the blood pump at pump power level 8 (P-level-8). Curve graph 921 includes: an x-axis 922 showing time in seconds, a y-axis 924 showing volume in milliliters, a measured ("true") trace 926 of the measured left ventricular volume, and an estimated ("predicted") trace 928 of the left ventricular volume predicted by the example model based on the irregular waveform.

Fig. 9D shows a fourth curve diagram 931 showing the measured heart beat volume and the estimated heart beat volume during the occurrence of irregular waveforms in the heart and during operation of the blood pump at P level 8. The curve pattern 931 includes: an x-axis 932 showing time in seconds, a y-axis 834 showing heart beat volume, a measured ("true") trace 936 of the measured heart beat volume, and an estimated ("predicted") trace 938 of the heart beat volume predicted by the example model based on irregular waveforms. The x-axis of the four graph plots of fig. 9A-9D are identical, and the four graph plots in fig. 9A-9D are all based on aortic pressure and pump flow as inputs to the model. Fig. 9A-9D show that the example model is accurate in predicting cardiac parameters even for irregular waveforms.

By creating a model that correlates blood pump parameters to cardiac parameters based on a first patient population, and applying the model to a second patient population, cardiac parameters can be accurately estimated in the second patient population without the use of additional measurement catheters or other diagnostic devices. Estimating cardiac parameters without the use of additional devices may be more efficient and also safer for some patients, as the additional devices may occupy additional space in the vascular system and/or interfere with the operation of cardiac support devices, such as blood pumps. Machine learning algorithms may be used to construct models of measured cardiac parameters and one or more measurable parameters of a blood pump or other medical device based on data from a large number of patients having various characteristics. By considering a wide range of characteristics in the model, an accurate model can be developed that helps predict the cardiac parameters of subsequent patients. For example, characteristics such as gender, weight, disease state, cardiac outcome and diagnosis may be taken into account in the development of the model.

Various systems may be configured to perform the steps of developing and applying models as described above. For example, the model may be developed and/or implemented in a controller of a blood pump. For example, one or more models derived as described above may be stored in the memory of the controller. The controller may include one or more processors configured to drive and control the blood pump and to provide information to and/or receive information from a healthcare professional via a display. The one or more processors may access a model stored in the memory, receive blood pump parameter measurements from the blood pump as inputs, and extract estimated cardiac parameters from the model using the blood pump parameters. The one or more processors may then display the estimated cardiac parameters and other health information on a display.

The model describes cardiac parameters in terms of measurable pump parameters, such as pump speed, flow or pressure, and enables interpretation of details of pump function in the heart to understand cardiac function of the heart without the need for additional diagnostic tools, such as additional catheters. The pump performance and pressure signals measured at the blood pump may be used to estimate cardiac output based on the model. This allows the left ventricular volume or other cardiac parameters of a patient to be understood and predicted based on pump parameters that are readily extracted from a blood pump device that provides cardiac support.

The foregoing merely illustrates the principles of the disclosure and these devices may also be practiced in other embodiments than those described, which are presented for purposes of illustration and not of limitation. It should be understood that the methods disclosed herein, while shown for use with an automated ventricular assist system, may also be applied to systems that will be used with other automated medical systems.

Variations and modifications will occur to those skilled in the art upon review of the present disclosure. The disclosed features may be implemented in any combination and subcombination (including multiple dependent combinations and subcombinations) 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. In addition, certain features may be omitted or not implemented.

Examples of changes, substitutions, and alterations are ascertainable by one skilled in the art and could be made without departing from the scope of the information disclosed herein. All references cited herein are incorporated by reference in their entirety and as part of this application.