A method for implantable ventricular assist system failure prediction using data envelopment analysis

Abstract

The application discloses a kind of to be implanted with data envelopment analysis and carry out ventricular assist system failure prediction method.The method is based on the weekly operation state of implantable ventricular assist system and constructs independent decision unit, obtains the original multidimensional data covering equipment, physiology and clinical dimension.After the original data are carried out missing value filling, abnormal value correction and standardization preprocessing, input by BCC model and super-efficiency DEA model constitute combination model and carry out efficiency measurement.BCC model is used to preliminarily screen out abnormal decision unit, and the optimal super-efficiency value is calculated by super-efficiency DEA model.Based on the super-efficiency value and its change trend along with monitoring period, combined with the preset two-stage early warning threshold, envelope analysis is carried out, and graded failure early warning signal is generated.Projection analysis is carried out on the unit triggering early warning, and the input index redundancy rate is calculated to sort and locate the failure cause.The application realizes early and accurate prediction and cause quantification analysis of LVAD failure.

Description

Technical Field

[0001] This invention relates to the field of medical device failure prediction technology, and in particular to a method for failure prediction of implantable ventricular assist systems using data envelopment analysis. Background Technology

[0002] End-stage heart failure is the final stage of various cardiovascular diseases, affecting a large number of patients in my country, while heart transplant donors are scarce. Implantable ventricular assist devices (LVADs), as implantable artificial hearts, can significantly prolong patient survival. However, LVADs operate at high speeds within the body for extended periods, facing various potential failure risks, with a high failure rate within one year post-operation.

[0003] Traditional LVAD fault prediction methods rely on fault tree analysis (FTA) to accurately understand complex mechanisms, which is difficult to achieve in a post-implantation environment. Failure Mode and Effects Analysis (FMEA) depends on subjective expert ratings, which have low reliability. Machine learning-based methods (SVM, LSTM, etc.) have stringent data quality requirements and are black-box models, which are prone to misjudgments such as "normal indicators but abnormal system", and cannot assess the operational status of LVAD from the overall system level.

[0004] Data Envelopment Analysis (DEA), as a non-parametric efficiency evaluation method, is still in its infancy in the field of LVAD. The main bottlenecks are: lack of a multi-dimensional indicator system adapted to the operating characteristics of LVAD, lack of preprocessing methods for clinical data, and lack of scientific graded early warning thresholds. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting faults in implantable ventricular assist systems using data envelopment analysis, in order to solve the problems mentioned in the background art.

[0006] This invention provides a method for predicting faults in implantable ventricular assist systems using data envelopment analysis, comprising the following steps: An independent decision-making unit is constructed based on the weekly operating status of each implantable ventricular assist system (LVAD), and the raw multi-dimensional data of the decision-making unit during the monitoring period is obtained. The raw multi-dimensional data includes input indicator data and output indicator data. The original multidimensional data is preprocessed to obtain standard multidimensional data; The standard multi-dimensional data is input into a preset DEA combined model for efficiency calculation. The DEA combined model includes a first-stage BCC model and a second-stage super-efficiency DEA model. First, the BCC model is used to perform an initial screening of operational efficiency to identify abnormal decision units. Then, the abnormal decision units are input into the super-efficiency DEA model to calculate the optimal super-efficiency value. Based on the optimal super-efficiency value and its changing trend with the monitoring period, envelope analysis is performed in conjunction with the preset two-level early warning thresholds to generate fault early warning signals of different levels. Projection analysis is performed on the decision unit that triggers the fault warning signal to calculate the redundancy rate of each input indicator, and the fault causes are sorted and located based on the magnitude of the redundancy rate.

[0007] Optionally, the input index data covers three dimensions: equipment, physiology, and clinical, specifically including: average motor operating current, blood pump speed fluctuation rate, average equipment power consumption, inlet / outlet pressure fluctuation rate, patient blood viscosity, tissue temperature at the implantation site, patient heart rate fluctuation rate, mean arterial pressure, anticoagulant drug dosage, equipment calibration cycle duration, frequency of clinical maintenance operations, and patient activity level. The output indicators include: effective pumping flow rate compliance rate, operational stability score, and percentage of fault-free duration.

[0008] Optionally, the data preprocessing specifically includes: Calculate the missing data rate for each indicator in the original multidimensional data; When the data missing rate is less than 2%, linear interpolation is used to supplement the missing values; When the data missing rate is between 2% and 5%, the K-nearest neighbor interpolation method is used to supplement the missing values; When the data missing rate is higher than 5%, the corresponding decision unit is removed. Outliers in the original multidimensional data are identified using the 3σ principle, and the outliers are corrected using the median replacement method. The corrected data is dimensionless by using the Z-score standardization method to obtain the standard multidimensional data.

[0009] Optionally, the initial screening of operational efficiency using the BCC model specifically involves: An input-oriented BCC model is constructed, and the overall technical efficiency value (PTE) of the decision-making unit is calculated by solving a linear programming problem. If the overall technical efficiency value (PTE) is less than 1, the decision-making unit is determined to be an abnormal decision-making unit and enters the second stage.

[0010] Optionally, calculating the optimal superefficiency value specifically involves taking the first value to be evaluated... The decision units are removed from the reference set by solving the following linear programming objective function: The constraints are: in, The function is for finding the minimum value; The first one to be solved The superefficiency value of each decision-making unit; It is a non-Archimedean infinitesimal; For the first Slack variables of input indicators; For the first Slack variables for output indicators; and These are the quantities of input and output indicators, respectively. The total number of abnormal decision-making units that have entered the second stage; For the first Weight variables for each decision-making unit; and These are the current numbers to be evaluated. The first decision-making unit The input indicator value and the first Output indicator value; For the first The first reference decision unit Input indicator values; For the first The first reference decision unit Output indicator value.

[0011] Optionally, the preset two-level warning thresholds include: Level 1 warning threshold: When the optimal super-efficiency value is... A Level 1 warning is triggered when the curve of change with the monitoring cycle is in the range of 0.7 to 1 for three consecutive monitoring cycles, or when a single monitoring cycle reaches the range of 0.5 to 0.7. Level 2 warning threshold: when the optimal super-efficiency value If the curve of the change with the monitoring cycle is in the range of 0.5 to 0.7 for three consecutive monitoring cycles, or if a single monitoring cycle is less than 0.5, a level II warning is triggered.

[0012] Optionally, the projection analysis includes calculating the current trigger warning number. The first decision-making unit Target values ​​for each input indicator and redundancy : The target value ; The redundancy ; in, For the first The first decision-making unit The actual observed values ​​of the input indicators; The optimal superefficiency value of the decision unit calculated by the DEA combined model; The first DEA combination model calculated for the DEA combination model The optimal slack variable value for each input indicator.

[0013] Optionally, the specific formula for calculating the redundancy rate of each of the input indicators is as follows: in, For the first Redundancy rate of input indicators.

[0014] Optionally, the sorting and locating of fault causes includes: Each of the aforementioned input indicators is calculated according to the aforementioned redundancy rate. The absolute values ​​are sorted to form a fault cause ranking table; The input indicator with the largest absolute value of redundancy in the fault cause ranking table is selected as the biggest cause of the fault, and the fault cause ranking table and the redundancy amount are output as clinical intervention data.

[0015] Optionally, the method further includes a parameter closed-loop adaptive optimization step based on virtual DEA calculations, specifically including: Object identification step: Determine whether the input indicator with the largest absolute value of redundancy in the fault cause ranking table belongs to the equipment dimension. If it belongs to the equipment dimension, activate the equipment parameter self-correction program and mark the input indicator as an object to be optimized. Virtual construction step: Based on the target value corresponding to the object to be optimized Compared with the actual observed values The difference between the values ​​is used to calculate the theoretical recommended value of the object to be optimized, combined with the preset safety damping coefficient. The theoretical recommended value is then used to construct a virtual decision-making unit. The data of the other dimensions in the virtual decision-making unit, except for the object to be optimized, are consistent with the current monitoring cycle. Simulation verification steps: Input the virtual decision-making unit as the unit to be evaluated into the DEA combined model for secondary calculation to obtain the virtual superefficiency value of the virtual decision-making unit; Closed-loop control steps: Determine whether the virtual super-efficiency value is greater than the optimal super-efficiency value of the current monitoring period; if so, generate a control command based on the theoretically recommended value and send it to the implantable ventricular assist system to adjust the operating parameters of the implantable ventricular assist system; if not, block the adjustment and generate an artificial intervention request signal.

[0016] The present invention has achieved the following beneficial effects: This invention overcomes the limitations of traditional fault prediction methods that focus only on single device parameters and ignore human-machine coupling effects by constructing a comprehensive indicator system covering three dimensions: device operation, patient physiology, and clinical maintenance. It objectively reflects the true operational status of the implantable ventricular assist system from a holistic system perspective. By introducing a targeted data preprocessing mechanism, this invention solves the common problems of missing data and noise interference in clinical follow-up data, eliminates dimensional differences between indicators, and ensures the quality and consistency of the assessment data. Compared to fault tree analysis relying on subjective experience or machine learning "black box" models lacking interpretability, this invention employs a two-stage measurement strategy combining the BCC model and the ultra-efficient DEA model. This not only achieves quantitative grading of system operational health but also accurately calculates the redundancy rate of each input indicator through projection analysis, thereby intuitively analyzing the specific causes of faults and effectively solving the clinical misjudgment problem of "indicators appearing normal but the system being abnormal." Furthermore, the two-level early warning mechanism, which is based on the trend of the optimal superefficiency value, closely matches the gradual nature of fault development and can sensitively capture weak signals such as those in the early stages of thrombosis. This provides clinicians with precise decision support that includes risk level and specific causes, significantly improving the safety and versatility of postoperative management.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for predicting faults in implantable ventricular assist systems using data envelopment analysis, provided in an embodiment of the present invention. Figure 2 A schematic diagram of the input-output index system architecture for LVAD fault prediction DEA constructed in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the trend of the superefficiency value change of a patient before and after the occurrence of a fault, and the triggering of two-level early warnings in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 As shown, the method for predicting faults in implantable ventricular assist systems using data envelopment analysis provided by this invention specifically includes the following steps: Step S1: Construct an independent decision unit (DMU) for each implanted artificial heart. Using the weekly LVAD operating status of each patient as an independent decision unit, acquire raw multidimensional data based on the decision unit.

[0022] In this embodiment, before constructing an independent decision-making unit based on each implantable artificial heart, the method further includes selecting end-stage heart failure patients implanted with third-generation magnetically levitated centrifugal LVADs (such as HeartMate 3 or HeartWare) as research subjects. It is understood that the weekly operating status of each patient's LVAD (with a one-week cycle) is used as an independent decision-making unit. The reason for choosing a weekly monitoring period is that the development of potential LVAD malfunctions (such as tubing blockage or motor wear) is gradual, and daily data is easily affected by factors such as patient diet, activity, emotions, and temporary medication, resulting in large fluctuations (coefficient of variation of approximately 15%–20%), making it difficult to reflect the true operating status. Weekly data can effectively filter out short-term random interference, more accurately reflect the long-term trend of device operation (coefficient of variation reduced to 5%–8%), and is highly consistent with the routine clinical follow-up cycle (weekly outpatient follow-up).

[0023] In this embodiment, acquiring raw multi-dimensional data specifically includes constructing an input indicator system and an output indicator system covering three dimensions: equipment, physiology, and clinical aspects. The specific structure is as follows: Figure 2 As shown.

[0024] The input indicator data includes equipment-level indicators and physiological-level indicators. The equipment-level indicators primarily focus on the mechanical and electrical operating status of the LVAD itself, specifically including the average motor operating current, blood pump speed fluctuation rate, and inlet / outlet pressure fluctuation rate. The average motor operating current is calculated by collecting current data within a single DMU cycle using the LVAD's built-in current sensor at a fixed frequency (e.g., once / second) and taking the arithmetic mean. The blood pump speed fluctuation rate is calculated using the following formula: The physical meaning of this indicator lies in reflecting the rotational stability and smoothness of the mechanical load of the magnetic levitation rotor. Abnormal fluctuations are usually a direct manifestation of early thrombus adhesion leading to rotor dynamic imbalance or motor electrical faults. The rated speed is the standard value of the normal operating speed preset according to the model of the implantable ventricular assist system. The formula for calculating the inlet / outlet pressure fluctuation is as follows: The physical meaning of this indicator is the degree of oscillation of fluid pressure at the inlet and outlet of the LVAD pump body. It is used to measure the hemodynamic coupling stability between the pump and the heart. High volatility often indicates the risk of ventricular sucking effect (i.e., excessive negative pressure at the inlet) or outflow pipe twisting (sudden change in outlet resistance). The standard pressure value is the ideal fluid pressure constant preset by the equipment.

[0025] Among the physiological dimensions, the indicators include patient heart rate variability and mean arterial pressure (MAP). Patient heart rate variability is used to measure the regulatory capacity of the patient's autonomic nervous system and the stress reserve of the heart under mechanical assistance. Its calculation formula is as follows: Resting heart rate is the patient's baseline heart rate recorded during inactivity. Mean arterial pressure (MAP) measures the average driving pressure that delivers blood to peripheral tissues and organs over a cardiac cycle; it is a key indicator reflecting systemic organ perfusion, and its calculation formula is... .

[0026] The output metrics include the effective pumping flow rate compliance rate, operational stability score, and fault-free duration percentage. The formula for calculating the effective pumping flow rate compliance rate is as follows: The physical meaning of this indicator lies in assessing whether the pumping output of the LVAD continuously meets the patient's metabolic needs, directly reflecting the hemodynamic effectiveness of the auxiliary system. The clinical standard flow rate is the lowest effective flow rate threshold set based on the patient's body surface area and equipment parameters. Operational stability is quantified using a weighted scoring method, selecting three secondary indicators: speed fluctuation rate (weight 0.4), pressure fluctuation rate (weight 0.3), and current fluctuation rate (weight 0.3). Scores are given based on the ratio of actual fluctuation values ​​to a safety threshold; the safety threshold is the maximum safe fluctuation range allowed for the corresponding indicator, set based on clinical expert experience. The formula for calculating the percentage of fault-free duration is as follows: The physical meaning of this indicator lies in quantifying the continuity of equipment operation and its anti-interference ability, reflecting the overall reliability and robustness of the system under complex physiological environments.

[0027] In this way, by constructing a comprehensive indicator system, the system operation status of LVAD can be fully reflected, overcoming the limitations of traditional methods that only consider equipment parameters.

[0028] Step S2: Data preprocessing.

[0029] In this step, data preprocessing is performed on the original multidimensional data. This preprocessing includes missing value handling, outlier handling, and dimensionless processing, resulting in standard multidimensional data. Specifically, missing value handling includes: calculating the missing value rate for each indicator. If the missing value rate is less than 2% (e.g., mean arterial pressure, PT, Scr), linear interpolation is used to supplement missing values. This method estimates missing values ​​based on the continuity of the data and the linear relationship between adjacent data points. If the missing value rate is between 2% and 5% (e.g., motor vibration frequency), K-nearest neighbor interpolation (KNN) is used to supplement missing values. The indicator data of the K patients (preferably K=5) whose physiological characteristics are most similar to the target patient are selected as references, and their average value is calculated as the missing value. If the missing value rate is greater than 5%, the decision unit is directly removed to ensure data quality.

[0030] Outlier handling includes: employing The principle is to identify outliers, that is, for index data that follows a normal distribution, values ​​exceeding the mean will be outliers. Data points exceeding one standard deviation are identified as outliers. For identified outliers, a median replacement method is used for correction; that is, the median of this indicator across all decision-making units is used to replace the outlier, thus avoiding interference from outliers in efficiency measurement results. For example, in this embodiment, 127 outliers were identified and corrected. Dimensionless processing includes: using the Z-score standardization method to process all input and output indicators; the standardization formula is: In the formula, The standardized index value, These are the original index values. This is the arithmetic mean of the indicator across all decision-making units. This represents the standard deviation of the indicator across all decision-making units.

[0031] In this way, comprehensive preprocessing can eliminate the dimensional differences between different indicators (such as rotational speed r / min and energy consumption Wh), ensuring that all indicators have equal weight in the DEA model.

[0032] Step S3: Efficiency calculation based on the DEA combined model.

[0033] The DEA combined model employs a method that combines the BCC model with the super-efficient DEA model. In this embodiment, the efficiency calculation is divided into two progressive stages: Phase 1: Initial screening of LVAD running efficiency based on the BCC model; In this embodiment, an input-oriented BCC model (variable scale return assumption) is selected to accommodate the varying scale of LVAD operation under different patient physiological states. The pure technical efficiency (PTE) is calculated by solving a constrained linear programming problem.

[0034] Specifically, regarding the first One LVAD operation decision unit ( ), construct the objective function .

[0035] The constraints include: in, The current problem to be solved is the... Decision-making unit The overall technical efficiency value, The total number of decision-making units; For the first Weight variables of each decision unit ( ); and The first The first decision-making unit The input indicator value and the first Output indicator value; and The first to be evaluated The first decision-making unit The input indicator value and the first Output indicator value; For the number of input indicators, This refers to the quantity of output indicators.

[0036] Based on the above solution, if This indicates that Purely technically effective; if This indicates that If purely technically ineffective, it is determined to be an abnormal DMU. After passing the first stage, abnormal DMUs with a PTE < 1 are selected and proceed to the second stage.

[0037] Phase Two: Fault severity classification based on the super-efficient DEA model; In this embodiment, the abnormal DMUs selected in the first stage are substituted into the super-efficient DEA model. The core of the super-efficient DEA model is to evaluate the... Remove from the reference set, and use the rest. A linear combination of DMUs is used to construct the effective frontier.

[0038] Specifically, regarding the first For each abnormal decision-making unit, construct the objective function: in, The first one to be solved The optimal superefficiency value of an abnormal decision-making unit. The total number of abnormal decision-making units that have entered the second stage; For the first The weight variables of each decision unit, and in the summation operation China satisfies That is, the unit to be evaluated is included in the calculation. Removed from the reference set; For the first Item input slack variable, For the first Output slack variables; and These are the quantities of input and output indicators, respectively; For non-Archimedean infinitesimals. The subscripts in the formula. Index of input indicators ( ), subscript Index of output indicators ( ), subscript For reference, the index of the centralized decision-making unit ( and ).

[0039] The constraints are: Through the above solution, the optimal superefficiency value of each anomalous DMU is obtained. and optimal input slack variables .

[0040] Step S4: Fault risk classification and early warning.

[0041] Based on the optimal superefficiency value selected in the previous step The system plots the weekly data curve of the superefficiency value for each decision-making unit within each weekly unit. Based on the pre-set two-level early warning thresholds, it performs envelope analysis of the superefficiency value curve, conducts hierarchical management of fault risk, and generates fault early warning signals.

[0042] In this embodiment, the graded early warning threshold is set as follows: Based on its optimal superefficiency value Fault prediction and classification based on size: A value between 0.7 and 1 indicates a mild abnormality (a precursor to a fault); A score of 0.5-0.7 indicates moderate abnormality. This indicates a severe abnormality.

[0043] Furthermore, based on statistical analysis of historical clinical data and considering the progressive nature of LVAD malfunction development, two levels of early warning thresholds were established: Level 1 warning threshold (potential risk warning): If the efficiency value curve is exceeded... If the value is in the mildly abnormal range of 0.7-1 for three consecutive monitoring cycles, or reaches the moderately abnormal range of 0.5-0.7 for a single monitoring cycle, a Level 1 warning is triggered. At this time, a yellow warning signal is generated and pushed to the clinician's workstation, indicating that there is a potential fault in LVAD (such as early thrombosis) and the fault is in an early stage of development.

[0044] Level 2 warning threshold (high-risk warning): If the efficiency value curve is exceeded... If the value is within the moderately abnormal range of 0.5-0.7 for three consecutive monitoring cycles, or reaches the severely abnormal range of <0.5 for a single monitoring cycle, a level-two warning is triggered. At this time, a red warning signal is generated, and doctors are notified via SMS or other means, indicating a high risk of malfunction and the need for immediate intervention.

[0045] For example, in an empirical study based on HeartMate 3 LVAD, taking patient A (male, 64 years old, ischemic cardiomyopathy) as an example, the early warning process and the trend of change in excess efficiency value are as follows: Figure 3 As shown, the details are as follows: Scenario 1: Stable operation period. 1-15 months post-surgery, exceeding efficiency value. The system remained stable within the 1.10-1.25 range, with normal operating parameters (speed 3200-3300 r / min, energy consumption 55-65 Wh / week) and no adverse events occurring.

[0046] Scenario 2: Level 1 warning triggered. 15 months post-surgery, 3rd week. The blood glucose level decreased from 1.15 to 1.02 in week 1 (a decrease of 11.3%); in week 4 it decreased to 0.98 (a cumulative decrease of 14.8%); and in week 5 it decreased to 0.97 (a cumulative decrease of 15.7%), meeting the condition of being within the 0.7-1 range for three consecutive monitoring periods with a cumulative decrease of ≥15%. The system generated a yellow warning signal. Clinical response: The physician increased the warfarin dose from 150 mg / week to 167.5 mg / week and recommended a low-fat diet.

[0047] Scenario 3: Level 2 warning triggered. 16 months post-surgery, 2nd week. The value plummeted from 0.97 to 0.72, a single drop of 24.2% (meeting the condition of a single drop ≥ 20% or < 0.5). The system generated a red warning signal.

[0048] Scenario 4: Diagnosis and Intervention. Subsequent clinical examination confirmed thrombosis and obstruction of the LVAD outflow duct (55% stenosis, energy consumption increased to 92Wh / week). Intensive anticoagulation therapy (warfarin 180mg / week) + low molecular weight heparin calcium thrombolysis was implemented.

[0049] Results: After 2 weeks of treatment, The blood flow output recovered to 0.92, and energy consumption dropped to 70Wh / week, avoiding the need for a second surgery.

[0050] Step S5: Locating the cause of the fault.

[0051] Projection analysis is performed on the decision-making units that trigger the early warning signal to calculate the target value, redundancy, and redundancy rate of each input indicator, and the indicators are sorted according to the absolute value of the redundancy rate to form a fault cause ranking table.

[0052] In this embodiment, the specific calculation formula for the projection analysis is as follows: Target input value: ; Redundancy input: ; Redundancy rate: ; in, For the first The actual observed values ​​of the inputs, The optimal input slack variables are determined. The redundancy rate of all 12 input indicators is calculated. Calculate them and sort them by absolute value; the largest one is the biggest contributing factor.

[0053] In this embodiment, a fault cause determination rule base is established based on redundancy analysis to achieve accurate fault type localization. Specifically, this includes: Fault type 1: Tube blockage / thrombosis.

[0054] Judgment criteria: If the average power consumption of the equipment is used for redundancy And the average blood output insufficiency rate Furthermore, the pressure difference between the inlet and outlet of the turbine pump is greater than 25 mmHg.

[0055] Empirical Case: Patient A (mentioned above) Projection analysis showed that the energy redundancy rate was 28%, the pressure redundancy rate was 22%, and the flow insufficiency rate was 32%, which fully met the rule, and the clinical diagnosis was outflow duct thrombosis.

[0056] Fault type 2: Wear and tear on motor components.

[0057] Judgment criteria: If the motor vibration frequency incorporates redundancy. Furthermore, the PTE remained below 0.7, accompanied by fluctuations in rotational speed. .

[0058] Empirical Case: Patient B ( The vibration frequency redundancy was 25%, and the speed fluctuation was 15%, which was clinically diagnosed as motor bearing wear.

[0059] Fault type 3: Patient's physiological state is unbalanced.

[0060] Judgment criteria: If the mean arterial pressure input redundancy Furthermore, the scale efficiency is <0.8, accompanied by an INR deviation from the target value >0.5.

[0061] Empirical Case: Patient C ( The patient had a blood pressure redundancy of 18% and a PT redundancy of 16%, and was clinically diagnosed with hypertensive crisis complicated by coagulation abnormalities.

[0062] Fault type 4: Poor device-patient fit.

[0063] Judgment criteria: If the LVEF output is insufficient... Furthermore, the 6MWD output shortfall rate .

[0064] Empirical Case: Patient D ( With LVEF less than 20% and 6MWD less than 26%, the patient was clinically diagnosed with deterioration of cardiac function due to excessively high engine speed setting.

[0065] Furthermore, personalized intervention suggestions are generated based on the above fault location: For tubal obstruction, we recommend enhanced anticoagulation therapy combined with thrombolytic therapy, with an empirical efficacy rate of 92.5%. For motor wear, it is recommended to reduce the LVAD operating speed by 5%–8% and shorten the monitoring cycle, with an empirical effectiveness rate of 88.3%. For physiological imbalances, it is recommended to adjust the dosage of anti-heart failure drugs and optimize blood pressure control (target MAP 75-85 mmHg), with an empirical efficacy rate of 90.7%.

[0066] In addition, to verify the universality of the method, a cross-device model validation study based on the HeartWare-type LVAD was conducted. Twenty-five patients (a total of 2680 DMUs) who had received HeartWare-type magnetic levitation centrifugal LVADs were selected as subjects. The normal ranges of some device parameters were adjusted (e.g., the normal operating speed range is 1500-3000 r / min, and the normal energy consumption range is 40-80 Wh / week).

[0067] The verification results show that the average superefficiency value is 1.06 ± 0.14. The sensitivity of the first-level warning threshold is 80.56% (AUC = 0.882), the sensitivity of the second-level warning threshold is 86.11% (AUC = 0.921), and the fault cause localization accuracy rate is 96.77%. This indicates that the present invention is applicable to different models of third-generation magnetic levitation centrifugal LVADs and has good versatility and universality.

[0068] Furthermore, in order to address the energy efficiency drift problem caused by the lack of active adaptation in implantable ventricular assist systems during long-term operation, as well as the technical challenges of delayed manual intervention, experience-dependent parameter adjustment, and high risk in traditional open-loop control modes, this invention, based on the aforementioned fault cause localization, further introduces and executes a parameter closed-loop adaptive optimization step calculated by virtual data envelopment analysis (DEA).

[0069] In this embodiment, the parameter closed-loop adaptive optimization step based on virtual DEA calculation is implemented using a high-performance heterogeneous dual-core processing unit built into the implantable ventricular assist system. This processing unit includes a main processor (e.g., based on an ARM Cortex-M7 architecture) responsible for logic control and safety monitoring, and a digital signal coprocessor (DSP) dedicated to matrix operations and linear programming solutions. Data exchange between the main processor and the coprocessor occurs via a high-speed internal bus, ensuring that the execution of the optimization algorithm does not consume hardware resources required for life support functions. Logically, this step closely follows the aforementioned fault cause localization step and specifically includes four progressively layered core sub-processes: object identification and authorization verification, virtual decision unit construction and parameter inversion, simulation verification and boundary assessment, and closed-loop control command issuance and circuit breaker protection.

[0070] The first step is the object identification and permission verification subprocess. After the system generates the fault cause ranking table through the aforementioned steps, the memory temporarily stores various input indicators that lead to low system efficiency in the current monitoring period, along with their corresponding redundancy weights. However, those skilled in the art will understand that the input indicator system of the DEA model is a mixed set encompassing equipment physical parameters, patient physiological parameters, and clinical maintenance parameters, and not all dimensions of the indicators possess online electrical controllability. For example, if the fault cause ranking table shows that the maximum cause is the patient's blood viscosity or mean arterial pressure, these two indicators are physiological properties of the organism itself, and the controller of the implantable device cannot directly physically adjust the patient's blood viscosity or vascular resistance by changing the motor voltage or frequency; conversely, if the maximum cause is the average current of the motor or the fluctuation rate of the blood pump speed, these are operating parameters of the device itself, and have a physical basis for adjustment through electrical commands.

[0071] Therefore, to construct a logically rigorous and secure control loop, the controller first reads the input indicator with the largest absolute value of redundancy in the fault cause ranking table through the main processor and extracts its unique feature identification code. Subsequently, the processor accesses the controllable parameter mapping database pre-stored in the on-chip read-only memory (ROM). This database defines the attribute labels for each monitoring indicator in the form of a lookup table: indicators involving motor current, speed, levitation magnetic field bias, power consumption, etc., are all marked as device domain - read-write attributes; indicators involving blood pressure, heart rate, body temperature, coagulation indicators, etc., are all marked as physiological domain - monitoring-only attributes. If the comparison results show that the main cause belongs to the physiological or clinical dimension, the system determines that the current inefficiency of the system stems from uncontrollable external factors. In this case, to avoid the risks of ineffective or even erroneous adjustments, the system will automatically block subsequent parameter correction procedures and directly jump to the manual intervention process, generating a clinical decision support message containing detailed fault data, and sending it to the external monitoring terminal via the radio frequency telemetry module. If the comparison results show that the main cause belongs to the equipment dimension, such as detecting a significant positive redundancy in the average power consumption of the equipment, the system determines that automatic optimization is technically feasible, then activates the equipment parameter self-correction thread, marks the input indicator as the object to be optimized, and locks the actual observed value of the indicator in the current cycle as the benchmark data for subsequent calculations.

[0072] The virtual construction and parameter inversion sub-processes are then executed. After identifying the object to be optimized, the system does not directly adjust the device parameters to the theoretical target value calculated by DEA projection analysis. This is because the target value calculated by DEA projection is based on the theoretical optimal solution derived from mathematical programming, representing the ideal state after eliminating all redundancy. However, in the engineering practice of implantable artificial organs, directly adjusting the operating parameters from the actual value to the theoretical target value in a stepwise manner (e.g., instantly reducing the rotation speed by 500 rpm) is extremely risky. Such drastic parameter changes may lead to abrupt changes in the flow field shear force, inducing red blood cell fragmentation (hemolysis) or causing ventricular wall suction effects. Therefore, this embodiment introduces a dynamic safety damping coefficient (denoted as λ) in the controller algorithm. This coefficient is calculated in real time by the system based on the patient's current activity level and heart rate variability (HRV), and its value is strictly limited to the range of 0.05 to 0.25. Specifically, when the accelerometer detects that the patient is in a resting sleep state, to prevent disturbing the patient's rest and avoid interference from changes in vagal tone, λ is set to a small value (e.g., 0.05) to ensure extremely smooth regulation; when the patient is in a conscious and stable state, λ can be appropriately increased (e.g., 0.2) to improve regulation efficiency. The system is based on the formula: The processor calculates the intermediate recommended value for this adjustment. Then, the main processor allocates a separate computation buffer in SRAM to construct a virtual decision unit. In this virtual data vector, the processor forcibly replaces the values ​​of the corresponding dimension of the object to be optimized with the theoretical recommended value just calculated, thereby strictly locking and copying the data of all other dimensions (including the patient's real-time blood pressure, heart rate, drug concentration, and other device parameters) to ensure complete consistency with the current actual sampled values.

[0073] The simulation verification and boundary assessment sub-processes are then executed. After the virtual decision-making unit is constructed, the system treats it as a real sample to be evaluated and sends it to the DSP coprocessor via the internal bus. The DSP coprocessor calls the pre-built DEA combined model algorithm (including the BCC model and the hyper-efficiency model) to perform a secondary efficiency calculation on the virtual decision-making unit. It should be noted that in this calculation, the reference set used as the comparison benchmark still consists of standard healthy samples from the historical database and remains unchanged. The DSP coprocessor obtains the virtual hyper-efficiency value of the virtual decision-making unit by quickly solving a linear programming problem. This value represents the expected comprehensive technical efficiency level of the system under the assumed optimization conditions. At this point, the system compares and judges this virtual hyper-efficiency value with the optimal hyper-efficiency value of the current actual monitoring period. If the virtual super-efficiency value does not show a significant increase or even decreases, it indicates that the current parameter adjustment scheme cannot bring global benefits under complex multi-dimensional coupling effects (for example, reducing the speed may reduce power consumption, but it may lead to a significant decrease in effective pumping flow, thus dragging down overall efficiency). In this case, the system will decisively determine that the optimization is ineffective, terminate the adjustment, and release the locked state of the object to be optimized. If the virtual super-efficiency value is significantly higher than the current actual value, and the increase exceeds the preset minimum gain threshold (e.g., 0.5%), it indicates that the adjustment direction is correct and effective. The system then enters the boundary security review stage. The controller will call the underlying security logic library to verify whether the theoretically recommended value falls within the absolute safety range allowed by the device hardware (for example, for a magnetic levitation centrifugal pump, the speed must not be lower than 2400 rpm to prevent backflow, nor higher than 9000 rpm to prevent hemolysis), and will also verify whether the expected parameter change rate exceeds the motor's response limit. Only when the virtual super-efficiency value increases and passes all hardware boundary reviews will the theoretically recommended value be finally confirmed as a valid control command parameter.

[0074] Finally, the closed-loop control command issuance and circuit breaker protection sub-processes are executed. Once the optimization scheme passes all verifications, the system will enter the substantive physical intervention stage. Based on the confirmed theoretical recommended values, the main processor generates hardware control commands that conform to the underlying communication protocol. For example, if the object to be optimized is the blood pump speed, the processor will convert the recommended speed value into the commutation frequency command of the brushless DC motor (BLDC) or the duty cycle of the pulse width modulation (PWM) signal. To prevent the control commands from being tampered with by the internal electromagnetic environment during transmission, this embodiment sets up a handshake communication protocol with CRC32 checksum and command readback mechanism. After receiving the command, the motor drive module needs to send the command value back to the main processor for secondary comparison. Only after confirming that it is correct will it start execution. During execution, the motor does not change its state instantaneously, but uses an S-curve smooth ramping algorithm to flexibly transition the operating parameters to the target value within a time window of 30 to 60 seconds. More importantly, during the entire parameter adjustment period, the system's safety circuit breaker mechanism is in the highest alert state. The watchdog timer monitors key output metrics (such as flow rate, pulsation index, and motor power fluctuations) in real time at a frequency of 100Hz. If any unexpected abnormal feedback is detected (e.g., a sudden drop in flow rate below the warning threshold, indicating a possible aspiration effect, or an abnormal spike in motor current), the system will trigger a high-priority hardware interrupt, immediately and forcibly terminating the current adjustment process. It will then roll back the parameters to their initial safe state as quickly as possible, and simultaneously send a red adaptive adjustment failure alarm signal to the external controller, requesting immediate medical intervention.

[0075] The present invention also provides an implantable ventricular assist system fault prediction device, comprising: Data acquisition module: used to build an independent decision unit (DMU) for each implantable artificial heart, and to collect raw data from three dimensions: device, physiological and clinical.

[0076] Preprocessing module: Used to perform missing value interpolation (linear / KNN), outlier correction, and Z-score normalization.

[0077] Efficiency Calculation Module: Used to run the BCC model for initial screening and the DEA model for super-efficiency to calculate the optimal super-efficiency value. .

[0078] Early warning and location module: used to generate early warning signals based on two levels of thresholds, calculate the redundancy rate of indicators through projection analysis, and output a fault cause ranking table.

[0079] Referring to the accompanying drawings, embodiments of the present invention also provide an electronic device, comprising: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the method for predicting faults in implantable ventricular assist systems using data envelopment analysis disclosed in any of the foregoing embodiments.

[0080] The solution provided by the present invention has been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0081] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for implantable ventricular assist system failure prediction using data envelopment analysis, characterized in that, Includes the following steps: An independent decision-making unit is constructed based on the weekly operating status of each implantable ventricular assist system (LVAD), and the raw multi-dimensional data of the decision-making unit during the monitoring period is obtained. The raw multi-dimensional data includes input indicator data and output indicator data. The original multidimensional data is preprocessed to obtain standard multidimensional data; The standard multi-dimensional data is input into a preset DEA combined model for efficiency calculation. The DEA combined model includes a first-stage BCC model and a second-stage super-efficient DEA model. First, the BCC model is used to perform an initial screening of operational efficiency to identify abnormal decision units. Then, the abnormal decision unit is input into the super-efficiency DEA model to calculate the optimal super-efficiency value; Based on the optimal super-efficiency value and its changing trend with the monitoring period, envelope analysis is performed in conjunction with the preset two-level early warning thresholds to generate fault early warning signals of different levels. Projection analysis is performed on the decision unit that triggers the fault warning signal to calculate the redundancy rate of each input indicator, and the fault causes are sorted and located based on the magnitude of the redundancy rate. The input index data covers three dimensions: equipment, physiology, and clinical, specifically including: average motor operating current, blood pump speed fluctuation rate, average equipment power consumption, inlet / outlet pressure fluctuation rate, patient blood viscosity, tissue temperature at the implantation site, patient heart rate fluctuation rate, mean arterial pressure, anticoagulant dosage, equipment calibration cycle duration, frequency of clinical maintenance operations, and patient activity level. The output indicators include: effective pumping flow rate compliance rate, operational stability score, and percentage of fault-free duration. The preliminary screening of operational efficiency using the BCC model specifically involves: An input-oriented BCC model is constructed, and the overall technical efficiency value (PTE) of the decision-making unit is calculated by solving a linear programming problem. If the overall technical efficiency value PTE is less than 1, the decision-making unit is determined to be an abnormal decision-making unit and enters the second stage; The projection analysis includes calculating the first warning currently triggered. The first decision-making unit Target values ​​for each input indicator and redundancy : The target value ; The redundancy ; in, For the first The first decision-making unit The actual observed values ​​of the input indicators; The optimal superefficiency value of the decision unit calculated by the DEA combined model; The first DEA combination model calculated for the DEA combination model The optimal slack variable value for each input indicator.

2. The method for predicting faults in implantable ventricular assist systems using data envelopment analysis according to claim 1, characterized in that, The data preprocessing specifically includes: Calculate the missing data rate for each indicator in the original multidimensional data; When the data missing rate is less than 2%, linear interpolation is used to supplement the missing values; When the data missing rate is between 2% and 5%, the K-nearest neighbor interpolation method is used to supplement the missing values; When the data missing rate is higher than 5%, the corresponding decision unit is removed. Outliers in the original multidimensional data are identified using the 3σ principle, and the outliers are corrected using the median replacement method. The corrected data is dimensionless by using the Z-score standardization method to obtain the standard multidimensional data.

3. The method for predicting faults in implantable ventricular assist systems using data envelopment analysis according to claim 1, characterized in that, The calculation of the optimal superefficiency value specifically involves evaluating the first... The decision units are removed from the reference set by solving the following linear programming objective function: ； The constraints are: ； ； ； ； in, The function is for finding the minimum value; The first one to be solved The superefficiency value of each decision-making unit; It is a non-Archimedean infinitesimal; For the first Slack variables of input indicators; For the first Slack variables for output indicators; and These are the quantities of input and output indicators, respectively. The total number of abnormal decision-making units that have entered the second stage; For the first Weight variables for each decision-making unit; and These are the current numbers to be evaluated. The first decision-making unit The input indicator value and the first Output indicator value; For the first The first reference decision unit Input indicator values; For the first The first reference decision unit Output indicator value.

4. The method for predicting faults in implantable ventricular assist systems using data envelopment analysis according to claim 3, characterized in that, The preset two-level early warning thresholds include: Level 1 warning threshold: When the optimal super-efficiency value is... A Level 1 warning is triggered when the curve of change with the monitoring cycle is in the range of 0.7 to 1 for three consecutive monitoring cycles, or when a single monitoring cycle reaches the range of 0.5 to 0.

7. Level 2 warning threshold: when the optimal super-efficiency value If the curve of the change with the monitoring cycle is in the range of 0.5 to 0.7 for three consecutive monitoring cycles, or if a single monitoring cycle is less than 0.5, a level II warning is triggered.

5. The method for predicting faults in implantable ventricular assist systems using data envelopment analysis according to claim 1, characterized in that, The specific formula for calculating the redundancy rate of each input indicator is as follows: ； in, For the first Redundancy rate of input indicators.

6. The method for predicting faults in implantable ventricular assist systems using data envelopment analysis according to claim 5, characterized in that, The process of sorting and locating the causes of failure includes: Each of the aforementioned input indicators is calculated according to the aforementioned redundancy rate. The absolute values ​​are sorted to form a fault cause ranking table; The input indicator with the largest absolute value of redundancy in the fault cause ranking table is selected as the biggest cause of the fault, and the fault cause ranking table and the redundancy amount are output as clinical intervention data.

7. The method for predicting faults in implantable ventricular assist systems using data envelopment analysis according to claim 6, characterized in that, The method further includes a parameter closed-loop adaptive optimization step based on virtual DEA calculations, specifically including: Object identification step: Determine whether the input indicator with the largest absolute value of redundancy in the fault cause ranking table belongs to the equipment dimension. If it belongs to the equipment dimension, activate the equipment parameter self-correction program and mark the input indicator as an object to be optimized. Virtual construction step: Based on the target value corresponding to the object to be optimized Compared with the actual observed values The difference between the values ​​is used to calculate the theoretical recommended value of the object to be optimized, combined with the preset safety damping coefficient. The theoretical recommended value is then used to construct a virtual decision-making unit. The data of the other dimensions in the virtual decision-making unit, except for the object to be optimized, are consistent with the current monitoring cycle. Simulation verification steps: Input the virtual decision-making unit as the unit to be evaluated into the DEA combined model for secondary calculation to obtain the virtual superefficiency value of the virtual decision-making unit; Closed-loop control steps: Determine whether the virtual super-efficiency value is greater than the optimal super-efficiency value of the current monitoring period; if so, generate a control command based on the theoretically recommended value and send it to the implantable ventricular assist system to adjust the operating parameters of the implantable ventricular assist system; if not, block the adjustment and generate an artificial intervention request signal.

Images