Training data for classification of position of catheter relative to septum

Abstract

The present disclosure relates to position monitoring of medical devices, and more particularly to techniques for enabling automatic monitoring of the position of a catheter relative to a septum. Aspects of the present disclosure include determining training data to be used to train a machine learning algorithm to classify a position of a catheter relative to a septum of a patient. Other aspects of the present disclosure include using a trained machine learning algorithm to classify a position of a catheter relative to a septum of a patient.

Description

Technical Field

[0001] The present disclosure relates to position monitoring of medical devices, and more particularly to techniques for enabling automated monitoring of the position of a catheter relative to a septum. Background Art

[0002] A catheter is a medical device that can deliver substances into the body or remove substances from the body. Specifically, a nasogastric catheter (also called a nasogastric tube) can deliver nutrition, fluids, and medications directly into a patient's stomach.

[0003] Proper positioning of such nasogastric tubes is critical to avoiding complications such as aspiration of delivered contents into the lungs when the tube is not inserted deeply enough, or physical penetration of the stomach wall when the tube is inserted too far.

[0004] The nasogastric tube can be equipped with sensors to detect bioelectrical signals in the body. These sensors can measure electrocardiographic and myoelectric signals from nearby muscles in the body. These signals can help assess whether the nasogastric tube is positioned correctly.

[0005] Typically, evaluating the positioning of a nasogastric tube is performed by a clinician. The clinician can perform a nose-earlobe-xiphoid measurement on the patient to estimate the length of the tube that should be inserted into the patient. Once the tube is inserted into the esophagus, the clinician can monitor the detected ECG and EMG signals to determine if the tube is positioned correctly and fine-tune the position as needed.

[0006] Accurately assessing catheter positioning is particularly difficult when no EMG signals are detected; this is often the case when the patient is sedated and therefore unable to breathe on their own (often requiring intubation). Detection of EMG signals is further complicated by the fact that ECG signals are typically more significant, e.g., about 1000 times stronger, relative to EMG signals. In such scenarios where EMG signals are difficult to detect, interpretation of the ECG signals is required. However, inexperienced clinicians may have difficulty finding and maintaining correct catheter position because ECG signals are more difficult to visually identify and often lack the precision required to fine-tune catheter placement. Furthermore, in all cases, monitoring of position relies on the clinician's continued attention.

[0007] Improved methods for assessing catheter positioning are needed. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A schematic diagram showing a position monitoring system for a catheter of a patient;

[0009] Figure 2 schematically illustrates a position monitoring system according to some examples;

[0010] Figure 3 is a flow chart detailing the steps performed by the location monitoring system;

[0011] Figure 4 is a flow chart detailing the steps performed by a location monitoring system in preparing training data for training a machine learning algorithm;

[0012] Figure 5 An example of a bioelectric signal detected by a catheter is shown;

[0013] Figure 6 and Figure 7 An example of labeled training data is shown;

[0014] Figure 8 An example of a machine learning architecture for a location monitoring system is shown;

[0015] Fig. 9 is a flow chart detailing the steps performed by the location monitoring system when deploying the machine learning algorithm;

[0016] Fig.10 is a flow chart detailing the steps performed by a position monitoring system in determining a positioning status of a catheter for a patient;

[0017] Fig.11 is a first example of a pattern recognition function for determining positioning status;

[0018] Fig.12 is a second example of a pattern recognition function for determining positioning status;

[0019] Fig.13 A display device for displaying information about the position of a catheter to a clinician is schematically shown. DETAILED DESCRIPTION

[0020] The object of the present disclosure is to determine the correct placement of an esophageal catheter in a patient. According to the present invention, in some embodiments of the present invention, this object is achieved by a method and system for training and deploying a machine learning algorithm that can be used to locate a catheter relative to the patient's diaphragm. Specifically, some embodiments relate to methods and systems for determining training data for the machine learning algorithm, and some embodiments relate to methods and systems for classifying the position of a catheter relative to the patient's diaphragm.

[0021] In a first aspect, a method for determining training data is provided, the training data to be used to train a machine learning algorithm to classify the position of a catheter relative to a patient's diaphragm. The method comprises the following steps: receiving a set of bioelectric signals detected by a catheter, the catheter carrying a plurality of electrodes at corresponding positions along the length of the catheter so that the electrodes are located at corresponding different distances from the patient's diaphragm, the plurality of electrodes being divided into a plurality of electrode pairs, each signal being detected by one of the plurality of electrode pairs, each signal comprising an electrocardiographic (ECG) component; identifying one or more first bioelectric signals from the set of bioelectric signals, the first bioelectric signals being detected by an electrode pair on the catheter determined to be correctly positioned relative to the diaphragm; dividing the set of bioelectric signals into at least two subsets of bioelectric signals, each bioelectric signal subset comprising: A set includes one or more bioelectric signals and corresponds to a corresponding group of electrodes associated with the detection of the one or more bioelectric signals in the bioelectric signal subset, wherein the division is performed so that each group of electrodes is a sequence of electrodes placed continuously along the length of the catheter; labeling each subset of the multiple subsets, wherein the labeling includes: labeling a subset that includes at least one of the first bioelectric signals as a signal subset detected from a correctly positioned electrode, and labeling a subset that does not include any of the first bioelectric signals as a signal subset detected from an incorrectly positioned electrode; and, including the bioelectric signal subsets and their corresponding labels in the training data.

[0022] Typically, the available bioelectric signal set is recorded with the catheter in the correct position, so the dataset is unbalanced due to the lack of data in the wrong position. In addition, this type of biological data tends to be high-dimensional and scarce, so the risk of overfitting is also increased.

[0023] The inventors have recognized that a subset of signals from a set of bioelectric signals recorded with the catheter in the correct position may anyway represent a mispositioned catheter because at least some of the signals in the set are typically recorded from electrodes that are in the wrong position (i.e., above or below the diaphragm). The method divides the complete set of signals into several smaller groups (subsets) of signals. Each subset is classified according to its relative position relative to a defined "correct position" (i.e., relative to one or more first signals). A subset includes one or more bioelectric signals. The fewer bioelectric signals there are in each subset, the greater the amount of different training data. Including more than one bioelectric signal in a subset can advantageously exploit dependencies between bioelectric signals in the training data by taking into account information provided by the neighbors of a particular bioelectric signal, which information is generated by the physical arrangement of the electrodes on the catheter.

[0024] Using the techniques described herein, it is possible to increase the amount of training data for each possible classification while balancing the training data.

[0025] In a second aspect, one or more non-transitory computer-readable media are provided, the one or more non-transitory computer-readable media storing instructions executable by one or more processors. The instructions, when executed, cause the one or more processors to perform operations substantially mapped to the steps of the method described in the first aspect.

[0026] In a third aspect, a system is provided. The system includes one or more processors and one or more non-transitory computer-readable media, the one or more non-transitory computer-readable media storing first computer-executable instructions. The non-transitory computer-readable media stores the first computer-executable instructions, which, when executed by the one or more processors, cause the system to perform actions substantially mapped to the operations caused by the instructions according to the second aspect or the steps of the method according to the first aspect.

[0027] The features of the steps of the method according to the first aspect also exist in the instructions stored in the non-transitory computer-readable medium according to the second aspect and the actions performed by the system according to the third aspect. The embodiments of the method can be understood to correspond to the embodiments of the non-transitory computer-readable medium and the embodiments of the system where appropriate.

[0028] In some embodiments, the step of marking a subset of bioelectric signals that does not include the first bioelectric signal as a subset of bioelectric signals detected from an incorrectly positioned electrode includes: determining whether the electrode associated with detecting the one or more bioelectric signals in the subset is positioned above the diaphragm or below the diaphragm; wherein, when it is determined that the electrode is positioned above the diaphragm, the subset of bioelectric signals is marked as a subset of bioelectric signals detected from an electrode located above the diaphragm, and when it is determined that the electrode is positioned below the diaphragm, the subset of bioelectric signals is marked as a subset of bioelectric signals detected from an electrode located below the diaphragm. This allows for a more specific classification of the positioning, i.e., distinguishing a subset having an incorrect position above the diaphragm from a subset having an incorrect position below the diaphragm.

[0029] In some embodiments, the method further comprises: enhancing each bioelectric signal in the subset, wherein enhancing the bioelectric signal comprises at least one of the following: stretching the bioelectric signal over time; compressing the bioelectric signal over time; or changing the amplitude of the bioelectric signal. Advantageously, overfitting of the machine learning algorithm can be avoided.

[0030] In some embodiments, at least one bioelectric signal includes an electromyographic (EMG) component, and the method further includes: applying a filtering algorithm to each bioelectric signal in the set of bioelectric signals, wherein the filtering algorithm is configured to at least reduce the corresponding electromyographic (EMG) component from the corresponding bioelectric signal. For example, a stopband filter can be applied to reduce the EMG component. In other examples, portions of the bioelectric signal including the EMG component are removed, and the bioelectric signal is truncated so that the remaining bioelectric signal does not contain or has a reduced EMG component. This can reduce the likelihood that the trained machine learning algorithm over-relies on detecting the EMG component to accurately classify the position of the catheter, because, for example, the EMG component may not be present in a patient who is unable to breathe on his own. This enables the trained machine learning algorithm to be deployed in a wider range of clinical scenarios.

[0031] In some embodiments, applying the filtering algorithm to the bioelectric signal includes: identifying a plurality of sub-portions of the bioelectric signal, each sub-portion including data detected during a heartbeat of the patient; and calculating an average bioelectric signal based on the plurality of sub-portions of the bioelectric signal. For example, calculating the average value can be a simple method and a relatively computationally inexpensive method to reduce the presence of the EMG component that may only exist in a limited number of sub-portions.

[0032] In some embodiments, each bioelectric signal from the set of bioelectric signals includes data detected during multiple heartbeats of the patient, wherein the method further comprises: identifying, in each bioelectric signal from the set of bioelectric signals, data detected in an intermediate period between two consecutive heartbeats in the multiple heartbeats; and deleting the identified data from the bioelectric signal. This can delete data that carries no information or limited information from which positioning information can be determined, which can improve the training of the machine learning algorithm.

[0033] In some embodiments, the step of identifying the one or more first bioelectric signals detected by the electrode pair on the catheter (the electrode pair is the electrode pair positioned closest to the diaphragm among the multiple electrode pairs) from the bioelectric signal set includes: detecting the presence and size of an electromyographic (EMG) component in at least one bioelectric signal from the bioelectric signal set; and, selecting at least one bioelectric signal as the one or more first bioelectric signals based on the size of the corresponding EMG component. For example, selecting the maximum EMG component, or selecting an EMG component above a threshold size. For example, this can provide an accurate method to identify the closest electrode and the first bioelectric signal. For example, the root mean square value (RMS) or peak amplitude can be used to determine the size of the EMG component.

[0034] In some embodiments, the step of marking a subset that does not include any of the first bioelectric signals further includes marking the subset that does not include any of the first bioelectric signals using a distance measured between an electrode associated with the subset and the electrode in the electrode pair associated with the correctly positioned subset of bioelectric signals. By providing a more fine-grained classification (marking) of the location of the subset associated with the first bioelectric signal, the usefulness of the output of the machine learning algorithm can be further enhanced, for example, because a clinician can receive an indication of how far the catheter is positioned from the optimal position. This can enable the clinician to more accurately position the catheter.

[0035] In some embodiments, the step of marking a subset that does not include the first bioelectric signal also includes: marking the subset that does not include the first bioelectric signal using multiple intermediate electrodes between the electrode associated with the subset and the electrode in the electrode pair on the catheter, the electrode pair being an electrode pair among the multiple electrode pairs positioned closest to the diaphragm. By providing a more granular classification (marking) of the location of the subset related to the first bioelectric signal, the machine learning algorithm can indicate the number of electrodes between a given electrode or a given electrode pair and the electrode pair positioned closest to the diaphragm. For example, this can further infer the distance based on the arrangement of the electrodes on the catheter. This can enable clinicians to receive more useful positioning information and more accurately position the catheter.

[0036] In some embodiments, the partitioning of the bioelectric signal set is performed so that at least two bioelectric signal subsets partially overlap, so that electrodes associated with detecting one or more bioelectric signals in the first subset are also associated with detecting one or more bioelectric signals in the second subset. This can generate more bioelectric signal subsets from a given bioelectric signal set, thereby increasing the amount of available training data. In addition, since a larger number of electrode positions relative to the diaphragm are considered, the overlap increases the granularity of the training data, which can improve the accuracy of the machine learning algorithm.

[0037] In some embodiments, the bioelectric signal set is divided into at least two bioelectric signal subsets, so that each subset is associated with a predetermined number of electrodes. This can ensure that each subset is of the same size. For example, the same size of each subset can reduce the deviation in the training data set. In an example, each subset is associated with a pair of electrodes. In an example, each subset is associated with a paired combination of 3 electrodes, 4 electrodes, or 5 electrodes.

[0038] In some embodiments, the set of bioelectric signals is divided into the at least two subsets of bioelectric signals so that the number of subsets determined to be associated with the correctly positioned electrode and marked as the subset of signals detected from the correctly positioned electrode is in a predetermined ratio to the number of subsets determined to be associated with the electrode positioned above or below the diaphragm and marked as the subset of signals detected from the electrode located above or below the diaphragm. This can enable the training data to include a target number ratio of correctly positioned markers to incorrectly positioned markers (i.e., markers above the diaphragm or below the diaphragm), which can improve the accuracy of the machine learning algorithm because the training data is balanced according to the requirements of the implementation method using the machine learning algorithm. In an embodiment, the predetermined ratio is 1:1. Advantageously, the number of correctly positioned markers is equal to or substantially similar to the number of incorrectly positioned markers, thereby providing a fully balanced training data set.

[0039] In some embodiments, the bioelectric signal includes voltage timing data.

[0040] In some embodiments, receiving the set of bioelectric signals includes receiving bioelectric signals from the catheter during use by the patient. Advantageously, the training data may be continually expanded using real-time data.

[0041] In some embodiments, receiving the bioelectric signal set includes receiving pre-recorded bioelectric signals. In other words, the bioelectric signals can be from historical clinical data.

[0042] In some embodiments, the machine learning algorithm is a neural network.

[0043] In some embodiments, the plurality of electrodes are equally spaced along the length of the catheter. The machine learning algorithm can become more accurate because the electrodes have a regular positional relationship to each other, which increases the flexibility of dividing the bioelectric signal set into subsets. In addition, equally spaced electrodes can simplify the assessment of the position of the catheter compared to, for example, electrodes randomly distributed along the length of the catheter.

[0044] In some embodiments, when the first bioelectric signal is identified as being associated with an electrode that is the farthest or closest electrode of the plurality of electrodes relative to the length of the catheter, the bioelectric signal set is not used as training data. Advantageously, for example, bioelectric signals derived from catheters that are positioned too low or too high overall can be discarded, thereby improving the effectiveness of the training data.

[0045] In some embodiments, each electrode pair is a pair of adjacent electrodes. That is, the electrodes are adjacent in the electrode sequence on the catheter and are nearest neighbors.

[0046] In some such embodiments, the one or more first bioelectric signals include a single first bioelectric signal detected by an electrode pair of the plurality of electrode pairs on the catheter that is positioned closest to the septum.

[0047] In a fourth aspect, a method for classifying a position of a catheter relative to a septum, the septum being a septum of a patient, is provided. The method comprises the following steps: (a) receiving a first set of bioelectric signals detected by a catheter, the catheter carrying a plurality of electrodes at corresponding positions along the length of the catheter so that the electrodes are located at corresponding different distances from the diaphragm of the patient, the plurality of electrodes being divided into a plurality of electrode pairs, each signal being detected by one of the plurality of electrode pairs, each signal comprising an electrocardiographic (ECG) component; (b) dividing the first set of bioelectric signals into at least two first bioelectric signal subsets, each first bioelectric signal subset comprising one or more bioelectric signals and corresponding to a corresponding set of electrodes associated with the detection of the one or more bioelectric signals in the bioelectric signal subset, wherein the dividing is performed such that each set of electrodes is a sequence of electrodes placed continuously along the length of the catheter; (c) inputting the first bioelectric signal subsets into a machine learning algorithm, the machine learning algorithm being trained to classify each subset into a plurality of classes, the classes comprising one or more classes for mispositioned electrodes and classes for correctly positioned electrodes; and (d) receiving a plurality of input classes from the machine learning algorithm. (e) inputting the plurality of input classifications into a pattern recognition function, the pattern recognition function being configured to classify the position of the catheter; and (f) using an output classification from the pattern recognition function to classify the position of the catheter.

[0048] The method described in accordance with the fourth aspect can allow processing of bioelectric signals received from the catheter and generating an output classification to classify the position of the catheter. In some examples, such output classification can be a binary classification (e.g., "correct placement" or "wrong placement"), and can allow a clinician to quickly assess whether the catheter needs to be repositioned, or whether the catheter can remain in its current position. In other examples, a more fine-grained output classification can enable additional levels of information about the positioning of the catheter.

[0049] When multiple input classifications are received from the machine learning algorithm and input into a pattern recognition function to determine an output classification, classifications from multiple instances of the received bioelectric signal may be aggregated to assign a collective classification.

[0050] By aggregating individual classifications into collective classifications, the decision-making process can be made more transparent and explainable. Stakeholders can investigate the contribution of each instance to the overall classification, which facilitates trust and understanding, especially in fields that require explanations, such as healthcare. For example, if an error occurs, it may be more straightforward to trace back through the aggregate process to identify and correct the source of the error at the instance level, rather than trying to diagnose a monolithic machine learning algorithm.

[0051] Additionally, the approach benefits from modularity, as the pattern recognition functionality can be updated without retraining the entire machine learning algorithm. This can improve the uptime of systems employing the approach, as time spent offline for maintenance or updates can be reduced.

[0052] Furthermore, training a model to predict an aggregate class may require a large amount of aggregate-level labeled data, which may be costly or time-consuming to collect. Aggregate methods can leverage more readily available instance-level data or human experts and can reduce the need for large amounts of labeled data. Furthermore, if the aggregate class is rare or unbalanced in the dataset, the machine learning algorithm may struggle because it does not have enough examples to accurately predict such edge case scenarios. In contrast, appropriate output classifications for rare or unbalanced aggregate classes can be directly identified and considered by pattern recognition algorithms.

[0053] In a fifth aspect, one or more non-transitory computer-readable media are provided, the one or more non-transitory computer-readable media storing instructions executable by one or more processors. The instructions, when executed, cause the one or more processors to perform operations substantially mapped to the steps of the method described in the fourth aspect.

[0054] In a sixth aspect, a system is provided. The system includes one or more processors and one or more non-transitory computer-readable media, the one or more non-transitory computer-readable media storing first computer-executable instructions. The non-transitory computer-readable media stores the first computer-executable instructions, which, when executed by the one or more processors, cause the system to perform actions substantially mapped to the operations caused by the instructions according to the fifth aspect or the steps of the method according to the fourth aspect.

[0055] The features of the steps of the method according to the fourth aspect also exist in the instructions stored in the non-transitory computer-readable medium according to the fifth aspect and the actions performed by the system according to the sixth aspect. The embodiments of the method can be understood to correspond to the embodiments of the non-transitory computer-readable medium and the embodiments of the system where appropriate.

[0056] In some embodiments, the pattern recognition function is configured to output an output confidence value associated with the output classification, the output confidence value indicating the probability of a correct output classification. In addition to the output classification, the clinician can also use the output confidence value to evaluate the action taken on the catheter, such as keeping the catheter in its current position or repositioning the catheter.

[0057] In some embodiments, the pattern recognition function is configured to output a plurality of output classifications, wherein each output classification is associated with a corresponding output confidence value. Thus, a clinician may receive the most likely output classification, the second most likely classification, etc. For example, this may improve the scope of information provided to the clinician to make their assessment and decision-making process regarding the catheter position more accurate.

[0058] In some embodiments, the pattern recognition function is configured to compare the plurality of input classifications to a predetermined list, the predetermined list comprising a plurality of candidate combinations of the input classifications, wherein each candidate combination is associated with a corresponding candidate output classification, and wherein the method further comprises: determining the output classification based on the predetermined list. For example, having such a predetermined list can enable straightforward processing, error diagnosis and analysis, and explanation of the decision-making process effectively performed by the pattern recognition function. For example, updating the predetermined list (e.g., by consulting a clinical expert) can be a quick way to correct, for example, classification errors that may occur, compared to retraining a machine learning algorithm.

[0059] In some embodiments, the method further comprises: calculating the corresponding variance of the input classification according to each candidate combination of the input classifications in the predetermined list, and identifying the candidate combination of the input classifications that produces the smallest variance; wherein the output classification is set to the candidate output classification of the identified candidate combination. For example, calculating the variance of the input classification according to the candidate combination can be a simple process that can be clearly explained to understand the origin of the output classification.

[0060] In some embodiments, calculating the corresponding difference of the input classification according to each candidate combination includes: calculating the difference of each input classification according to the corresponding corresponding candidate input classification of the candidate combination of input classifications; and aggregating the calculated differences to determine the corresponding difference of the input classification. In this way, the input classification and the candidate combination can be compared instance by instance, for example, can be directly summed to determine the corresponding difference. For example, this can also be easily explained to understand the origin of the output classification.

[0061] In some embodiments, the method further comprises calculating an output confidence value of the output classification based on the minimum difference.

[0062] In some embodiments, the predetermined list includes all possible permutations of the input classifications as candidate combinations of the input classifications, wherein the method further comprises: identifying from the list the candidate combination of input classifications that are the same as the input classifications, and wherein the output classification is set to the candidate output classification of the identified candidate combination. The predetermined list may be viewed as a lookup table.

[0063] In some embodiments, the first bioelectric signal set is a first set of a plurality of bioelectric signal sets, each set of the plurality of sets corresponding to bioelectric signals detected in a corresponding time period; and wherein the method further comprises: performing steps (b) to (d) for each bioelectric signal set, thereby receiving a plurality of input classifications associated with each bioelectric signal set from the machine learning algorithm; forming an aggregation of input classifications, the aggregation of input classifications comprising the plurality of input classifications received for each bioelectric signal set; and, inputting the aggregation of input classifications into the pattern recognition function, and receiving the output classifications based on the aggregation of input classifications as outputs from the pattern recognition function. This can effectively achieve a time-averaged output classification, which can reduce sensitivity to, for example, transient or short-term classification changes, such as those that may be caused by a set of bioelectric signals in the plurality of bioelectric signals having a different set of classifications than the rest of the bioelectric signals in the plurality of bioelectric signals. This can be a more useful output for clinicians because it can prevent unnecessary changes in the catheter position that may be caused by transient classification changes.

[0064] In some embodiments, the pattern recognition function calculates an intermediate output classification for each plurality of input classifications in the aggregation of input classifications, and calculates the output classification based on the majority of intermediate output classifications. This can be a simple method of calculating the output classification that is still time-averaged and resistant to transient changes in the output classification, but does not necessarily require more complex computational steps, for example.

[0065] In some embodiments, the method further includes: calculating a corresponding intermediate confidence value for each intermediate output classification; and calculating an overall confidence value for the output classification based on the intermediate confidence value. For example, similar to the overall confidence value calculated for a single bioelectric signal set, the overall confidence value can provide clinicians with an indication of the reliability of the output classification and the probability of possible error.

[0066] In some embodiments, the first bioelectric signal set is a first set of a plurality of bioelectric signal sets, each set of the plurality of sets corresponding to bioelectric signals detected in a corresponding time period; and wherein the method further comprises: performing steps b) to d) for each bioelectric signal set, thereby receiving a plurality of input classifications associated with each bioelectric signal set from the machine learning algorithm; forming an aggregation of input classifications, the aggregation of input classifications comprising the plurality of input classifications received for each bioelectric signal set; and, calculating an average input classification set based on the aggregation of input classifications; and, inputting the average classification set to the pattern recognition function, and receiving the output classification based on the average classification as an output from the pattern recognition function. This can allow, for example, the use of a pattern recognition function that is configured to be used with a single input classification set by effectively time averaging or otherwise aggregating the input classifications into an average input classification set. This can reduce the sensitivity of the output classification to transient or short-term changes in the input classification, for example, which can help a clinician determine whether the catheter is properly placed.

[0067] In some embodiments, the pattern recognition function is configured to output one of the following: a first output classification, the first output classification corresponding to the catheter being positioned too low relative to the diaphragm of the patient; a second output classification, the second output classification corresponding to the catheter being correctly positioned relative to the diaphragm of the patient; a third output classification, the third output classification corresponding to the catheter being positioned too high relative to the diaphragm of the patient. In an example, for example, the pattern recognition function is configured to output a distance to the correct positioning of the catheter. In an example, for example, the pattern recognition function is configured to output a plurality of electrodes based on the correct positioning of the catheter. For example, the pattern recognition function may output one or more of the above output classifications.

[0068] The above embodiments may allow a catheter to be inserted into a patient and the position of the catheter to be continuously and automatically monitored. This may make it easier to find and / or maintain the correct position of the catheter. Automatically monitoring the catheter may avoid the need to perform other procedures (e.g., x-rays) to confirm the position, which are often time consuming and expensive. For example, reducing the risk of incorrectly positioning the catheter may reduce the risk of a feeding tube puncturing the duodenum, or reduce the risk of aspiration, which may lead to pneumonia.

[0069] These embodiments also benefit devices that operate in conjunction with the catheter. For example, in clinical scenarios employing ventilation techniques such as neurally-adjusted ventilatory assistance (NAVA), a reliable, continuous source of bioelectric signals from the catheter is desired. Accurately positioning the catheter through an embodiment allows the reliable source of bioelectric signals to be provided to the ventilation system. In addition, false triggering due to signal leakage caused by incorrect positioning of the catheter may result in false alarms, inaccurate monitoring values ​​of respiratory rate and / or tidal volume, and false triggering of the ventilation system, resulting in the patient being out of sync with the ventilation system. Accurately positioning the catheter can address these problems.

[0070] The present invention will now be described more fully hereinafter with reference to the accompanying drawings, in which embodiments of the invention are shown.

[0071] Figure 1 Schematically shown is an overview of a patient 50 having an esophageal catheter 30 extending along the patient's esophagus 18 into the patient's stomach 15 and recording bioelectric signals 25 , such as electrocardiogram (ECG) signals from the heart 22 and electromyography (EMG) signals from the diaphragm 20 .

[0072] The catheter 30 includes a plurality of electrodes, labeled e1 to e9, at respective locations along the length of the catheter 30. In this example, there are nine electrodes, but it will be appreciated that in other examples, there may be a different number of electrodes. Figure 1 In the example, the electrode e 1-9 The electrodes are arranged as a linear array of equally spaced electrodes, so that each electrode is spaced substantially the same distance from its two adjacent electrodes. Each electrode is located at a different distance from the diaphragm 20. When the electrodes are directed toward the esophagus of the patient, for example, when the electrodes are located Figure 1 The electrode is considered to be located above the diaphragm when the electrode is above the indicated axis A in FIG. 1 . Similarly, when the electrode is directed toward the patient's stomach, for example, when the electrode is located Figure 1 When the electrode is below the indicating axis B in the diagram, the electrode can be considered to be below the diaphragm. That is, above and below should be understood as the degree to which the catheter is inserted into the stomach relative to the position of the diaphragm. Similarly, when the electrode is between the indicating axis A and the indicating axis B, the electrode can be considered to be aligned with the diaphragm, or located near the diaphragm.

[0073] Use paired electrodes n and e x≠n and adjacent pairs of electrodes (e.g., e n and e n-1 or e n and e n+1) to perform bipolar measurements and generate bioelectric signals. The following discussion considers bioelectric signals generated by nearest neighbor pairs, but more generally, non-neighbor pairs can also or alternatively be used.

[0074] Therefore, the nine electrodes generate a bioelectric signal set, which in this example is based on 8 channels of voltage timing data of 8 adjacent pairs. The bioelectric signal set can be understood as a plurality of signals below, wherein each signal in the plurality of signals corresponds to a corresponding pair of electrodes of the catheter 30. Figure 5 An example of a collection of bioelectric signals detected by the electrodes of the catheter 30 is shown. Figure 5 In the example of , the bioelectric signal is recorded as voltage time series data with a time resolution of 2kHz and an amplitude resolution of 20 bits. An example of the ECG component of the signal is labeled 25a, and an example of the EMG component is labeled 25b.

[0075] Bioelectric signal 25 typically includes contributions from both ECG signal and EMG signal. Each time the patient's heart beats, heart 22 will send an ECG signal. Given the proximity of heart 22 to diaphragm 20, ECG can be a reliable source of information from which the position of the diaphragm can be inferred. ECG signals include a PQRST complex that describes the pattern of bioelectric activity of the heart during the cardiac cycle. The initial peak (P wave) represents the propagation of a bioelectric pulse initiated by the sinoatrial node located in the right atrium, passing through the two atria of the heart and causing them to contract. The Q, R, and S waves together depict the subsequent ventricular contraction triggered by the P wave. The terminal peak (T wave) indicates that the ventricles are recovering and ready for the next cardiac cycle. EMG signals are emitted during diaphragm contraction (e.g., during respiratory effort); that is, when the patient breathes spontaneously, the diaphragm contracts and emits EMG signals.

[0076] The position monitoring system 100 receives the signals detected by the electrodes, thereby monitoring the position of the catheter 30 in the patient 50. Figure 2 The position monitoring system 100 is described in more detail. However, in general, the position monitoring system is responsible for determining whether the catheter 30 is correctly positioned or incorrectly positioned and notifying the clinician thereof. In this example, the position monitoring system 100 outputs positioning information describing whether the catheter 30 is correctly positioned or incorrectly positioned to the display unit 200. Fig.13 Outputting positioning information and display of such information is described in more detail. Substance delivery device 210 administers fluids, nutrients, and medications to a patient via catheter 30. For example, substance delivery device 210 may wait for proper positioning of catheter 30 before delivering fluids, nutrients, or medications. For example, proper positioning of the catheter may be confirmed by position monitoring system 100 and / or by a clinician. It should be understood that Figure 1Other components not shown in the drawings may similarly use positioning information from position monitoring system 100. For example, if it is determined that the catheter is mispositioned, the ventilation control system may temporarily inhibit ventilation to reduce the risk of falsely triggered breaths, which may be caused, for example, by noise in the received signal due to the position error.

[0077] Now go to Figure 2 , the position monitoring system 100 will be described in more detail. The position monitoring system 100 includes a processor 140 that interacts with and executes data and program routines stored in a memory 110. In this example, the memory 110 includes a data storage device 111 for storing clinical data 112 and training data 114, and stores program routines 115 that allow execution of a clinical data to training data process 116, a machine learning algorithm training process 118, a deployed machine learning algorithm process 120, and a position determination process 122.

[0078] As an overview, the position monitoring system 100 receives clinical data 112 including bioelectric signals detected by a catheter, processes the clinical data 112 to create training data 114 for training a machine learning algorithm, trains the machine learning algorithm based on the training data 114, and deploys the machine learning algorithm 120. Once deployed, the trained machine learning algorithm 120 can be used to monitor the position of the catheter 30 while the catheter 30 is inserted into the patient 50 based on the bioelectric signals 25 detected by the catheter 30. This overview is provided by Figure 3 1 , data 112 , 114 and processes 116 , 118 , 120 , 122 will be described in more detail later on based on this figure.

[0079] More generally, the location monitoring system 100 may include any suitable arrangement or configuration of one or more computing devices that allows for storage of clinical data 112 and training data 114, and execution of a clinical data to training data process 116, a machine learning algorithm training process 118, a deployed machine learning algorithm process 120, and a location determination process 122. For example, the storage device may include remotely accessible cloud storage or locally accessible solid-state disk storage or hard disk storage, or storage on a local networked server, for example. For example, the instruction programs for the aforementioned processes may also be stored on the same storage device, or in their own respective storage devices. Similarly, for example, execution of the aforementioned processes may be performed on a local processor (e.g., a local processor of a personal computer located in a clinical environment), may be performed on a local networked server, or may be performed on a cloud-based server. Each process may be executed on a correspondingly different processor, or may be executed on the same processor. The location monitoring system 100 may include circuitry configured to implement the functionality described herein (using one or more non-transitory computer-readable media). For example, processors suitable for executing instruction programs include general-purpose microprocessors and special-purpose microprocessors, as well as the sole processor of any type of computer or one of multiple processors or cores. The processor may be supplemented by an ASIC (application-specific integrated circuit) or incorporated therein. It should be understood by those skilled in the art that the above exemplary embodiments may be implemented by any suitable software, hardware or hardware configuration or a combination thereof. Exemplary hardware platforms for implementing exemplary embodiments may include, for example, an Intel x86-based platform with a compatible operating system, Windows OS, Mac platform and MAC OS, and mobile devices with operating systems such as iOS, Android, etc. In yet another example, an exemplary embodiment of the described method may be embodied as a program comprising lines of code stored on a non-transitory computer-readable storage medium, which may be executed on a processor or microprocessor when compiled.

[0080] Figure 2The position monitoring system 100 includes a catheter data input module 130. The catheter data input 130 receives bioelectric signal data from the catheter 30 that has detected the EMG and / or ECG signals 25 of the patient 50. The catheter data input 130 can be arranged so that the catheter 30 is docked with, for example, a personal computing device of the position monitoring system 100, and a program running on the personal computing device of the position monitoring system 100 processes the data received from the catheter 30. In other examples, an intermediate peripheral device (e.g., an oscilloscope) receives data from the catheter 30, and an onboard processor (e.g., a field programmable gate array (FPGA)) can perform processing on the received data. For example, the catheter data input 130 can perform data normalization or data cleaning on the data received from the catheter. The catheter data input 130 can then provide the received data to the processor 140 for storage, for example as clinical data 112, or for use by the process of the position monitoring system 100. The use of the data received by the catheter 30 will be described in more detail later.

[0081] Figure 2 The position monitoring system 100 includes a display data output module 124. The functionality of the display data output module 124 will be described in more detail later, but generally, the display data output module 124 receives information to output to the display unit 200. Thus, the display data output module 124 may include graphics processing capabilities that may, for example, render graphics related to outputs from processes of the position monitoring system (e.g., the position determination process 122). The information output by the display data output module 124 and the display unit 200 may be used by a clinician.

[0082] An example of the overall process flow performed by the position monitoring system 100 will now be described. Steps S1 to S6 of the overall process flow are as follows: Figure 3 shown.

[0083] At item S1, clinical data 112 is obtained, and the clinical data 112 includes data collected by a catheter (e.g., Figure 1 A collection of bioelectric signals detected by electrodes of the catheter 30) described in FIG.

[0084] Typically, the purpose of item S1 is to obtain data that can be used to train a machine learning algorithm, the purpose of which is described below. The clinical data 112 can be data recently collected from a patient, or it can be historical data collected previously. In some embodiments, data is collected only from the adult population, thereby defining adults as the target population for the use of a machine learning algorithm. In other embodiments, for example, data can only be collected from the pediatric population, thereby defining children as the target population for the use of a machine learning algorithm. In a further embodiment, the clinical data includes a mixture of adults and children, so that a single model can be trained and applied independently of the patient's age. In a further embodiment, for example, data recorded when used on animals can be used as training data for use on human patients, or for training and use of veterinary target populations. In a further embodiment, clinical data 112 is collected during the use of the catheter 30. For example, the clinical data collected during use can be added to an existing historical clinical data 112 collection, or the clinical data can be used to form a new clinical data 112 collection based on, for example, specific training needs. For example, clinical data may be obtained from a federated computing system, in which data is shared among nodes in a network in a standardized or coordinated manner that protects patient privacy.

[0085] At item S2, the clinical data is processed to generate training data corresponding to Figure 2 , the clinical data to training data process 116 shown in . Both clinical data and training data include a set of bioelectric signals detected by the electrodes of the catheter. As used herein, clinical data refers to bioelectric signals before being processed to form training data. The inventors have observed that the clinical data set 112 acquired at item S1 (e.g., a historical data set from an intensive care unit) is typically recorded when the catheter is in a correctly placed position. Therefore, from a machine learning perspective, the clinical data set may be unbalanced due to the lack of data recorded when the catheter is in an incorrectly placed position. In general, the purpose of item S2 is to improve the training data available to the machine learning algorithm trained at item S3. The steps taken at item S2 to achieve this are as follows: Figure 4 The figure is shown in the figure, and will be described in more detail later based on the figure.

[0086] At item S3, the machine learning algorithm is trained using the training data, corresponding to Figure 2 1 . In general, the purpose of item S3 is to train the machine learning algorithm so that given the bioelectric signals detected by the electrodes of the catheter 30, the machine learning algorithm can determine or classify whether the catheter is misplaced or correctly placed relative to the septum. The output of item S3 is the trained machine learning algorithm.

[0087] At item S4, the trained machine learning algorithm is deployed, corresponding to Figure 2 1. In general, the purpose of item S4 is to use the bioelectric signals detected by the electrodes of the catheter 30 and received by the position monitoring system 100 at the catheter data input module 130 as input to the trained machine learning algorithm, and output position information through the machine learning algorithm.

[0088] At item S5, the position information output by the machine learning algorithm is used to determine the positioning state of the catheter, corresponding to Figure 2 The position determination process 122 depicted in FIG. 1 is generally intended to determine how the output of the machine learning algorithm should be used to determine the positioning state. The steps taken at item S5 are as follows: Fig.10 As shown in FIG. 1 , a more detailed description will be given later based on the accompanying drawing.

[0089] At item S6, the positioning status is output to the user. Generally, the purpose of item S6 is to output the positioning status to the clinician and corresponds to the processor 140 in combination with the positioning status. Figure 2 The process performed by the display data output module 140 and the display unit 200. For example, the positioning status can indicate that the catheter is correctly placed or incorrectly placed. In other examples, the positioning status can indicate that the catheter is incorrectly placed (due to being too high or too low). In further examples, the positioning status can indicate that the catheter is incorrectly placed and, for example, indicate an estimated distance from the correct position.

[0090] Steps S1 to S6 do not need to be performed in strict order. For example, steps S1 and S2 may occur multiple times before step S3 is performed. For example, steps S1, S2, and S3 may occur multiple times before step S4 is performed. Steps S1 to S6 do need to be performed in a sequential manner. For example, a first step (e.g., step S1) may occur at a first time, while a second step (e.g., step S2) may occur at a significantly later second time (e.g., hours, weeks, or months later).

[0091] Processing clinical data

[0092] The process of processing clinical data to prepare training data will now be described, corresponding to Figure 2 Item S3 in . Figure 4 A flow chart detailing the steps taken to process clinical data and prepare training data therefrom is shown. Items S1 and S3 are included in the flow chart to help illustrate the processes preceding and following item S2.

[0093] At item S1, as described above, clinical data including a set of bioelectric signals detected by electrode pairs on the catheter is obtained. As described above, the inventors have recognized that such data generally over-represents correct placement of the catheter and under-represents incorrect placement of the catheter. The steps of item S2 described below can help solve this problem.

[0094] In step S202, a data preparation process is performed on the bioelectric signal. Generally speaking, this may include some form of data normalization and / or data cleaning, for example, to ensure a degree of consistency in the format of the bioelectric signal under consideration, or to optimize the data for use as training data, for example by reducing the size of the data, or by removing portions of the data that contribute relatively little to the training process.

[0095] In an example of the process, the bioelectric signal is segmented into individual heartbeats. This can be performed based on, for example, identification of a PQRST complex or constituent components of a PQRST complex. For example, an R wave can be detected and used to identify a specific heartbeat. Thus, the bioelectric signal can include bioelectric activity associated with one heartbeat detected by the electrode pair. In other examples, the bioelectric signal includes multiple heartbeats and either segmentation does not occur or is not segmented into individual heartbeats.

[0096] In an example of this process, data from a time window around the R wave (e.g., within a window of ±300 ms) is retained, while data detected in the intervening period between two consecutive heartbeats in a plurality of heartbeats is deleted. This can ensure that the entire PQRST complex within the ECG signal is captured, while removing the signal between heartbeats, which contain relatively little (if any) useful data related to the location of the electrodes. This can occur in conjunction with segmentation, such that individual heartbeats are isolated and irrelevant data outside of the PQRST complex is deleted, or not in conjunction with segmentation, such that multiple heartbeats are recorded, but irrelevant data between PQRST complexes is deleted.

[0097] In another example of the process 202, a filtering algorithm is applied to reduce the presence of an EMG component within the bioelectric signal. The filtering algorithm may include calculating a moving average of a number of heartbeats, wherein the EMG component is present only on a subset of the number of heartbeats. Thus, the moving average reduces the impact of the EMG component. Other filtering algorithms may be used to reduce the EMG component, such as applying a band-stop filter in the EMG spectrum (frequencies around 60–100 Hz), or, for example, completely excluding heartbeats in which EMG components are detected. Removing the EMG component can ensure that the machine learning algorithm learns how to predict catheter positioning based solely on the ECG component, rather than relying on the EMG component. Thus, the machine learning algorithm can correctly determine the position of a catheter inserted into a patient who is taking sedatives and is therefore unable to breathe on his own (usually requiring intubation).

[0098] In yet another example of the process 202, data augmentation is applied to at least some of the bioelectric signals. For example, enhancing the bioelectric signal includes at least one of: stretching the bioelectric signal over time; compressing the bioelectric signal over time; or changing the amplitude of the bioelectric signal. Advantageously, the number of samples for data augmentation can be synthetically increased. Synthetic data augmentation can be critical for machine learning classification, especially for biological data, which is often high-dimensional and scarce. Data augmentation can also be useful in order to avoid overfitting and make machine learning algorithms more robust.

[0099] At item S203, a first bioelectric signal associated with an electrode pair that is properly positioned relative to the diaphragm is identified from the set of bioelectric signals. As used herein, in the context of a first bioelectric signal, "first" is used as a label and does not imply any type of positional information about the bioelectric signal. In general, multiple first bioelectric signals can be identified, each of which is associated with a corresponding electrode pair that is considered to be properly positioned relative to the diaphragm. For example, the electrode pair closest to the diaphragm can be identified, but other electrode pairs can be considered to be positioned close enough to the diaphragm so that they also represent properly positioned electrodes. For clarity of illustration, the following discussion considers a single first bioelectric signal.

[0100] In the example, identifying the first bioelectric signal can be performed by detecting the presence of an EMG component in the bioelectric signal. As an illustrative example, an EMG component can be detected in the bioelectric signal of channels 2-4, but an EMG component cannot be detected in channels 1, 5, 7 or 8. The sizes of the detected EMG components are compared, and the largest EMG component is identified. For example, as an illustrative example, channel 5 has the largest EMG component. For example, the size of the EMG component can be an amplitude measurement, such as a peak-to-peak measurement or a root mean square amplitude. The associated channel is identified as being detected from the electrode pair positioned closest to the diaphragm. When channel 5 is identified as having the largest EMG component, electrodes e5 and e6 are identified as a pair of electrodes closest to the diaphragm. In the example, the method is extended to identify multiple first bioelectric signals. For example, a bioelectric signal with an EMG component higher than a threshold size can be considered to be detected by a correctly placed electrode. For example, the threshold size can be relative to the absolute measurement of the EMG component, or relative to the size of the EMG component of other bioelectric signals. For example, an EMG component within 10% of the maximum EMG component detected can be considered to be associated with a correctly positioned electrode. The threshold size can be determined by the precise arrangement of the multiple electrodes on the catheter or the characteristics of the electrodes. In other examples, the identification process of one or more first bioelectric signals can include checking which of the received bioelectric signals is associated with a marker or flag that identifies it as a bioelectric signal detected by an electrode pair on the catheter determined to be correctly positioned relative to the diaphragm.

[0101] This identification process using the EMG component occurs before the aforementioned filtering algorithm of, for example, item S201, so that the first bioelectric signal can be identified and then the EMG component can be removed. In some examples, a further filtering process can be implemented on the bioelectric signal to improve the visibility of the EMG component, thereby identifying one or more first bioelectric signals. For example, a bandpass filter can be applied at the characteristic frequency of the EMG component, and a bandstop filter can be applied at the characteristic frequency of the ECG component, thereby reducing the ECG component and relatively enhancing the EMG component.

[0102] In an alternative embodiment, one of the bioelectric signals in the set of bioelectric signals may be marked (e.g., in the historical clinical data) as a signal detected by the electrode pair closest to the diaphragm. In this embodiment, no further analysis may be required to determine the first bioelectric signal other than identifying which bioelectric signal in the historical clinical data includes the mark.

[0103] At item S204, the bioelectric signal set is divided into at least two bioelectric signal subsets, each of which includes signals associated with continuously placed electrodes. For example, channels 1-3 corresponding to electrode pairs e1 and e2, e2 and e3, and e3 and e4, respectively, can form a subset including signals associated with continuously placed electrodes. Similarly, channels 2-4, 3-5, 4-6, 5-7, and 6-8 can be grouped together. Accordingly, subsets can overlap so that channels are represented by multiple subsets. Allowing overlap can increase the number of subsets that can be derived from the bioelectric signal set. This can improve the number of electrode positions represented by each subset, which can allow the position of the catheter to be more accurately estimated by a trained machine learning algorithm, because the dependency between adjacent (e.g., channel 2 is adjacent to channels 1 and 3) bioelectric signals can be learned and utilized by the machine learning algorithm. When a subset includes signals from multiple channels, the classification problem is effectively converted into a multi-channel time series classification problem. The data enhancement process discussed above can generally be applied to each bioelectric signal in the subset.

[0104] In the example of using non-adjacent electrode pairs (e.g., e1 and e3, e2 and e4, e3 and e5, etc.) to detect bioelectric signals, signal subsets associated with a group of continuously placed electrodes can still be prepared when dividing the bioelectric signal set. For example, a subset may include bioelectric signals detected by electrode pairs e1 and e3 and bioelectric signals detected from electrode pairs e2 and e4. Therefore, the group of electrodes associated with the subset is e1, e2, e3, and e4, which is a sequence of electrodes placed continuously along the length of the catheter. The second subset may include bioelectric signals detected from electrode pairs e5 and e7 and bioelectric signals detected from electrode pairs e6 and e8. Therefore, the group of electrodes associated with the second subset is e5, e6, e7, and e8, which is also a sequence of electrodes placed continuously along the length of the catheter. In both cases of using adjacent electrode pairs or non-adjacent electrode pairs to detect bioelectric signals, these subsets may overlap with each other so that electrodes or electrode pairs are found in multiple subsets.

[0105] At item S205, each subset of the plurality of subsets is labeled according to whether the associated electrode is correctly positioned relative to the diaphragm or incorrectly positioned relative to the diaphragm.

[0106] In one example of the labeling process, the subset that includes the first biopotential signal is labeled as detected from correctly positioned electrodes. That is, the electrode closest to the diaphragm is considered to be correctly positioned. The subset that does not include the first biopotential signal is labeled as detected from incorrectly positioned electrodes. This can provide a simple way to label all training data because only the closest electrode pairs need to be identified. Figure 6 This is shown schematically. Figure 6In the example of , the diaphragm is closest to electrodes e4 and e5, which correspond to channel 4. Therefore, for example, channel 4 is identified as the first biopotential signal based on having the largest EGM component. Any subset that includes channel 4 is marked as detected from a correctly positioned electrode, i.e., in this case, subsets 2, 3, and 4 are marked "C". Subsets 1, 5, and 6 that do not include the first biopotential signal are marked as being misplaced, indicated by the mark "NC".

[0107] In another example of the labeling process, subsets that do not include the first bioelectric signal are labeled based on whether the subsets are associated with electrodes positioned below or above the diaphragm. Figure 7 Use with Figure 6 The same data shown schematically illustrates this. Similarly, channel 4 is associated with the electrode closest to the diaphragm, and subsets 2, 3, and 4 are identified as being correctly positioned. Subset 1, which includes channels 1-3, is labeled "H" because it is associated with the electrode above the diaphragm. Subsets 5 and 6, which include channels 5-7 and 6-8, respectively, are labeled "L" because they are associated with the electrode below the diaphragm. The physical arrangement of the electrodes can be used to determine whether the electrodes are above or below the diaphragm. For example, based on the catheter and electrode design, it is known that electrode 1 is toward the proximal end of the catheter relative to electrodes 4 and 5, so it is known that electrode 1 can be inferred to be above the diaphragm when electrodes 4 and 5 are correctly positioned. Similarly, electrode 9 is toward the distal end of the catheter and is inferred to be below the diaphragm. Therefore, in the example, the configuration of the electrodes of the catheter is used to determine the labeling strategy, such as determining whether an electrode or electrode pair is above or below the diaphragm.

[0108] In some embodiments, when the first bioelectric signal is identified as being associated with an electrode that is the farthest or closest electrode of the plurality of electrodes relative to the length of the catheter, the bioelectric signal set is not used as training data. Figure 5 In the example in , if the first bioelectric signal corresponds to channel 1 or 8, the entire set of signals (channels 1 to 8) can be ignored when determining the training data. Advantageously, the set of bioelectric signals recorded from a catheter placed completely below or above the diaphragm can be ignored, which in turn can improve the training data because the signal from a catheter positioned in such a manner may not represent a correctly positioned electrode. Specifically, if the EMG component is used to determine the first signal, one of channels 1 or 8 is typically determined to have the largest EMG component (because that channel is closest to the diaphragm). However, since the entire catheter is positioned above (in the example of channel 8 having the largest EMG component) or below (in the example of channel 1 having the largest EMG component), none of these channels can truly represent a correctly positioned electrode pair, and thus such data sets can be ignored.

[0109] Importantly, the set of bioelectric signals representing correctly positioned catheters has been divided into subsets of bioelectric signals, some of which are labeled as mispositioned because all electrodes associated with these subsets are located outside the vicinity of the septum (either above or below). This increases the number of bioelectric signals that can be considered mispositioned relative to the original clinical dataset. This can improve the effectiveness of the data used to train machine learning algorithms.

[0110] It will be appreciated that the labeling strategy may be modified depending on, for example, the physical arrangement of the electrodes (eg, the spacing of the electrodes), the number of electrodes and the regularity of the electrode spacing, the size of the subsets, or the degree of overlap of the subsets.

[0111] In some examples, knowledge of the physical arrangement of the catheter and electrodes can be used as an additional marker during the marking process to indicate whether the position is correct or incorrect. For example, in the case where a subset of bioelectric signals does not include a first bioelectric signal, the subset can be marked using the distance measured between the electrode associated with the subset and the electrode pair associated with the first signal. For example, the distance can be between the geometric midpoint of the electrode pair associated with the first signal and the geometric midpoint of the closest electrode pair of the subset that does not include the first bioelectric signal, thereby giving an indication of the distance between the electrode associated with the subset that does not include the first bioelectric signal and the correct positioning. The geometric midpoint can be considered to be located in the middle position between the corresponding electrodes along the length of the catheter. Other metrics are also possible, such as the distance between the geometric midpoint of all associated electrodes and the geometric midpoint of the correctly positioned electrode pair, or the distance between the nearest electrode in the associated electrodes and the nearest electrode in the correctly positioned electrode pair.

[0112] At item S206, the subset and its corresponding labels are stored as training data.A number of processes may be performed during the storage phase.

[0113] In some examples, downsampling can be performed on the bioelectric signal data. For example, when the original sampling rate is 2kHz, the signal can be downsampled to 400Hz. A rate of 400Hz has been found to be a safe upper limit because frequencies below 400Hz have been shown to capture key information of the PQRST complex within the ECG signal. This reduction can reduce the size of the input data, thereby increasing the training speed without reducing or reducing the accuracy of the trained machine learning algorithm within a tolerable range.

[0114] In some examples, data balancing techniques can be performed. Data balancing of labeled data can help match larger classes with smaller classes. For example, after labeling is performed, more subsets of bioelectric signals may be labeled as "H / placed too high" compared to "L / placed too low" or "placed correctly". Therefore, the size of "H / placed too high" can be reduced to match "L / placed too low" or "placed correctly", such as by drawing random samples from larger classes. Similarly, for example, all classes can be downsampled to match the size of the smallest class. By ensuring that each class has enough examples (represented by the corresponding labels) for machine learning to effectively learn how to recognize, generating a flat distribution of class populations can help improve the accuracy of trained machine learning algorithms.

[0115] In some examples, data set reduction can be performed based on the self-similarity of bioelectric signals, or by intermittently sampling bioelectric signals without calculating the self-similarity of bioelectric signals. It has been observed that consecutive heartbeats within the same set of bioelectric signals are often highly similar, while heartbeats separated in time show greater variation. For example, the change between the heartbeat at the beginning of a 10-minute recording and the heartbeat at the end of the 10-minute recording is often greater than the previous or next consecutive heartbeat. For example, discarding similar consecutive heartbeats while retaining heartbeats spaced apart in time can reduce the size of training data without accuracy or with a tolerable loss of accuracy because the reduced training data still captures different sets of heartbeats. In some embodiments, for example, autocorrelation techniques are used to identify particularly different heartbeats within bioelectric signals to increase the diversity of heartbeats. The reduced training data size can increase processing speed, such as reducing the training time required before a trained machine learning algorithm can be deployed.

[0116] Use training data to train machine learning algorithms

[0117] An example of training a machine learning algorithm using the training data prepared at item S2 will now be described, corresponding to item S3. In general, the machine learning algorithm to be trained can be an algorithm suitable for solving a time series classification problem. The general task of a time series classification problem is to predict the label of the entire time series sequence based on the temporal pattern within the time series sequence. More specifically, the position monitoring system 100 uses a machine learning algorithm that is capable of outputting a position label based on the input bioelectric signal of the catheter 30, such as being correctly positioned, incorrectly positioned, too high or too low.

[0118] In the example, decision trees, k-nearest neighbors, and support vector machines can be suitable for solving time series classification problems. In other examples, the machine learning algorithm to be trained is a deep learning model (e.g., a convolutional neural network), which is particularly suitable for solving time series classification problems because the deep learning model can extract relevant features from the original ECG signal without manual feature engineering, and can capture long-term dependencies in the time series. Specifically, variants of convolutional neural networks can be used, such as the known random convolution kernel transform (RandOm Convolution Kernal Transform, referred to as ROCKET) and the minimum random convolution kernel transform (Minimally RandOm Convolution Kernel Transform, referred to as Mini-ROCKET). For example, an algorithm that utilizes a residual network architecture can also be used.

[0119] In some examples, transfer learning methods can be employed to leverage pre-trained neural networks that have been trained on large datasets to perform conceptually similar tasks. For example, an existing time series signal classifier can be retrained using the training data provided at item S2. This can, for example, reduce the overall computational expense of training a machine learning algorithm.

[0120] Figure 8 The architecture of a machine learning algorithm for training according to training data according to one example is schematically shown. The architecture comprises a feature extractor 302 followed by a linear classifier 304 that outputs a classification 306.

[0121] Training can be performed in a software environment. For example, one such software environment is a Jupyter notebook.

[0122] exist Figure 8 In the example of , the training data 114 including the set of bioelectric signals and associated labels is initially passed to the feature extractor 302. In the example, this is a ROCKET or MiniROCKET convolutional model, which uses a predetermined number of convolution kernels to extract features. In other examples, other feature extraction models can be used.

[0123] In the example using the ROCKET model, in the generation phase, convolution kernels are generated based on random sampling of kernel properties (including kernel length, weight, bias, dilation, and padding) according to a predefined distribution, and the kernels are convolved with the training data to extract the features of each kernel. In the example using the MiniROCKET model, in the generation phase, there is no need to randomly sample the kernel parameters, but the dilation, length, and padding parameters are fixed, the weights are restricted to two possible values, and the bias is derived directly from the convolution output.

[0124] In the example using the ROCKET model, in the transformation phase, each generated kernel is applied to each input training data time series, resulting in a corresponding feature map. Two aggregate output features are calculated based on the feature map: the proportion of positive values ​​(PPV) and the maximum or global maximum pooling. The PPV indicator can be considered to capture the frequency at which the time series training data exhibits a positive response to the kernel. The maximum value can be considered to capture the highest response of the time series to the corresponding kernel, indicating the area with the greatest similarity. Therefore, for k kernels, 2k features are generated. In the example using the MiniROCKET model, in the transformation phase, the maximum value feature is not relied on, and only the PPV value is calculated.

[0125] In some examples, the feature extractor 302 can be trained using a machine learning framework that employs stochastic gradient descent (such as PyTorch, etc.), and in some examples batch gradient descent or mini-batch gradient descent can be used. In an example, the training can be performed using the following hyperparameter configuration:

[0126] ●Loss function: CrossEntropyLoss

[0127] ●Optimizer: Adam algorithm

[0128] ●Scheduler: ReduceLROnPlateau

[0129] Batch size: 256

[0130] Learning rate: 0.0001

[0131] ●Tolerance times: 5

[0132] Hyperparameter optimization can be performed through the Tree Parzen Estimator (TPE), a recognized method for automatic hyperparameter tuning, to further optimize the training process. The execution of TPE can cause the value of the hyperparameter to be updated beyond its initial setting value. For example, although the tolerance value is initially set to 5, after TPE, the value may be updated to 3.

[0133] In an example, a cache-based feature extraction method may be used to allow Mini-ROCKET to perform better with the size of the training data set 114. Specifically, the training data 114 may be divided into several smaller chunks, which are then iterated over by Mini-ROCKET to extract features that are subsequently used in the training process.

[0134] In this example, 9996 features corresponding to 9996 cores are extracted from the feature extractor 302, but it should be understood that the number of cores, and therefore the number of features extracted, is a hyperparameter that may be different in other examples.

[0135] exist Figure 8 In the example of FIG. 1 , the features extracted by the feature extractor 302 are then input to a linear classifier 304, which in this example includes a linear classifier composed of nodes X1, X2, X3 . . . X n A dense layer represented by , followed by a softmax activation function. A linear classifier can be trained on the extracted features using, for example, ridge regression or logistic regression. The output of the softmax activation function is the following label or classification 306: "0": too high; "1": correct; or "2": too low.

[0136] It should be understood that the above machine learning algorithm is only one example, and other architectures are possible. The benefits of the training data prepared at item S2 described previously are not limited to use with the specific example of the above machine learning algorithm.

[0137] Deploy machine learning algorithms

[0138] At item S4, the machine learning algorithm that has been trained at item S3 is deployed. Fig. 9 A flowchart is shown that details the steps taken when deploying and using a deployed trained machine learning algorithm. Items S3 and S5 are included in the flowchart to help illustrate the process before and after item S4.

[0139] At item S3, as previously described, the machine learning algorithm is trained using the training data prepared at item S2. However, more generally, any machine learning algorithm that has been trained to classify each subset of bioelectric signals into multiple categories (wherein one or more categories are used for mispositioned electrodes and at least one category is used for correctly positioned electrodes) can be used to process the output of the machine learning algorithm described below. For example, the machine learning algorithm may have been trained on a separate, different system and imported into the current position monitoring system 100, and may not have been trained according to item S3.

[0140] At item S400, the trained machine learning algorithm prepared at item S2 is deployed. In general, this means that the machine learning algorithm is operable to receive new data, rather than training data, which includes unseen data without known ground truth labels, and the machine learning algorithm will generate predictions for it. The data can be considered real-time data because it usually comes from the catheter used by the patient. In some examples, once the training phase of item S2 is over and the machine learning algorithm is deployed, any parameters, weights or other variables of the algorithm optimized during the training process are fixed. In other examples, for example, even once deployed, further training and optimization steps can be performed to continuously improve or otherwise update the machine learning algorithm. Deployment can mean that the algorithm is available in a local computing unit (e.g., a desktop computer that can directly access the algorithm), or available on a server (e.g., a cloud server or local server that can remotely access the algorithm via a network). It should be understood that the precise manner in which the trained machine learning algorithm is computed is not important to the present invention and can vary between embodiments.

[0141] At item S401, catheter data including bioelectric signals detected from a patient is received. Figure 1 and 2 In the example of the present invention, these bioelectric signals are detected from the catheter 30 and arrive via the catheter data input module 130. The bioelectric signal corresponds to a time period, which can vary between embodiments. For example, the bioelectric signal can correspond to a single heartbeat detected by the catheter. In other examples, the bioelectric signal can correspond to several consecutive heartbeats detected by the catheter.

[0142] Similar to the processing of historical clinical data described with respect to item S202, the catheter data input module 130 can perform a data preparation process on the bioelectric signal, which data preparation process can include some form of data normalization and / or data cleaning, for example, to ensure a degree of consistency in the format of the bioelectric signal under consideration. In some examples, for example, the processing can prepare the data in a manner that is most suitable for use with the deployed machine learning algorithm, such as preparing the data into a specific file format, or removing specific components of the data that are not needed for classification purposes, or otherwise undermine the performance of the deployed machine learning algorithm. For example, the machine learning algorithm may have been trained based on a single heartbeat, and the received bioelectric signal corresponds to a number of consecutive heartbeats. For example, the received bioelectric signal can be divided into single heartbeats before being input into the deployed machine learning model.

[0143] At item S403, the bioelectric signal set is divided into bioelectric signal subsets, where each subset includes signals associated with consecutively placed electrodes. This process may be substantially similar to the process described for item S204.

[0144] At item S405, a subset of the bioelectric signals is input to a deployed training machine learning algorithm, and at item S407, the deployed training machine learning algorithm generates a prediction tag for each subset of the bioelectric signals. The prediction tag is a form of positional information, because given the subset of bioelectric signals, the machine learning algorithm is prepared to predict whether the electrodes associated with each subset of bioelectric signals are correctly positioned. As previously described, the positional information can be a binary classification, such as "correctly positioned" or "incorrectly positioned", or it can have a finer granularity, such as "too high", "correctly positioned" or "incorrectly positioned". Depending on the configuration of the machine learning algorithm, other predictions can be generated instead of or in addition to the above classifications, such as indicating the number of electrodes between the pairs associated with the subset and the correctly positioned pairs, or the estimated distance between the pairs associated with the subset and the correctly positioned pairs.

[0145] The position information output by the machine learning algorithm and received at item S407 forms a plurality of input classifications to be used at item S5, which will be based on Fig.10 Further description.

[0146] Processing machine learning algorithm output to determine location status

[0147] At item S5, the positioning state is determined using the position information output by the machine learning algorithm. The machine learning algorithm generates prediction classifications based on the bioelectric signals detected by the catheter, the prediction classifications representing the position information. Fig.10 A flow chart detailing the steps taken to determine the position state based on the position information output by the machine learning algorithm is shown. Items S4 and S6 are included in the flow chart to help illustrate the process preceding and following item S5.

[0148] At item S4, as described above, the deployed machine learning algorithm is deployed to receive bioelectric signal data detected by a catheter in the patient's body and generate a predicted classification for a subset of the bioelectric data. Therefore, the output of item S4 is a plurality of classifications, which can then be used as inputs to items S501 to S505 and are referred to as input classifications hereinafter.

[0149] At item S501, multiple input categories are input into a Figure 2 Pattern recognition functionality that is part of the position determination process 122.

[0150] At item S503, the pattern recognition function determines the location status of the catheter, which includes an output classification that classifies the location of the catheter. Fig.11 and Fig.12An example of a pattern recognition function is described in more detail. In general, the pattern recognition function aggregates instance-level classifications (e.g., in the form of input classifications) into fewer higher-level classifications (e.g., a single classification that classifies the overall location of a catheter), where each instance-level classification corresponds to a subset of the bioelectric signals classified by the machine learning algorithm. In general, the classifications produced by the pattern recognition process can match or differ from the latent classes generated by the machine learning algorithm.

[0151] At item S505, a positioning status and a constituent output classification are received from the pattern recognition function, and at item S6, the positioning status is output to the user.

[0152] Example of pattern recognition function

[0153] Fig.11 A first example of a pattern recognition function is schematically shown, in which an output classification 126 is generated from an input classification 121 generated by a machine learning algorithm 120 using a lookup table 123a.

[0154] As shown in the figure, in an example, the input classification 121 generated by the machine learning algorithm 120 may include electrodes corresponding to the classification. n A list of associated classifications "C". In this example, the input classification 121 has classes including "0", "1", and "2", "0" indicating that the associated electrode is too high, "1" indicating that the associated electrode is correctly positioned, and "2" indicating that the associated electrode is too low. In this example, for clarity of illustration, the electrodes are arranged in pairs, but as previously described, other groupings are also possible. In some examples, the input classification 121 includes a list of classifications, but the corresponding associated electrodes are not explicitly stored, and the corresponding associated electrodes can be implicitly encoded in the position of the classification. For example, it may be known that the first entry in the input classification 121 always corresponds to electrodes e1 and e2, in which case it may not be necessary to explicitly record the electrode identification.

[0155] In this example, the lookup table 123a lists each permutation of possible input classifications, hereinafter referred to as candidate combinations of the input classifications. For example, in this example, there are 7 classifications, each with 3 categories, resulting in 2187 different combinations, each of which is represented by a corresponding candidate combination in the lookup table 123a. Each candidate combination is stored in association with a corresponding candidate output classification, stored in Fig.11 The candidate output classifications may be defined by expert clinicians, or may be determined, for example, by analysis of historical clinical data. In other examples, an algorithmic approach may be employed to determine the candidate output classifications.

[0156] Given any input classification 121, the pattern recognition function searches the lookup table 123a and finds a matching candidate combination and an associated candidate output classification. The candidate output classification can then be set as the output classification 126 of the pattern recognition function.

[0157] In this example, the lookup table 123a includes every permutation of the input classification, but in other examples may include a subset, such as a majority of every permutation of the input classification, but still not include combinations that may be rejected in advance, for example, because they do not produce a meaningful output classification. Instead, the pattern recognition function may, for example, display "wrong" classifications and only produce a classification related to positioning when the input classification matches a candidate combination in the lookup table 123a. For example, this can help diagnose a malfunction condition of the catheter, as contrasted with an erroneous positioning. For example, it should be understood that the lookup table may include enough permutations of the input classification for use by a clinician, and that edge case combinations (e.g., combinations that do not correspond to a reasonable description of the catheter position even taking into account the inherent errors of the machine learning algorithm classification) may be removed from the lookup table to save storage space.

[0158] In some examples, the lookup table 123a may additionally include a confidence value associated with each candidate combination, stored in Fig.11 The confidence value of the matching candidate combination can be output as an output confidence value 127 in association with the output classification 126. For example, the confidence value represents the confidence that the output classification is correct and can be considered to correspond to the probability that the output classification is correct. Therefore, the clinician can understand the likelihood that the catheter is positioned as indicated by the output classification. Fig.11 In the example of , the confidence value ranges from 0 to 100, but it should be understood that any suitable representation is possible. The confidence value can be manually predefined by a clinician, for example, to evaluate candidate combinations. In other examples, the confidence value can be determined algorithmically, for example, by penalizing combinations that do not follow an expected pattern.

[0159] The precise contents of the lookup table 123a may vary depending on factors such as the patient's height, the design of the catheter, and more generally, factors that may affect the mapping of the classification of constituent electrode pairs to the overall classification of catheter positions. For example, if a first catheter design is being used on a child, a first lookup table including suitable candidate combinations and candidate output classifications may be used, while for a second catheter design being used on an adult, a second lookup table may be used that includes more suitable candidate combinations than the first lookup table. For example, the catheter design may dictate the total number of electrodes and their relative spacing, while the patient's height or age may determine the constituent contents of a properly positioned catheter based on, for example, the electrode signals.

[0160] Fig.12 A pattern recognition function according to a second example is schematically shown, in which the pattern recognition function comprises a predetermined list 123b of classified reference candidate combinations, the predetermined list being similar to Fig.11 Instead of using a lookup table, an algorithmic approach is used to determine the variance based on a predetermined list of reference candidate combinations.

[0161] The pattern recognition function includes a predetermined list 123b of reference candidate combinations. Each reference candidate combination is a set of reference candidate classifications, where each reference candidate classification corresponds to a subset of electrodes of the catheter that is similar to the set of input classifications 121. Thus, the reference candidate combination can represent an expected combination of input classifications resulting from the use of the catheter. Thus, for example, the reference candidate combination can represent a combination of input classifications resulting from a correctly placed catheter and input classifications resulting from the catheter being positioned too low or too high relative to the septum. Thus, for example, each reference candidate combination includes candidate output classifications based on whether it represents that the catheter is correctly positioned, positioned too low, or too high.

[0162] The pattern recognition function compares the input classification 121 set with each reference candidate combination and calculates the difference of the input classification according to each reference candidate combination. This process can be regarded as calculating the similarity of the input classification with each reference candidate combination respectively.

[0163] In the example shown, this is performed by considering the difference between the reference candidate classification stored in the "Ref-1" row and the corresponding input classification 121 stored in the "Input" row to calculate the channel-by-channel difference. The channel-by-channel difference between the classifications is stored in the "Var" row, and in this example, the size of the difference is taken so that the difference is stored as a positive value. The difference is a channel-by-channel difference meaning that a difference calculation is performed for each input classification and the corresponding reference candidate classification. That is, the input classification of electrodes e1-e2 is compared with the reference candidate classification of electrodes e1-e2, and the difference of e1-e2 is calculated; the input classification of electrodes e2-e3 is compared with the reference classification of electrodes e2-e3, and the difference of e2-e3 is calculated, and so on. This applies to all reference classifications for a given reference candidate combination. Then, in this example, the channel-by-channel difference is aggregated by summing the difference values ​​and stored in the "SumVar" column to represent the difference of the input classification as a whole according to the reference candidate combination. In other examples, for example, weighted aggregation may be used, where some difference values ​​are weighted more highly than other difference values ​​in the aggregation. Fig.12 The process is repeated (not shown).

[0164] Then, the output classification of the pattern recognition function is set as the candidate output classification corresponding to the reference candidate combination with the lowest difference in the input classification. Fig.12 In the example of FIG. 1 , according to the first reference candidate combination “Ref-1”, the difference in input classification is 4; according to “Ref-2”, the difference in input classification is 1. Therefore, the output classification of the pattern recognition function is set to the candidate output classification of “Ref-2” (i.e., 1), corresponding to the catheter being correctly positioned.

[0165] As described above, by computing channel-by-channel differences in the classifications and aggregating these differences, the differences of the input classifications can be computed based on the reference candidate combinations. Other approaches are also possible. For example, only a subset of the channel-by-channel differences can be aggregated or even computed initially. In the above examples, all electrode pairs contribute equally to the final aggregated differences, but in some examples, the farthest electrode ( Figure 1 The position of the nearest electrode (e7-e9 in the example) can be considered as Figure 1e1-e3 in the example of , for example, to reduce the risk of the catheter tip contacting the duodenum. Therefore, the method of calculating the difference can be weighted for the farthest electrode rather than the closest electrode. In other examples, the weight of the most central electrode located between the closest electrode and the farthest electrode may be less than the closest electrode and the farthest electrode because the inventors have observed that it may be more difficult to classify the electrode closest to the diaphragm (e.g., which is usually the most central electrode).

[0166] In the above example, the difference is a simple subtraction between two classes, but in other examples, more complex functions may be used, such as nonlinear functions that penalize larger differences more severely than smaller differences. The method used may depend on the classes used; in the above example, there are three classes: "0", "1", and "2", but in other examples, a larger number of classes may be used, or continuous quantities such as distance metrics may be used, so different methods of calculating the difference may be used.

[0167] Similar to Fig.11 An exemplary pattern recognition function may output a confidence value, for example, in some examples, the confidence value may be based on the calculated difference. For example, the confidence value may be calculated by an equation (e.g., Confidence = 100 - (SumVar × SF), where SumVar is stored in Fig.11 The calculated difference in the "SumVar" column of the predetermined list 123b of , SF is a scaling factor that determines the impact of SumVar on the confidence value. For example, SF can be 10; if SumVar is always positive and does not exceed 10, the confidence value will be a value between 0 and 100. It should be understood by the skilled person that other measures of confidence values ​​are possible and do not necessarily need to be positive values ​​between 0 and 100. In other examples, some other function can be used, such as Confidence = 100-SumVar 2 , which has the effect of disproportionately penalizing confidence values ​​for larger aggregate differences.

[0168] In an example of a pattern recognition function, the pattern recognition function may be configured to output a number of classifications, each with a corresponding associated confidence value. For example, the pattern recognition function may output a first output classification associated with a first output confidence value, and output a second output classification associated with a second output confidence value. For example, the output of the pattern recognition function may be "1: 90%; 0: 10%", which indicates that there is a 90% probability of outputting the classification "1" and a 10% probability of outputting the classification "0". For example, this approach can be easily extended to three or more output classifications, each with an accompanying confidence value.

[0169] For example, in the case of a lookup table, the pattern recognition function may output an output classification associated with a candidate combination that matches the input classification, but may further consider the candidate combination within a single channel-by-channel change and consider whether the output classification changes as a result. For example, the pattern recognition function may check whether, in the case where channel e4-e5 changes from classification 1 to 0, the new matching candidate combination will have a candidate output classification that is different from the candidate output classification currently output. This may represent a scenario where the catheter is at the boundary between two output classifications, and both may be output with a corresponding corresponding confidence. Thus, the clinician is able to obtain a more accurate or more nuanced impression of the catheter positioning.

[0170] In an example, the pattern recognition function is aware of previous output classifications, which can be viewed as a memory capability. The pattern recognition function can be constrained or otherwise biased toward output classifications based on predetermined relationships between output classifications. For example, after a "too low" classification, the pattern recognition function may be constrained or biased toward outputting "too low" or "correctly positioned" rather than "too high" during subsequent time periods. This can prevent erratic jumps from "too low" to "too high," which may be unanticipated but may be due to misclassification by the machine learning algorithm. Thus, for example, the output classifications are more fluid and easier for the clinician to follow. For example, the degree of constraint can be related to the rate at which the pattern recognition function generates new output classifications and the rate at which the catheter position is expected to change.

[0171] exist Fig.11 and Fig.12 In the example of , the pattern recognition function outputs one of the categories of "0", "1" and "2". This corresponds to an example of a classification output by a machine learning algorithm trained at item S3. In other examples, the machine learning algorithm may output a more fine-grained classification, such as "0", "1", "2", "3", "4" and "5", corresponding to, for example, too low, slightly too low, correctly positioned, slightly too high, too high. Alternatively, for example, the machine learning algorithm may output a predicted distance (in millimeters) from the correct position, in which case the input classification may be, for example, a number between 0 and 50. However, the pattern recognition function may process these classifications as outputs of, for example, one of "0", "1" and "2". In this way, the pattern recognition function reduces the number of instance-level categories when generating an output classification. In other examples, the situation may be the opposite: the pattern recognition function may receive an input classification from a machine learning algorithm with categories "0", "1" and "2", and generate a more fine-grained output classification according to, for example, "0", "1", "2", "3", "4" and "5".

[0172] In an example, the pattern recognition function is configured to identify abnormal behavior. For example, if a specific bioelectric signal subset / electrode pair and therefore the corresponding input classification is generally different from the rest of the set, the corresponding input classification can be identified and excluded from the calculation of the output classification. For example, an input classification different from the closest matching reference candidate combination is marked as an outlier. For example, channels e2-e3 can be marked as outlier channels because the input classification is different from the closest reference candidate combination. For example, if this situation occurs a predetermined threshold number of times in a row, the channel can be marked as having a fault and ignored when determining the output classification in the future. In other examples, the position system 100 is more generally, or in some examples specifically the catheter data input module 130, configured to detect a noisy signal from a catheter before the signal is passed to the deployed machine learning algorithm, where the noisiness of the signal is expected to cause, for example, misclassification of the machine learning algorithm. For example, signal processing or catheter recalibration can be performed in such scenarios to smooth the signal and reduce the noise.

[0173] For example, item S5 may include a step that allows for temporal averaging, which may make the output classification more stable and less prone to short-term switching between classifications.

[0174] For example, the multiple input classifications received from the machine learning model at item S501 can correspond to bioelectric signals from a single heartbeat detected by the catheter. The catheter data input module 130 can collect multiple bioelectric signals from different corresponding times, thereby corresponding to multiple heartbeats. Multiple bioelectric signals from different corresponding times can be provided to the machine learning algorithm to generate a set of multiple input classifications, wherein each multiple input classification corresponds to a bioelectric signal from a corresponding time period. The set of multiple input classifications can then be input to a pattern recognition function. For example, the pattern recognition function can calculate the potential output classification of each input classification in the multiple input classifications, and determine the overall output classification corresponding to the majority of the potential output classifications in these potential output classifications. In this way, the pattern recognition function is configured to process input classifications corresponding to bioelectric signals corresponding to different corresponding time periods.

[0175] This approach can also be used when the bioelectric signals received from the catheter 30 capture multiple heartbeats, rather than just a single heartbeat as described above. For example, a first set of bioelectric signals received from the catheter can capture the first five heartbeats, and the machine learning algorithm can generate a first input classification set based on the first five heartbeats. A second set of bioelectric signals can be received from the catheter that captures the next five heartbeats and provided to the machine learning algorithm, which generates a second input classification set based on these next five heartbeats. The pattern recognition function can consider the first input classification set in conjunction with the second input classification set to determine the output classification, as described above. In this example, the first set of bioelectric signals that capture the first five heartbeats corresponds to a first time period, and the second set of bioelectric signals that capture the last five heartbeats corresponds to a second time period, the second time period being after the first time period. In other examples, these time periods can partially overlap with each other to allow calculation of a moving average output classification.

[0176] Similarly, the confidence value can be based on the degree of change between the corresponding time periods. For example, if there is a large change between the input classification of the first time period and the second (subsequent) time period, this may reduce the confidence value in the output classification. Smaller changes may increase the confidence value in the output classification because, for example, it may indicate that the behavior or position is stable.

[0177] In other examples, the input classifications undergo an averaging process before being passed to the pattern recognition function. For example, a direct numerical average can be calculated, or the majority classification for each channel can be calculated.

[0178] Monitoring Mode

[0179] In some examples, the position monitoring system 100 can disable position monitoring by the position determination process 122 during a procedure in which a catheter is being inserted by a clinician, because it can be assumed that the catheter will be mispositioned for most of the procedure. For example, upon detecting a trigger signal from an electrode indicating proximity to the septum, the position determination process 122 can be automatically resumed. The trigger signal can be a bioelectric signal that the machine learning algorithm determines originated from, for example, a correctly positioned electrode pair with a high degree of confidence. For example, when the clinician determines that the catheter is close enough to the septum to begin the position determination process 122, the position determination process 122 can be manually resumed.

[0180] In some examples, the position monitoring system 100 can enter an approach mode and a steady-state mode that affect the position determination process 122. The approach mode can shorten the time averaging period previously described so that the user receives more frequent position updates. For example, more frequent position updates may be useful when the catheter is initially moved into place, so as to, for example, avoid excessive insertion of the catheter and collision with the patient's body. Once it is determined that the catheter is correctly positioned, the position monitoring system can enter a steady-state mode, which can increase the time average, for example, to reduce short-term classification changes. In some examples, the steady-state mode includes a clinician absent mode. For example, this can be used when the clinician does not actively reposition the catheter, so it is assumed that any change in the position state of the catheter is due to some other reasons such as patient movement or accidental repositioning of the catheter. Therefore, compared with the mode in which the clinician is expected to actively reposition the catheter, when in the clinician absent mode, the data display output module 124 can take different actions. For example, considering that the movement of the catheter is not expected, the position monitoring system 100 may be more sensitive to the movement of the catheter or the movement of the catheter above the threshold size, for example, in the clinician absent mode.

[0181] Output positioning status to the user

[0182] At item S6, the positioning status of the catheter 30 is output to the user. Fig.13 An example of a graphical user interface 400 is schematically shown, which is displayed on the display unit 200 and is generally used to display information related to the catheter position to the clinician. The display unit 200 is controlled by the display data output module 124, which may include a graphics processing unit and generally communicates with the processor 140 to determine the information displayed by the display unit 200.

[0183] In this example, the graphical user interface 400 includes a bioelectric signal portion 402, a channel-by-channel position portion 404 that displays position status information, and an overall catheter position portion 406. The bioelectric signal portion 402 displays the bioelectric signal data received from the catheter 30, and the bioelectric signal data can be interpreted by a clinician independently of the position status information. For example, the channel-by-channel position portion 404 displays the classification generated by the machine learning model, thereby indicating, for example, which channels and associated electrodes are correctly positioned, positioned too low, or positioned too high, channel by channel. The overall catheter position portion 406 displays the output classification of the pattern recognition function, such as indicating to the clinician whether the catheter as a whole is correctly positioned, too low, or too high. For example, the information displayed by the channel-by-channel position portion 404 and the overall catheter position portion 406 can be symbolically implemented. In other examples, text or numbers corresponding to the classification can be displayed. In the example, additionally or alternatively, based on the output classification, the display can involve visual effects such as color, brightness, flickering, or luminescence to improve the visibility of the clinician. For example, additionally, the bioelectric signal portion 402 can be enhanced with these visual effects.

[0184] In some embodiments, auditory effects may be used in conjunction with or in place of visual effects. For example, an alarm may sound when the location is wrong. The alarm may further indicate in what manner the location is wrong, i.e., for example, a first alarm sound may indicate too high, while a second alarm sound may indicate too low. For example, the tone or volume may vary depending on the location state. For example, the auditory effects may be managed by the display data output module 124.

[0185] As described above, the display data output module may also include a component for notifying the clinician based on a location status update depending on whether the clinician is present. For example, when in a steady state, clinician-absent mode, the display data output module may notify the clinician via an electronic communication device or add a sound or visual effect so that the clinician can be notified without actively monitoring the patient.

[0186] For example, the display unit 200 can be an LCD, LED, OLED, CRT, or any other type of display panel capable of displaying visual content, which can be part of a television or computer monitor. In an example, the display unit 200 can be an augmented reality, mixed reality, or virtual reality headset, for example, worn by a clinician.

[0187] Other embodiments

[0188] The above embodiments should be understood as illustrative examples of the present invention. It should be understood that any feature described with respect to any one embodiment can be used alone or in combination with other features described, and can also be used with one or more features of any other embodiment or any combination of any other embodiment. In addition, equivalents and modifications not described above may also be adopted without departing from the scope of the present invention as defined in the appended claims.

Claims

1. A method for determining training data to be used to train a machine learning algorithm to classify the position of a catheter relative to a patient's septum, the method comprising the steps of: receiving a set of bioelectric signals detected by a catheter, the catheter carrying a plurality of electrodes at respective locations along the length of the catheter so that the electrodes are located at respective different distances from a diaphragm of the patient, the plurality of electrodes being divided into a plurality of electrode pairs, each signal being detected by one of the plurality of electrode pairs, each signal comprising an electrocardiographic (ECG) component; identifying, from the set of bioelectric signals, one or more first bioelectric signals, each first bioelectric signal being detected by an electrode pair on the catheter determined to be properly positioned relative to the diaphragm; dividing the bioelectric signal set into at least two bioelectric signal subsets, each bioelectric signal subset including one or more bioelectric signals and corresponding to respective electrode groups associated with the detection of the one or more bioelectric signals in the bioelectric signal subset, wherein the dividing is performed such that each electrode group is a sequence of electrodes placed continuously along the length of the catheter; marking each subset in the plurality of subsets, wherein the marking comprises: marking a subset including at least one of the first bioelectric signals as a subset of signals detected from correctly positioned electrodes, and marking a subset of the first bioelectric signals that does not include any of the first bioelectric signals as a subset of signals detected from an incorrectly positioned electrode; as well as The subset of bioelectric signals and their corresponding labels are included in the training data.

2. The method according to claim 1, wherein the step of marking the bioelectric signal subset that does not include the first bioelectric signal as a bioelectric signal subset detected from an incorrectly positioned electrode comprises: Determine whether the electrode associated with detecting the one or more bioelectric signals in the subset is positioned above the diaphragm or below the diaphragm, wherein, when it is determined that the electrode is positioned above the diaphragm, the bioelectric signal subset is marked as a bioelectric signal subset detected from the electrode located above the diaphragm, and when it is determined that the electrode is positioned below the diaphragm, the bioelectric signal subset is marked as a bioelectric signal subset detected from the electrode located below the diaphragm.

3. The method according to any one of claims 1 or 2, further comprising: Enhance each bioelectric signal in the subset, wherein enhancing the bioelectric signal comprises at least one of the following: stretching the bioelectric signal over time; compressing the bioelectric signal over time; or The amplitude of the bioelectric signal is changed.

4. The method according to any one of claims 1 to 3, wherein at least one bioelectric signal comprises an electromyographic (EMG) component, and the method further comprises: A filtering algorithm is applied to each bioelectric signal in the set of bioelectric signals, wherein the filtering algorithm is configured to at least reduce a corresponding electromyographic (EMG) component from the corresponding bioelectric signal.

5. The method of claim 4, wherein applying the filtering algorithm to the bioelectric signal comprises: identifying a plurality of sub-portions of the bioelectrical signal, each sub-portion comprising data detected during a heartbeat of the patient; as well as An average bioelectric signal is calculated based on the multiple sub-portions of the bioelectric signal.

6. The method according to any one of claims 1 to 5, wherein each bioelectric signal from the set of bioelectric signals comprises data detected during a plurality of heartbeats of the patient, wherein the method further comprises: identifying, in each bioelectric signal from the bioelectric signal set, data detected in an intermediate period between two consecutive heartbeats in the plurality of heartbeats; as well as The identified data is deleted from the bioelectric signal.

7. The method according to any one of claims 1 to 6, wherein the step of identifying the one or more first bioelectric signals from the bioelectric signal set comprises: detecting the presence and magnitude of an electromyographic (EMG) component in at least one bioelectric signal from the set of bioelectric signals; And, at least one bioelectric signal is selected as the one or more first bioelectric signals based on the magnitude of the corresponding EMG component.

8. The method according to any one of claims 1 to 7, wherein the step of marking a subset of the first bioelectrical signals that does not include any of the first bioelectrical signals further comprises: A subset of first bioelectric signals that does not include any of the first bioelectric signals is marked using a distance between electrodes associated with the subset and electrodes associated with the correctly located subset of bioelectric signals.

9. The method according to any one of claims 1 to 8, wherein: The partitioning of the set of bioelectric signals is performed such that at least two subsets of bioelectric signals partially overlap, such that electrodes associated with detecting one or more bioelectric signals in a first subset are also associated with detecting one or more bioelectric signals in a second subset.

10. A method according to claim 2 or any one of claims 3 to 9 when dependent on claim 2, wherein: The bioelectric signal set is divided into the at least two bioelectric signal subsets so that the number of subsets determined to be associated with the correctly positioned electrode and marked as signal subsets detected from the correctly positioned electrode is in a predetermined proportion to the number of subsets determined to be associated with the electrode positioned above or below the diaphragm and marked as signal subsets detected from the electrode positioned above or below the diaphragm.

11. A method according to any one of claims 1 to 10, wherein when a first bioelectric signal is identified as being associated with an electrode that is the farthest or closest electrode among the multiple electrodes relative to the length of the catheter, the set of bioelectric signals is not included in the training data.

12. A method according to any preceding claim, wherein each electrode pair is a pair of adjacent electrodes.

13. The method of claim 12, wherein the one or more first bioelectric signals include a single first bioelectric signal detected by an electrode pair on the catheter, the electrode pair being an electrode pair among the plurality of electrode pairs positioned closest to the diaphragm.

14. One or more non-transitory computer-readable media storing instructions executable by one or more processors, wherein the instructions, when executed, cause the one or more processors to perform operations comprising: receiving a set of bioelectric signals detected by a catheter, the catheter carrying a plurality of electrodes at respective locations along the length of the catheter so that the electrodes are located at respective different distances from a diaphragm of the patient, the plurality of electrodes being divided into a plurality of electrode pairs, each signal being detected by one of the plurality of electrode pairs, each signal comprising an electrocardiographic (ECG) component; identifying, from the set of bioelectric signals, one or more first bioelectric signals detected by a pair of electrodes on the catheter, each first bioelectric signal being determined to be properly positioned relative to the diaphragm; dividing the bioelectric signal set into at least two bioelectric signal subsets, each bioelectric signal subset including one or more bioelectric signals and corresponding to respective electrode groups associated with the detection of the one or more bioelectric signals in the bioelectric signal subset, wherein the dividing is performed such that each electrode group is a sequence of electrodes placed continuously along the length of the catheter; marking each subset in the plurality of subsets, wherein the marking comprises: marking a subset including at least one of the first bioelectric signals as a subset of signals detected from correctly positioned electrodes, and marking a subset of the first bioelectric signals that does not include any of the first bioelectric signals as a subset of signals detected from an incorrectly positioned electrode; as well as The subset and its corresponding labels are included in training data for training a machine learning algorithm to classify the position of the catheter relative to the septum.

15. A system, comprising: one or more processors; as well as One or more non-transitory computer-readable media storing first computer-executable instructions that, when executed by the one or more processors, cause the system to perform actions including: receiving a set of bioelectric signals detected by a catheter, the catheter carrying a plurality of electrodes at respective locations along the length of the catheter so that the electrodes are located at respective different distances from a diaphragm of the patient, the plurality of electrodes being divided into a plurality of electrode pairs, each signal being detected by one of the plurality of electrode pairs, each signal comprising an electrocardiographic (ECG) component; identifying, from the set of bioelectric signals, one or more first bioelectric signals, each first bioelectric signal being detected by an electrode pair on the catheter determined to be properly positioned relative to the diaphragm; dividing the bioelectric signal set into at least two bioelectric signal subsets, each bioelectric signal subset including one or more bioelectric signals and corresponding to respective electrode groups associated with detection of the one or more bioelectric signals in the bioelectric signal subset, wherein the dividing is performed such that each electrode group is a sequence of electrodes placed continuously along the length of the catheter; marking each subset in the plurality of subsets, wherein the marking comprises: marking a subset including at least one of the first bioelectric signals as a subset of signals detected from correctly positioned electrodes, and marking a subset of the first bioelectric signals that does not include any of the first bioelectric signals as a subset of signals detected from an incorrectly positioned electrode; as well as The subset and its corresponding labels are included in training data for training a machine learning algorithm to classify the position of the catheter relative to the septum.

Images