Multimodal waveform feature fusion machine learning nasogastric tube blockage risk prediction method

Abstract

The application discloses a nasal feeding tube blockage risk prediction method based on multi-modal waveform feature fusion machine learning, which comprises the following steps: preprocessing original pressure time sequence waveform samples of a nasal feeding tube to obtain standardized time sequence data; multi-modal feature extraction is performed on the preprocessed time sequence data to fuse time domain, frequency domain and nonlinear features, and sample multi-modal feature vectors are obtained; the sample multi-modal feature vectors are taken as input features, and the corresponding infusion liquid types and the opening rate of the nasal feeding tube are taken as data labels to construct a multi-modal feature dataset; an optimal classification model is obtained by training and optimization based on the dataset; the opening rate of the nasal feeding tube is used to divide the blockage risk grade; real-time pressure waveform samples of the nasal feeding tube are collected, and after preprocessing and multi-modal feature extraction, the optimal model is inputted, the model outputs the opening rate prediction value of the current nasal feeding tube, and risk judgment is performed in combination with the constructed experimental system. The technical problem that the existing method cannot effectively predict the nasal feeding tube blockage risk is solved.

Description

Technical Field

[0001] This invention relates to the field of early prediction technology for the risk of nasogastric tube blockage, and in particular to a method for predicting the risk of nasogastric tube blockage by fusing multimodal waveform features with machine learning. Background Technology

[0002] Enteral nutrition is a crucial nutritional support method for critically ill patients and those unable to eat orally. The nasogastric tube, as a key channel for nutrient delivery, directly impacts treatment efficacy and patient safety. According to the Hagen-Poiseuille law in fluid mechanics, under constant flow conditions, the pressure within the nasogastric tube is inversely proportional to the fourth power of the tube's effective radius. The physical nature of tube blockage is the continuous accumulation of deposits on the inner wall of the tube, leading to localized narrowing and a non-uniform reduction in the effective flow cross-section. This, in turn, causes a dynamic evolution process in which the pressure within the tube rises rapidly and superlinearly.

[0003] Currently, there is a lack of effective predictive and early warning methods for nasogastric tube obstruction in clinical practice. The judgment of obstruction mainly relies on the passive observation and experience of medical staff, such as a decrease in infusion rate, abdominal distension and discomfort in the patient, and increased resistance to tubing aspiration. When the above phenomena occur, the obstruction has usually already occurred or is close to complete, resulting in a serious delay in intervention.

[0004] While existing enteral nutrition infusion devices possess some monitoring and alarm functions, significant shortcomings remain. Some devices rely solely on a single instantaneous pressure threshold for alarms, while others depend on drip rate sensors to detect infusion interruptions and determine tube blockage. Both methods suffer from high false alarm rates, delayed warnings, and an inability to predict blockage trends in advance, making it difficult to meet the needs of early clinical intervention.

[0005] Traditional techniques for preventing and managing nasogastric tube blockage also have significant drawbacks. The incidence of blockage is high, and the effectiveness of prevention is highly dependent on the standardization of manual operation, easily leading to failure due to improper execution. Furthermore, blockage detection relies heavily on manual inspections, making it difficult to detect early, minor blockages in a timely manner. In terms of managing blockages, warm water flushing is the primary method of unblocking, with a low success rate; methods such as using guidewires can easily damage the tube. Severe blockages usually require tube removal and repositioning, increasing patient suffering and medical costs, and significantly increasing the workload of medical staff.

[0006] Therefore, existing nasogastric feeding systems cannot identify the gradual process of slow pressure rise before blockage occurs, making it difficult to predict and warn of nasogastric tube blockage in the early stages. Clinically, there is an urgent need for a nasogastric tube blockage prediction technology that can monitor in real time and predict the blockage trend in advance. Summary of the Invention

[0007] This invention provides a method for predicting the risk of nasogastric tube blockage by fusing multimodal waveform features with machine learning, in order to overcome the above-mentioned technical problems.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A method for predicting the risk of nasogastric tube obstruction by fusing multimodal waveform features with machine learning, specifically including the following steps: S1: Based on the constructed experimental system, a nasogastric tube blockage simulation experiment was carried out. At a set fixed infusion rate, several infusion fluids were collected, and the original pressure time-series waveform samples of the nasogastric tube were collected under different tube opening rates. S2: Perform data preprocessing on the original pressure time-series waveform samples of the nasogastric feeding tube to obtain preprocessed sequence samples; S3: Perform multimodal feature extraction on the preprocessed sequence samples to obtain the sample stress feature vector; Furthermore, the multimodal feature extraction involves extracting time-domain features, frequency-domain features, and nonlinear features from the preprocessed samples. The time-domain features include data mean, data standard deviation, data maximum value, data minimum value, data fluctuation amplitude, data coefficient of variation, data kurtosis, and data skewness. The frequency-domain features include the proportion of low-frequency energy and the proportion of high-frequency energy obtained by performing a fast Fourier transform on the preprocessed sequence samples. The nonlinear feature is the sample entropy. S4: Use the pressure feature vector of the preprocessed sequence sample as feature data, and use the corresponding infusion fluid type and the opening rate of the nasogastric tube as data labels to obtain a multimodal feature vector dataset. S5: Train the pre-set classification model based on the multimodal feature vector dataset to obtain the optimal classification model; set different risk thresholds based on the opening rate of the nasogastric tube; obtain the latest nasogastric tube pressure waveform samples; repeat S2 to S3 to obtain the multimodal feature vector at the current moment; input the multimodal feature vector at the current moment into the optimal classification model; and output the current infusion fluid type and the opening rate of the nasogastric tube. Based on the set risk thresholds of different levels, predict the risk of nasogastric tube blockage based on the current opening rate of the nasogastric tube; and issue warnings through the constructed experimental system.

[0009] Furthermore, the experimental system constructed in S1 includes a main control module, a power supply module, a peristaltic pump assembly, a drip rate detection sensor, a human-machine interaction module, an alarm module, an infusion tube fixing clamp, a connection interface, a signal processing module, a real-time pressure detection sensor, an infusion pipeline, a Y-connector, a nasogastric tube, and a roller-type flow regulator. The input terminals of the main control module are connected to the output terminals of the power supply module and the signal processing module, respectively; the output terminals of the main control module are connected to the alarm module and the peristaltic pump; the main control module has a bidirectional communication connection with the human-machine interface module; the input terminal of the signal processing module is connected to the real-time pressure detection sensor through a connection interface; one end of the infusion pipeline is connected to the nutrient solution storage bag, and the other end of the infusion pipeline is sequentially connected to a dropper, a Y-connector, and one end of the nasogastric tube; the other end of the nasogastric tube is inserted into the nasal cavity; the peristaltic pump assembly acts on the infusion pipeline to push the nutrient solution delivery; the roller-type flow regulator is installed on the nasogastric tube and is used to change the flow cross-sectional area of ​​the nasogastric tube by squeezing it with a fixed-position mechanical roller to simulate the current blockage state of the nasogastric tube, thereby controlling the flow. The patency rate of the nasogastric tube; the real-time pressure detection sensor is installed on the infusion pipeline via a Y-connector, adjacent to one end of the nasogastric tube; the infusion tube fixing clamp is installed on the infusion pipeline and located on both sides of the peristaltic pump assembly; the drip rate detection sensor is used to detect the fluid velocity of the infused liquid in the infusion pipeline; the real-time pressure detection sensor is used to detect the pipeline pressure of the infused liquid in the infusion pipeline; the signal processing module is used to amplify and convert the pipeline pressure corresponding to the preset sampling period to obtain the corresponding nasogastric tube pressure waveform sample; the human-computer interaction module is used to display the risk level corresponding to the output result of the optimal classification model deployed in the main control module; the alarm module is used to provide audible and visual warnings for the risk level.

[0010] Furthermore, step S2 specifically includes the following steps: S21: Perform data cropping on the original pressure time-series waveform sample of the nasogastric feeding tube to obtain the cropped effective pressure time-series waveform sample; The data trimming refers to removing a preset number of start and end sampling point sequences from the original pressure time-series waveform samples of the nasogastric tube. S22: The effective pressure time-series waveform sample after clipping is expanded using the sliding window method to obtain effective sub-samples of the nasogastric tube pressure waveform. S23: Use a moving average filter to denoise the effective subsamples and obtain the filtered effective subsamples; S24: Perform data normalization on the filtered effective sub-samples to obtain the final processed sample, i.e., the preprocessed sequence sample.

[0011] Furthermore, step S3 specifically includes the following steps: performing multimodal feature extraction on the preprocessed sequence samples to obtain the sample stress feature vector. Furthermore, the multimodal feature extraction involves extracting temporal features, frequency domain features, and nonlinear features from the preprocessed samples; the temporal features include the data mean. Data standard deviation Maximum data value Minimum value of data Data fluctuation range Coefficient of variation Data kurtosis and data skewness The frequency domain features include the proportion of low-frequency energy obtained by performing a fast Fourier transform on the preprocessed sequence samples. With high frequency energy ratio The nonlinear characteristic is the sample entropy. The sample pressure feature vector for: .

[0012] Furthermore, the methods for obtaining the optimal classification model in S5 specifically include: S51: Based on the set partitioning strategy, the multimodal feature vector dataset is divided into a training set and a validation set; the partitioning strategy includes, but is not limited to, the 5-fold hierarchical cross-validation method; S52: Train a pre-set classification model using a training set to obtain a trained classification model; the classification model includes, but is not limited to, a random forest model. S53: Use the mean squared error function as the loss function and validate the trained classification model using the validation set; that is, determine whether the output of the trained classification model has converged. If the output of the trained classification model converges, then the model is confirmed to be the optimal classification model. Otherwise, based on the backpropagation method, the weight parameters of the trained classification model are adaptively adjusted, and step S51 is repeated until the weight parameters of the trained classification model that has converged are confirmed to be the optimal weight parameters, and the classification model is reconstructed to obtain the optimal classification model.

[0013] This invention provides a method for predicting the risk of nasogastric tube blockage by fusing multimodal waveform features with machine learning, with the following beneficial effects: (1) By introducing a time-domain, frequency-domain, and nonlinear feature coupling analysis mechanism for pressure time-series data, and fusing the features in a multimodal manner, the time-domain features are used as the core features, directly corresponding to the physical process of increased fluid resistance due to increased deposits in the pipe cavity. Furthermore, by allowing the classification model to simultaneously learn the joint distribution of time-domain, frequency-domain, and nonlinear dynamic features, it can capture the differences in the pressure rise mode in terms of change mode, frequency structure, and dynamic essence, thereby accurately identifying the precursors of pipe blockage. This invention, by combining time-domain, frequency-domain, and nonlinear features, and tailoring specific feature engineering for the physical process of pipe blockage, effectively distinguishes between legitimate pressure fluctuations caused by speed regulation and illegal pressure rises caused by pipe blockage, minimizing the false alarm rate.

[0014] (2) This invention proposes using the pressure waveform characteristics of the nasogastric tube as the basis for determining blockage. Simultaneously, it employs a 5-fold stratified verification method combined with a classification model to construct a high-precision prediction model, namely the optimal classification model. This enables early warning of nasogastric tube blockage risk, effectively addressing the pain point of frequent nasogastric tube blockage in clinical practice and improving the safety and continuity of nasogastric feeding treatment. The optimal classification model in this invention has a dynamic adaptive early warning function for blockage. The early warning system is no longer limited by a fixed pressure threshold, can understand the pressure context, and achieve personalized and scenario-based accurate prediction.

[0015] (3) Through a machine learning model, i.e., an optimal classification model, it is possible to prospectively predict the risk of nasogastric tube blockage and the type of infusion fluid, output a probability value of blockage risk, and establish a dynamic hierarchical early warning system based on this. This transforms the nursing model from passive treatment to proactive prevention, improving the quality of medical care. This invention classifies and identifies different types of infusion fluids. Since the types of commonly used nutritional solutions in clinical practice are limited, the model can establish a prediction mechanism for each type of infusion fluid, thereby significantly improving the accuracy and clinical applicability of blockage prediction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the nasogastric tube obstruction risk prediction method based on multimodal waveform feature fusion machine learning according to the present invention; Figure 2 This is a schematic diagram of the architecture and connection of the experimental system in this embodiment; Figure 3 This is a comparison of the pressure waveforms of physiological saline under different tubing patency rates in this embodiment; Figure 4 This is a comparison of the pressure waveforms of 5% glucose injection under different pipeline opening rates in this embodiment; Figure 5 This is a comparison of the pressure waveforms of Abbott's TPF-FOS nutrient solution under different pipeline opening rates in this embodiment. Figure 6 This is a comparison chart of the opening rate and liquid type prediction model performance indicators in this embodiment; Figure 7 This is a confusion matrix diagram of the activation rate prediction results in this embodiment; Figure 8 This is a confusion matrix diagram of the liquid type prediction results in this embodiment; Figure 9 This is a block diagram of the core technology of the nasogastric tube obstruction risk prediction method based on multimodal waveform feature fusion and machine learning in this embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This embodiment provides a method for predicting the risk of nasogastric tube blockage by fusing multimodal waveform features with machine learning. Figure 1 and Figure 9 As shown, the specific steps include: S1: Based on the constructed experimental system, a nasogastric tube blockage simulation experiment was conducted. At a fixed infusion rate (120 ml / h), several samples of the original nasogastric tube pressure waveforms were collected at different tube opening rates. Specifically, such as... Figure 2As shown, the constructed experimental system includes a main control module (1), a power supply module (2), a peristaltic pump assembly (3), a drip rate detection sensor (4), a human-machine interaction module (5), an alarm module (6), an infusion tube fixing clamp (7), a connection interface (8), a signal processing module (9), a real-time pressure detection sensor (10), an infusion pipeline (11), a Y-connector (12), a nasogastric tube (13), a roller-type flow regulator (14), and a nutrient solution storage bag (15). The input end of the main control module is connected to the output end of the power supply module and the output end of the signal processing module, respectively. The output end of the main control module is connected to the alarm module and the peristaltic pump. The main control module has a bidirectional communication connection with the human-machine interaction module. The input end of the signal processing module is connected to the real-time pressure detection sensor through the connection interface. The drip rate detection sensor is set on the infusion pipeline outside the drip tube. One end of the infusion pipeline is connected to the nutrient solution storage bag, and the other end of the infusion pipeline is connected to the drip tube, the Y-connector, and one end of the nasogastric tube in sequence. The other end of the tube is inserted into the nasal cavity, and the peristaltic pump assembly acts on the infusion tubing to drive the delivery of nutrient solution. The roller-type flow regulator is installed on the nasogastric tube and is used to change the flow cross-sectional area of ​​the nasogastric tube by squeezing it with a fixed-position mechanical roller to simulate the current blockage state of the nasogastric tube, thereby controlling the opening rate of the nasogastric tube. The real-time pressure detection sensor is installed on the infusion tubing via a Y-connector and adjacent to one end of the nasogastric tube. The infusion tube fixing clamp is installed on the infusion tubing and located on both sides of the peristaltic pump assembly. The drip rate detection sensor is used to detect the fluid velocity of the infused liquid in the infusion tubing. The real-time pressure detection sensor is used to detect the tubing pressure of the infused liquid in the infusion tubing. The signal processing module is used to amplify and convert the tubing pressure corresponding to the preset sampling period to obtain the corresponding nasogastric tube pressure waveform sample. The human-computer interaction module is used to display the risk level corresponding to the output result of the optimal classification model deployed in the main control module. The alarm module is used to provide audible and visual warnings for the risk level.

[0020] Nasogastric feeding is a core nutritional support method for patients with swallowing dysfunction or inability to eat independently, such as those suffering from stroke, traumatic brain injury, or critical illness. It is crucial for maintaining nutritional supply, promoting recovery, and preventing complications such as malnutrition. Traditional manual injection or gravity drip methods are burdensome for nursing staff and cannot meet the needs for precise and safe nutritional support in clinical practice. Nasogastric feeding pumps, as standardized and intelligent infusion devices, have become an important tool for providing nutrition to patients requiring nasogastric feeding in clinical practice, and are widely used in various departments and among various patient groups requiring nasogastric feeding. Because nasogastric feeding solutions typically have a high protein and fat content, the lumen gradually narrows during infusion. If the tube is not flushed promptly, it can easily become blocked. Currently, the mainstream clinical approach to managing nasogastric tube blockage is to replace the tube, a process often requiring painful endoscopy, which not only causes significant pain for the patient but also increases medical expenses and the overall healthcare burden due to the high cost of the nasogastric tube itself and frequent replacements. Therefore, conducting research on predicting nasogastric tube blockage, providing early warnings of blockage risks, and implementing targeted preventative measures are of significant practical importance. Figure 1This is a schematic diagram of the overall architecture of the system described in this embodiment: The core of the system is a main control module equipped with a control chip for unified coordination, and the power supply module provides power to each module; during operation, the nutrient solution in the nutrient solution storage bag is transported through a disposable infusion tube, which is fixed by an infusion tube clamp; the velocity of the fluid in the tube and the pressure in the tube are monitored by an infrared through-beam drip rate sensor and a real-time pressure sensor, respectively, and the data is transmitted through a detachable connection interface. The interface type can be a common medical device interface type, such as a USB interface, an Abbott interface, or a Mindray interface. The purpose of the connection interface is to prevent incorrect wiring and facilitate flushing of the pressure sensor. The output data of the real-time pressure sensor... Since the signal is analog, a signal processing module is needed to amplify and convert the signal from analog to digital before transmitting the data to the main control module. The inlet of the real-time pressure sensor is connected to the Y-connector of the disposable infusion tubing via a Luer lock connector. Specifically, the real-time pressure sensor uses a male Luer lock connector, and the Y-connector is a female Luer lock connector. The end of the disposable infusion tubing is connected to the beginning of the nasogastric tube via a clinically standardized multi-column connector. This symmetrical and standardized interface design allows the real-time pressure sensor to be seamlessly integrated into existing enteral nutrition infusion tubing as an independent monitoring unit without altering the original tubing structure. Installation and disassembly are convenient, greatly facilitating clinical use and subsequent flushing and maintenance. The nutrient solution storage bag serves as the storage container for the enteral nutrition solution and is the storage location for the infusion fluid in the entire system. A roller-type flow regulator is installed near the end of the nasogastric tube. By using a fixed-position mechanical roller to compress the tubing and change the flow cross-sectional area, it simulates the current blockage state of the nasogastric tube and controls the opening rate. The main control module embeds a machine learning algorithm, namely the optimal classification model mentioned below, to perform fusion analysis on the received multimodal data and calculate the probability of pipe blockage risk in real time. Once the risk is determined to exceed the threshold, the main control module immediately drives the alarm module to issue an audible and visual alarm and can send instructions to the peristaltic pump component to adjust or suspend the infusion, thereby realizing closed-loop control from intelligent perception, risk decision-making to active intervention.

[0021] In this embodiment, an experimental system was built to complete the basic hardware and software framework. Medical nutrition infusion tubing was installed, and a real-time pressure sensor was installed within the tubing. A roller-type flow regulator was deployed in the nasogastric tube passage to control the tube's patency. To eliminate interference from the infusion rate on the pressure waveform and ensure a single experimental variable, the infusion rate of the nasogastric solution was set to a fixed value. The type of infused fluid and the patency of the nasogastric tube were set according to the experimental variables, and each experiment was repeated 5 times. Specifically, the following steps were taken: S11: In this embodiment, by setting the type of infusion fluid, the commonly used clinical fluids such as physiological saline, 5% glucose injection, and Abbott Cialis TPF-FOS nutrient solution are selected according to the experimental requirements to clarify the type of infusion fluid and record its physical properties. S12: In this embodiment, the infusion rate is precisely set to the clinically recommended rate of 120 ml / h through the human-computer interaction module, and the flow control conforms to the infusion rate below the maximum safe rate of 150 ml / h for clinical adults, ensuring that the liquid flows at a constant rate and controlling the consistency of variables; S13: This embodiment sets seven levels of pipeline opening rates, namely 0%, 20%, 40%, 50%, 60%, 80%, and 100%. By adjusting the position of the roller of the roller-type flow regulator, different pipeline opening rates are set corresponding to different positions of the roller, thereby simulating pipeline blockage scenarios of different degrees, defining experimental conditions of different opening levels, and being used to study the pressure change law and fluid behavior characteristics under conditions of different opening rates; S14: In this embodiment, data is collected at a fixed sampling frequency. The main control module collects real-time pressure time-series data, and through the real-time pressure detection sensor installed on the infusion pipeline, continuously collects the pressure data of a fixed time in the pipeline to form the original pressure time-series waveform, that is, the original nasogastric tube pressure waveform sample. . In this example, the pressure data in the pipeline within 50 s is continuously collected at a fixed frequency of 50 Hz to form the original pressure time-series waveform sample P = { }, n = 2500. Under the experimental conditions of the same infusion liquid type and the opening rate of the same nasogastric tube pipeline, the experiment is repeated 5 times, and finally 105 samples can be obtained.

[0022] This embodiment uses an in vitro simulation experiment platform to carry out relevant experiments. The enteral nutrition infusion pump and the clinical standard nasogastric tube are used as the core experimental devices, and three liquids, namely normal saline, 5% glucose injection, and Abbott Jevity TPF-FOS nutrient solution, are selected as the infusion media. During the experiment, by connecting a high-precision pressure sensor in the nasogastric tube pipeline, the real-time acquisition of the pressure time-series signal is realized. To ensure consistent experimental conditions and improve the repeatability of data, this study sets the nasogastric fluid infusion rate to a fixed value to exclude the interference of flow rate changes on the pressure baseline, ensure the comparability of the pressure waveforms collected under different opening rates of the nasogastric tube pipeline, and make the difference in the pressure signal only determined by the degree of pipeline blockage, thereby improving the reliability and accuracy of model training. In the experiment, the on-off state of the pipeline is controlled by the roller-type flow regulator supporting the nasogastric tube. Adjusting the opening and closing degree of the roller clamp can change the pipeline flow resistance, and thus achieve precise adjustment of seven levels of opening rates, so as to simulate different degrees of blockage states of clinical nasogastric tubes. The larger the pipeline opening, the smaller the liquid flow resistance, the higher the opening rate, and the lighter the blockage degree; the higher the clamping degree of the roller clamp, the smaller the pipeline opening, the greater the flow resistance, the lower the opening rate, and the heavier the blockage degree. In this embodiment, the infusion liquid type, pipeline opening rate, and complete pressure waveform are synchronously recorded during the experiment, and a nasogastric tube blockage pressure signal correlation dataset containing multiple types of nutrient solutions and multiple blockage levels is constructed to provide data support for subsequent feature extraction and prediction model training.

[0023] S2: Perform data preprocessing on the original pressure time-series waveform samples from the nasogastric feeding tube to obtain preprocessed sequence samples. Specific steps include: S21: Perform data cropping on the original pressure time-series waveform sample of the nasogastric tube to obtain the cropped effective pressure time-series waveform sample; the data cropping is to remove a preset number of start sampling point sequences and end sampling point sequences from the original pressure time-series waveform sample of the nasogastric tube. Specifically, based on the data acquisition duration and sampling frequency, a total of 2500 original pressure time-series waveform samples P are obtained. To eliminate invalid fluctuations at the beginning and end of the data, the data is cropped, deleting the first 5 seconds and the last 5 seconds of sampling points, retaining the valid data in the middle 40 seconds of each sample, thus obtaining the effective pressure time-series waveform samples. }, m=2000, For each effective pressure time-series waveform sample The number of sampling points in the sample; S22: The effective pressure time-series waveform sample after clipping is expanded using the sliding window method to obtain effective sub-samples of the nasogastric tube pressure waveform. Specifically, the effective pressure timing waveform sample of the trimmed nasogastric tube. The sliding window method was used to expand the sample size. The window duration was set to 10 seconds and the sliding step size to 5 seconds. A single 40-second effective pressure time series waveform was uniformly truncated from its starting point, with a 10-second continuous pressure waveform segment extracted every 5 seconds as one effective subsample. The effective subsamples were obtained using the sliding window method. Where r is the number of data points sampled in the effective subsample. After the above sample expansion, the single-line pruning sample Able to extract 7 valid subsamples Ultimately, 735 standardized pressure waveform training samples were obtained, achieving a 7-fold increase in sample size; S23: Use a moving average filter to denoise the effective subsamples and obtain the filtered effective subsamples; Specifically, for all the above valid subsamples Use a window size of A moving average filter is used for initial smoothing to remove sharp impulse noise. Based on the fluid dynamics of the nasogastric tube, pressure changes in laminar flow are continuous, smooth, and respond slowly to flow rate and tube diameter. That is, the pressure change caused by tube blockage is gradual rather than abrupt. To preserve the pressure details to the greatest extent, a moving average filter is used. 1; the filtered first sampling points The calculation formula is:

[0024] in, Indicates the sequence number of the current sampling point to be filtered. This indicates traversing the index within the window. Indicates half the width of the window. Indicates the first in the window 1 effective subsample sampling points were obtained .

[0025] S24: Perform data normalization on the filtered effective sub-samples to obtain the final sample, i.e., the preprocessed sequence sample; Specifically, this embodiment will include all sampling points Perform min-max normalization to linearly map the sample points to the [0,1] interval, resulting in normalized sample points. :

[0026] in These are the effective subsamples after filtering. The minimum and maximum values ​​in the sequence are determined. The final preprocessed sequence sample is then obtained. .

[0027] S3: Perform multimodal feature extraction on the preprocessed sequence samples to obtain the sample stress feature vector; specifically including the following steps: extracting multimodal features from 735 preprocessed sequence samples. Perform multimodal feature extraction to obtain sample stress feature vectors. Furthermore, the multimodal feature extraction involves extracting temporal features, frequency domain features, and nonlinear features from the preprocessed samples; the temporal features include the data mean. Data standard deviation Maximum data value Minimum value of data Data fluctuation range Coefficient of variation Data kurtosis and data skewness The time-domain characteristics directly reflect the numerical variation of the pressure signal and are the core feature set for pipe blockage prediction; the frequency-domain characteristics include the proportion of low-frequency energy. With high frequency energy ratio Frequency domain features are obtained by analyzing pressure sequences. The signal is obtained by performing a Fast Fourier Transform to capture changes in its frequency components; the nonlinear feature is the sample entropy of the signal quantified through nonlinear characteristics. This helps improve prediction accuracy; the sample pressure feature vector for:

[0028] S4: Use the pressure feature vector of the preprocessed sequence sample as feature data, and use the corresponding infusion fluid type and the opening rate of the nasogastric tube as data labels to obtain a multimodal feature vector dataset. Specifically, in this embodiment, each preprocessed sequence sample has two labels corresponding to its multimodal feature vector. , The infusion fluids used in the sample experiments are indicated by 1, 2, and 3, which represent physiological saline, 5% glucose, and Abbott Cialis TPF-FOS nutrient solution, respectively. This indicates the patency rate of the nasogastric tube in the sample experiment. These represent the current patency level of the nasogastric tube. Five risk levels are used to assess the risk of pipe blockage; the lower the value, the more severe the blockage. The final result is a feature vector. and their corresponding tags These features are combined to form a multimodal feature vector dataset for model training.

[0029] S5: Train the pre-set classification model using the multimodal feature vector dataset to obtain the optimal classification model. This includes the following steps: S51: Based on the set partitioning strategy, the multimodal feature vector dataset is divided into a training set and a validation set; the partitioning strategy includes, but is not limited to, the 5-fold hierarchical cross-validation method; S52: Train a pre-set classification model using a training set to obtain a trained classification model; the classification model includes, but is not limited to, a random forest model. S53: Use the mean squared error function as the loss function and validate the trained classification model using the validation set; that is, determine whether the output of the trained classification model has converged. If the output of the trained classification model converges, then the model is the optimal classification model. Otherwise, based on the backpropagation method, the weight parameters of the trained classification model are adaptively adjusted, and step S51 is repeated until the weight parameters of the trained classification model that has converged are confirmed to be the optimal weight parameters, and the classification model is reconstructed to obtain the optimal classification model.

[0030] In this embodiment, the multimodal feature vector dataset D is partitioned: the preprocessed pressure feature dataset is partitioned according to a 5-fold stratified cross-validation rule, with the stratification based on infusion fluid type 1, 2, and 3, and the activation rate. The combination of categories ensures that the proportion of samples from each category in each fold of data is completely consistent with the distribution of the 735 preprocessed sequence samples, avoiding the loss or imbalance of samples with specific infusion conditions in a certain fold of data due to random partitioning. A random forest model is constructed: the number of decision trees in the random forest is set to 100, the decision tree splitting criterion is the Gini coefficient, the above 11 core stress features are selected, and the maximum depth of a single decision tree is controlled to 10 layers to prevent model overfitting. This embodiment performs 5-fold iterative training: 4 folds of data are used as the training set and 1 fold as the validation set, and the random forest model is iteratively trained. After each round of training, the stress prediction accuracy and mean squared error of the validation set are calculated, and the model performance indicators for each fold are recorded. Model evaluation: the average performance indicators of the 5-fold validation are calculated: average accuracy, average mean squared error, and confusion matrix. If the performance deviation of a certain fold exceeds 5%, the feature subset or the number of decision trees in the random forest is readjusted until the 5-fold validation results are stable, which is then used as the final classification model M. Different risk thresholds are set based on the patency rate of the nasogastric tube. By obtaining the latest nasogastric tube pressure waveform samples, S2 to S3 are repeated to obtain the multimodal feature vector at the current moment. The multimodal feature vector at the current time. The data is input into the optimal classification model, which outputs the current infusion fluid type and the patency rate of the nasogastric tube. Based on different risk thresholds, the risk of nasogastric tube blockage is predicted according to the current patency rate. Simultaneously, an experimental system is used for alerts. In this embodiment, the model outputs a 7-level patency rate to achieve intelligent risk-free, low-risk, medium-risk, high-risk, and ultra-high-risk tiered alarms. 100% and 80% patency rates correspond to unobstructed and risk-free operation; 60% corresponds to low risk; 50% and 40% correspond to medium risk; 20% corresponds to high risk; and 0% corresponds to ultra-high risk. The fluid type output by the model is used to verify the universality of the risk level classification method, ensuring that the correspondence between patency rate and risk level remains stable under different fluid conditions. This embodiment's dynamic tiered early warning and active intervention: The main control module assigns the risk level corresponding to the classification results. No risk: The main control module drives the human-computer interaction module to display a green box that says "No risk".

[0031] Low risk: The main control module drives the human-computer interaction module to display a blue box that says "Low risk".

[0032] Medium risk: The main control module drives the human-machine interaction module to display a yellow box indicating "medium risk", reminding users to increase the flushing frequency.

[0033] High risk: The main control module drives the human-machine interaction module to display an orange "High risk" prompt, and simultaneously executes the pipe flushing reminder module, i.e., the alarm module, to issue an audible and visual reminder, and the interface displays the message "High risk pipe blockage warning, please intervene immediately!"

[0034] Extremely high risk: The main control module drives the human-machine interaction module to display a red box indicating "extremely high risk", and simultaneously executes the pipe blockage alarm, that is, the alarm module issues an audible and visual reminder, and the interface displays the message "High risk pipe blockage warning, please intervene immediately!"

[0035] This embodiment also includes the following experimental results: Figures 3 to 5 The images show pressure waveforms of physiological saline, 5% glucose injection, and Abbott Cialis TPF-FOS nutrient solution under different tubing patency rates. The amplitude changes and waveform characteristics were analyzed by collecting pressure signals from the nasogastric tube. Figures 3 to 5 The three graphs show that the pressure waveform change for the same infusion fluid is related to the current patency of the nasogastric tube, i.e., it is significantly correlated with the degree of blockage of the nasogastric tube. Furthermore, due to the different physical properties of different fluids, the pressure waveform can also be used to distinguish the current infusion fluid type. This verifies the feasibility and effectiveness of the present invention in predicting nasogastric tube blockage and infusion fluid type using pressure signals.

[0036] Figure 6 The chart shows a comparison of the performance metrics of the open rate and liquid type prediction models. The results show that the random forest model achieved high prediction accuracy in both tasks, with average 5-fold precision of 0.9138 and 0.9324, respectively. The weighted precision, recall, and F1 score were all stable in the range of 0.91 to 0.94, and the standard deviation of the precision was all below 0.04, indicating that the model has good classification ability and generalization stability.

[0037] Figures 7 to 8 The confusion matrix for the prediction results of openness rate (blockage degree) and liquid type. Figure 7 The confusion matrix of the opening rate shows that the model's accuracy in identifying each congestion level category is above 0.86, which can effectively distinguish different congestion levels. Figure 8 The liquid type confusion matrix shows that the model achieves a classification accuracy exceeding 0.91 for all three liquid categories, with only minor inter-category confusion, demonstrating extremely high classification consistency. These results fully validate the effectiveness and reliability of the proposed model in both access rate grading and liquid type identification tasks.

[0038] The method described in this embodiment integrates time-domain, frequency-domain, and nonlinear features to provide machine learning models with a basis for distinguishing the physical origins of different pressure fluctuations. This multi-dimensional, context-aware comprehensive analysis mechanism theoretically avoids the shortcomings of the single threshold method, which tends to interpret information out of context, and is expected to significantly reduce alarm fatigue caused by false alarms. This invention is the first to propose using nasogastric tube pressure waveform features as the core basis for determining nasogastric tube obstruction. Based on nasogastric tube pressure waveform feature extraction and a 5-fold hierarchical cross-validation strategy, an obstruction prediction model is constructed to achieve early warning of nasogastric tube obstruction risk.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the risk of nasogastric tube blockage using multimodal waveform feature fusion and machine learning, characterized in that, The specific steps include: S1: Based on the constructed experimental system, a nasogastric tube blockage simulation experiment was carried out. At a set fixed infusion rate, several infusion fluids were collected, and the original pressure time-series waveform samples of the nasogastric tube were collected under different tube opening rates. S2: Perform data preprocessing on the original pressure time-series waveform samples of the nasogastric feeding tube to obtain preprocessed sequence samples; S3: Perform multimodal feature extraction on the preprocessed sequence samples to obtain the sample stress feature vector; Furthermore, the multimodal feature extraction involves extracting time-domain features, frequency-domain features, and nonlinear features from the preprocessed sequence samples. The time-domain features include data mean, data standard deviation, data maximum value, data minimum value, data fluctuation amplitude, data coefficient of variation, data kurtosis, and data skewness. The frequency-domain features include the proportion of low-frequency energy and the proportion of high-frequency energy obtained by performing a fast Fourier transform on the preprocessed sequence samples. The nonlinear feature is the sample entropy. S4: Use the pressure feature vector of the preprocessed sequence sample as feature data, and use the corresponding infusion fluid type and the opening rate of the nasogastric tube as data labels to obtain a multimodal feature vector dataset. S5: Train the pre-set classification model based on the multimodal feature vector dataset to obtain the optimal classification model; set different risk thresholds based on the opening rate of the nasogastric tube; obtain the latest nasogastric tube pressure waveform samples; repeat S2 to S3 to obtain the multimodal feature vector at the current moment; input the multimodal feature vector at the current moment into the optimal classification model; and output the current infusion fluid type and the opening rate of the nasogastric tube. Based on the set risk thresholds of different levels, predict the risk of nasogastric tube blockage based on the current opening rate of the nasogastric tube; and issue warnings through the constructed experimental system.

2. The method for predicting the risk of nasogastric tube blockage using multimodal waveform feature fusion machine learning according to claim 1, characterized in that, The experimental system built in S1 includes a main control module, a power supply module, a peristaltic pump assembly, a drip rate detection sensor, a human-machine interaction module, an alarm module, an infusion tube fixing clamp, a connection interface, a signal processing module, a real-time pressure detection sensor, an infusion pipeline, a Y-connector, a nasogastric tube, and a roller-type flow regulator. The input terminals of the main control module are connected to the output terminals of the power supply module and the signal processing module, respectively; the output terminal of the main control module is connected to the alarm module and the peristaltic pump; the main control module has a bidirectional communication connection with the human-machine interface module; the input terminal of the signal processing module is connected to the real-time pressure detection sensor through a connection interface; one end of the infusion pipeline is connected to the nutrient solution storage bag, and the other end of the infusion pipeline is sequentially connected to a dropper, a Y-connector, and one end of the nasogastric tube, the other end of which is inserted into the nasal cavity; the peristaltic pump assembly acts on the infusion pipeline to push the nutrient solution delivery; the roller-type flow regulator is installed on the nasogastric tube and is used to change the flow cross-sectional area of ​​the nasogastric tube by squeezing it with a fixed-position mechanical roller to simulate the current blockage state of the nasogastric tube, thereby controlling the flow. The system controls the opening rate of the nasogastric tube; the real-time pressure detection sensor is installed on the infusion pipeline via a Y-connector, adjacent to one end of the nasogastric tube; the infusion tube fixing clamp is installed on the infusion pipeline and located on both sides of the peristaltic pump assembly; the drip rate detection sensor is used to detect the fluid velocity of the infused liquid in the infusion pipeline; the real-time pressure detection sensor is used to detect the pipeline pressure of the infused liquid in the infusion pipeline; the signal processing module is used to amplify and convert the pipeline pressure corresponding to the preset sampling period to obtain the corresponding nasogastric tube pressure waveform sample; the human-computer interaction module is used to display the risk level corresponding to the output result of the optimal classification model deployed in the main control module; the alarm module is used to provide audible and visual warnings for the risk level.

3. The method for predicting the risk of nasogastric tube blockage using multimodal waveform feature fusion machine learning according to claim 2, characterized in that, S2 specifically includes the following steps: S21: Perform data cropping on the original pressure time-series waveform sample of the nasogastric feeding tube to obtain the cropped effective pressure time-series waveform sample; The data trimming refers to removing a preset number of start and end sampling point sequences from the original pressure time-series waveform samples of the nasogastric tube. S22: The effective pressure time-series waveform sample after clipping is expanded using the sliding window method to obtain effective sub-samples of the nasogastric tube pressure waveform. S23: Use a moving average filter to denoise the effective subsamples and obtain the filtered effective subsamples; S24: Perform data normalization on the filtered effective sub-samples to obtain the final processed sample, i.e., the preprocessed sequence sample.

4. The method for predicting the risk of nasogastric tube blockage using multimodal waveform feature fusion machine learning according to claim 3, characterized in that, S3 specifically includes the following steps: Multimodal feature extraction is performed on the preprocessed sequence samples to obtain the sample stress feature vector. Furthermore, the multimodal feature extraction involves extracting time-domain features, frequency-domain features, and nonlinear features from the preprocessed data. The time-domain features include the data mean. Data standard deviation Maximum data value Minimum value of data Data fluctuation range Coefficient of variation Data kurtosis and data skewness The frequency domain features include the proportion of low-frequency energy obtained by performing a fast Fourier transform on the preprocessed sequence samples. With high frequency energy ratio The nonlinear characteristic is the sample entropy. The sample pressure feature vector for: 。 5. The method for predicting the risk of nasogastric tube blockage using multimodal waveform feature fusion machine learning according to claim 4, characterized in that, The methods for obtaining the optimal classification model in S5 include: S51: Based on the set partitioning strategy, the multimodal feature vector dataset is divided into a training set and a validation set; the partitioning strategy includes, but is not limited to, the 5-fold hierarchical cross-validation method; S52: Train a pre-set classification model using a training set to obtain a trained classification model; the classification model includes, but is not limited to, a random forest model. S53: Use the mean squared error function as the loss function and validate the trained classification model using the validation set; that is, determine whether the output of the trained classification model has converged. If the output of the trained classification model converges, then the model is confirmed to be the optimal classification model. Otherwise, based on the backpropagation method, the weight parameters of the trained classification model are adaptively adjusted, and step S51 is repeated until the weight parameters of the trained classification model that has converged are confirmed to be the optimal weight parameters, and the classification model is reconstructed to obtain the optimal classification model.

Images