KR20240135146A - A postoperative hypoxia prediction system through real-time analysis of intraoperative spirometry signals - Google Patents

Abstract

The present invention provides a postoperative hypoxia prediction system through real-time analysis of intraoperative spirometry signals, characterized by including a training module for extracting first characteristic information from a planned spirometry signal and implementing a hypoxia learning model based on the results of machine learning and deep learning of the extracted first characteristic information, and a hypoxia prediction module for extracting second characteristic information from a spirometry signal input in real time during surgery and predicting hypoxia based on the results predicted through the machine-learned hypoxia first learning model in the training module and the deep-learned hypoxia second learning model in the training module, respectively.

Description

{A postoperative hypoxia prediction system through real-time analysis of intraoperative spirometry signals}

The present invention relates to a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery, and more specifically, to a system for predicting hypoxia in a patient after surgery using spirometry signals.

Laparoscopic or robotic-assisted surgery is often used for lower abdominal surgery because it generally has advantages in terms of postoperative pain and postoperative recovery. Robot-assisted surgery has very small scars, a low possibility of bleeding and blood transfusion, less pain, and a low risk of infection, so the satisfaction with the surgical results is high. Accordingly, the demand and importance of robotic surgery are steadily increasing.

However, in the case of laparoscopic or robotic-assisted surgery, the Trendelenburg position is required for the procedure of the pelvic organs, and the position changes the respiratory mechanics by causing the diaphragm to move toward the head and increase the thoracic pressure. In addition, hypoxia is induced by the formation of atelectasis due to paralyzed muscles due to anesthesia and surgery, which increases the respiratory-perfusion mismatch. It has been reported that postoperative hypoxia occurred in 37% of patients undergoing laparoscopic or robotic-assisted surgery. Hypoxia refers to a state in which the partial pressure of oxygen in the body is low, and the normal oxygen saturation in room air is 95% or higher.

It has been reported that prolonged hypoxia is associated with poor prognosis, such as brain damage and postoperative pulmonary embolism. Postoperative hypoxia can also be used as an indicator of postoperative pulmonary complications. However, postoperative intervention after postoperative hypoxia occurs has the disadvantage of increasing medical costs, such as oxygen therapy and increased length of hospital stay. If postoperative hypoxia can be predicted during surgery, it will be helpful in preventing postoperative hypoxia through active intervention during invasive airway management.

However, there is currently no technology to predict postoperative hypoxia in real time.

In particular, various pulmonary protective anesthesia strategies have been proposed to prevent postoperative pulmonary complications, but it is still questionable whether uniform intervention without stratifying patients improves clinical prognosis. In other words, if it is possible to predict postoperative hypoxia in patients and perform stratified interventions, and if it is possible to re-evaluate patients through spirometry that changes according to the intervention, it can be said that the clinical utility is very high.

(Patent Document 1) Republic of Korea Publication Patent No. 10-2022-0108944 (August 4, 2022)

The purpose of the present invention to solve the above problems is to provide a real-time post-operative hypoxia prediction system through analysis of intraoperative spirometry signals, which constructs a post-operative hypoxia learning model by using the first characteristic information extracted from the first spirometry signal acquired in a planned manner, and then applies the second characteristic information extracted from the second spirometry signal input during an actual patient's surgery to the hypoxia learning model to output the presence or absence of hypoxia in the patient, a CAM image, and the first characteristic information, thereby stratifying the patient in need of intervention at the anesthesia maintenance stage and improving the patient's prognosis, and preventing hypoxia in the patient in advance.

The technical problems to be achieved by the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present invention belongs from the description below.

In order to achieve the above purpose, the present invention provides a postoperative hypoxia prediction system through real-time analysis of spirometry signals during surgery, characterized by including: a training module for extracting first characteristic information from a planned and acquired spirometry signal and implementing a hypoxia learning model based on machine learning and deep learning, respectively, using the extracted first characteristic information; and a hypoxia prediction module for extracting second characteristic information from a spirometry signal input in real time during surgery and predicting hypoxia based on results predicted through the hypoxia first learning model learned by machine learning in the training module and the hypoxia second learning model learned by deep learning in the training module, respectively.

In an embodiment of the present invention, the first characteristic information includes first machine learning characteristic information and first deep learning characteristic information, and the training module may include a characteristic extraction unit that smoothes the acquired spirometry signal, corrects missing values, and preprocesses it, and then extracts the first machine learning characteristic information including basic signal characteristics, variability, correlation between lung capacity and intra-airway pressure, and differences and correlations between characteristic information extracted from the Supine posture and the Trendelenburg posture, respectively; and a machine learning unit that learns the first machine learning characteristic information through machine learning to construct a hypoxia primary learning model and performs a primary analysis of the first machine learning characteristic information.

In an embodiment of the present invention, the training module may further include an image conversion unit that converts the first deep learning characteristic information, which is a lung capacity signal for lung capacity extracted from the acquired spirometry signal and an airway pressure signal for airway pressure extracted from the acquired spirometry signal, into a two-dimensional image; and a deep learning unit that deep-learns the two-dimensional image to construct a hypoxia secondary learning model and performs a secondary analysis of the two-dimensional image.

In an embodiment of the present invention, the training module further includes a hypoxia prediction algorithm unit that constructs a hypoxia learning model by integrating the hypoxia first learning model and the hypoxia second learning model, and the hypoxia prediction algorithm unit may be characterized by integrating the hypoxia first learning model and the hypoxia second learning model through an ensemble technique (hard voting, weighted voting, soft voting).

In an embodiment of the present invention, the second characteristic information includes second machine learning characteristic information and second deep learning characteristic information, and the hypoxia prediction module may be characterized by including: a real-time characteristic extraction unit that smoothes the spirometry signal, corrects missing values, and preprocesses it, and then extracts the second machine learning characteristic information including basic signal characteristics, variability, correlation between lung capacity and intra-airway pressure, and differences and correlations between characteristic information extracted from the Supine posture and the Trendelenburg posture, respectively; and a hypoxia primary prediction unit that applies the second machine learning characteristic information to the hypoxia primary learning model to first predict the hypoxia.

In an embodiment of the present invention, the hypoxia prediction module may further include a real-time image conversion unit that converts the second deep learning characteristic information, which is a lung capacity signal for lung capacity extracted from the spirometry signal and an airway pressure signal for airway pressure extracted from the spirometry signal, into a real-time two-dimensional image; and a hypoxia secondary prediction unit that applies the real-time two-dimensional image to the hypoxia secondary learning model to make a second prediction of the hypoxia.

In an embodiment of the present invention, the hypoxia prediction module may further include a hypoxia comprehensive prediction unit that predicts hypoxia by synthesizing the first predicted result through the hypoxia first learning model and the second predicted result through the hypoxia second learning model.

In an embodiment of the present invention, the hypoxia prediction module may further include a CAM analysis unit that analyzes a characteristic map for the real-time two-dimensional image to generate a CAM image.

In an embodiment of the present invention, the hypoxia prediction module may further include a display unit that displays the presence or absence of hypoxia predicted by the hypoxia comprehensive prediction unit, the CAM image, and the first characteristic information.

The effect of the present invention according to the above configuration is that after constructing a hypoxia learning model using the first characteristic information extracted from the first spirometry signal acquired, the second characteristic information extracted from the second spirometry signal input during surgery of an actual patient is applied to the hypoxia learning model to output whether the patient has hypoxia, a CAM image, and the first characteristic information, thereby stratifying the patient in need of intervention at the anesthesia maintenance stage, improving the patient's prognosis, and preventing hypoxia in the patient in advance.

The effects of the present invention are not limited to the effects described above, and should be understood to include all effects that can be inferred from the detailed description of the present invention or the composition of the invention described in the claims.

FIG. 1 is a block diagram showing a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention.
FIGS. 2(a), (b), (c), (d), and (e) are graphs showing that the feature extraction unit of a system for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals according to one embodiment of the present invention preprocesses the supine posture and the Trendelenburg posture through peak inspiratory pressure (PIP).
FIGS. 3(a), (b), (c), and (d) are drawings showing the conversion of spirometry signals acquired by an image conversion unit of a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention into two-dimensional images.
FIGS. 4(a) and 4(b) are drawings showing a CAM analysis unit provided in a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention, which analyzes a characteristic map for the real-time two-dimensional image to generate a CAM image.
FIG. 5 (a) and (b) are drawings showing a display unit provided in a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention, which outputs actual machine learning classification performance and importance by signal characteristic.
FIG. 6 is a flowchart illustrating a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention.
FIG. 7 is a block diagram showing the flow of a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention.
FIG. 8 is a flowchart showing detailed steps of a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention.

Hereinafter, the present invention will be described with reference to the attached drawings. However, the present invention can be implemented in various different forms, and therefore is not limited to the embodiments described herein. In addition, in order to clearly describe the present invention in the drawings, parts that are not related to the description are omitted, and similar parts are assigned similar drawing reference numerals throughout the specification.

Throughout the specification, when a part is said to be "connected (connected, contacted, joined)" to another part, this includes not only the case where it is "directly connected" but also the case where it is "indirectly connected" with another member in between. Also, when a part is said to "include" a certain component, this does not mean that other components are excluded, unless otherwise specifically stated, but that other components can be included.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present invention. The singular expression includes the plural expression unless the context clearly indicates otherwise. As used herein, the terms "comprises" or "has" and the like are intended to specify the presence of a feature, number, step, operation, component, part or combination thereof described in the specification, but should be understood to not exclude in advance the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts or combinations thereof.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the attached drawings.

1. Postoperative hypoxia prediction system through real-time analysis of spirometry signals during surgery (300)

Hereinafter, a system for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals according to one embodiment of the present invention will be described with reference to FIGS. 1 to 4.

FIG. 1 is a block diagram showing a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention.

Referring to FIG. 1, a system (300) for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention includes a training module (100) and a hypoxia prediction module (200).

The training module (100) extracts first characteristic information from the acquired spirometry signal and implements a hypoxia learning model based on machine learning and deep learning, respectively, using the extracted first characteristic information.

Here, the first characteristic information refers to information about some of the multiple signals (=factors) for determining whether hypoxia occurs or not according to lung dynamics.

Specifically, the first feature information includes first machine learning feature information and first deep learning feature information.

The first machine learning feature information includes the basic signal characteristics, variability, correlation between lung capacity and airway pressure, and differences and correlations between feature information extracted from the Supine and Trendelenburg positions, respectively, after preprocessing the acquired spirometry signal by smoothing it and correcting missing values.

In addition, the above-mentioned acquired spirometry signals are spirometry signals acquired from various patients (hereinafter referred to as 'first spirometry signals') as data for deriving a hypoxia first learning model through machine learning and a hypoxia second learning model through deep learning.

Additionally, the present invention utilizes a spirometry signal obtained through a pulmonary function test, which is a method of examining the function of the respiratory process that takes place in the lungs.

In this regard, the lungs play a role in taking in oxygen and expelling carbon dioxide produced in the body. In a narrow sense, respiration refers to the gas exchange process that takes place in the lungs, which includes ventilation, which refers to the flow of air entering and leaving the lungs, perfusion, which refers to the flow of blood that reaches the lungs, and diffusion, which refers to the process in which oxygen binds to red blood cells in the capillaries in the alveoli.

The above-mentioned pulmonary function tests are tools that objectively evaluate the functional aspects of the lungs. Pulmonary function tests include spirometry, bronchodilator tests, diffusing capacity tests, lung volume tests, and nonspecific bronchial provocation tests. These tests do not specifically diagnose a specific disease, but they show unique patterns depending on the disease, so they help diagnose the disease early or evaluate the progression of the disease by stage. In addition, changes in the measured values reflect changes in the severity of the disease, suggesting solutions to clinical situations in various pulmonary diseases, and can help determine the effectiveness before and after treatment or the reversibility of treatment.

The types of pulmonary function tests mentioned above include spirometry, bronchodilator tests, diffusing capacity tests, lung volume tests, and nonspecific bronchial provocation tests.

Spirometry is the most basic pulmonary function test method, which measures the amount of air a subject can exhale after inhaling as much as possible, and is the method used in the present invention.

Next, the bronchodilator test measures the change in lung capacity after inhaling a short-acting bronchodilator that expands the airway. When obstructive ventilation disorders are observed, it helps diagnose and evaluate asthma and chronic obstructive pulmonary disease by measuring the reversibility of bronchodilator-induced obstructive pulmonary disease. It is performed to assess the degree of dyspnea, the severity of chronic obstructive pulmonary disease, and the risk before surgery, and is also performed to assess and grade disability in occupational lung disease. It is measured at time intervals to monitor the progress of the disease.

Next, the diffusion capacity test measures the ability of the lung to transmit carbon monoxide to the alveolar capillaries by inhalation. It is performed to determine the progression of interstitial lung disease, evaluate emphysema, pulmonary edema, primary pulmonary hypertension, acute pulmonary embolism, lung involvement of systemic diseases (rheumatoid arthritis, lupus), use of drugs that cause lung damage, pulmonary hemorrhage, and evaluation before lung resection surgery.

Next, lung volume tests are necessary in addition to spirometry to accurately assess the physiological nature of the disease, such as the severity of the disease, the degree of functional impairment, the course of the disease, and the response to treatment.

Next, nonspecific bronchial provocation test is a disease characterized by airway hyperresponsiveness in which the airway narrows due to airway inflammation response and various provoking stimuli. The method of measuring airway hyperresponsiveness can be broadly divided into a method of observing the improvement of airflow limitation by bronchodilators and a bronchial provocation test in which airflow limitation is induced by applying pharmacological or non-pharmacological stimuli. There are direct bronchial provocation tests and indirect bronchial provocation tests in bronchial provocation tests.

The training module (100) includes a feature extraction unit (110), a machine learning unit (120), an image conversion unit (130), a deep learning unit (140), and a hypoxia prediction algorithm unit (150).

FIGS. 2(a), (b), (c), (d), and (e) are graphs showing that the feature extraction unit of a system for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals according to one embodiment of the present invention preprocesses the supine posture and the Trendelenburg posture through peak inspiratory pressure (PIP).

The feature extraction unit (110) smoothes the acquired spirometry signal (hereinafter referred to as the 'first spirometry signal') and preprocesses it by correcting missing values, and then extracts first machine learning feature information including basic signal features, variability, correlation between lung volume and airway pressure, and differences and correlations between feature information extracted from the supine and Trendelenburg postures, respectively.

First, referring to (a), (b), (c), (d), and (e) of FIG. 2, the feature extraction unit (110) preprocesses the first spirometry signal to distinguish between the supine posture and the Trendelenburg posture through peak inspiratory pressure (PIP).

The supine position is a position where the patient lies upright on an operating table parallel to the ground, and is commonly used in general surgery.

Meanwhile, the Trendelenburg position is a position in which the patient lies on their back on a bed or operating table, lowers their head 30 to 40 degrees, and bends the bed or operating table under their knees. This position is used during laparoscopic surgery.

Although it varies from patient to patient, the spirometry signals obtained in the supine and Trendelenburg positions are different, and these are used to distinguish the patient's position.

Generally, when performing the first surgery, the patient starts in the supine position and changes to the Trendelenburg position during the surgery. At this time, the peak inspiratory pressure (PIP) value increases significantly, and the two positions can be distinguished based on an algorithm through the change and value of the peak inspiratory pressure (PIP).

Methods for distinguishing between the Supine and Trendelenburg postures by tracking changes in the peak inspiratory pressure (PIP) value include a change point detection method for the peak inspiratory pressure (PIP) signal and a rule-based method. In the present invention, the rule-based method was used.

Specifically, referring to (a), (b), (c), (d), and (e) of FIG. 2, the characteristic extraction unit (110) determines that the supine position has been entered when the peak inspiratory pressure (PIP) value is 0.3 or higher and the difference between consecutive peak inspiratory pressure (PIP) values is 0.1 or lower for 5 minutes, and thereafter, when the peak inspiratory pressure (PIP) value exceeds 0.7 twice in a row, the supine position is determined to have ended.

Next, referring to (a), (b), (c), (d), and (e) of FIG. 2, the characteristic extraction unit (110) determines that the Trendelenburg position has been entered when a peak inspiratory pressure (PIP: Peak Inspiratory Pressure) value of 0.7 or higher continues for 5 minutes, and thereafter, when a peak inspiratory pressure (PIP: Peak Inspiratory Pressure) value of 0.5 or lower is continuously recorded, the Trendelenburg position is determined to have ended.

When using the Raw Peak Inspiratory Pressure (Raw PIP) signal, accurate position separation may not be achieved due to the large fluctuations in the signal itself. Therefore, position separation is performed after smoothing the values through Gaussian smoothing.

In the case of such position separation, information obtained from either the Supine position or the Trendelenburg position may have a more significant effect, so that more diverse characteristics can be obtained, such as by considering the correlation between extracted data from the two positions.

Specifically, the surgical time may vary for each patient, and applying windowing rather than extracting one value from the entire signal interval can select more features, and stability and reliability of the extracted signal can also be secured.

Therefore, the feature extraction unit (110) applies a windowing shift to the Supine and Trendelenburg sections obtained for each patient to secure 20 sections (total window number).

Each window consists of 2000 data points (window size).

Since the sampling frequency of the signal acquired from VitalDB is 15 Hz, values for approximately 133 seconds are included in the window.

For the window overlap size, it is defined by [Mathematical Formula 1] below so that the entire section can be used as evenly as possible by adjusting it to the entire patient surgical length.

[Mathematical formula 1]

Overlap size = (total surgery time - window size) / total window number

Therefore, in the present invention, 20 windows are extracted for each of the supine position and the Trendelenburg position, and among the first spirometry signals, the signals for lung volume, airway pressure, and peak inspiratory pressure (PIP) are analyzed.

Accordingly, in the case of signal characteristics extracted for each data and window in the feature extraction unit (110), three types of characteristics are extracted: characteristics of the signal itself, correlation, and variance, and based on the characteristics of the signal itself, maximum, minimum, average, center, skewness, kurtosis, deviation, and moment are extracted.

Specifically, in the case of continuous signal analysis, the feature extraction unit (110) can divide the interval into specific time intervals and extract the average, deviation, kurtosis, skewness, maximum, minimum, and median within the interval, and for two or more signals, can measure the correlation (Pearson correlation, coherence, DTW, rank correlation coefficient).

In the case of correlation, correlation analysis is possible between lung capacity information and airway pressure information in the same position, and correlation analysis is possible between different positions (Supine position) and (Trendelenburg position). In the case of correlation measurement, the correlation coefficient (Pearson correlation) for linear correlation measurement, Kendall Tau correlation, which is a type of rank correlation coefficient for nonlinear analysis, dynamic time warping (DTW), and coherence, which are mainly used to identify signal similarity in the medical field, are used.

In addition, the representative biosignals included in the first characteristic information mentioned above, the heart rate (ECG) signal and pulse (PPG) signal, have heart rate variability (HRV) as their main characteristic and are used as a measure of resistance to disease, mental stress, and cardiovascular disease.

The variability features include (Time domain: mean peak, SDNN, HRV-index, pNN50, RMSSD, SDSD, Frequency domain: VLF, LF, HF, TP, norm LF, norm HF, LF/HF, etc.).

Additionally, in addition to the time-domain signal characteristics, various mutation analyses are possible using the frequency-domain signal characteristics.

For signal analysis, 20 windows were previously divided, and in order to use them as inputs for machine learning performed in the machine learning unit (120), each extracted window can be used as input, and the average value of the 20 windows can also be input. In this way, dividing windows and analyzing them has the advantage of allowing analysis by time zone.

For example, the initial signal that enters the supine position may be important, and the signal data from the latter part of the Trendelenburg that lasts for a certain period of time may be important. Such detailed analysis is possible by dividing the window.

The machine learning unit (120) constructs a hypoxia primary learning model by machine learning the first machine learning characteristic information and performs a primary analysis of the first machine learning characteristic information.

Specifically, the machine learning unit (120) performs machine learning-based learning to primarily predict hypoxia using extracted signal characteristics as input, and a feature selection process for selecting key characteristics may be included in relation to this.

For the feature selection process, filter-based and wrapper-based methods can be used in parallel.

For this purpose, the machine learning unit (120) uses a support vector machine (SVM), a tree random forest (RF), and a boosting-based algorithm (e.g., XGBoost, LGBoost, CatBoost).

Based on the results of the machine learning unit (120), the importance of each feature and the threshold value from the predicted hypoxia can be assumed and used as an explanatory factor provided to the clinician in the hypoxia prediction module (200) in the future, and boosting, feature importance of the tree model, and important variable analysis (SHAP value, Shapley additive explanations value) can be used.

FIGS. 3(a), (b), (c), and (d) are drawings showing the conversion of spirometry signals acquired by an image conversion unit of a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention into two-dimensional images.

The image conversion unit (130) converts the lung capacity signal for lung capacity extracted from the acquired spirometry signal and the airway pressure signal for airway pressure extracted from the acquired spirometry signal into two-dimensional images.

For this purpose, the image conversion unit (130) converts the signal into a two-dimensional image using various methods such as GASF (Garamian Angular Summation Field) based on each coordinate, MTF (Markov Transition Field) based on spatial coordinate, scalogram and spectrogram based on time-frequency domain, and connectivity adjacent matrix through signal correlation analysis.

In addition, it can be performed based on actual imaging results that can be confirmed by actual clinicians during surgery. As an example, there is a method of creating a 2D image of a Volume-AWP curve with lung capacity and airway pressure values as the X-Y axes, respectively.

To this end, preprocessing of image signals suitable for converting the above-mentioned signals into two-dimensional images must be performed in parallel, and in the case of image preprocessing, the process of unifying the image size and normalizing (Z-normalization) is included.

The image conversion unit (130) can largely visualize the first deep learning characteristic information using a window or visualize the first deep learning characteristic information using a loop.

First, the method of visualizing the first deep learning characteristic information using a window is similar to the analysis through machine learning. It divides it into 20 sections, creates one image for each window, and then creates a deep learning model (=hypoxia secondary learning model) that inputs 20 inputs.

Meanwhile, the method of visualizing the first deep learning characteristic information using a loop inputs 20 inputs for GASF, MTF, and scalogram and performs voting, and when drawing a loop, a method of making the entire section into one image is applied, and in the present invention, the method of visualizing the first deep learning characteristic information using a loop was used.

In addition, the method of imaging the first deep learning feature information using a loop largely images signals in the supine position and the Trendelenbrug position depending on the patient's position during surgery.

Specifically, for the loop, the reason the entire surgical section (one supine section and one trendelenburg section) was imaged is that as the loop is drawn repeatedly, the section where the loop is repeated becomes thicker, and the image is acquired reflecting this repetition.

The deep learning unit (140) constructs a hypoxia secondary learning model by deep learning a two-dimensional image and performs a secondary analysis of the two-dimensional image.

Specifically, the deep learning unit (140) trains a deep learning diagnosis model (=hypoxia secondary learning model) by utilizing a two-dimensional image of a unified size through a preprocessing process.

The platform for training the classification model applied in the deep learning department (140) may be Tensorflow or PyTorch.

Additionally, for classification models, CNN-based algorithms (VGG, ResNet) and Transformer-based algorithms (ViT, Swin transformer) can be applied.

The hypoxia secondary learning model learned in the above-mentioned deep learning unit (140) includes a window learning model and a loop learning model, and the loop learning model includes a Supine learning model and a Trendelenburg learning model.

The hypoxia prediction algorithm section (150) constructs a hypoxia learning model by integrating the hypoxia first learning model and the hypoxia second learning model.

For this purpose, the hypoxia prediction algorithm part (150) integrates the hypoxia first learning model and the hypoxia second learning model through an ensemble technique (hard voting, weighted voting, soft voting).

The hypoxia prediction module (200) extracts second characteristic information from a spirometry signal input in real time during surgery, and predicts hypoxia based on the results predicted by the machine-learned hypoxia first learning model in the training module (100) and the deep-learned hypoxia second learning model in the training module (100).

Here, the second characteristic information is information about some of the multiple signals (=factors) for determining the presence or absence of hypoxia according to the dynamics of the lung, and is information extracted from the patient's spirometry signal measured in real time during surgery.

Specifically, the second characteristic information includes second machine learning characteristic information and second deep learning characteristic information.

The second machine learning feature information includes the basic signal characteristics, variability, correlation between lung capacity and airway pressure, and differences and correlations between the feature information extracted from the Supine and Trendelenburg positions, respectively, after smoothing and correcting missing values for the patient's spirometry signal measured in real time during surgery.

Here, the spirometry signal input in real time during surgery is a signal obtained in real time through a patient's pulmonary function test during surgery (hereinafter referred to as the 'second spirometry signal').

The hypoxia prediction module (200) for this purpose includes a real-time feature extraction unit (210), a hypoxia primary prediction unit (220), a real-time image conversion unit (230), a hypoxia secondary prediction unit (240), a hypoxia comprehensive prediction unit (250), a CAM analysis unit (260), and a display unit (270).

The real-time feature extraction unit (210) smoothes the second spirometry signal, corrects missing values, and preprocesses the spirometry signal, and then extracts second machine learning feature information including basic signal features, variability, correlation between lung capacity and airway pressure, and differences and correlations between feature information extracted from the supine and Trendelenburg positions, respectively.

The hypoxia first prediction unit (220) applies second feature information (=second machine learning) for the second spirometry signal to the hypoxia first learning model to first predict hypoxia.

The real-time image conversion unit (230) converts a lung capacity signal for lung capacity extracted from the second spirometry signal, which is a spirometry signal, and an airway pressure signal for airway pressure extracted from the second spirometry signal, which is a spirometry signal, into real-time two-dimensional images.

The hypoxia secondary prediction unit (240) applies real-time two-dimensional images to the hypoxia secondary learning model to make secondary predictions of hypoxia.

The hypoxia comprehensive prediction unit (250) finally predicts hypoxia by synthesizing the first predicted result through the hypoxia first learning model and the second predicted result through the hypoxia second learning model.

FIGS. 4(a) and 4(b) are drawings showing a CAM analysis unit provided in a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention, which analyzes a characteristic map for the real-time two-dimensional image to generate a CAM image.

The CAM analysis unit (260) analyzes the characteristic map for the real-time two-dimensional image converted into an image by the real-time image conversion unit (230) to generate a CAM image, and the deep learning CAM (Class Activation Map) result for the spariometry curve of an actual patient is shown in FIG. 4.

In Figures 4 (a) and (b), the vertical axis represents volume and the horizontal axis represents pressure (awp).

As shown in (a) and (b) of Figure 4, the loop is repeated multiple times, and the more frequently the section is repeated, the darker and thicker the value is expressed.

Also, the loop is drawn from top (inspiration) to bottom (expiration), and you can see that the beginning where it goes up and the end where it comes down are expressed in red.

Therefore, it can be seen that the initial slope of the loop in inspiration and the initial slope of the loop in expiration are important.

Additionally, the slopes in (a) and (b) of Fig. 4 represent the change in volume (y-axis) according to the same pressure (x-axis).

Presenting explanatory power through this loop may not have much meaning to the general public, but it can have a great effect on clinicians.

FIG. 5 (a) and (b) are drawings showing a display unit provided in a system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention, which outputs actual machine learning classification performance and importance by signal characteristic.

The display unit (270) displays the presence or absence of hypoxia predicted by the hypoxia comprehensive prediction unit, the CAM image, and the first characteristic information.

Specifically, it functions as a criterion for determining hypoxia based on the key features selected by the machine learning unit (120) and the critical points of the features, and the values output in real time by the display unit (270).

In addition, the CAM image extracted from the deep learning unit (140) provides the location information of deep learning judgment to the clinician, thereby providing transparency and reliability of the present invention.

Referring to Figures 5 (a) and (b), various features were extracted from Supine and Trendelenburg, and approximately 188 features were used for analysis.

From the perspective of machine learning, selecting and using important features rather than using all features can improve accuracy and efficiency.

Therefore, we used wrapper feature selection to determine how many features can be used in the analysis to obtain the highest accuracy (see (b) of Fig. 5), and then used only that number of features in the analysis, and calculated the SHAP importance to interpret the meaning of the features used in the analysis.

For SHAP importance, the redder the value, the smaller the value, and the further to the right it is, the higher the probability of hypoxia occurrence. Therefore, it enables clinical interpretation. For example, pip_tren_mean, which had the highest importance, means the average value of the peak inspiratory pressure (PIP) signal in the Trendelenburg position, and a larger average value means that the occurrence of hypoxia can significantly increase.

2. Method for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals

Hereinafter, a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to one embodiment of the present invention will be described with reference to FIGS. 1 to 8.

FIG. 6 is a flowchart showing a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to an embodiment of the present invention. FIG. 7 is a block diagram showing the flow of a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to an embodiment of the present invention. FIG. 8 is a flowchart showing detailed steps of a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to an embodiment of the present invention.

Referring to FIGS. 6 to 8, a method for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery according to an embodiment of the present invention comprises: (a) a step of extracting first characteristic information from a spirometry signal acquired by a training module (S100); (b) a step of implementing a hypoxia learning model based on the results of machine learning and deep learning of the acquired spirometry signals by the training module (S200); (c) a step of extracting second characteristic information from a spirometry signal input in real time during surgery by the hypoxia prediction module (S300); (d) a step of predicting hypoxia based on the results predicted by the hypoxia prediction module through the first hypoxia learning model acquired by machine learning in the training module and the second hypoxia learning model acquired by deep learning in the training module (S400); (e) a step of generating a CAM image by analyzing a characteristic map for the spirometry signal by the training module (S500); and (f) a step of generating a CAM image by the training module based on the patient's It includes a step (S600) of displaying the presence or absence of hypoxia, CAM image, and first characteristic information.

According to the present invention as described above, a hypoxia learning model is constructed by integrating a hypoxia first learning model and a hypoxia second learning model, which are machine-learned and deep-learned, respectively, with first machine-learning feature information and first deep-learning feature information included in first feature information extracted from a spirometry signal acquired in advance, and by applying second machine-learning feature information and second deep-learning feature information included in second feature information extracted from a spirometry signal of a patient acquired in real time during an actual surgery, thereby predicting a result of the possibility of hypoxia occurring (= probability of hypoxia occurring or prediction of hypoxia) and displaying it to a clinician, thereby supporting the performance of medical measures before hypoxia occurs in a patient during surgery.

The above description of the present invention is for illustrative purposes, and those skilled in the art will understand that the present invention can be easily modified into other specific forms without changing the technical idea or essential characteristics of the present invention. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. For example, each component described as a single component may be implemented in a distributed manner, and likewise, components described as distributed may be implemented in a combined form.

The scope of the present invention is indicated by the claims described below, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention.

100: Training Module
110: Feature Extraction Unit
120: Machine Learning Department
130: Image conversion section
140: Deep Learning Department
150: Hypoxia prediction algorithm section
200: Hypoxia prediction module
210: Real-time feature extraction unit
220: Hypoxia Primary Prediction Unit
230: Real-time image conversion section
240: Hypoxia 2nd Prediction Unit
250: Hypoxia Comprehensive Prediction Unit
260: CAM Analysis Department
270: Display section
300: Postoperative Hypoxia Prediction System Using Real-Time Analysis of Intraoperative Spirometry Signals

Claims (9)

A training module that extracts first characteristic information from the acquired spirometry signal and implements a hypoxia learning model based on the results of machine learning and deep learning of the extracted first characteristic information; and
A system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery, characterized by including a hypoxia prediction module for extracting second characteristic information from a spirometry signal input in real time during surgery and predicting hypoxia based on the results predicted by the machine-learned hypoxia primary learning model in the training module and the deep-learned hypoxia secondary learning model in the training module,
In the first paragraph,
The above first characteristic information includes first machine learning characteristic information and first deep learning characteristic information,
The above training module,
A feature extraction unit that preprocesses the above-mentioned acquired spirometry signal by smoothing it and correcting missing values, and then extracts the first machine learning feature information including the basic signal characteristics, variability, correlation between lung capacity and airway pressure, and the difference and correlation between the feature information extracted from the Supine posture and the Trendelenburg posture, respectively; and
A system for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals, characterized by including a machine learning unit that constructs a hypoxia primary learning model by performing machine learning on the first machine learning characteristic information and performs a primary analysis on the first machine learning characteristic information.
In the second paragraph,
The above training module,
An image conversion unit that converts the first deep learning characteristic information, which is a lung capacity signal for lung capacity extracted from the above-mentioned acquired spirometry signal and an airway pressure signal for airway pressure extracted from the above-mentioned acquired spirometry signal, into a two-dimensional image; and
A system for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals, characterized in that it further includes a deep learning unit that constructs a hypoxia secondary learning model by deep learning the above two-dimensional image and performs a secondary analysis of the above two-dimensional image.
In the third paragraph,
The above training module further includes a hypoxia prediction algorithm section that constructs a hypoxia learning model by integrating the hypoxia first learning model and the hypoxia second learning model;
A postoperative hypoxia prediction system through real-time analysis of intraoperative spirometry signals, characterized in that the hypoxia prediction algorithm section integrates the hypoxia first learning model and the hypoxia second learning model through an ensemble technique (hard voting, weighted voting, soft voting).
In the first paragraph,
The above second characteristic information includes second machine learning characteristic information and second deep learning characteristic information,
The above hypoxia prediction module is,
A real-time feature extraction unit that preprocesses the spirometry signal by smoothing it and correcting missing values, and then extracts the second machine learning feature information including the basic signal characteristics, variability, correlation between lung capacity and airway pressure, and the difference and correlation between the feature information extracted from the Supine and Trendelenburg postures, respectively; and
A system for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals, characterized by including a hypoxia primary prediction unit that first predicts hypoxia by applying the second machine learning characteristic information to the hypoxia primary learning model.
In clause 5,
The above hypoxia prediction module is,
A real-time image conversion unit that converts the second deep learning characteristic information, which is a lung capacity signal for lung capacity extracted from the spirometry signal and an airway pressure signal for airway pressure extracted from the spirometry signal, into a real-time two-dimensional image; and
A system for predicting postoperative hypoxia through real-time analysis of intraoperative spirometry signals, characterized in that it further includes a hypoxia secondary prediction unit that applies the real-time two-dimensional image to the hypoxia secondary learning model to predict the hypoxia for the second time.
In Article 6,
A system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery, characterized in that the hypoxia prediction module further includes a hypoxia comprehensive prediction unit that predicts hypoxia by synthesizing the first predicted result through the hypoxia first learning model and the second predicted result through the hypoxia second learning model.
In Article 6,
A system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery, characterized in that the hypoxia prediction module further includes a CAM analysis unit that analyzes a characteristic map for the real-time two-dimensional image to generate a CAM image.
In the first paragraph,
A system for predicting postoperative hypoxia through real-time analysis of spirometry signals during surgery, characterized in that the hypoxia prediction module further includes a display unit that displays the presence or absence of hypoxia predicted by the hypoxia comprehensive prediction unit, the CAM image, and the first characteristic information.

Images