TWI795928B - System and method for prediction of intradialytic adverse event and computer readable medium thereof - Google Patents

Abstract

Provided are a system and a method for prediction of an intradialytic adverse event, where a machine learning model of two-class classification is utilized to predict intradialytic adverse events in quasi-real time, such that features extracted in an ongoing hemodialysis process in real time can have the hemodialysis session alerted for forthcoming adverse events. Therefore, clinicians can be warned to take necessary actions and adjust the hemodialysis machine settings in advance. In addition, a computer readable medium thereof is also provided.

Description

System, method and computer readable medium for predicting adverse events in dialysis

The present disclosure relates to medical monitoring applications, and more particularly to systems, methods and computer readable media for real-time prediction of adverse events in dialysis.

Hemodialysis (HD) therapy plays an important role in care management. Due to oliguria or even anuria, most patients with renal failure must remove fluid during hemodialysis treatment (HD course) to maintain emovolemia. The volume-dependent element of hypertension can be corrected by fluid removal, but the ultrafiltration process exposes hemodialysis patients to the risk of hemodynamic instability, which can lead to fatal consequences such as cardiac arrest. Intradialytic hypotension is the most common complication during hemodialysis and has been identified as a cause of reduced efficacy of hemodialysis. Acutely, adverse events in dialysis may be fatal; in the long term, frequent adverse events in dialysis will increase the morbidity and long-term all-cause mortality of patients.

Devices (e.g., the Crit-Line monitor developed by Fresenius Medical Care) have been developed for the non-invasive monitoring of instantaneous hematocrit, oxygen saturation, and in-dialysis volume status by means of light transmission. Assist in liquid removal during ultrafiltration. Although uncontrolled studies have shown that this The Crit-Line setting reduces symptoms during dialysis and helps assess target weight, but an open-label randomized controlled trial showed that the Crit-Line group had a higher rate of hospitalization than the control group.

Artificial intelligence is also applied to hemodialysis patients to assist clinical practice, such as urea clearance rate, dietary protein intake, volume status, erythropoietin stimulating agent response, iron supplementation response, hemoglobin value, hemodialysis quality, mortality etc. forecast. Although artificial intelligence has also been applied to predict the risk of hypotension in dialysis, previous research on this application still lacks the consideration of time series data input.

Therefore, how to consider time series data in machine learning methods and predict adverse events in dialysis more than hypotension risk in an unbiased way is an unmet need in this technical field.

In view of this, the present disclosure provides a system for predicting adverse events in dialysis, which includes: a feature extraction module configured to collect and process data about patients' hemodialysis sessions; and a model building and optimization module a set configured to build a machine learning model for predicting the in-dialysis adverse event during the hemodialysis session based on the data.

The present disclosure further provides a method for predicting adverse events in dialysis, which includes: configuring a feature extraction module to collect and process data about a patient's hemodialysis course; and configuring a model building and optimization module based on the data A machine learning model for predicting adverse events in dialysis during the course of hemodialysis is established.

In at least one embodiment of the present disclosure, the information about the hemodialysis course of the patient includes: one or more of demographic information, physiological data, dialysis data, and registered adverse events in dialysis.

In at least one embodiment of the present disclosure, the data includes a data set having a plurality of records with measurements at corresponding time stamps.

In at least one embodiment of the present disclosure, the feature extraction module collects and processes the data about the patient's hemodialysis session by: deriving from the measurements a mean value, a standard deviation of the mean value, a coefficient of variation At least one of linear regression slope and linear regression R square is used as the feature of the plurality of records.

In at least one embodiment of the present disclosure, the measurements include venous pressure and transmembrane pressure, and wherein the feature extraction module collects and processes the data about the patient's hemodialysis session by: The measurement of pressure and the transmembrane pressure derives at least one of a maximum, minimum, and average value, and a second differential, of a rate of change to characterize the plurality of records.

In at least one embodiment of the present disclosure, the machine learning model is established based on a first dimension related to a target adverse event in dialysis to be predicted, and a second dimension related to a target period of time during the dialysis session to be predicted.

In at least one embodiment of the present disclosure, the machine learning model is trained with the data labeled with outcomes related to the adverse event in dialysis.

In at least one embodiment of the present disclosure, the machine learning model is trained by combining key features extracted from the data.

In at least one embodiment, the system described herein further includes a data storage module configured to store the data, and the method described herein further includes configuring the data storage module to store the data.

The present disclosure further provides a computer-readable medium, which stores computer-executable codes, and implements the above-mentioned method after the computer-executable codes are executed.

1: system

10: Feature extraction module

20: Data storage module

30:Model building and optimization module

S1~S4: steps

The present disclosure can be more fully understood by reading the following description of the embodiments, with reference to the accompanying drawings, in which:

FIG. 1 is a schematic diagram showing an exemplary structure of a system for predicting adverse events in dialysis according to an embodiment of the present disclosure;

2 is a flowchart showing the steps of the decision-making process for selecting patients and hemodialysis sessions according to an embodiment of the present disclosure;

Figure 3A and Figure 3B are schematic diagrams showing the performance of the machine learning model used to predict adverse events belonging to Group 1 according to an embodiment of the present disclosure; ctr: control; f72, f76, f77, f78, f82: in Table 4 and Table 5 The number of the indicated feature; average △ (UF rate): the average value of the ultrafiltration rate change;

4 is a schematic diagram showing the performance of a machine learning model for predicting adverse events belonging to group 2 according to an embodiment of the present disclosure; ctr: control; mean Δ(UF rate): mean value of change in ultrafiltration rate; mean UF volume : average value of ultrafiltration volume;

5 is a schematic diagram showing the performance of a machine learning model for predicting adverse events belonging to group 3 according to an embodiment of the present disclosure; ctr: control; and

FIG. 6 and FIG. 7 are schematic diagrams showing the consistency of the predicted probability of an adverse event based on a machine learning model over time according to an embodiment of the present disclosure.

The following examples serve to illustrate the present disclosure. Those skilled in the art can easily understand the advantages and effects of the present disclosure after reading the present disclosure, and can also be implemented or applied in other different embodiments. Accordingly, any element or method within the scope of the disclosure set forth herein may be combined with any other element or method disclosed in any embodiment of the disclosure.

The proportions, structures, dimensions and other features shown in the drawings of the disclosure are only used to illustrate the embodiments described herein, so that those skilled in the art can read and understand the disclosure content and is not intended to limit the scope of this disclosure. Any changes, modifications or adjustments to the above-mentioned features shall fall within the scope of the technical content of the present disclosure without affecting the purpose and effect of the present disclosure.

In this article, when it is described that an object "comprises", "comprises" or "has" technical features, unless otherwise specified, other elements, components, structures, regions, components, devices, systems, steps, connections, etc. may be included in addition , and other features should not be excluded.

In this article, ordinal terms such as "first" and "second" are only used to describe or distinguish technical features such as elements, components, structures, regions, components, devices, systems, etc., and are not intended to limit the present disclosure. It is not intended to limit the spatial sequence among such technical features. Furthermore, terms in the singular such as "a (a)", "an" and "the" also refer to the plural unless otherwise stated, and terms such as "or" and "and/or" are interchangeable use.

As used herein, the terms "subject," "individual" and "patient" are used interchangeably and may refer to animals, eg, mammals including humans. Unless a gender is specifically indicated, the term "subject" can refer to both males and females.

As used herein, the terms "comprises", "includes", "includes", "includes", "has", "has", "includes", "covers" or any other variation thereof are intended to cover non-exclusive Include. For example, a composition, mixture, process, or method comprising a list of elements is not necessarily limited to those elements, but may include other elements not expressly listed, or inherent to such composition, mixture, process, or method.

In this context, the term "at least one" in relation to a series of one or more elements should be understood to mean at least one element selected from any one or more elements in the series of elements, but not necessarily including the At least one of each element listed in a list of elements, without excluding any combination of elements in said list of elements. This definition also allows for the optional presence of elements other than the indicated element in the list of elements to which the term "at least one" refers, whether related or unrelated to the indicated element. therefore, As a non-limiting example, "at least one of A and B" (or equivalently, "at least one of A or B", or equivalently, "at least one of A and/or B") may: In one embodiment, means at least one, optionally including more than one A and no B (and optionally including elements other than B); in another embodiment, means at least one, optionally including More than one B and no A (and optionally including elements other than A); in yet another embodiment, means at least one, optionally including more than one A, and at least one, optionally including more than A B (and optionally other elements).

Herein, the terms "one or more" and "at least one" may have the same meaning and include one, two, three or more.

As used herein, the terms "measure" and "measure" are interchangeable with "determine", "assess", "determine", "detect" and the like, which designate both quantitative and qualitative determinations. In the case of quantitative determination, terms such as "measurement quantity" may be used. In the case of qualitative or quantitative determination, the term "measurement level" or "judgment level" can be used.

Referring to FIG. 1 , it shows a system 1 for predicting adverse events in dialysis, including: a feature extraction module 10 , a data storage module 20 and a model building and optimization module 30 . The elements of the system 1 can be connected to each other by any suitable wired or wireless means, and the present disclosure is not limited thereto.

In some embodiments, the feature extraction module 10 can be coupled to or implemented in a hemodialysis (HD) machine (not shown), so that any records obtained from a patient during a hemodialysis session can be collected and used for on-site features extract.

In some embodiments, the data storage module 20 is configured to maintain the data received and processed by the feature extraction module 10 for later data inspection and/or model building. The data storage module 20 can be realized by any suitable data storage device, system, database, cloud storage, or the like, and the present disclosure is not limited thereto.

In some embodiments, the model building and optimization module 30 is configured to build a machine learning model for predicting adverse events in a patient's dialysis, and/or based on improvements from a patient's hemodialysis treatment Improve the performance of machine learning models by improving data quality. In at least one embodiment, the model establishment and optimization module 30 can establish more than one machine learning model to correspond to the data of various characteristics collected by the feature extraction module 10 . For example, a machine learning model can be built on two dimensions: the target adverse event in dialysis to be predicted (e.g., blood pressure increase, muscle cramps, and all events except blood pressure increase); and the hemodialysis course to be predicted The target time period of the period (eg, different 30-minute time periods during a hemodialysis session). However, the number of established machine learning models and prediction targets are not intended to limit the scope of this disclosure, and can be changed in any appropriate way according to planning requirements.

In some embodiments, elements of the system may be individually implemented as any suitable computing device, device, application, system, or the like, but the disclosure is not limited thereto. For example, any two or three of the feature extraction module 10 , the data storage module 20 and the model building and optimization module 30 can be integrated together instead of being implemented as three independent units. In some embodiments, the three requirements can also be integrated and implemented in a cloud computing environment. However, without departing from the operating philosophy of the present disclosure, the configuration of the elements in the system may be implemented in any suitable form, and should not limit the scope of the present disclosure.

The operational relationship among the aforementioned elements of system 1 is further described in FIG. 1 , indicated by arrows (described herein as "steps") and described in detail herein.

In some embodiments, the step S1 means that the feature extraction module 10 will collect and process the data related to the patient's record in real time during the hemodialysis treatment, and store the data in the data storage module 20 .

In some embodiments, step S2 means that the model building and optimization module 30 will use the data stored in the data storage module 20 to build and/or optimize a machine learning model for predicting adverse events in dialysis patients. For example, the machine learning model can be based on any one of linear models, random forest support vector regression, XGBoost, LASSO regression, ensemble methods, deep learning, or the like or any combination of the above, and the present disclosure is not limited thereto .

In some embodiments, step S3 means that for the machine learning model established and fully trained by the model building and optimization module 30, the feature extraction module 10 can also send the data about the patient's records during the hemodialysis course to the machine learning model. A machine learning model for immediate prediction of adverse events in dialysis.

In some embodiments, step S4 means that after the machine learning model completes its prediction, the prediction result will be sent back to the feature extraction module 10 (or the hemodialysis machine paired with it) to inform the patient of the risk of adverse events during dialysis.

In some embodiments, there is also a computer-readable medium storing computer-executable code configured to implement the above-discussed steps of the present disclosure after being executed.

Starting from here, how to plan the working mechanisms of the feature extraction module 10 , the data storage module 20 and the model building and optimization module 30 will be described in detail.

method

Research Protocol and Study Objects

In the practice of a retrospective observational study conducted at a single institution, the records of all patients receiving maintenance hemodialysis at Changhua Christian Hospital were reviewed. Figure 2 shows the decision-making process of how patients and hemodialysis sessions were selected for the study according to the examples described herein. For example, 108 of 129 eligible patients completed the three-month study during the three-month period, and said 108 patients were excluded from hemodialysis sessions based on the following reasons: (1) Interruption of a course of treatment due to dialysis machine replacement; (2) interruption of more than one course of treatment due to urination or defecation of the patient; and/or (3) course of treatment during which the patient is unable to freely express their discomfort. Ultimately, a total of 4221 hemodialysis sessions from the 108 patients, each receiving 39 or 40 hemodialysis sessions during the 3-month study period, were used to build a machine learning model to predict adverse events in dialysis Dialysis sessions.

Dialysis and Physiological Data Collection

In the embodiments described herein, the feature extraction module 10 collects several types of data for building a machine learning model, such as demographic information, physiological data, dialysis data, and registered adverse events in dialysis. The above-mentioned types of data can be collected manually (eg, measured by medical personnel) or automatically (eg, measured by the hemodialysis machine itself), but the present disclosure is not limited thereto.

In at least one embodiment, demographic information can be derived from medical records and can include information such as age, gender, and years of dialysis treatment for any patient.

In at least one embodiment, physiological data is measured and recorded at approximately every 30 to 60 minute intervals during each hemodialysis session of the participating patient (approximately 4 hours per hemodialysis session).

In at least one embodiment, dialysis data is collected from a hemodialysis machine during each hemodialysis session of a participating patient. Examples of hemodialysis machine readings associated with dialysis data are shown in Table 1 below.

In at least one embodiment, the logged adverse dialysis events are recorded based on measured physiological data or patient complaints. Examples of dialysis adverse events recorded for participating patients are shown in Table 2 below.

Refer to Tables 3-1 to 3-2 below, which disclose examples of data collected by the feature extraction module 10 . In the practice of studying 108 patients with a total of 4221 hemodialysis courses, the data set HD _i collected by the feature extraction module 10 for each hemodialysis course i (i=1 to 4221) is composed of multiple records {Y _j,k ,T _k }, where j (see Table 3-1, which ranges from 1 to 9) is the index value of the category of data from dialysis and physiological measurements; k is the index of the time when the measurement occurred value; and Y _j,k is the value of j measured at time T _k . Refer to the records marked with thick boundaries in Table 3-1, which show the data points Y for all classes measured at the time stamp "T ₄ =9:35:44" (ie, {Y _j,4 ,T ₄ }) _j,4 . According to the manufacturer's default settings, once the value of venous pressure (VP) or transmembrane pressure (TMP) changes and is different from the previous measurement at T=T _k-1 , the dialysis data is usually automatically recorded from the hemodialysis machine and / or physiological data. Therefore, the time interval T _k -T _k-1 between any two consecutive records may not be equal.

Continuing from Table 3-2, each data set HD _i also includes additional time-invariant patient-specific information Y _j (j=10 to 13), which respectively represent the age, sex, blood Information such as years of dialysis and weight before dialysis.

It should be understood that when the system is ready for practical use (for example, when a fully trained machine learning model is developed), the dataset _HDi described in Table 3-1 can be directly used for feature extraction and prediction of adverse events in dialysis . However, adverse events during dialysis can also be registered in the data set HD _i as training data for building a machine learning model.

feature extraction

After the data is collected, the feature extraction module 10 continues to extract the data from the data set HD _i from the hemodialysis treatment session as features for analysis. In the embodiments described herein, feature extraction is ideally performed using the AWK program, but any suitable program or application may be used, as the disclosure is not limited thereto.

In order to avoid artifacts at the beginning of the data set HD _i due to the different operating procedures for setting and starting dialysis in each hemodialysis session, if there is blood flow velocity at the beginning of each data set HD _i (that is, data point Y _5,k ) changes between time stamps T ₁ and T ₂ , then the first data point Y _j,1 of this data set HD _i (ie record {Y _j,1 ,T ₁ }) will be excluded. Any record {Y _j,k ,T _k } at a given time stamp T _k also has a blood flow rate equal to or lower than zero due to a dialysis interruption (e.g., dialysis machine replacement or patient urination/defecation). will be excluded. If a hemodialysis session is interrupted more than once, the entire hemodialysis session (full dataset HD _i ) will be excluded from the feature extraction process.

For the training data, the complete set of records {Y _j,k ,T _k } from the HD _i data set of the hemodialysis sessions in the absence of registered adverse events in dialysis (such as any of the adverse events listed in Table 2 above) will be included for feature extraction. On the other hand, for hemodialysis courses that have registered adverse events in dialysis, only the records {Y _j,k ,T _k } of the corresponding data set HD _i before the first occurrence of adverse events will be included for feature extraction , meaning that the hemodialysis session lasted less than 4 hours.

Regression analysis is challenging due to the time interval between two adjacent records {Y _i,k ,T _k } and the duration of hemodialysis sessions, and the temporal characteristics of the measured variables need to be included for classification in the analysis. To this end, the feature extraction module 10 derives the mean, the standard deviation of the mean and the coefficient of variation from the records {Y _j,k ,T _k } of the dialysis and physiological measurements, together with the slope and R-square of the linear regression, for analysis The characteristics. In addition, the feature extraction module 10 also derives the maximum value, minimum value and average value (first differential) of the rate of change of venous pressure (VP) and transmembrane pressure (TMP), together with the second differential as features for analysis.

Refer to Table 4 and Table 5 below, which disclose examples of features extracted by the feature extraction module 10 from hemodialysis sessions. For example, a total of 84 extracted features {X _h } (h=1 to 84, which refers to the feature numbers shown below, where "#" represents "number"), including _the original Measured characteristics and temporal characteristics derived therefrom. However, based on the content of the dataset HD _i in practice, the total number of features may be greater or less than 84, and the present disclosure is not limited thereto.

illustrate:

A: Time interval≡T _i -T _i-1 , i=1~n

B: K _i =△X _i /(T _i -T _i-1 )=(X _i -X _i-1 )/(T _i -T _i-1 )

C: △X _i ≡X _i -X _i-1

D: Average* is weighted by duration and divided by total recording time:

Σ((X _i -X _i-1 )×(T _i -T _i-1 ))/Σ(T _i -T _i-1 )=Σ((X _i -X _i-1 )×(T _i - T _i-1 ))/(T _n -T ₀ )

After feature extraction, the feature extraction module 10 stores the extracted features in the data storage module 20 for later use (for example, for establishing a machine learning model), or directly sends them to the model building and optimization module 30 for predicting Adverse events during dialysis. As mentioned above, {Y _j,k ,T _k } of the HD _i data set is recorded when the venous or transmembrane pressure changes. Thus, any measurement at time T _p between two adjacent measurement timestamps T _k and T _k-1 can be specified as {Y _j,k ,T _p }={Y _j,k ,T _{k- 1} }. That is, when the feature extraction module 10 is running in real time, the feature extraction of the data set HD _i in the hemodialysis session can be terminated at any time (for example, T _p ) for storage or prediction.

Result markers for model building

During training, outcomes related to hemodialysis sessions may be labeled before the corresponding data set HD _i and/or its extracted features are used for model building. For example, among the 4221 hemodialysis sessions studied, hemodialysis sessions with one or more adverse events were marked as 1, while hemodialysis sessions with no adverse events were marked as 0. A negative control set was also created by randomly relabeling the 4221 hemodialysis sessions without considering the true outcome and maintaining the same 0 to 1 ratio as the experimental set.

After the results are marked, the model building and optimization module 30 performs its machine learning model building process, including: (1) building a binary classification model (for example, using an algorithm such as ensemble or perceptron), which is based on the given input Dataset HD _i outputs a label of 0 or 1; and (2) evaluates the binary classification model by quadruple cross-validation (eg, using the Azure service developed by Microsoft). For each machine learning model, its model building is performed by repeating the above-mentioned building process at least three times by introducing different random numbers.

Pick the best performing features

In addition to labeling results, machine learning models can also be built based on key features selected from the full 84 features. For example, the model building and optimization module 30 is also configured to implement an algorithm during machine learning model building to select key features associated with the occurrence of adverse events in dialysis. By determining the key features established by the model, the computing load can be effectively reduced without losing the prediction accuracy. Further, after the machine learning model predicts adverse events in dialysis based on key features, the key features can be As a reference for adjusting the parameters of the hemodialysis machine.

For example, in order to find out which features are more important than other features in predicting adverse events in target dialysis, key features (from all 84 features) can be selected and used in the model building of the model building and optimization module 30, so that It is possible to compare the difference between the selected key features and the prediction results of the machine learning model built using all 84 features. Thus, key feature selection may be performed using MATLAB (eg, MATrixLABoratory, developed by MathWorks Inc.), although any suitable program or application may be utilized without limiting the scope of the present disclosure. The selected key features can then be used to build a machine learning model as discussed above (eg, a binary classification model using ensemble random downsampled boosting trees and evaluated by quadruple cross-validation). Then, the established machine learning model can be given a score by summing the percentages of true positives and true negatives from the prediction results of the established machine learning model for comparison in a later stage.

The process of key feature selection is further detailed here. First, the first set of machine learning models is built by sequentially using a single feature in the 84 features, and each of the first set of machine learning models is given according to its prediction results (based on the percentage of true positives and true negatives). Give a rating. Then, the best binomial feature combination is selected from the binomial feature combination pool to establish a second group of machine learning models. The selected best features (eg, the features that contributed the highest score to the machine learning model) and each of the remaining 83 features were individually combined to build. Next, binomial feature combinations selected for the second set of machine learning models with higher scores than the highest scores of the first set of machine learning models are retained for use in the next step. Similarly, the best three feature combinations can be selected from the three feature combination pools to build a third group of machine learning models, and the three feature combination pools are selected from the scores of the second group of machine learning models. The best binomial feature combination (e.g., the binomial feature combination that contributed the highest score to the machine learning model) and each of the remaining 82 features The combination of the three features established together and selected for the third group of machine learning models with a score higher than the highest score of the second group of machine learning models is retained for the next step. This process was repeated until the best 20 feature combinations were selected, and the features that most frequently appeared in these 20 feature combinations were defined as key features.

result

Demographics of Study Participants

Table 6 below shows a table summarizing the characteristics of the 108 patients involved, according to the examples described herein. For example, among the above 108 patients, the average age was 63.6 years; 60 patients (55.6%) were male; the average age of hemodialysis was 7.7 years; 47 patients (43.5%) had diabetes; 10 patients (63.9%) had hypertension; 11 patients (10.2%) had coronary artery disease; 12 patients (11.1%) had congestive heart failure; 7 patients (6.5%) had History of stroke; chronic obstructive pulmonary disease in 3 patients (2.8%); peripheral vascular disease in 2 patients (1.9%); and malignancy in 2 patients (1.9%).

Data are presented as mean ± standard deviation or percentage as appropriate.

In addition, according to the occurrence of adverse events in dialysis (see Table 2) recorded from these 108 patients, there were more than 3 adverse events in dialysis in 4 courses of hemodialysis; 3 adverse events in 19 courses of hemodialysis Events; 106 hemodialysis sessions had two adverse events and 276 hemodialysis sessions had one adverse event. Of a total of 4221 hemodialysis sessions, 406 hemodialysis sessions had adverse events.

Performance of predictive models

To increase the ratio of 1 to 0 outcomes (i.e., a hemodialysis session with an adverse event is marked as 1 and a hemodialysis session without an adverse event is marked as 0), the 27 adverse events listed in Table 2 were divided into three groups, with Used to build machine learning models. The first group corresponds to all adverse events in dialysis except blood pressure increase, vascular access obstruction and vascular access embolism, and a total of 323 hemodialysis sessions were assigned to this group (group 1). Adverse events during dialysis included muscle cramps in the second group, to which 138 hemodialysis sessions were assigned (group 2). Adverse events during dialysis included elevated blood pressure in the third group, to which 108 hemodialysis sessions were assigned (group 3).

Group 1: All events except elevated blood pressure

Figure 3A and Figure 3B and Table 7 describe the performance of the machine learning model for predicting adverse events belonging to Group 1, and curves a to k represent machine learning models built using different feature combinations. In this scenario, a binary categorical average perceptron with a learning rate of 20 and a maximum iteration of 20 is used for model building.

For the 84-item feature model (curve a), the mean area under the curve (AUC) was 0.83, the standard deviation (SD) was 0.03, the F1 score was 0.53, the sensitivity was 0.53, and the specificity was 0.96. Compared to the negative control (curve b) (mean AUC 0.50, SD 0.04, and F1 score 0.15), the 84-item feature model of the binary classification average perceptron reasonably predicted adverse events. The predictions of other algorithms were also tested. For example, a support vector machine (SVM) for binary classification yielded a mean AUC of 0.83 (SD 0.02), an F1 score of 0.55, a sensitivity of 0.53, and a specificity of 0.96. This result is similar to that obtained by the average perceptron. Compared with the average perceptron and SVM algorithms, binary logistic regression and decision forests are not Can predict adverse events well. The average AUC obtained by logistic regression was 0.82 (SD 0.02), and the F1 score was 0.48, and the average AUC obtained by decision forest was 0.83 (SD 0.02), and the F1 score was 0.46. Additionally, the within-patient partition (mean AUC of 0.83, SD 0.03, and mean F1 score of 0.53, SD 0.02) and the between-patient partition (mean AUC of 0.82, SD 0.04, and mean F1 score of 0.50) used for sampling , SD 0.06) showed no significant difference in prediction.

Ultrafiltration rate and ultrafiltration volume are parameters related to hemodialysis. However, the performance of the models in this paper showed that adverse events could not be adequately predicted using a single feature, for example, maximum ultrafiltration volume (feature 78, see curve k) or mean change in ultrafiltration rate (feature 77, see curve j). The model established by the maximum ultrafiltration volume (defined as the ultrafiltration volume recorded at the last time point) had an AUC of 0.48 and an F1 score of 0.15, similar to the results of the negative control. On the other hand, the model built from the mean of the change in ultrafiltration rate during a hemodialysis session had an AUC of 0.70 and an F1 score of 0.28. Combining the two ultrafiltration-related features (curve h) also did not predict adverse events. After prediction using up to 6 features related to ultrafiltration volume (features 78 to 83, see curve f), AUC increased from 0.48 to 0.82, while F1 score increased from 0.15 to 0.46. The model with 14 ultrafiltered features (features 70 to 83, see curve e) had an AUC of 0.83 and an F1 score of 0.52.

Next, select the 21 features with the highest frequency among the 20 feature combinations for evaluation (curve c). A binary classification averaged perceptron model based on these top-performing 21-item features, but omitting ultra-filtered relevant features, showed a mean AUC of 0.82 (SD 0.02) and an F1 score of 0.45. However, adding one or two features to this model did not significantly enhance predictions (e.g., the best 23-item feature model only resulted in a mean AUC of 0.82, SD 0.02, and an F1 score of 0.46). Compared to a model based on all features excluding ultra-filtered relevant features (cf. curve d, which uses 70 features and has an AUC of 0.81, F1 score 0.45), the results of the 21 best feature models (excluding ultra-filtered features) showed that using a quarter of all 84 features was sufficient for predicting adverse events.

To be precise, the 21 features selected from the total 84 features are age, maximum transmembrane pressure, minimum systolic blood pressure (SBP), minimum diastolic blood pressure (DBP), minimum pulse pressure, minimum blood flow rate, mean SBP, mean Venous pressure, mean transmembrane pressure, SBP linear regression slope, DBP linear regression slope, pulse pressure linear regression slope, pulse rate linear regression slope, transmembrane pressure linear regression slope, blood flow rate mean standard deviation, pulse pressure linear regression R Squared, and related parameters of the second differential of venous pressure (see features 2, 5, 6, 8, 11, 14, 17, 20, 21, 26, 29, 31, 36, 47 to 52, 57 and 59).

Group 2: Muscle cramps

Figure 4 and Table 8 describe the performance of machine learning models for predicting adverse events belonging to Group 2, curves a to k represent machine learning models built using different feature combinations.

As seen in Figure 4 and Table 8, the model based on 14 ultra-filtered features (refer to curve d) had an average AUC of 0.85 (SD 0.04) in predicting the occurrence of muscle spasm, and an F1 score of 0.45, which was similar to that of 84 features model (cf. curve a, with mean AUC of 0.82, SD of 0.04, and F1 score of 0.42), and outperformed models based on all features excluding ultrafiltration-related features (see curve c, with mean AUC of 0.79 , SD is 0.04, and the F1 score is 0.30). However, a single ultrafiltration-related feature (see curves i and k) does not adequately predict spasticity. The combination of two ultrafiltration-related features also failed to predict muscle spasticity (see curve f, which includes features 70 to 77, and has an AUC of 0.79, and an F1 score of 0.29; or curve e, which includes features 78 to 83, and has AUC was 0.84 and F1 score was 0.37). The above results suggest that ultrafiltration-related features contribute more to the prediction of muscle spasticity than other features.

Group 3: Elevated blood pressure

Figure 5 and Table 9 describe the performance of machine learning models for predicting adverse events belonging to group 3, curves a to d represent machine learning models built using different combinations of features.

As seen in Figure 5 and Table 9, the model based on all 84 features to predict the occurrence of hypertension (refer to curve a) had an average AUC of 0.93 (SD 0.02), and an F1 score of 0.41. Compared with the model based on 14 ultrafiltration-related features (refer to curve c, its AUC is 0.72, and its F1 score is 0.22), the results show that ultrafiltration parameters do not play an important role in predicting hypertension in dialysis. Although the AUC of the model based on 24 blood pressure-related features (cf. curve d, with AUC of 0.92, SD of 0.03, and F1 score of 0.38) was higher than 0.9, features other than blood pressure could lead to additional improvements in F1 score.

Consistency of Predicted Probability of Adverse Events Over Time

Fig. 6 and Fig. 7 and table 10 have described the performance of the machine learning model that predicts the adverse event in dialysis based on time series feature, and the curve that is expressed as 0, 5, 10, 15, 20 and 60 minutes represents the described machine learning model in Predictive ability at the time indicated before the occurrence of any adverse event in dialysis.

In the examples described herein, the time series feature is used throughout the hemodialysis session from the beginning of the hemodialysis session until the time point before the adverse event is recorded, or the end of the hemodialysis session collected during the time point (i.e., no adverse events were recorded). In this case, the end of feature collection will be defined as 0 minutes if its end time point is just before the occurrence of recorded adverse events. In addition to 0 minutes, the cut-off end time points of feature collection were also set as 5, 10, 15, 20, and 60 minutes before the occurrence of adverse events to evaluate the prediction accuracy.

As shown in Figure 6 and Table 10, the machine learning model used to predict all adverse events in dialysis (except blood pressure increase) showed that compared with the performance learned from the characteristics of earlier cut-off time points, even from the The AUC scores of the characteristics of the 5, 10, 15 and 20 minutes cut-off time points before the end of the adverse event or the end of the hemodialysis course without adverse events were about 0.80, and their F1 scores were all lower than 0.5. The feature leads to the best AUC and F1-score (see the 0-minute curve with AUC of 0.83 and F1-score of 0.53). This result suggests that information falling in the 20-minute time window preceding the indexed adverse event is valuable, but information falling in the 5-minute time window preceding the indexed adverse event has a greater impact on event prediction.

See Figure 7. In order to further understand the dependence of the prediction accuracy on the cut-off end time point, 500 hemodialysis sessions were randomly selected to compare with the 84 features obtained at 0, 5, 10, 15 and 20 minutes before the occurrence of recorded adverse events Comparison of predicted probabilities of adverse events. As shown by the circled data points, there are 5 hemodialysis courses that use the features extracted based on different cut-off time points to predict the probability of adverse events with strong consistency, and these 5 hemodialysis courses have adverse events . Since adverse events should occur in about 40 hemodialysis sessions out of 500 randomly selected hemodialysis sessions, the results show that at least one in ten hemodialysis sessions with adverse events could be predicted as early as 20 minutes in advance and can be predicted by Further confirmation was made by real-time machine learning using features from subsequent cut-off end time points.

Although none of the 84 features contained explicit time series information, the linear and difference analyzes used for feature extraction may be affected by the length of hemodialysis sessions. Therefore, hemodialysis sessions without adverse events (negative) were truncated and their predictions were compared with those of non-truncated hemodialysis sessions. Since the mean length of hemodialysis sessions with adverse events (positive) was 3.3 hours, negative hemodialysis sessions were truncated and randomized to end points between 3 and 3.5 hours in time (T _end ), while the end points of positive sessions were kept constant. The {Y _j,k ,T _k } record at the end point T _end is defined in the same way as the {Y _j,k ,T _k } record at any time T _p . In terms of results, the mean AUC was 0.89 (SD 0.019), the F1 score was 0.55, the sensitivity was 0.52, and the specificity was 0.97. Alternatively, when the end point was set at exactly 3.3 hours, its AUC was 0.86 and the F1 score was 0.55. Compared to the original results obtained with an untruncated negative HD session lasting approximately 4 hours (AUC 0.83, F1 score 0.53, sensitivity 0.53, and specificity 0.96), the end point was set at an earlier The forecast is better. In fact, when the end point of negative hemodialysis sessions was randomized between 2.5 and 3.5 hours, the AUC was 0.92, the F1 score was 0.62, the sensitivity was 0.61, and the specificity was 0.98.

Contributions and key observations

According to the findings of the examples described herein, an algorithm combining linear and difference analysis with binary classification machine learning has a high AUC for predicting adverse events in dialysis. When trying to find the features that contributed the most to predicting all adverse events except hypertension (group 1) from the full 84 features, only features 76 and 82 of the best 23 features were correlated with ultrafiltration (ultrafiltration rate change The number of times and the linear regression slope of the ultrafiltration volume) are related. After excluding the above two ultra-filtering related features, it was found that the remaining 21 features were sufficient for accurate prediction and good discrimination, and its AUC dropped slightly from 0.83 (84 features) to 0.82 (21 features). The model established by 14 ultra-filtered related features also has a good AUC of 0.83. Therefore, select the best 21 features that are not related to ultra-filtering or integrate a total of 14 features that are related to ultra-filtering, instead of All 84 features are included in the model building, which can reduce the computational load. The results shown in Figures 3A and 3B also suggest that the two population characteristics (curves c and e) may embed similar factors leading to the occurrence of adverse events.

In findings according to the examples described herein, muscle cramps were the most frequently occurring adverse events during hemodialysis sessions. Muscle cramps are a common adverse event during hemodialysis treatment, with a prevalence of 28% across all hemodialysis treatments. Muscle cramps are caused by ischemia of skeletal muscle tissue, represent an early sign of hypotension, and may lead to premature interruption of hemodialysis. Tissue ischemia during hemodialysis was positively correlated with ultrafiltration rate. In an attempt to identify the features that contributed most to the prediction of muscle spasticity (group 2), the model built from 14 ultra-filtered relevant features had an AUC of 0.85 for predicting the occurrence of muscle spasticity in this study. When excluding all ultrafiltration-related features (including ultrafiltration rate and ultrafiltration volume) to test the prediction accuracy, the AUC dropped from 0.82 (84 features) to 0.79 (70 features), indicating that ultrafiltration-related features are important for prediction Muscle cramps are associated but not required. Machine learning results showed that features unrelated to ultrafiltration also helped predict muscle cramps in dialysis.

In general, symptomatic hypotension occurs in 20 to 30 percent of hemodialysis sessions. There are two main pathophysiological mechanisms of hypotension during dialysis. First, blood volume decreases when the rate at which plasma fluid is removed by ultrafiltration exceeds the rate at which plasma refills blood vessels. At the same time, hypotension can occur if the cardiovascular and neurohormonal systems fail to compensate for the acute loss of vascular volume during ultrafiltration. Frequent episodes of hypotension during dialysis may lead to decreased ultrafiltration, insufficient "dry body mass", increased preload, and impaired cardiac function, which eventually lead to more episodes of hypotension, thus creating a vicious cycle. At the same time, during frequent dialysis, hypotension undermines dialysis efficiency and efficacy, which is associated with higher morbidity and mortality, which partly explains why cardiovascular disease is one of the major causes of morbidity and mortality in hemodialysis patients. main cause. Existing Products recently developed an intelligent system to predict hypotension in dialysis. However, according to the machine learning model of the embodiments described herein, not only can the features related to ultrafiltration be further ruled out, but also the Adverse events, rather than just focusing on hypotensive events.

For example, as shown in Table 11 below, the best 16 features that are mainly helpful for predicting muscle spasm are listed, including patient characteristics, venous pressure, transmembrane pressure, ultrafiltration, blood flow rate, and pulse pressure. Here, the venous pressure minimum and the transmembrane pressure mean were the features with the most hits (20 and 17, respectively), and thus the top two features. The two previous features used to predict muscle spasticity were derived from hemodialysis machine output parameters, suggesting that it is possible to integrate the algorithms discussed in this paper into hemodialysis machine software to alert clinicians and adjust hemodialysis machine settings in advance. However, unlike the prediction of adverse events in group 1 (only 2 of its best 23 features were related to ultrafiltration), for the prediction of muscle spasms, 8 of its best 16 features were related to ultrafiltration, This indicates that ultrafiltration-related parameters are correlates of muscle spasm.

Among several binary classification models (Bayzier point machine, boosted decision tree, and SVM) built according to the embodiments described herein, the model built by the binary classification average perceptron had the best AUC and F1 score. Clinicians are facing a new era of artificial intelligence, and the integration of computer science and dialysis medicine is considered the first step to comprehensively improve the quality of care for hemodialysis patients. The examples described herein demonstrate the feasibility of such integration. In addition, combining machine learning with hemodialysis machines and adjusting algorithms in real time through cloud computing and data collection accumulation is expected to further improve predictive performance.

Regarding the early prediction of adverse events in dialysis, given that the agreement of the predicted probability of adverse events using features based on different cut-off end time points is about one-tenth that of the predicted hemodialysis sessions with adverse event predictions (see Figures 6 and 7 and Table 10), it would be expected that increasing the number of hemodialysis sessions with adverse events in model training would improve unbalanced data and possibly advance the timing of alerts. Furthermore, since most adverse events occurred in the second half of hemodialysis, it can be expected that if more hemodialysis data sets were included, the second half of hemodialysis data sets would be sufficient for prediction. Furthermore, if more hemodialysis sessions with adverse events were enrolled, it is expected that individual predictive models for each adverse event would be achievable without the need to group adverse events to reduce unbalanced data outcomes.

In the examples described herein, a binary classification machine learning model with an AUC higher than 0.8 was built to predict adverse events in dialysis nearly instantaneously. Consistency in the predicted probabilities of adverse events in dialysis obtained from features extracted in real time from ongoing hemodialysis sessions can provide an alert for impending adverse events in a hemodialysis session. This approach, enabled by cloud computing, can alert clinicians Take the necessary measures and adjust the hemodialysis machine settings in advance.

The present disclosure clarifies its characteristics and functions by means of exemplary embodiments, but is not intended to limit the scope of the present disclosure. Various changes and modifications may be made to this disclosure by those skilled in the art without departing from the scope thereof. However, any equivalent changes and modifications realized according to the present disclosure should be deemed to be included in the scope of the present disclosure. The scope of this disclosure shall be limited by the appended claims.

1: system

10: Feature extraction module

20: Data storage module

30:Model building and optimization module

S1~S4: steps

Claims (17)

A system for predicting adverse events in dialysis, comprising: a feature extraction module configured to collect and process data about a patient's hemodialysis sessions; and a model building and optimization module configured to based on the The data is used to establish a machine learning model for predicting the adverse event in dialysis during the course of hemodialysis; wherein, the machine learning model is based on the first dimension of the target adverse event in dialysis to be predicted and the blood to be predicted The second dimension of the target time period during the dialysis session was established. The system according to claim 1, wherein the data about the hemodialysis course of the patient includes one or more of demographic information, physiological data, dialysis data, and registered adverse events in dialysis. The system of claim 1, wherein the data comprises a data set having a plurality of records having measurements at corresponding time stamps. The system of claim 3, wherein the feature extraction module collects and processes the data about the patient's hemodialysis session by: deriving from the measurements a mean value, a standard deviation of the mean value, At least one of coefficient of variation, linear regression slope, and linear regression R-square is used as the characteristic of the plurality of records. The system of claim 3, wherein the measurements include venous pressure and transmembrane pressure, and wherein the feature extraction module collects and processes the data about the patient's hemodialysis session by: The measurement of venous pressure and the transmembrane pressure derives at least one of a maximum, minimum, and average value, and a second differential, of rates of change to characterize the plurality of records. The system of claim 1, wherein the machine learning model is trained by using the data labeled with outcomes related to the adverse event in dialysis. The system as claimed in claim 1, wherein the machine learning model is trained by combining key features extracted from the data. The system as described in claim 1 further includes a data storage module configured to store the data. A method for predicting adverse events in dialysis, which includes: configuring a feature extraction module to collect and process data about the patient's hemodialysis course; A machine learning model of the adverse event in dialysis during the hemodialysis course; wherein, the machine learning model is based on the first dimension of the target adverse event in dialysis to be predicted and the target period of the hemodialysis course to be predicted Created by the second dimension. The method according to claim 9, wherein the data about the hemodialysis course of the patient includes one or more of demographic information, physiological data, dialysis data, and registered adverse events in dialysis. The method of claim 9, wherein the data comprises a data set having a plurality of records having measurements at corresponding time stamps. The method of claim 11, wherein the feature extraction module collects and processes the data about the patient's hemodialysis session by: deriving from the measurements a mean value, a standard deviation of the mean value, a variance At least one of coefficient, linear regression slope, and linear regression R-square is used as the feature of the plurality of records. The method of claim 11, wherein the measurements include venous pressure and transmembrane pressure, and wherein the feature extraction module collects and processes the data about the patient's hemodialysis session by: The measurement of venous pressure and the transmembrane pressure derives at least one of a maximum, minimum, and average value, and a second differential, of rates of change to characterize the plurality of records. The method according to claim 9, wherein the machine learning model is trained by using the data labeled with outcomes related to the adverse events in the dialysis. The method according to claim 9, wherein the machine learning model is trained by combining key features extracted from the data. The method described in Claim 9 further includes configuring a data storage module to store the data. A computer-readable medium storing computer-executable codes, the computer-executable codes are executed to implement the method described in Claim 9.

Images