Sepsis organ dysfunction early warning method based on sparse data incremental encoding
By analyzing patient data using sparse data incremental coding technology, an early warning model is constructed, which solves the problem of insufficient sensitivity in the early diagnosis of sepsis, and realizes efficient sepsis risk identification and early intervention, applicable to a variety of medical environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2024-11-13
- Publication Date
- 2026-07-24
AI Technical Summary
Existing methods for diagnosing sepsis lack sensitivity in early diagnosis and rely on multiple clinical parameters and laboratory tests, leading to diagnostic delays, especially in resource-limited environments where early and accurate identification is difficult.
A method based on sparse data incremental coding is adopted. By analyzing patients' clinical diagnostic data, treatment record data, laboratory historical test data and past medical history data, machine learning technology is used to build an early warning model, extract sparse patterns, identify sepsis risk indicators and output early warning information.
It improves the accuracy and timeliness of early diagnosis of sepsis, reduces reliance on physician experience and resources, is applicable to various medical environments, and significantly improves patient prognosis, especially in resource-constrained areas.
Smart Images
Figure CN119581047B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to an early warning method for sepsis-related organ dysfunction based on sparse data incremental coding. Background Technology
[0002] Sepsis is a complex clinical syndrome caused by infection, characterized by a systemic inflammatory response and rapidly progressing multi-organ dysfunction. This condition is often caused by bacterial, viral, fungal, or other microbial infections, and its clinical symptoms are diverse, potentially including fever, chills, rapid heartbeat, tachypnea, and altered mental status. Due to the high heterogeneity of sepsis' clinical manifestations and its rapid onset, early diagnosis is extremely difficult, but crucial for patient survival and recovery.
[0003] The high mortality rate and rapid progression of sepsis pose a significant challenge to healthcare systems. Statistics show that sepsis is the leading cause of non-cardiac death in hospitals worldwide, with a mortality rate reaching 30% to 50%. Organ dysfunction secondary to sepsis leads to even higher mortality rates. Furthermore, sepsis and its organ dysfunction can cause long-term physical and psychological sequelae, severely impacting patients' quality of life. Therefore, the medical community places great emphasis on the early identification and timely treatment of sepsis and its organ dysfunction.
[0004] Early diagnosis of sepsis and its organ dysfunction is crucial for improving patient survival rates. Currently, this process relies heavily on clinicians' meticulous identification of patient symptoms and professional interpretation of laboratory test results. However, because sepsis symptoms often resemble those of other diseases and lack specificity, it places high demands on physicians' experience and diagnostic skills. Furthermore, laboratory tests typically take time to complete, which may delay timely diagnosis and treatment of sepsis.
[0005] In resource-constrained healthcare settings, the early diagnosis of sepsis faces greater challenges due to a lack of sufficient medical equipment and professional personnel. These factors collectively contribute to delays in sepsis diagnosis, increasing the health risks for patients.
[0006] While existing methods for diagnosing sepsis play an important role in clinical practice, they still have some significant limitations in early diagnosis. First, clinical scoring systems such as SOFA and qSOFA, although providing a quantitative assessment of sepsis severity, are not sensitive enough in identifying early-stage sepsis. These scoring systems typically require multiple clinical parameters, which may not be apparent in the early stages of sepsis, leading to delays in diagnosis.
[0007] Secondly, laboratory tests, such as blood cultures and the measurement of inflammatory markers like C-reactive protein (CRP) and procalcitonin (PCT), while crucial for diagnosing and assessing the severity of sepsis, often take time to complete, potentially delaying treatment in emergency situations. Furthermore, the accuracy of laboratory tests can be affected by sample collection, processing, and analysis methods. Moreover, existing diagnostic methods are highly dependent on medical resources. In resource-constrained environments, the necessary laboratory equipment and personnel may be lacking to perform these tests. This limits the widespread availability and accessibility of early sepsis diagnosis.
[0008] Machine learning technology, by learning potential patterns in disease development, holds promise for early warning of sepsis. Classical machine learning methods are based on time-series data (data that has a time sequence). Machine learning techniques capture the characteristics of data changes over a period of time to predict future trends. However, due to the high heterogeneity of patients and the strong sparsity of data in sepsis, conventional time-series models struggle to handle these issues. Most existing sepsis early warning methods are cross-sectional. Cross-sectional methods struggle to capture complex time series data, thus still facing the problem of poor early warning effectiveness. Some methods that use time series analysis for early warning (such as CN116580847B) use traditional imputation methods such as mean or median imputation for missing values when processing time-series data. These methods perform poorly on highly sparse data. Summary of the Invention
[0009] Therefore, this application discloses the following technical solution:
[0010] The first aspect of this application provides a method for early warning of sepsis-related organ dysfunction based on sparse data incremental coding, including:
[0011] Obtain event sequence data and non-event sequence data of the target object. The event sequence data includes at least one of clinical diagnostic data, treatment record data, and laboratory historical test data. The non-event sequence data includes at least past medical history data and target vital sign data.
[0012] Incremental encoding is performed on the event sequence data to obtain the sparse pattern of the event sequence data;
[0013] The sparse pattern, the event sequence data, and the non-event sequence data are processed according to the pre-built early warning model to obtain the sepsis risk index of the target object;
[0014] When the sepsis risk indicators meet the preset warning conditions, target warning information is output.
[0015] Optionally, the event sequence data includes multiple data points corresponding to multiple timestamps;
[0016] The incremental encoding of the event sequence data to obtain the sparse pattern of the event sequence data includes:
[0017] The timestamps in the event sequence data that are not empty are determined.
[0018] The number of timestamps in the event sequence data that are not null is counted to obtain the sparse pattern of the event sequence data.
[0019] Optionally, obtain the event sequence data of the target object, including:
[0020] Obtain the target object's historical medical records and historical test results;
[0021] Clinical diagnostic data and treatment record data are obtained from the historical medical record texts;
[0022] Historical test data of the laboratory is obtained based on the historical test text.
[0023] Optionally, obtaining the target vital sign data includes:
[0024] Establish a communication connection with the target monitoring equipment;
[0025] The target monitoring device obtains the target vital signs data of the target object at multiple times within the target time period.
[0026] Optionally, the method for constructing the early warning model includes:
[0027] Multiple training samples are obtained, each training sample corresponding to a reference object. The training sample includes event sequence data, non-event sequence data, and sepsis label values corresponding to the reference object.
[0028] The multiple training samples are processed according to the initial early warning model to be trained, and the model loss of the initial early warning model is obtained.
[0029] The initial warning model is updated based on the model loss, and the process of processing the multiple training samples based on the initial warning model to be trained is returned until the model loss meets the convergence condition, so as to obtain the warning model.
[0030] A second aspect of this application provides an early warning system for sepsis-related organ dysfunction based on sparse data incremental coding, comprising:
[0031] The acquisition unit is used to acquire event sequence data and non-event sequence data of the target object. The event sequence data includes at least one of clinical diagnostic data, treatment record data and laboratory historical test data. The non-event sequence data includes at least past medical history data and target vital sign data.
[0032] An encoding unit is used to incrementally encode the event sequence data to obtain the sparse pattern of the event sequence data.
[0033] The processing unit is used to process the sparse pattern, the event sequence data and the non-event sequence data according to the pre-built early warning model to obtain the sepsis risk index of the target object;
[0034] The output unit is used to output target warning information when the sepsis risk index meets the preset warning conditions.
[0035] Optionally, the event sequence data includes multiple data points corresponding to multiple timestamps;
[0036] When the encoding unit performs incremental encoding on the event sequence data to obtain the sparse pattern of the event sequence data, it is used for:
[0037] The timestamps in the event sequence data that are not empty are determined.
[0038] The number of timestamps in the event sequence data that are not null is counted to obtain the sparse pattern of the event sequence data.
[0039] Optionally, when the obtaining unit obtains the event sequence data of the target object, it is used to:
[0040] Obtain the target object's historical medical records and historical test results;
[0041] Clinical diagnostic data and treatment record data are obtained from the historical medical record texts;
[0042] Historical test data of the laboratory is obtained based on the historical test text.
[0043] Optionally, when the obtaining unit obtains the target vital sign data, it is used to:
[0044] Establish a communication connection with the target monitoring equipment;
[0045] The target monitoring device obtains the target vital signs data of the target object at multiple times within the target time period.
[0046] Optionally, the processing unit is used to construct the early warning model, specifically for:
[0047] Multiple training samples are obtained, each training sample corresponding to a reference object. The training sample includes event sequence data, non-event sequence data, and sepsis label values corresponding to the reference object.
[0048] The multiple training samples are processed according to the initial early warning model to be trained, and the model loss of the initial early warning model is obtained.
[0049] The initial warning model is updated based on the model loss, and the process of processing the multiple training samples based on the initial warning model to be trained is returned until the model loss meets the convergence condition, so as to obtain the warning model.
[0050] The beneficial effects of this application are as follows:
[0051] This solution can automatically identify potential risks of sepsis by analyzing patients' medical records (such as clinical diagnostic data, treatment records, and past medical history) and physiological data (such as historical laboratory test data and target vital signs data), utilizing advanced data analysis and machine learning technologies. In this way, the algorithm can assist doctors in quickly identifying high-risk patients and enabling early intervention.
[0052] Furthermore, the solution is highly scalable and adaptable, enabling deployment in diverse healthcare environments, including resource-constrained areas. By reducing reliance on physician experience and resource availability, the algorithmic system promises to significantly improve the accuracy and timeliness of diagnosing sepsis and its organ dysfunction, thereby improving patient outcomes and alleviating the burden on the healthcare system.
[0053] Furthermore, this solution addresses the sparsity of event sequence data by employing incremental encoding to extract sparse patterns. These sparse patterns allow the early warning model to pay special attention to non-missing locations in the data during processing, ensuring that the model can accurately capture early signals in sparse data and achieve rapid and efficient sepsis risk assessment. This solution significantly improves the ability and accuracy of identifying early sepsis indicators, overcoming the challenges posed by data sparsity. Attached Figure Description
[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0055] Figure 1 This is a flowchart of an early warning method for sepsis-related organ dysfunction based on sparse data incremental coding, provided in an embodiment of this application.
[0056] Figure 2 This is a flowchart illustrating a method for constructing an early warning model provided in an embodiment of this application;
[0057] Figure 3 This is a schematic diagram of the structure of an early warning system for sepsis organ dysfunction based on sparse data incremental coding, provided in an embodiment of this application. Detailed Implementation
[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0059] This application provides an embodiment of a sepsis organ dysfunction early warning method based on sparse data incremental coding. Please refer to [link to relevant documentation]. Figure 1 The method may include the following steps.
[0060] S101, obtain event sequence data and non-event sequence data of the target object. The event sequence data includes at least one of clinical diagnostic data, treatment record data and laboratory historical test data. The non-event sequence data includes at least past medical history data and target sign data.
[0061] S102, Incremental encoding is performed on the event sequence data to obtain the sparse pattern of the event sequence data.
[0062] S103, based on the pre-built early warning model, processes sparse patterns, event sequence data and non-event sequence data to obtain sepsis risk indicators for the target object.
[0063] S104 outputs target warning information when the sepsis risk indicators meet the preset warning conditions.
[0064] The beneficial effects of this embodiment are as follows:
[0065] This solution can automatically identify potential risks of sepsis by analyzing patients' medical records (such as clinical diagnostic data, treatment records, and past medical history) and physiological data (such as historical laboratory test data and target vital signs data), utilizing advanced data analysis and machine learning technologies. In this way, the algorithm can assist doctors in quickly identifying high-risk patients and enabling early intervention.
[0066] Furthermore, the solution is highly scalable and adaptable, enabling deployment in diverse healthcare environments, including resource-constrained areas. By reducing reliance on physician experience and resource availability, the algorithmic system promises to significantly improve the accuracy and timeliness of diagnosing sepsis and its organ dysfunction, thereby improving patient outcomes and alleviating the burden on the healthcare system.
[0067] Furthermore, this proposal addresses the sparsity of event sequence data by employing incremental encoding to extract sparse patterns. These sparse patterns allow the early warning model to focus on non-missing locations in the data, ensuring accurate capture of early signals within sparse data and enabling rapid and efficient sepsis risk assessment. Compared to traditional methods, the proposed algorithm fully utilizes useful information from incomplete data, significantly improving the accuracy and stability of early diagnosis of sepsis and its organ dysfunction. It exhibits stronger robustness and sensitivity, particularly in cases of atypical symptoms or significant data gaps, thus overcoming the challenges posed by data sparsity.
[0068] In step S101, the target object can be any object requiring sepsis early warning. The event sequence data of the target object can include at least one of clinical diagnostic data, treatment record data, and laboratory historical test data. Each type of event sequence data can include multiple data points corresponding to multiple timestamps within a pre-defined historical time period.
[0069] The historical time period can be set as needed, such as the last 30 days, the last 10 days, etc. Each type of event sequence data can include the corresponding data for each day within that historical time period. For example, an event sequence composed of clinical diagnostic data can include the clinical diagnostic data of the target subject for each day within the last 30 days; an event sequence composed of treatment record data can include the treatment record data of the target subject for each day within the last 30 days; and an event sequence composed of laboratory historical test data can include the laboratory historical test data of the target subject for each day within the last 30 days.
[0070] Specifically, the clinical diagnostic data of the target subject can be obtained for each day within the last 30 days. If no clinical diagnosis is performed on a certain day, the clinical diagnostic data for that day can be empty. If a clinical diagnosis is performed on a certain day, the clinical diagnostic data for that day can represent the clinical diagnosis result for that day. Thus, an event sequence composed of clinical diagnostic data from the last 30 days can be obtained.
[0071] Similarly, we can obtain the daily treatment record data of the target object for the past 30 days. If the target object is not treated on a certain day, the treatment record data for that day can be empty. If the target object is treated on a certain day, the treatment record data for that day can represent the treatment measures implemented on that day, such as the type and dosage of drugs used, as well as other non-drug treatment methods, thereby obtaining an event sequence composed of the treatment record data for the past 30 days.
[0072] It is possible to obtain the daily laboratory historical test data of the target object for the past 30 days. If no laboratory test is performed on the target object's physiological sample (such as blood sample, urine sample, etc.) on a certain day, then the laboratory historical test data for that day can be empty. If the target object is subjected to the above laboratory tests on a certain day, then the laboratory historical test data for that day can represent the test results of that day, such as the values of various blood indicators detected on that day. Thus, an event sequence composed of laboratory historical test data for the past 30 days can be obtained.
[0073] For any target object, the above event sequence data can be obtained based on various text records of the target object within a historical time period. For example, the event sequence data can be obtained in the following way:
[0074] Obtain the target object's historical medical records and historical test results;
[0075] Clinical diagnostic data and treatment record data are obtained from historical medical records.
[0076] Historical test data of the laboratory is obtained from historical test texts.
[0077] The aforementioned historical medical records and historical test texts can be in paper or electronic form. If they are in paper form, the paper text can be scanned to obtain the corresponding electronic text.
[0078] After obtaining the electronic text, we can first use character recognition and natural language understanding technologies to identify information related to clinical diagnosis and treatment in the historical medical records. Then, for each day's clinical diagnosis-related information, we can use any existing encoding technology to comprehensively encode and integrate it, obtaining standardized and structured clinical diagnosis data. For each day's treatment-related information, we can encode it in the same way to obtain standardized and structured treatment record data.
[0079] Similarly, historical test texts can be identified and analyzed to obtain information related to the test results. Then, information related to the test results for each day can be encoded to obtain standardized and structured historical laboratory test data.
[0080] Encoding the original text yields corresponding clinical diagnostic data, treatment record data, and laboratory historical test data, ensuring that the original text data is structured and standardized while providing a rich and high-quality data source for subsequent steps.
[0081] The method for obtaining the target subject's past medical history data can be to obtain the target subject's past health record data, identify and encode the health record data in the manner described above, and obtain past medical history data that characterizes the target subject's past medical history.
[0082] The target vital sign data can be obtained in the following ways:
[0083] Establish a communication connection with the target monitoring equipment;
[0084] The target monitoring device obtains the target's vital signs data at multiple times within the target time period.
[0085] The target monitoring device may include any one or more devices for monitoring the target subject's vital signs. Specifically, it may be a portable wearable device or a fixed desktop device. The target vital signs may be one or more pre-specified physiological signs, including but not limited to heart rate, blood oxygen concentration, blood pressure, and vital capacity.
[0086] The target time period can be any time period that can be set as needed, such as the most recent 1 hour, the most recent 30 minutes, etc. Target vital sign data at multiple times can include the average value of the target object's target vital sign data per minute.
[0087] For example, target vital sign data may include the target subject’s average heart rate per minute over the most recent hour.
[0088] In this embodiment, the target vital signs data can also be obtained by encoding the raw data collected by the target monitoring device, and the encoding method can be the same as the aforementioned encoding method.
[0089] In this embodiment, all event sequence data can be represented by a single T*C. I A matrix X of dimension I This indicates that T can be the number of timestamps within a historical time period. For example, if the timestamp is a date and the historical time period is the most recent 30 days, then T can be equal to 30. I This represents the number of types of event sequence data, i.e., a total of C. I In the case of event sequence data, such as event sequence data including clinical diagnostic data, treatment record data, and historical laboratory test data, C I It can be equal to 3. Any element in this matrix can be denoted as x. t , c , representing the c-th type of event sequence data corresponding to the t-th timestamp, for example, x2，3 This can represent the treatment record data of the target object on day 2 within a historical time period, and each element can be a vector representing the corresponding information, such as x. 2，2 It can be a vector obtained by encoding the treatment record data of the target object on the second day of the historical time period according to the aforementioned encoding method.
[0090] All non-event sequence data can be used with T*C nI A matrix X of dimension nI It means that C nI This represents the number of types of non-event sequence data, i.e., a total of C. nI For non-event sequence data, such as non-event sequence data including past medical history data and target sign data, C nI It can be equal to 2. The T parameter for non-event sequence data can be independent of the timestamp. This parameter can be set when encoding non-event sequence data. Setting it to T is to align with event sequence data so that the early warning model can process it.
[0091] In step S102, the event sequence data can be encoded using an event encoding strategy to obtain the coefficient pattern of the event sequence data. This process can be represented by the following formula (1).
[0092] I = Inc(X) I (1)
[0093] Where Inc() represents incremental encoding. This encodes the event sequence X. I Encoding as a sparse pattern I can indicate the clinical level of attention given to the target object. Here, I can be a dimension and X. I The same matrix. Formula (1) can be understood as, for X I Each element is encoded according to the following formula (2) to obtain the element i at the corresponding position in I. t，c .
[0094]
[0095] NaN represents empty data, for example, event sequence X. I In this context, c equals 3, representing historical laboratory testing data. If the historical testing data for day 5 within the historical time period is empty, then x... 5，3 x5,3 is not equal to empty data NaN. If the laboratory historical test data for day 5 within the historical time period is not empty, then x5,3 is not equal to empty data NaN.
[0096] 1() represents an indicator function, which can be defined by the following formula (3).
[0097]
[0098] For example, for 1(x) t，c (≠NaN), if the c-th type of data on day t is empty, the value of the indicator function is 0; if the c-th type of data on day t is not empty, the value of the indicator function is 1.
[0099] In summary, the incremental encoding method for the event sequence data in step S102 can be summarized as follows:
[0100] Determine the timestamp in the event sequence data that is not null;
[0101] The number of timestamps in the event sequence data that are not null is counted to obtain the sparse pattern of the event sequence data.
[0102] By encoding sparse event sequence data using the methods described above, this embodiment can remove irrelevant or redundant information, correct errors, standardize data format, and extract key medical features. Through this encoding process, this solution can effectively capture indicative information in sparse data, providing accurate input for the training and processing of early warning models.
[0103] The early warning model used in this embodiment can be pre-trained based on multiple training samples. This early warning model can be based on a deep neural network or a support vector machine; there is no limitation on the type of model.
[0104] The early warning model processes sparse patterns, event sequence data, and non-event sequence data to obtain a sepsis risk index for the target object. The sepsis risk index can characterize the probability of the target object having sepsis. The larger the value of the index, the higher the probability, and the smaller the value, the lower the probability.
[0105] In step S104, if the sepsis risk indicator does not meet the preset warning conditions, no information may be output. If the sepsis risk indicator meets the preset warning conditions, target warning information may be output. This warning information indicates that the target subject has a high probability of developing sepsis and should be intervened or treated in advance.
[0106] The warning condition can be that the value of the sepsis risk indicator is greater than a set threshold.
[0107] Target warning information can be output in various ways, such as voice, text display, vibration, and ringing, without limitation.
[0108] The target warning information can be output to the target object or to the medical staff responsible for the diagnosis or care of the target object.
[0109] Optionally, in the early warning method of this embodiment, it can be based on... Figure 2The method shown pre-builds an early warning model, and then implements the early warning method of the above embodiment based on the built early warning model.
[0110] S201, obtain multiple training samples, each training sample corresponds to a reference object, and the training samples include event sequence data, non-event sequence data and sepsis label values of the corresponding reference object.
[0111] S202, process multiple training samples based on the initial early warning model to be trained, and obtain the model loss of the initial early warning model.
[0112] S203, update the initial warning model based on the model loss, and return to process multiple training sample steps based on the initial warning model to be trained until the model loss meets the convergence condition to obtain the warning model.
[0113] In step S201, multiple reference objects can be selected. The reference objects can be any objects known to have or not have sepsis.
[0114] For each reference object, the event sequence data and non-event sequence data of the reference object can be obtained in the manner described in step S101 above, which will not be repeated here.
[0115] For each reference subject, the sepsis label value can be determined based on the range of sepsis risk indicators and the actual disease status of the reference subject. For example, the upper limit of the sepsis risk indicator is 1 and the lower limit is 0. For a reference subject known to have sepsis, its label value can be set to be equal to the lower limit of 0, or set to a value close to the lower limit of 0, such as 0.01. For a reference subject known not to have sepsis, its label value can be set to be equal to the upper limit of 1, or set to a value close to the upper limit of 1, such as 0.99.
[0116] For each reference object, a training sample consisting of the event sequence data, non-event sequence data, and sepsis label value of that reference object can be obtained in the manner described above.
[0117] The initial early warning model to be trained can be obtained by determining a model structure, such as a deep neural network model structure, and then determining the parameter values of all model parameters in the model structure through random initialization. The resulting model can be used as the initial early warning model to be trained.
[0118] In step S202, for each training sample, the sparse pattern of the event sequence data in the training sample can be obtained first, as in steps S102 and S103. Then, the event sequence data, non-event sequence data and sparse pattern of the training sample are processed by the initial early warning model to be trained to obtain the predicted value of the training sample.
[0119] The predicted value of a training sample can be considered as the sepsis risk index of the reference object corresponding to that training sample, output by the initial early warning model to be trained.
[0120] After obtaining the predicted values of multiple training samples, the model loss of the initial warning model to be trained can be determined based on the deviation of multiple training samples. The deviation of a training sample can be represented by the difference between the predicted value and the label value of that training sample. The larger the absolute value of the deviation of the training sample and the more training samples with large absolute values of deviation, the greater the model loss. Conversely, the smaller the absolute value of the deviation of the training sample and the fewer training samples with large absolute values of deviation, the smaller the model loss.
[0121] In S202, the absolute values of the deviations of multiple training samples can be averaged to obtain the model loss. Alternatively, the deviations of multiple training samples can be processed in other ways to obtain the model loss. For details, please refer to the relevant existing technologies, which will not be elaborated here.
[0122] In step S203, it can be first determined whether the current model loss meets the convergence condition. If it does, the training ends, and the initial warning model at this time is output as... Figure 1 If the early warning model used in the corresponding embodiment does not meet the requirements, the parameter values of each model parameter in the initial early warning model at this time are updated according to the model loss. After the update is completed, S202 can be executed again, and then the update is performed again in the manner of S203, until the model loss obtained after executing S202 once meets the convergence condition.
[0123] The method for updating model parameters based on model loss can be found in relevant existing technologies and will not be elaborated here.
[0124] The convergence condition can be defined as the model loss being less than a preset loss threshold. That is, if the model loss is less than the loss threshold, the convergence condition is met; if it is greater than or equal to the loss threshold, the convergence condition is not met.
[0125] In some embodiments, several updated samples may be collected periodically in the same manner as the training samples obtained in S201, and these updated samples may be processed using the early warning model to update the early warning model based on the processing results.
[0126] The method of using the early warning model to process these updated samples can be found in step S202, and the method of updating the early warning model based on the processing results can be found in step S203, which will not be elaborated here.
[0127] By extracting the sparse patterns of event sequence data during model training and updates, non-missing data from sparse event sequence data can be gradually integrated into the early warning model in the face of missing data and imbalanced samples, thereby optimizing the information extraction process.
[0128] This application also provides an early warning system for sepsis-related organ dysfunction based on sparse data incremental coding. Please refer to [link to relevant documentation]. Figure 3 This is a schematic diagram of the system structure, which may include the following units.
[0129] The acquisition unit 301 is used to acquire event sequence data and non-event sequence data of the target object. The event sequence data includes at least one of clinical diagnostic data, treatment record data and laboratory historical test data. The non-event sequence data includes at least past medical history data and target sign data.
[0130] Encoding unit 302 is used to incrementally encode the event sequence data to obtain the sparse pattern of the event sequence data;
[0131] Processing unit 303 is used to process sparse pattern, event sequence data and non-event sequence data according to a pre-built early warning model to obtain sepsis risk indicators of the target object;
[0132] The output unit 304 is used to output target warning information when the sepsis risk indicators meet the preset warning conditions.
[0133] Optionally, the event sequence data includes multiple data points corresponding to multiple timestamps;
[0134] When encoding unit 302 performs incremental encoding on event sequence data to obtain the sparse pattern of the event sequence data, it is used for:
[0135] Determine the timestamp in the event sequence data that is not null;
[0136] The number of timestamps in the event sequence data that are not null is counted to obtain the sparse pattern of the event sequence data.
[0137] Optionally, when obtaining the event sequence data of the target object, the obtaining unit 301 is used for:
[0138] Obtain the target object's historical medical records and historical test results;
[0139] Clinical diagnostic data and treatment record data are obtained from historical medical records.
[0140] Historical test data of the laboratory is obtained from historical test texts.
[0141] Optionally, when obtaining target vital sign data, the obtaining unit 301 is used for:
[0142] Establish a communication connection with the target monitoring equipment;
[0143] The target monitoring device obtains the target's vital signs data at multiple times within the target time period.
[0144] Optionally, the processing unit 303 is used to construct the early warning model, specifically for:
[0145] Multiple training samples are obtained, each training sample corresponds to a reference object, and the training samples include event sequence data, non-event sequence data and sepsis label values of the corresponding reference object;
[0146] The model loss of the initial early warning model is obtained by processing multiple training samples based on the initial early warning model to be trained.
[0147] The initial warning model is updated based on the model loss, and then multiple training sample steps are processed based on the initial warning model to be trained until the model loss meets the convergence condition to obtain the warning model.
[0148] The sepsis organ dysfunction early warning system based on sparse data incremental coding provided in this embodiment works by referring to the relevant steps of the sepsis organ dysfunction early warning method based on sparse data incremental coding provided in this application embodiment, and will not be repeated here.
[0149] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0150] For ease of description, the above systems or devices are described separately by function, divided into various modules or units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.
[0151] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.
[0152] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0153] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A method for early warning of organ dysfunction in sepsis based on sparse data incremental coding, characterized in that, include: Obtain event sequence data and non-event sequence data of the target object. The event sequence data includes at least one of clinical diagnostic data, treatment record data, and laboratory historical test data. The non-event sequence data includes at least past medical history data and target vital sign data. The event sequence data includes multiple data corresponding to multiple timestamps. The timestamps in the event sequence data that are not empty are determined. The number of timestamps in the event sequence data that are not null values is counted to encode the event sequence data into a sparse pattern; wherein, the sparse pattern indicates the clinical attention level of the target object; the method of obtaining the sparse pattern is as follows: using an indicator function to determine whether each type of event sequence data under each timestamp is null, the value of the indicator function is 1 when it is not null, and the value of the indicator function is 0 when it is null; the number of timestamps with an indicator function value of 1 under each timestamp is counted to obtain the number of timestamps that are not null values, so as to remove irrelevant or redundant information in the event sequence data, standardize the data format, encode and extract key medical features, capture the indicative information in the event sequence data, and obtain the sparse pattern; The sparse pattern, the event sequence data, and the non-event sequence data are processed according to the pre-built early warning model to obtain the sepsis risk index of the target object; When the sepsis risk indicators meet the preset warning conditions, target warning information is output; the target warning information indicates that the target object has a high probability of developing sepsis and should be intervened or treated in advance.
2. The method according to claim 1, characterized in that, Obtain the event sequence data of the target object, including: Obtain the target object's historical medical records and historical test results; Clinical diagnostic data and treatment record data are obtained from the historical medical record texts; Historical test data of the laboratory is obtained based on the historical test text.
3. The method according to claim 1, characterized in that, Obtaining the target vital sign data includes: Establish a communication connection with the target monitoring equipment; The target monitoring device obtains the target vital signs data of the target object at multiple times within the target time period.
4. The method according to claim 1, characterized in that, The method for constructing the early warning model includes: Multiple training samples are obtained, each training sample corresponding to a reference object. The training sample includes event sequence data, non-event sequence data, and sepsis label values corresponding to the reference object. The multiple training samples are processed according to the initial early warning model to be trained, and the model loss of the initial early warning model is obtained. The initial warning model is updated based on the model loss, and the process of processing the multiple training samples based on the initial warning model to be trained is returned until the model loss meets the convergence condition, so as to obtain the warning model.