Multi-modal data-based sepsis monitoring method and system

Abstract

The invention relates to the technical field of sepsis monitoring, and discloses a sepsis monitoring method and system based on multi-modal data, and the method comprises the steps: setting a plurality of sepsis patient types based on influence factors, and generating a plurality of disease course period sequences which comprise a plurality of disease course periods, each disease course period is mapped with a corresponding multi-modal data set; comparing and analyzing the multi-modal data sets of the plurality of disease course cycle sequences to obtain feature data of each disease course cycle, change features of each feature data and a weight coefficient, and constructing a sepsis monitoring model; the real-time characteristic data of the to-be-monitored patient is obtained and input into the sepsis monitoring model, the monitoring result is obtained, whether the early warning instruction is generated or not is judged, the complexity of sepsis and the individual difference of the patient are fully considered, the monitoring accuracy is improved, and powerful support is provided for timely diagnosis and treatment of the sepsis patient.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sepsis monitoring, in particular to a sepsis monitoring method and system based on multi-modal data. BACKGROUND

[0002] Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection, with rapid onset, rapid progression and high mortality. Timely and accurate monitoring and diagnosis is crucial to reduce the mortality of sepsis patients.

[0003] In the prior art, sepsis monitoring methods rely on local features of single modal data, which is difficult to capture the complex dynamic changes of the disease, and uses fixed thresholds for judgment, which lacks adaptability to individual differences of patients, resulting in insufficient sensitivity and specificity of monitoring, and easy occurrence of missed diagnosis or misdiagnosis, which delays the best treatment opportunity. SUMMARY

[0004] To solve the above technical problems, the present application provides a sepsis monitoring method and system based on multi-modal data, which sets several sepsis patient categories and determines the feature data of each sepsis patient category in different disease periods, the change characteristics of the feature data and the weight coefficients, thereby constructing a sepsis monitoring model, fully considering the complexity of sepsis and individual differences of patients, improving the monitoring accuracy and providing strong support for timely diagnosis and treatment of sepsis patients.

[0005] In some embodiments of the present application, a sepsis monitoring method based on multi-modal data is provided, comprising: determining an influence factor, setting several sepsis patient categories based on the influence factor, and generating several disease course period sequences for each sepsis patient category, wherein the disease course period sequence includes several disease course periods, and each disease course period is mapped with a corresponding multi-modal data set; comparing and analyzing the multi-modal data sets of several disease course period sequences of the same sepsis patient category to obtain several feature data of each disease course period, change characteristics of each feature data and weight coefficients; performing neural network training on the several feature data of all disease course periods of all sepsis patient categories, the change characteristics of each feature data and the corresponding weight coefficients to obtain a sepsis monitoring model; obtaining real-time feature data of a patient to be monitored and inputting it into the sepsis monitoring model to obtain a monitoring result and determine whether to generate an early warning instruction.

[0006] In some embodiments of the present application, several sepsis patient categories are set based on the influence factor, and several disease course period sequences for each sepsis patient category are generated, comprising: determining a plurality of influence factors, each of the influence factors comprising a plurality of influence factor intervals; randomly selecting one of the influence factor intervals of each of the influence factors, and combining to obtain an influence factor set; sequentially combining to obtain a plurality of influence factor sets; setting each of the influence factor sets as a sepsis patient category, and setting a plurality of sepsis patient categories according to all of the influence factor sets; determining a plurality of monitored patients of each of the sepsis patient categories; obtaining a historical course log of each of the monitored patients, and dividing the historical course log in combination with a plurality of preset risk coefficient intervals to obtain a plurality of course periods of each of the monitored patients; wherein each of the course periods is mapped with multi-modal data of the corresponding monitored patient in the historical course log and a corresponding historical period length; constructing a course period sequence according to the plurality of course periods of each of the monitored patients; generating a plurality of course period sequences of all of the monitored patients of each of the sepsis patient categories; wherein each of the course period sequences comprises a plurality of course periods, and each of the course periods is mapped with a multi-modal data set of the corresponding monitored patient, the multi-modal data set comprising a plurality of dimensional data.

[0007] In some embodiments of the present application, a plurality of feature data of each of the course periods is obtained, comprising: randomly selecting one of the course periods as a target course period; comparing the multi-modal data set of the target course period with the multi-modal data sets of other course periods in the same course period sequence to obtain a plurality of difference coefficients of each of the dimensional data in the multi-modal data set of the target course period; previously setting a preset difference coefficient threshold; if all of the difference coefficients of the same dimensional data of the target course period are greater than the preset difference coefficient threshold, setting the corresponding dimensional data as the pending feature data of the target course period; sequentially generating a plurality of pending feature data of each of the course periods; comparing the pending feature data of the same course period of different monitored patients in the same sepsis patient category to obtain a credibility coefficient of each of the pending feature data; setting the pending feature data with a credibility coefficient greater than a preset credibility coefficient threshold as the feature data of the corresponding course period; sequentially generating the feature data of each of the course periods.

[0008] In some embodiments of the present application, the change feature of each of the feature data comprises: a preset monitoring time interval for each disease course cycle is set in advance, and a plurality of acquisition nodes corresponding to the disease course cycle are generated according to the preset monitoring time interval; a historical period length of each disease course cycle is taken as a time reference line; characteristic data of each disease course cycle is collected according to each acquisition node and mapped to the corresponding time reference line, so as to obtain a characteristic data change curve segment of each characteristic data in the corresponding disease course cycle and the historical period length; all characteristic data change curve segments of the same disease course cycle of the same sepsis patient category are constructed to obtain a comparison curve graph of the corresponding disease course cycle, and the comparison curve graph includes a plurality of comparison curve segment groups; Each comparison curve segment group includes a characteristic data change curve segment of the same characteristic data in different historical period lengths of the same disease course cycle. According to each comparison curve segment group, the change characteristics of each characteristic data in the corresponding disease course cycle are determined, and the change characteristics include normal change characteristics and abnormal change characteristics.

[0009] In some embodiments of the present application, according to each comparison curve segment group, the change characteristics of the corresponding characteristic data in the corresponding disease course cycle are determined, including: The characteristic data change curve segments in the same comparison curve segment group are subjected to cluster analysis to obtain a plurality of similar curve segment clusters; The plurality of similar curve segment clusters are compared with a standard curve segment of the corresponding disease course cycle, and the characteristic curve segment cluster is determined according to the comparison result; The change characteristics of the characteristic curve segment cluster and each similar curve segment cluster are determined, and the change characteristics include shape, change trend, data interval and time length; The difference quantization value of each similar curve segment cluster and the characteristic curve segment cluster is calculated, and the abnormal level of the change characteristics of the corresponding similar curve segment cluster is set according to the difference quantization value; A preset difference quantization value threshold is set in advance; The change characteristics of the characteristic curve segment cluster and the change characteristics of the similar curve segment cluster with a difference quantization value less than the preset difference quantization value threshold are set as the normal change characteristics of the corresponding characteristic data in the disease course cycle, and the normal label is matched; The change characteristics of the similar curve segment cluster with a difference quantization value not less than the preset difference quantization value threshold are set as the abnormal change characteristics of the corresponding characteristic data in the disease course cycle, and the abnormal label is matched, and each abnormal change characteristic is mapped to a corresponding abnormal level.

[0010] In some embodiments of the present application, the weight coefficient includes: A plurality of difference coefficient difference values of the same characteristic data in the corresponding disease course cycle are calculated, and a difference coefficient difference value average of the corresponding characteristic data in the corresponding disease course cycle is calculated. generating a first weight coefficient of the corresponding feature data according to the difference value mean of the difference coefficient difference; calculating a confidence coefficient difference of the feature data; generating a compensation coefficient of the corresponding feature data according to the confidence coefficient difference; generating a weight coefficient of the corresponding feature data according to the compensation coefficient and the first weight coefficient.

[0011] In some embodiments of the present application, a sepsis monitoring model is obtained, comprising: taking each sepsis patient category, corresponding feature data and weight coefficient of the feature data as first training input data, and taking the corresponding disease course period and confidence as first training output data, performing neural network training to obtain a first monitoring model; taking the change feature of the feature data of each disease course period and the weight coefficient as second training input data, and taking the corresponding label as second training output data, performing neural network training to obtain a second monitoring model; obtaining a sepsis monitoring model according to the first monitoring model and the second monitoring model.

[0012] In some embodiments of the present application, before obtaining real-time feature data of a patient to be monitored and inputting the real-time feature data into the sepsis monitoring model, comprising: obtaining a plurality of real-time influence factor intervals of the patient to be monitored, and constructing a real-time influence factor set of the patient to be monitored; performing similarity analysis on the real-time influence factor set and the influence factor set of each sepsis patient category to obtain a similarity coefficient of the patient to be monitored and each sepsis patient category; sorting the sepsis patient categories according to the similarity coefficient, and setting the first sorted sepsis patient category as the sepsis patient category of the patient to be monitored.

[0013] In some embodiments of the present application, a monitoring result is obtained and it is judged whether to generate an early warning instruction, comprising: inputting the sepsis patient category of the patient to be monitored and the real-time feature data into the sepsis monitoring model to obtain a monitoring result, the monitoring result comprising a disease course period and a label; when the label is a normal label, no early warning instruction is generated; when the label is an abnormal label, a corresponding level of early warning instruction is generated according to the corresponding abnormal level.

[0014] In some embodiments of the present application, further comprising a sepsis monitoring system based on multi-modal data, characterized by comprising: The setting module is configured to determine an influence factor, set a plurality of sepsis patient categories based on the influence factor, and generate a plurality of disease course period sequences for each sepsis patient category, wherein each disease course period sequence comprises a plurality of disease course periods, and each disease course period is mapped to a corresponding multi-modal data set; The analysis module is configured to perform comparative analysis on the multi-modal data sets of the plurality of disease course period sequences of the same sepsis patient category, to obtain a plurality of feature data of each disease course period, a change feature of each feature data, and a weight coefficient; The construction module is configured to perform neural network training based on the plurality of feature data of all disease course periods of all sepsis patient categories, the change feature of each feature data, and the corresponding weight coefficient, to obtain a sepsis monitoring model. The early warning module is configured to obtain real-time feature data of a patient to be monitored, and input the real-time feature data into the sepsis monitoring model to obtain a monitoring result and determine whether to generate an early warning instruction.

[0015] Compared with the prior art, the sepsis monitoring method and system based on multi-modal data according to the embodiments of the present application have the following beneficial effects: By setting a plurality of sepsis patient categories and determining the feature data of each sepsis patient category in different disease periods, the change feature of the feature data, and the weight coefficient, a sepsis monitoring model is constructed, the complexity of sepsis and the individual differences of patients are fully considered, the monitoring accuracy is improved, and strong support is provided for timely diagnosis and treatment of sepsis patients. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 FIG. 1 is a flowchart of a sepsis monitoring method based on multi-modal data according to an embodiment of the present application. DETAILED DESCRIPTION

[0017] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application.

[0018] In the description of the present application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like indicate the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the present application and simplify the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0019] The terms "first", "second", "third", etc. are used only for descriptive purposes and do not connote or imply relative importance or an ordering between or among the indicated technical features. Thus, a feature defined with "first", "second", etc. can include one or more of the features implicitly or explicitly.

[0020] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connection" should be understood in a broad sense, for example, can be fixed connection, can also be detachable connection, or integral connection; can be mechanical connection, can also be electrical connection; can be direct connection, can also be indirect connection through an intermediate medium, can be internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0021] As shown in Figure 1 The sepsis monitoring method based on multi-modal data provided by the embodiment of the present application comprises: determining an impact factor, setting a plurality of sepsis patient categories based on the impact factor, and generating a plurality of disease course period sequences for each sepsis patient category, wherein each disease course period sequence comprises a plurality of disease course periods, and each disease course period is mapped with a corresponding multi-modal data set; comparing and analyzing the multi-modal data sets of the plurality of disease course period sequences of the same sepsis patient category to obtain a plurality of feature data of each disease course period, a change feature of each feature data, and a weight coefficient; performing neural network training on the plurality of feature data of all disease course periods of all sepsis patient categories, the change feature of each feature data, and the corresponding weight coefficient to obtain a sepsis monitoring model; obtaining real-time feature data of a patient to be monitored and inputting the real-time feature data into the sepsis monitoring model to obtain a monitoring result and determine whether to generate an early warning instruction.

[0022] In the present embodiment, the impact factor refers to a patient's own factor that has a greater impact on the sepsis monitoring process or disease course period, including but not limited to the patient's age, genetic background, underlying disease or medication history, etc., and is determined according to the historical monitoring logs of a plurality of sepsis patients.

[0023] In the present embodiment, when generating the disease course period sequence of each sepsis patient category, the dynamic change characteristics of the sepsis disease course are fully considered, and the disease course period includes early stage (latent stage), occurrence stage, development stage, and outcome stage.

[0024] In the embodiment, the monitoring result includes a disease course period and a label, and a monitoring time interval of real-time feature data is set according to the monitoring result. When the disease course period is closer to the outcome period and the label is a normal label, the monitoring time interval can be appropriately extended to reduce unnecessary monitoring operations and resource consumption. When the disease course period is in the early stage (the incubation period), the occurrence stage or the development stage, or the label is an abnormal label, the monitoring time interval is shortened to ensure that subtle changes in the disease condition can be captured in time. In this way, the monitoring time interval is dynamically adjusted, which further improves the efficiency and accuracy of monitoring, better adapts to the complex and changeable condition of the sepsis patient, and provides more scientific and effective support for the treatment and management of the sepsis patient.

[0025] In some embodiments of the present application, a plurality of sepsis patient categories are set based on influence factors, and a plurality of disease course period sequences of each sepsis patient category are generated, including: determining a plurality of influence factors, each influence factor including a plurality of influence factor intervals; randomly selecting an influence factor interval of each influence factor and combining to obtain an influence factor set; sequentially combining to obtain a plurality of influence factor sets; setting each influence factor set as a sepsis patient category, and setting a plurality of sepsis patient categories according to all influence factor sets; determining a plurality of monitored patients of each sepsis patient category; obtaining a historical disease course log of each monitored patient, and dividing the historical disease course log in combination with a plurality of preset risk coefficient intervals to obtain a plurality of disease course periods of each monitored patient; wherein each disease course period is mapped with multi-modal data of the corresponding monitored patient in the historical disease course log and a corresponding historical period length; constructing a disease course period sequence according to a plurality of disease course periods of each monitored patient; generating a plurality of disease course period sequences of all monitored patients of each sepsis patient category; wherein each disease course period sequence includes a plurality of disease course periods, and each disease course period is mapped with a multi-modal data set of the corresponding monitored patient, the multi-modal data set including a plurality of dimensional data.

[0026] In the embodiment, the influence factor intervals further subdivide different states of the influence factors. For example, when the influence factor is age, the influence factor intervals include 0-6, 7-18, 19-40, 41-60, 61 and above, etc. Through such subdivision, the patients can be classified more accurately according to age.

[0027] In the embodiment, a random selection of each influencing factor is combined into an influencing factor set, which can cover various possible patient feature combinations, scientifically and reasonably generate several sepsis patient categories, and lay a foundation for subsequent monitoring work.

[0028] In the embodiment, the disease course cycle sequence includes different disease course cycles such as early stage (latent stage), occurrence stage, development stage, and outcome stage, and different dimension data in the multi-modal data set include historical vital sign data, historical medical image feature data, historical laboratory test index data, and historical nursing record text data of the corresponding monitored patient in the corresponding disease course cycle In the embodiment, the preset risk coefficient interval is calculated according to the data difference between the preset data interval of the multi-modal data of each disease course cycle and the preset standard data (or the preset abnormal data corresponding to sepsis), a historical risk coefficient is calculated by calculating the data difference between the multi-modal data of each monitored patient in the historical disease course log and the corresponding preset standard data (or the preset abnormal data corresponding to sepsis), and the corresponding relationship between the historical risk coefficient and the preset risk coefficient interval is determined, so as to dynamically evaluate the disease course cycle and the risk state of the patient.

[0029] In the embodiment, when the disease course cycle changes from early stage, occurrence stage, development stage, and outcome stage, the corresponding preset risk coefficient interval becomes larger and larger.

[0030] In the embodiment, by constructing multiple sepsis patient categories and constructing several disease course cycle logs, the disease course change of different feature patient groups can be more comprehensively covered, and rich data support is provided for subsequent accurate sepsis monitoring.

[0031] In some embodiments of the present application, several feature data of each disease course cycle are obtained, including: A random disease course cycle is selected as a target disease course cycle; The multi-modal data set of the target disease course cycle is compared with the multi-modal data set of other disease course cycles in the same disease course cycle sequence, and several difference coefficients of each dimension data in the multi-modal data set of the target disease course cycle are obtained; A preset difference coefficient threshold is preset; If all difference coefficients of the same dimension data of the target disease course cycle are greater than the preset difference coefficient threshold, the corresponding dimension data is set as the pending feature data of the target disease course cycle; Several pending feature data of each disease course cycle are sequentially generated; The pending feature data of the same disease course cycle of different monitored patients in the same sepsis patient category are compared, and a credibility coefficient of each pending feature data is obtained. The pending feature data with a trust coefficient greater than a preset trust coefficient threshold is set as the feature data of the corresponding disease course period; The feature data of each disease course period is sequentially generated.

[0032] In the embodiment, the difference coefficient is obtained by overlapping comparison of the data interval of the same dimension data in the target disease course period and other disease course periods in the same sequence, the more the data intervals overlap, the smaller the difference coefficient, and vice versa. For example, the heart rate in the early stage is sixty to eighty times per minute, and the heart rate in the onset stage is seventy to one hundred times per minute, so the data difference coefficient is 30 / 40. Similarly, the data difference coefficients of different dimension data of the same patient are calculated according to the above method. The same dimension data refers to data belonging to the same category in the same disease course period or different disease course periods, such as physiological parameter data such as heart rate, blood pressure, body temperature, or laboratory test result data such as white blood cell count and C-reactive protein.

[0033] In the embodiment, the difference coefficient threshold refers to the minimum difference coefficient that can distinguish the same dimension data in different disease course intervals, which is 0.6 in this application. This value is obtained through a large amount of experimental data and clinical experience, and can accurately distinguish the same dimension data in different disease course periods, thereby screening out feature data with significant differences. In the embodiment, the feature data of each disease course period is not a single indicator, but a specific numerical interval of the corresponding indicator, for example, the feature data of the early stage is heart rate feature data (sixty to eighty times per minute), and the feature data of the onset stage is heart rate feature data (seventy to one hundred times per minute). Through this setting, the disease course period of the patient can be more accurately determined.

[0034] In the embodiment, the trust coefficient refers to the ratio of the number of the same pending feature data appearing in the same disease course period of all monitored patients in the same sepsis patient category to the number of all monitored patients. The trust coefficient threshold is set to 0.8. The trust coefficient threshold is set to further screen out feature data with high trustworthiness and reduce the inaccuracy of feature data caused by individual patient data anomalies or errors, to ensure that the final feature data can truly reflect the change characteristics of the disease course period of sepsis patients.

[0035] In the embodiment, the feature data obtained through the above series of steps can comprehensively and accurately reflect the data situation of sepsis patients in different disease course periods, providing high-quality data support for subsequent neural network training, and helping to improve the accuracy and reliability of the sepsis monitoring model, thereby better achieving real-time monitoring and early warning of sepsis patients.

[0036] In some embodiments of the present application, the change feature of each feature data comprises: a preset monitoring time interval for each disease course period is set, and a plurality of acquisition nodes for the corresponding disease course period are generated according to the preset monitoring time interval; a historical period length of each disease course period is taken as a time reference line; feature data of each disease course period is collected according to each acquisition node and mapped to the corresponding time reference line, so as to obtain a feature data change curve segment of each feature data in the corresponding disease course period and the historical period length; all feature data change curve segments of the same disease course period of the same sepsis patient category are constructed to form a comparison curve graph of the corresponding disease course period, wherein the comparison curve graph comprises a plurality of comparison curve segment groups; Each comparison curve segment group comprises a feature data change curve segment of the same feature data in different historical period lengths of the same disease course period. According to each comparison curve segment group, the change feature of each feature data in the corresponding disease course period is determined, and the change feature comprises normal change feature and abnormal change feature.

[0037] In this embodiment, the preset monitoring time interval is set according to the disease course characteristics of the sepsis patient and the actual clinical needs, for example, it can be set to every hour, every two hours, etc. By reasonably setting the preset monitoring time interval, the change of the feature data at different time points can be accurately captured.

[0038] In this embodiment, the historical period length is taken as the time reference line, so that the change trend of the feature data in the disease course period can be more intuitively observed, which is convenient for subsequent trend feature analysis.

[0039] In this embodiment, the comparison curve graph is constructed to clearly show the change of the same feature data in different historical period lengths. For example, for the feature data of heart rate, the heart rate change curve segments in different historical period lengths in the early disease course period can intuitively show the fluctuation trend of the heart rate in the early stage. By comparing the curve segments of different historical period lengths, the normal change feature and the abnormal change feature of the feature data can be determined.

[0040] In the embodiment, the normal change feature refers to a feature data change rule expected for a sepsis patient in a corresponding disease course period. For example, in the development period, the body temperature usually shows a gradually increasing trend. If the collected body temperature feature data change curve segment conforms to this trend, it is determined as a normal change feature. Otherwise, if a rapidly increasing or decreasing trend appears, it is an abnormal change feature. The abnormal change feature refers to a change deviating from the expected rule, which provides more detailed and targeted data information for subsequent neural network training, further improves the recognition ability of the sepsis monitoring model to the feature change in different disease course periods, and thus more accurately monitors and warns the sepsis patient.

[0041] In some embodiments of the present application, the change feature of the corresponding feature data in the corresponding disease course period is determined according to each comparison curve segment group, comprising: performing cluster analysis on the feature data change curve segments in the same comparison curve segment group to obtain a plurality of similar curve segment clusters; comparing the plurality of similar curve segment clusters with the standard curve segment of the corresponding disease course period, and determining the feature curve segment cluster according to the comparison result; determining the change feature of the feature curve segment cluster and each similar curve segment cluster, wherein the change feature comprises a shape, a change trend, a data interval, and a time length; calculating a difference quantization value of each similar curve segment cluster and the feature curve segment cluster, and setting an abnormal level of the change feature of the corresponding similar curve segment cluster according to the difference quantization value; pre-setting a preset difference quantization value threshold; setting the change feature of the feature curve segment cluster and the change feature of the similar curve segment cluster with a difference quantization value less than the preset difference quantization value threshold as the normal change feature of the corresponding feature data in the disease course period, and matching a normal label; setting the change feature of the similar curve segment cluster with a difference quantization value not less than the preset difference quantization value threshold as the abnormal change feature of the corresponding feature data in the disease course period, and matching an abnormal label, and each abnormal change feature is mapped with a corresponding abnormal level.

[0042] In the embodiment, the standard curve segment is a feature data change curve segment under the normal development rule of each disease course period, which is obtained by statistical analysis on a large number of normal sepsis patient disease course data, and can represent the ideal change trend of the feature data in the disease course period. However, the standard curve segment is not divided according to the patient category, so by comparing the standard curve segment with each similar curve segment cluster, the change feature of the similar curve segment cluster closest to the standard curve segment is set as the normal change feature of the feature data in the corresponding patient category and disease course period, thereby improving the judgment accuracy of the normal change feature of the feature data of different patient categories, and laying a foundation for subsequent construction of a monitoring model.

[0043] In the embodiment, each similar curve segment cluster includes at least one characteristic data variation curve segment, and the curve segments in the same similar curve segment cluster are similar, the difference quantization value is obtained by quantizing the time length difference, the trend difference, the shape difference, and the data interval difference of the characteristic curve segment cluster and the similar curve segment cluster, and the corresponding weight coefficients are 0.3, 0.3, 0.2, and 0.2 respectively, the difference quantization value is calculated by weight processing, and the greater the difference quantization value, the higher the corresponding abnormal level.

[0044] In the embodiment, the preset difference quantization value threshold is a key limit value for distinguishing normal variation characteristics and abnormal variation characteristics, which is set to 0.3 in the embodiment. When the difference quantization value is less than the threshold, it indicates that the variation characteristics of the similar curve segment cluster are close to the variation characteristics of the characteristic curve segment cluster, which is within the range of normal variation characteristics. When the difference quantization value is not less than the threshold, it means that there is a significant difference between the variation characteristics of the similar curve segment cluster and the characteristic curve segment cluster, which should be determined as abnormal variation characteristics, and the corresponding abnormal level is assigned according to the size of the difference quantization value, so as to more accurately evaluate and warn the condition of the sepsis patient in the subsequent process.

[0045] In some embodiments of the present application, the weight coefficient includes: A plurality of difference coefficient difference values of the same characteristic data in the corresponding disease course period are calculated, and a difference coefficient difference value average of the corresponding characteristic data in the corresponding disease course period is calculated; A first weight coefficient of the corresponding characteristic data is generated according to the difference coefficient difference value average; A confidence coefficient difference value of the characteristic data is calculated; A compensation coefficient of the corresponding characteristic data is generated according to the confidence coefficient difference value; The weight coefficient of the corresponding characteristic data is generated according to the compensation coefficient and the first weight coefficient.

[0046] In the embodiment, the difference coefficient difference value is the difference between the difference coefficient of the same monitored patient and the preset difference coefficient threshold, the plurality of difference coefficient difference values are the difference coefficient difference values of the same characteristic data in the corresponding disease course period of different monitored patients, and the difference coefficient difference value average is obtained by averaging all the difference coefficient difference values.

[0047] In the embodiment, the difference coefficient difference value average of all the characteristic data in the same disease course period is weighted, the greater the difference coefficient difference value average, the greater the first weight coefficient, and the higher the importance of the characteristic data in the corresponding disease course period.

[0048] In the embodiment, the difference value of the trust coefficient refers to the difference between the trust coefficient of the same feature data and the preset trust coefficient threshold, the value range of the compensation coefficient is (0.8, 1.2), when the difference value of the trust coefficient is large, it indicates that the trust degree of the feature data is low, at this time the compensation coefficient will appropriately reduce the first weight coefficient, so as to reduce the interference of the untrusted feature data on the subsequent neural network training; when the difference value of the trust coefficient is small, it indicates that the trust degree of the feature data is high, the compensation coefficient will appropriately increase the first weight coefficient, so as to highlight the role of the trusted feature data in the model training.

[0049] In the embodiment, the weight coefficient of the corresponding feature data is generated according to the compensation coefficient and the first weight coefficient, the weight coefficient comprehensively considers the importance and trust degree of the feature data, and can more accurately reflect the importance of the feature data in the sepsis monitoring. By reasonably assigning the weight coefficient to different feature data, the subsequent neural network training can pay more attention to the learning of important feature data, thereby improving the accuracy and reliability of the sepsis monitoring model, and better realizing the accurate monitoring and early warning of the sepsis patients.

[0050] In some embodiments of the present application, the sepsis monitoring model is obtained, including: The first training input data is obtained by taking each sepsis patient category, corresponding feature data and weight coefficient of the feature data as the first training input data, and the first training output data is obtained by taking the corresponding disease course period and trust degree as the first training output data, and the neural network is trained to obtain the first monitoring model; The second training input data is obtained by taking the change feature and weight coefficient of the feature data of each disease course period as the second training input data, and the second training output data is obtained by taking the corresponding label as the second training output data, and the neural network is trained to obtain the second monitoring model; The sepsis monitoring model is obtained according to the first monitoring model and the second monitoring model.

[0051] In the embodiment, the first neural network training is based on the feature data, weight coefficient and corresponding disease course period and trust degree of the sepsis patient category, aiming to initially construct a model framework capable of judging the disease course period of sepsis, and through learning and analysis of a large amount of data, the model has the recognition ability of different disease course periods of sepsis, and the trust degree is obtained by weight processing according to the weight coefficient and the corresponding trust coefficient of the feature data, and the recognition accuracy of the output disease course period.

[0052] In this embodiment, the second neural network training takes the change characteristics of the feature data in the disease course period as input and trains the label as output. The label is obtained by combining the normal label or abnormal label (and abnormal level) output by the multiple feature data and the corresponding weight coefficient after weight processing. In this way, it is more accurate to identify whether the patient's feature data deviates from the normal trend. Through the combination of the two neural network trainings, multiple factors such as patient categories, feature data, disease course periods, and change characteristics of feature data are comprehensively considered, thereby providing more reliable and detailed support for the monitoring and early warning of sepsis patients, which helps medical staff to take appropriate treatment measures in time and improves the treatment effect and survival rate of patients.

[0053] In some embodiments of the present application, before the real-time feature data of the patient to be monitored is obtained and input into the sepsis monitoring model, the following steps are included: Obtain a plurality of real-time influence factor intervals of the patient to be monitored, and construct a real-time influence factor set of the patient to be monitored; Perform similarity analysis on the real-time influence factor set and the influence factor set of each sepsis patient category to obtain a similarity coefficient of the patient to be monitored and each sepsis patient category; Sort the sepsis patient categories according to the similarity coefficients, and set the first sorted sepsis patient category as the sepsis patient category of the patient to be monitored.

[0054] In this embodiment, the similarity coefficient refers to a comprehensive quantitative value of the similarity degree of the real-time influence factor set of the patient to be monitored and the influence factor set of a certain sepsis patient category in each dimension. When calculating the similarity coefficient, the matching conditions of multiple influence factors are considered, such as patient's basic physical condition indicators, past medical history related factors, and current symptom performance factors. By comparing and analyzing these influence factors one by one, a specific similarity calculation algorithm such as cosine similarity algorithm is used to quantify the similarity degree of each influence factor, and finally a comprehensive similarity coefficient is obtained.

[0055] In this embodiment, the patient to be monitored is classified into a suitable sepsis patient category so as to more accurately monitor and warn it by using the corresponding sepsis monitoring model.

[0056] In some embodiments of the present application, after obtaining the monitoring result and determining whether to generate a warning instruction, the following steps are included: Input the sepsis patient category and real-time feature data of the patient to be monitored into the sepsis monitoring model to obtain a monitoring result, wherein the monitoring result includes a disease course period and a label; When the label is a normal label, no warning instruction is generated; When the label is an abnormal label, a warning instruction of a corresponding level is generated according to the corresponding abnormal level.

[0057] In this embodiment, through real-time identification and prediction by the sepsis monitoring model, the current disease course cycle in which the patient to be monitored is located and the change of the characteristic data can be quickly and accurately judged. If the monitoring result shows that the label is a normal label, it means that the changes of the characteristic data of the patient conform to the normal rules in the corresponding disease course cycle, and the development of the disease is within the controllable range. At this time, the system will not generate an early warning instruction, and the medical staff can continue to observe and treat the patient according to the routine diagnosis and treatment process. When the label is an abnormal label, it means that the changes of the characteristic data of the patient at the current or future period of the current disease course cycle deviate from the normal trend, which may mean that the disease has worsened or abnormally fluctuated. At this time, the system will generate a warning instruction of the corresponding level according to the abnormal level corresponding to the abnormal label.

[0058] In this embodiment, when the label in the monitoring result is an abnormal label, the system will immediately start the early warning mechanism. According to the different abnormal levels, the early warning instruction can be divided into multiple levels, such as mild abnormality, moderate abnormality and severe abnormality, etc. Each level corresponds to different early warning signals and countermeasures. Mild abnormality may only trigger the reminding function of the system, notifying the medical staff to pay attention to the changes of the patient's condition; moderate abnormality will further upgrade to an alarm, prompting the medical staff to check the patient in detail in time; and severe abnormality will directly trigger an emergency warning, requiring the medical staff to take immediate rescue measures to prevent the disease from further worsening. Through this hierarchical early warning mechanism, the system can more accurately reflect the patient's condition, providing timely and effective decision support for medical staff, thereby greatly improving the success rate of treatment of sepsis patients.

[0059] In some embodiments of the present application, a sepsis monitoring system based on multi-modal data is also included: The setting module is configured to determine an influence factor, set a plurality of sepsis patient categories based on the influence factor, and generate a plurality of disease course cycle sequences for each sepsis patient category, wherein each disease course cycle sequence includes a plurality of disease course cycles, and each disease course cycle is mapped with a corresponding multi-modal data set; The analysis module is configured to compare and analyze the multi-modal data sets of the plurality of disease course cycle sequences of the same sepsis patient category to obtain a plurality of characteristic data of each disease course cycle, a change feature of each characteristic data, and a weight coefficient; The construction module is configured to perform neural network training based on the plurality of characteristic data of all disease course cycles of all sepsis patient categories, the change feature of each characteristic data, and the corresponding weight coefficient to obtain a sepsis monitoring model; The early warning module is configured to obtain real-time characteristic data of a patient to be monitored and input the real-time characteristic data into the sepsis monitoring model to obtain a monitoring result and determine whether to generate an early warning instruction.

[0060] The above merely preferred embodiments of the present application, it should be noted that for those of ordinary skill in the art, without departing from the technical principles of the present application, can make several improvements and replacements, these improvements and replacements should also be considered as the protection scope of the present application.

Claims

1. A sepsis monitoring method based on multimodal data, characterized in that, include: The influencing factors are determined, and several sepsis patient categories are set based on the influencing factors. Several disease course cycle sequences are generated for each sepsis patient category. The disease course cycle sequence includes several disease course cycles, and each disease course cycle is mapped to a corresponding multimodal dataset. By comparing and analyzing multimodal datasets of several disease course sequences of patients in the same sepsis patient category, we can obtain several feature data of each disease course, the change characteristics of each feature data, and the weight coefficients. A sepsis monitoring model is obtained by training a neural network based on several characteristic data of all disease courses of all sepsis patient categories, the change characteristics of each characteristic data and the corresponding weight coefficients. The system acquires real-time characteristic data of the patients to be monitored and inputs it into the sepsis monitoring model to obtain monitoring results and determine whether to generate an early warning instruction.

2. The sepsis monitoring method based on multimodal data as described in claim 1, characterized in that, Based on the impact factor, several sepsis patient categories are defined, and several disease course cycle sequences are generated for each sepsis patient category, including: Several impact factors are identified, and each impact factor includes several impact factor intervals. Randomly select an impact factor interval for each impact factor and combine them to obtain an impact factor set; Several sets of influence factors are obtained by combining them sequentially; Each set of influencing factors is set as a sepsis patient category, and several sepsis patient categories are set based on the entire set of influencing factors. Identify several monitored patients for each sepsis patient category; The historical medical logs of each monitored patient are obtained, and the historical medical logs are divided into several pre-set risk coefficient intervals to obtain several medical cycles for each monitored patient. Each disease course cycle is mapped to the corresponding multimodal data of the monitored patients in the historical disease course log and the corresponding historical cycle duration; A disease course sequence was constructed based on several disease course cycles for each monitored patient; Generate several disease course cycle sequences for all monitored patients in each sepsis patient category; Each disease course sequence includes several disease courses, and each disease course is mapped to a corresponding multimodal dataset of monitored patients, which includes data in several dimensions.

3. The sepsis monitoring method based on multimodal data as described in claim 2, characterized in that, Several characteristic data points for each disease course cycle were obtained, including: A disease course cycle is randomly selected as the target disease course cycle; The multimodal dataset of the target disease course is compared with the multimodal datasets of other disease courses in the same disease course sequence to obtain several difference coefficients for each dimension of the multimodal dataset of the target disease course. Pre-set the threshold for the difference coefficient; If all the difference coefficients of the same dimension of data in the target disease course are greater than the preset difference coefficient threshold, the corresponding dimension data will be set as the undetermined feature data of the target disease course. Several undetermined feature data for each disease course are generated sequentially; By comparing the undetermined characteristic data of different monitored patients in the same sepsis patient category at the same disease course, the confidence coefficient of each undetermined characteristic data is obtained; Undefined feature data with a confidence coefficient greater than a preset confidence coefficient threshold are set as feature data for the corresponding disease course cycle. Feature data for each disease course cycle are generated sequentially.

4. The sepsis monitoring method based on multimodal data as described in claim 3, characterized in that, The variation characteristics of each feature data include: Pre-set a preset monitoring time interval for each disease course cycle, and generate several collection nodes for the corresponding disease course cycle according to the preset monitoring time interval; A historical cycle duration for each disease course is used as a time reference line; The feature data of each disease course cycle is collected by each acquisition node and mapped to the corresponding time reference line to obtain the feature data change curve segment of each feature data in the corresponding disease course cycle and the duration of the historical cycle. A comparison curve is constructed by taking all characteristic data change curve segments of the same disease course for the same sepsis patient category and constructing a comparison curve for the corresponding disease course. The comparison curve includes several comparison curve segment groups. Each of the comparison curve segment groups includes curve segments showing the change of characteristic data for the same characteristic data in different historical cycle durations within the same disease course. The variation characteristics of each feature data in the corresponding disease course are determined based on each comparison curve segment group. The variation characteristics include normal variation characteristics and abnormal variation characteristics.

5. The sepsis monitoring method based on multimodal data as described in claim 4, characterized in that, Based on each comparison curve segment group, the change characteristics of the corresponding feature data in the corresponding disease course are determined, including: Cluster analysis is performed on the characteristic data change curve segments in the same comparison curve segment group to obtain several similar curve segment clusters; Several clusters of similar curve segments are compared with the standard curve segments of the corresponding disease course, and the clusters of characteristic curve segments are determined based on the comparison results. Determine the variation characteristics of the characteristic curve segment cluster and each similar curve segment cluster, the variation characteristics including shape, variation trend, data range and time length; Calculate the difference quantification value between each cluster of similar curve segments and the cluster of characteristic curve segments, and set the anomaly level of the change characteristics of the corresponding cluster of similar curve segments based on the difference quantification value; Pre-set a threshold for the quantification of differences; The variation characteristics of the characteristic curve segment clusters and the variation characteristics of similar curve segment clusters with difference quantization values ​​less than the preset difference quantization value threshold are set as the normal variation characteristics of the corresponding feature data in the course of the disease, and matched with normal labels. The change features of similar curve segments whose difference quantification value is not less than the preset difference quantification value threshold are set as the abnormal change features of the corresponding feature data in the course of the disease, and abnormal labels are matched, and each abnormal change feature is mapped to a corresponding abnormal level.

6. The sepsis monitoring method based on multimodal data as described in claim 5, characterized in that, The weighting coefficients include: Calculate the difference of several difference coefficients for the same feature data in the corresponding disease course, and calculate the mean of the difference coefficients for the corresponding feature data in the corresponding disease course. The first weighting coefficient for the corresponding feature data is generated based on the mean of the difference coefficients. Calculate the difference in confidence coefficients of the feature data; Compensation coefficients for corresponding feature data are generated based on the difference in confidence coefficients. The weight coefficients of the corresponding feature data are generated based on the compensation coefficient and the first weight coefficient.

7. The sepsis monitoring method based on multimodal data as described in claim 6, characterized in that, A sepsis surveillance model was obtained, including: The first monitoring model is obtained by using each sepsis patient category, corresponding feature data and weight coefficients of feature data as the first training input data, and the corresponding disease course and credibility as the first training output data to train the neural network. The second monitoring model is obtained by using the change characteristics and weight coefficients of several feature data of each disease cycle as the second training input data and the corresponding labels as the second training output data to train the neural network. A sepsis monitoring model was constructed based on the first and second monitoring models.

8. The sepsis monitoring method based on multimodal data as described in claim 7, characterized in that, Before acquiring real-time characteristic data of the patients to be monitored and inputting it into the sepsis monitoring model, the following steps are included: Several real-time impact factor intervals for the patients to be monitored are obtained, and a set of real-time impact factors for the patients to be monitored is constructed. A similarity analysis was performed between the real-time set of influencing factors and the set of influencing factors for each sepsis patient category to obtain the similarity coefficient between the monitored patients and each sepsis patient category. Patients with sepsis were sorted according to their similarity coefficients, and the patient category with the highest similarity coefficient was selected as the patient category to be monitored.

9. The sepsis monitoring method based on multimodal data as described in claim 8, characterized in that, Obtaining monitoring results and determining whether to generate an early warning instruction includes: The sepsis patient category and real-time characteristic data of the patients to be monitored are input into the sepsis monitoring model to obtain monitoring results, which include disease course and labels; No warning instruction is generated when the label is a normal label; When the label is an abnormal label, an early warning instruction of the corresponding level is generated based on the corresponding abnormality level.

10. A sepsis monitoring system based on multimodal data, characterized in that, include: The module is used to determine the influencing factors, define several sepsis patient categories based on the influencing factors, and generate several disease course cycle sequences for each sepsis patient category. The disease course cycle sequence includes several disease course cycles, and each disease course cycle is mapped to a corresponding multimodal dataset. The analysis module is used to compare and analyze multimodal datasets of several disease course sequences of the same sepsis patient category to obtain several feature data of each disease course, the change characteristics of each feature data, and the weight coefficient. The module is used to train a neural network based on several feature data of all disease cycles of all sepsis patient categories, the change features of each feature data and the corresponding weight coefficients, so as to obtain a sepsis monitoring model. The early warning module is used to acquire real-time characteristic data of the patients to be monitored and input it into the sepsis monitoring model to obtain monitoring results and determine whether to generate an early warning command.

Images