Health index-based sepsis risk prediction method

Abstract

The invention relates to the technical field of artificial intelligence medical aid decision making, and particularly discloses a sepsis risk prediction method based on a health index. The method comprises the following steps: collecting health index data to construct a standardized feature vector, and training a cumulative health index scoring model; constructing an ordered logistic regression model based on the health index, and outputting sepsis grade probability distribution; introducing a teacher-student model to carry out probability transfer learning to obtain a first sepsis risk prediction result; and explaining and analyzing by adopting an SHAP method, and outputting a second prediction result. Compared with the technical problems that in the prior art, sensitivity is low mainly depending on a regular scoring system such as SIRS or SOFA, and particularly risk early warning is difficult to achieve under the condition that early clinical indexes are slightly abnormal, due to the fact that accumulative scoring modeling and ordered logic probability mapping are introduced, sepsis risk level continuous early warning and explanation are achieved, and the risk early warning efficiency is improved. And the early sepsis risk prediction effect is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence medical assistant decision-making, and particularly relates to a sepsis risk prediction method based on a health index. Background Art

[0002] Sepsis is a systemic inflammatory response syndrome caused by infection, characterized by insidious onset, rapid progression, and high mortality. Early identification and accurate risk prediction are of great significance for reducing the fatality rate and guiding interventions. Currently, in clinical practice, regular scoring methods such as SIRS (Systemic Inflammatory Response Syndrome) criteria or SOFA (Sequential Organ Failure Assessment) are often used for sepsis screening and evaluation, but there are still many deficiencies. For example, although the SIRS score is simple to use, the threshold setting is too broad, resulting in high sensitivity but poor specificity, and it is easy to cause over-warning or misjudgment; the SOFA score depends on relatively comprehensive laboratory test data and is not applicable to the situation of missing early admission data or low collection frequency. In addition, traditional scoring mechanisms mostly adopt a rigid segmentation method, which cannot reflect the non-linear influence and risk evolution trend between indicators and is difficult to identify high-risk patients in a marginal state of the disease. In recent years, some studies have introduced deep learning, knowledge graphs, and large language models (such as GPT) for sepsis prediction, improving the model accuracy, but also exposing problems such as strong dependence on training data, complex model structure, and lack of clinical interpretability, making it difficult to be popularized and applied in actual medical scenarios. Therefore, there is an urgent need for a new type of sepsis prediction method with data-driven ability, risk continuous expression ability, and output interpretability, which can still accurately identify and dynamically warn the sepsis risk level in the case of slightly abnormal indicators or atypical symptoms, so as to improve the efficiency of clinical assistant diagnosis and prediction credibility. Summary of the Invention

[0003] Aiming at the above-mentioned existing technical deficiencies, the purpose of the present invention is to propose a sepsis risk prediction method based on a health index, aiming to solve the technical problem that the existing technology mainly relies on regular scoring systems such as SIRS or SOFA, with low sensitivity, especially under the condition of slightly abnormal early clinical indicators, it is difficult to achieve risk warning.

[0004] To solve the above technical problems, the present invention adopts the following technical solutions: The present invention provides a sepsis risk prediction method based on a health index.

[0005] The sepsis risk prediction method based on a health index includes:

[0006] Step S10: Collect the health index data of the measured person within 24 hours, and construct a standardized input feature vector according to the health index data;

[0007] Step S20: Pre-construct and train a cumulative health index scoring model, and input the standardized input feature vector into the trained cumulative health index scoring model to obtain a continuous health index ;

[0008] Step S30: Process the continuous health index using an improved ordered logistic regression method to obtain the probability distribution of sepsis grades;

[0009] Step S40: Introduce a teacher-student model, and perform knowledge transfer supervised learning according to the probability distribution of sepsis grades combined with the teacher-student model to obtain the first sepsis risk prediction result;

[0010] Step S50: Based on the SHAP method, interpret and analyze the first sepsis risk prediction result to obtain the second sepsis risk prediction result.

[0011] Preferably, in step S10, the steps of collecting the health index data of the subject within 24 hours and constructing a standardized input feature vector according to the health index data specifically include:

[0012] Step S101: Collect the health index data of the subject within 24 hours, where the health index data includes heart rate HR, respiratory rate RR, body temperature T, systolic blood pressure SBP, blood oxygen saturation SpO2, lactate concentration Lac, serum creatinine Scr, and platelet count PLT;

[0013] Step S102: Perform standard deviation normalization processing on the health index data to obtain standardized health index data;

[0014] Step S103: Construct a standard input feature vector according to the standardized health index data.

[0015] Preferably, in step S20, the steps of pre-constructing and training a cumulative health index scoring model, and inputting the standardized input feature vector into the trained cumulative health index scoring model to obtain a continuous health index specifically include:

[0016] Step S201: Obtain the historical electronic medical record sample data of the subject, extract the diagnostic label data and disease condition scoring data from the electronic medical record sample data, and divide multi-level health grade labels according to the diagnostic label data and disease condition scoring data, including Grade 1: Clinically stable, Grade 2: Moderate warning, Grade 3: High warning;

[0017] Step S202: Construct a historical standardized input feature vector based on historical electronic medical record sample data. Use the historical standardized input feature vector as the input of the cumulative health index scoring model, and use multi-level health level labels as the output of the cumulative health index scoring model to train the cumulative health index scoring model;

[0018] Among them, during the training process, minimize the multi-level cross-entropy as the loss function of the cumulative health index scoring model. The loss function formula of the cumulative health index scoring model is: ;

[0019] Among them, L is the loss function of the cumulative health index scoring model, is the total number of training samples; is the number of categories of health level labels; is the label indicator variable indicating whether the sample j belongs to the k-th category; is the standardized input feature vector of the j-th sample; is the probability that the model predicts that the sample j belongs to the k-th category;

[0020] Step S203: Input the standardized input feature vector into the trained cumulative health index scoring model to obtain a continuous multi-level health level label output. Perform linear processing and smoothing processing on the continuous multi-level health level label output to obtain a continuous health index .

[0021] Preferably, in step S30, for the continuous health index adopt the improved ordered logistic regression method to process the steps of obtaining the sepsis level probability distribution, specifically including:

[0022] Step S301: Divide the sepsis risk into M levels in advance and set M - 1 risk level threshold parameters. The risk level threshold parameters are used to define the conversion interval between the health index and each level of risk;

[0023] Step S302: Adopt the improved ordered logistic regression method to process the continuous health index to obtain the sepsis level probability distribution. The probability in the sepsis level probability distribution is expressed as: ; ; ;

[0024] Among them, ( ) is the Sigmoid function, is the probability of belonging to the first-level sepsis risk when the health index is H, The probability of belonging to the sepsis risk of the nd level when the health index is H, The probability of belonging to the sepsis risk of the th level when the health index is H, is the preset first classification threshold, is the preset mth classification threshold, is the preset (M - 1)th classification threshold.

[0025] Preferably, in step S40, introducing a teacher - student model, and performing knowledge transfer supervised learning according to the sepsis level probability distribution in combination with the teacher - student model to obtain the first sepsis risk prediction result specifically includes: introducing a teacher - student model, the teacher - student model includes a sub - teacher model and a sub - student model, using the sepsis level probability distribution obtained in step S30 as the sub - teacher model to guide the sub - student model to learn the mapping from the standardized input feature vector to the sepsis level probability distribution, obtaining the trained sub - student model, and the trained sub - student model outputs the first sepsis risk prediction result.

[0026] Preferably, in step S40, during the training process, the loss function used is a joint loss function including the KL divergence between distributions, and the joint loss function: , where, is the joint loss function, is the weight balance coefficient between the cross - entropy term and the KL divergence term in the loss function, is the number of categories of the health level label, is the one - hot encoding indicator variable of the true label, is the probability label of the student model predicting the kth class, is the standardized input feature vector of the current sample, is the probability label of the kth class output by the teacher model.

[0027] Preferably, in step S50, based on the SHAP method, performing interpretive analysis on the first sepsis risk prediction result to obtain the second sepsis risk prediction result specifically includes: obtaining the first sepsis risk prediction result, introducing the SHAP method to perform interpretability analysis on the prediction process of the model, including: quantifying the positive and negative impacts of each input feature in the standardized input feature vector on the first sepsis risk prediction result to obtain the feature contribution result; correcting the parameters in the teacher - student model according to the feature contribution result, and outputting the second sepsis risk prediction result.

[0028] The present invention also provides a sepsis risk prediction system based on a health index, including:

[0029] A health index collection and standardization module, which is used to collect the health index data of the subject within 24 hours and construct a standardized input feature vector according to the health index data;

[0030] A health index score construction module, which is used to pre-construct and train an accumulative health index score model, and input the standardized input feature vector into the trained accumulative health index score model to obtain a continuous health index ;

[0031] A sepsis grade probability modeling module, which is used to process the continuous health index by using an improved ordered logistic regression method to obtain a sepsis grade probability distribution;

[0032] A teacher-student modeling and risk prediction module, which is used to introduce a teacher-student model, and perform knowledge transfer supervised learning according to the sepsis grade probability distribution combined with the teacher-student model to obtain a first sepsis risk prediction result;

[0033] A risk interpretation and result correction module, which is used to perform interpretation and analysis on the first sepsis risk prediction result based on the SHAP method to obtain a second sepsis risk prediction result.

[0034] The present invention also provides a sepsis risk prediction device based on a health index, including: a memory, a processor, and a sepsis risk prediction program based on a health index stored on the memory and executable on the processor. When the sepsis risk prediction program based on a health index is executed by the processor, it implements the sepsis risk prediction method based on a health index.

[0035] The present invention also provides a computer program product, including a sepsis risk prediction program based on a health index. When the sepsis risk prediction program based on a health index is executed by a processor, it implements the sepsis risk prediction method based on a health index described above.

[0036] The beneficial effects of the present invention are as follows: Compared with the prior art, which mainly relies on rule-based scoring systems such as SIRS or SOFA and has low sensitivity, especially under the condition of slightly abnormal early clinical indicators, it is difficult to achieve risk warning. Since the present application introduces accumulative scoring modeling and ordered logical probability mapping, it realizes continuous warning and interpretation of sepsis risk levels, and improves the early sepsis risk prediction effect. Description of the Drawings

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0038] Figure 1 It is a schematic flowchart of a sepsis risk prediction method based on a health index according to the present invention.

[0039] Figure 2 It is a schematic structural diagram of a sepsis risk prediction device based on a health index according to the present invention. Detailed implementation manners

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0041] Embodiment 1: As Figure 1 shown, it is a schematic flowchart of the sepsis risk prediction method based on a health index according to the present invention, and Embodiment 1 of the sepsis risk prediction method based on a health index according to the present invention is proposed.

[0042] In Embodiment 1, the sepsis risk prediction method based on a health index includes:

[0043] Step S10: Collect the health index data of the measured person within 24 hours, and construct a standardized input feature vector according to the health index data;

[0044] It should be noted that in Step S10, the steps of collecting the health index data of the measured person within 24 hours and constructing a standardized input feature vector according to the health index data specifically include:

[0045] Step S101: Collect the health index data of the measured person within 24 hours. The health index data includes heart rate HR, respiratory rate RR, body temperature T, systolic blood pressure SBP, blood oxygen saturation SpO2, lactate concentration Lac, serum creatinine Scr, and platelet count PLT;

[0046] Step S102: Perform standard deviation normalization processing on the health index data to obtain standardized health index data;

[0047] Step S103: Construct a standard input feature vector according to the standardized health index data.

[0048] It can be understood that this step is to convert multi-source health indicator data of different dimensions and different value ranges into a unified standard data vector format, so as to ensure the data comparability and convergence stability in the subsequent modeling stage. Through feature standardization, the model can be prevented from being dominated by certain indicators with large values ​​(such as Scr and SBP), and the model's perception of weak pathological signals can be enhanced, which is especially suitable for the actual needs of early identification of sepsis.

[0049] It should be understood that in clinical practice, the frequency and time of collecting health indicators may fluctuate. Therefore, this embodiment uses the latest or average value within 24 hours as the input source, taking into account data integrity and response timeliness. At the same time, in the process of feature standardization, it is preferred to calculate the mean and standard deviation based on historical large sample statistical data to avoid the problem of normalization imbalance in scenarios with very few samples. This processing method improves the adaptability to heterogeneous data sources and cross-department deployment.

[0050] For example, taking 10 suspected sepsis patients actually collected from an ICU ward as an example, the original lactate concentration ranged from 1.8 to 4.7 mmol / L. If it was directly input into the scoring model, it might lead to high numerical features dominating the output results. After standard deviation normalization, all indicators were compressed to a relatively symmetrical distribution interval, the model training accuracy was improved by about 12.5%, and the recognition rate of marginal cases was increased to 86.2%. This shows that the standardization step has a significant optimization effect in actual data modeling.

[0051] Step S20: Pre-construct and train a cumulative health index scoring model, input the standardized input feature vector into the trained cumulative health index scoring model, and obtain a continuous health index ;

[0052] It should be noted that in step S20, a cumulative health index scoring model is pre-built and trained, and the standardized input feature vector is input into the trained cumulative health index scoring model to obtain a continuous health index. The steps include:

[0053] Step S201: Obtain the historical electronic medical record sample data of the subject, extract the diagnostic label data and the condition score data from the electronic medical record sample data, and divide the multi-level health level labels according to the diagnostic label data and the condition score data, including level 1: clinical stability, level 2: moderate warning, and level 3: high warning;

[0054] Step S202: Construct a historical standardized input feature vector based on historical electronic medical record sample data. Use the historical standardized input feature vector as the input of the cumulative health index scoring model, and use multi-level health level labels as the output of the cumulative health index scoring model to train the cumulative health index scoring model;

[0055] Among them, during the training process, minimize the multi-level cross-entropy as the loss function of the cumulative health index scoring model. The loss function formula of the cumulative health index scoring model is: ;

[0056] Among them, L is the loss function of the cumulative health index scoring model, is the total number of training samples; is the number of categories of health level labels; is the label indicator variable indicating whether sample j belongs to the k-th category; is the standardized input feature vector of the j-th sample; is the probability that the model predicts that sample j belongs to the k-th category;

[0057] Step S203: Input the standardized input feature vector into the trained cumulative health index scoring model to obtain a continuous multi-level health level label output. Perform linear processing and smoothing processing on the continuous multi-level health level label output to obtain a continuous health index .

[0058] It can be understood that most traditional sepsis risk modeling methods rely on rule scoring or binary classification models, and it is difficult to express the "grading transition process" of the health status of the measured person. By introducing a cumulative model in this step, the relationship between each level is no longer independent, but has monotonic orderliness, which helps the model to automatically identify "critical samples between levels" during training. Especially when the health status fluctuates critically, this method can accurately capture the state evolution trend through continuous probability output, avoiding misclassification and jump problems caused by traditional hard classification.

[0059] It should be understood that the cumulative health index scoring model does not output a single label, but uses the cumulative probability of each health level as the modeling goal, thereby embedding the natural orderliness of health levels in the training process, which helps to optimize the level confusion common in multi-level prediction problems. In addition, the finally generated continuous health index is a weighted fusion of the probability results of multiple levels, with stronger fitting ability and generalization performance, and is also convenient for smooth connection with the subsequent ordinal regression prediction module.

[0060] For example, the cumulative health index scoring model trained based on the real medical record data of a hospital's ICU covers about 1,600 inpatients in the past three years, of which about 32% were diagnosed with sepsis. When using the model to predict newly admitted patients, the continuous health index output by the model is highly positively correlated with the clinical SOFA score (Pearson coefficient is 0.81), and for patients who do not meet the SIRS diagnostic criteria but have shown weak abnormal signals such as mild lactate increase and SpO2 decrease, the model predicts a health index score higher than 0.72. Subsequent follow-up found that about 68% of such patients developed sepsis within 48 hours, indicating that this method has a stronger ability to identify marginal risk groups in advance.

[0061] Step S30: Continuous health index The modified ordered logistic regression method was used to obtain the probability distribution of sepsis grade;

[0062] It should be noted that in step S30, the continuous health index The modified ordered logistic regression method was used to obtain the probability distribution of sepsis grade, including:

[0063] Step S301: pre-dividing sepsis risk into M levels, and setting M-1 risk level threshold parameters, where the risk level threshold parameters are used to define the conversion interval between the health index and each level of risk;

[0064] Step S302: Continuous health index The improved ordered logistic regression method was used to obtain the probability distribution of sepsis grade. The probability in the probability distribution of sepsis grade was expressed as: ; ; ;

[0065] in, ( ) is the Sigmoid function, is the probability of belonging to the first level of sepsis risk when the health index is H, When the health index is H, it belongs to The probability of grade 1 sepsis risk, When the health index is H, it belongs to The probability of grade 1 sepsis risk, is the preset first classification threshold, is the preset mth classification threshold, is the preset M-1th classification threshold.

[0066] It can be understood that different from the traditional classification model that regards the sepsis level as an independent label, the improved ordered logistic regression method adopted in this embodiment embeds a prior structure of "the risk levels have strict orderliness". By setting the threshold of the continuous health index, the model mathematically constructs a gradually increasing probability structure, which can express the risk gradual change process of patients in different health states. This method not only improves the medical rationality of the prediction output, but also avoids the common "level disorder" problem in the classification model, making the prediction results more in line with the clinical use logic.

[0067] It should be understood that the improved ordered logistic regression model introduced in this embodiment can adaptively learn the conversion thresholds between different risk levels, rather than relying on fixed segmentation rules. This design enables the model to be flexibly adjusted according to the data distribution characteristics of different populations or different institutions, thereby improving the adaptability and stability in the actual deployment process. At the same time, by outputting the complete level probability distribution instead of a single prediction result, the model can provide a richer risk basis for subsequent decision-making, realizing the transformation from "deterministic" prediction to "inferential" prediction.

[0068] For example, in an experiment using the data of 1,200 patients, the sepsis risk is divided into three levels: non-sepsis, suspected sepsis, and confirmed sepsis. After modeling with the improved ordered logistic regression model, the probabilities corresponding to the three levels can be output. For example, for patient A with a health index of 0.66, the risk predictions are "non-sepsis" 18.2%, "suspected sepsis" 49.7%, and "confirmed sepsis" 32.1%. In this sample set, compared with the traditional multi-classification model, the accuracy of this method in identifying critical samples is improved by about 14.8%, and the prediction misalignment between medium and high levels is significantly reduced, verifying the effectiveness and refined expression ability of the level probability modeling mechanism.

[0069] Step S40: Introduce a teacher-student model, and perform knowledge transfer supervised learning according to the sepsis level probability distribution in combination with the teacher-student model to obtain the first sepsis risk prediction result;

[0070] It should be noted that in step S40, the step of introducing a teacher-student model and performing knowledge transfer supervised learning according to the sepsis level probability distribution in combination with the teacher-student model to obtain the first sepsis risk prediction result specifically includes: introducing a teacher-student model, where the teacher-student model includes a sub-teacher model and a sub-student model, using the sepsis level probability distribution obtained in step S30 as the sub-teacher model to guide the sub-student model to learn the mapping from the standardized input feature vector to the sepsis level probability distribution, obtaining the trained sub-student model, and the trained sub-student model outputs the first sepsis risk prediction result.

[0071] In step S40, during the training process, the loss function adopted is a joint loss function including the KL divergence between distributions. The joint loss function: , where is the joint loss function, is the weight balance coefficient between the cross-entropy term and the KL divergence term in the loss function, is the number of categories of the health level labels, is the one-hot encoded indicator variable of the true label, is the probability label predicted by the student model for the k-th category, is the standardized input feature vector of the current sample, is the probability label of the k-th category output by the teacher model.

[0072] It can be understood that the purpose of introducing the teacher-student model in this step is to solve the problems of scarce sepsis sample labels and high annotation costs. By using the sepsis grade probability distribution calculated in the previous step as the output of the teacher model, the structural information of sepsis risk evolution can be transmitted to the student model without relying on a large number of artificial labels. The student model not only learns the mapping from features to risk levels but also inherits the probability expression method of the teacher model, so that it still has strong risk recognition ability in samples with incomplete early annotations or samples that have not been diagnosed by clinical interventions.

[0073] It should be understood that in the teacher-student structure, the sub-teacher model is calculated based on the health index and has good medical interpretability; the sub-student model is the core structure finally used for prediction deployment and has strong generalization ability. By introducing knowledge transfer, when the student model is optimized, it not only fits the true label but also actively approaches the output probability of the teacher model, thus realizing the "knowledge alignment" between model structures. This mechanism enables the model to reproduce the grading judgment logic of the teacher model even when only receiving the original health index input during actual deployment, thereby improving the prediction consistency and credibility under non-ideal label conditions.

[0074] For example, in the experiment of training the model with a mixture of partially labeled and partially unlabeled samples, the sepsis probability distribution generated in step S30 is used as the teacher output to guide the joint training of the student model. Compared with the traditional supervised model that only uses true labels, the prediction stability of this method on the validation set is improved by about 19%, and it can still maintain a high recognition accuracy in the case of missing true labels. At the same time, the student model can predict medium and high-grade probabilities earlier in samples with mild abnormalities in multiple indicator combinations but not meeting the traditional diagnostic criteria, and significantly reduce the false negative risk, proving the effective supplement of the "soft knowledge" transmitted by the teacher model to the model's prediction ability.

[0075] Step S50: Based on the SHAP method, interpret and analyze the first sepsis risk prediction result to obtain the second sepsis risk prediction result.

[0076] It should be noted that in step S50, the step of interpreting and analyzing the first sepsis risk prediction result based on the SHAP method to obtain the second sepsis risk prediction result specifically includes: obtaining the first sepsis risk prediction result, introducing the SHAP method to perform interpretability analysis on the prediction process of the model, including: quantifying the positive and negative impacts of each input feature in the standardized input feature vector on the first sepsis risk prediction result to obtain the feature contribution result; correcting the parameters in the teacher-student model according to the feature contribution result, and outputting the second sepsis risk prediction result.

[0077] It can be understood that in this step, by introducing the SHAP method, the model is no longer a "black box" output, but has the ability to explain the role played by each input feature in the current prediction. This feature contribution analysis can not only be used for visualizing the model judgment process, but also for identifying possible "bias sources" or "wrong focus points" in the model. In the case where the prediction result is inconsistent with clinical experience, the SHAP method can be used for reverse analysis of the model's judgment basis, and appropriate correction can be made to the model response or result accordingly.

[0078] It should be understood that the advantage of the SHAP method lies in its "local interpretability" and "global consistency", that is, it can not only make causal explanations for the predictions of each specific sample, but also identify the risk features that the model generally focuses on among multiple samples. For example, for key features such as increased lactic acid concentration and increased respiratory rate, the SHAP value can reflect whether they consistently push the model towards high-risk predictions in most patients. If it is found that the model has a large positive contribution to some features that are not significantly meaningful in medicine, the structure or parameters of the teacher-student model can be fine-tuned accordingly, so as to form a more interpretable risk prediction result in the second stage.

[0079] For example, in the process of performing SHAP analysis on a group of clinical samples, the model outputs a first prediction result of "moderate sepsis risk" for a certain patient, but the SHAP interpretation shows that the feature with the greatest contribution to the risk prediction is "normal body temperature", while the contributions to highly sensitive features such as "increased lactic acid concentration" and "thrombocytopenia" are relatively small. Combining with clinical judgment, this prediction is obviously biased. By referring to the SHAP value to fine-tune the model weights, the second prediction result is adjusted to "high risk", which is consistent with the result of the patient being diagnosed with sepsis within the subsequent 48 hours. This process effectively verifies the auxiliary role of the SHAP method in model risk correction, and significantly improves the practical usability and interpretability of the model prediction.

[0080] Embodiment 2: In addition, a sepsis risk prediction system based on a health index provided by the present invention adopts a sepsis risk prediction method based on a health index in the above embodiment, and can solve the technical problem of sepsis risk prediction based on a health index. Compared with the prior art, the beneficial effects of a sepsis risk prediction system based on a health index provided by the present invention are the same as those of a sepsis risk prediction method based on a health index provided in the above embodiment, and other technical features in the sepsis risk prediction system based on a health index are the same as the features disclosed in the method of the above embodiment, and will not be elaborated here.

[0081] Embodiment 3: The present invention provides a sepsis risk prediction device based on a health index. Please refer to Figure 2, A sepsis risk prediction device based on a health index includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute a sepsis risk prediction method based on a health index in the first embodiment above. The sepsis risk prediction device in the embodiments of the present invention may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description: tablet computers), PMPs (Portable Media Players), vehicle terminals (such as vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. A sepsis risk prediction device based on a health index is merely an example and should not impose any limitations on the functions and scope of use of the embodiments of the present invention. A sepsis risk prediction device may include a processor 1001 (such as a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM: Random Access Memory) 1004. In the RAM 1004, various programs and data required for the operation of a sepsis risk prediction device based on a health index are also stored. The processor 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow a sepsis risk prediction device based on a health index to communicate with other devices wirelessly or wiredly to exchange data. Although the figure shows a sepsis risk prediction device having various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be alternatively implemented or had.

[0082] Embodiment 4: The present invention further provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of a sepsis risk prediction method based on a health index as described above are implemented. The computer program product provided by the present invention can solve the technical problem of sepsis risk prediction based on a health index. Compared with the prior art, the beneficial effects of the computer program product provided by the present invention are the same as those of the sepsis risk prediction method based on a health index provided in the above embodiment, and will not be elaborated here.

[0083] Specifically, according to the embodiments disclosed by the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed by the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program codes for executing the methods shown in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by a processor 1001, the above functions defined in the methods of the embodiments disclosed by the present invention are executed.

[0084] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in a suitable manner in any one or more embodiments or examples.

[0085] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these modifications and variations.

Claims

1. A sepsis risk prediction method based on health index, characterized in that: Methods include: Step S10: collecting the health index data of the subject within 24 hours, and constructing a standardized input feature vector according to the health index data; Step S20: Pre-construct and train a cumulative health index scoring model, input the standardized input feature vector into the trained cumulative health index scoring model, and obtain a continuous health index ; Step S30: Continuous health index The modified ordered logistic regression method was used to obtain the probability distribution of sepsis grade; Step S40: introducing a teacher-student model, performing knowledge transfer supervised learning based on the probability distribution of sepsis levels in combination with the teacher-student model, and obtaining a first sepsis risk prediction result; Step S50: interpreting and analyzing the first sepsis risk prediction result based on the SHAP method to obtain a second sepsis risk prediction result.

2. A method for predicting sepsis risk based on health index according to claim 1, characterized in that: In step S10, the steps of collecting the health index data of the subject within 24 hours and constructing a standardized input feature vector according to the health index data specifically include: Step S101: collecting the health index data of the subject within 24 hours, the health index data including heart rate HR, respiratory rate RR, body temperature T, systolic blood pressure SBP, blood oxygen saturation SpO2, lactate concentration Lac, serum creatinine Scr, and platelet count PLT; Step S102: performing standard deviation normalization processing on the health index data to obtain standardized health index data; Step S103: constructing a standard input feature vector based on the standardized health indicator data.

3. The method for predicting sepsis risk based on health index according to claim 1, characterized in that: In step S20, a cumulative health index scoring model is pre-built and trained, and the standardized input feature vector is input into the trained cumulative health index scoring model to obtain a continuous health index The steps include: Step S201: Obtain the historical electronic medical record sample data of the subject, extract the diagnostic label data and the condition score data from the electronic medical record sample data, and divide the multi-level health level labels according to the diagnostic label data and the condition score data, including level 1: clinical stability, level 2: moderate warning, and level 3: high warning; Step S202: construct a historical standardized input feature vector based on historical electronic medical record sample data, use the historical standardized input feature vector as the input of the cumulative health index scoring model, use the multi-level health grade label as the output of the cumulative health index scoring model, and train the cumulative health index scoring model; Among them, during the training process, minimizing the multi-level cross entropy is used as the loss function of the cumulative health index scoring model. The loss function formula of the cumulative health index scoring model is: ； Among them, L is the loss function of the cumulative health index scoring model, is the total number of training samples; is the number of categories of health grade labels; Is the label indicator variable for whether sample j belongs to the kth category; is the standardized input feature vector of the jth sample; The model predicts the probability that sample j belongs to the kth class; Step S203: Input the standardized input feature vector into the trained cumulative health index scoring model to obtain a continuous multi-level health grade label output, perform linear processing and smoothing processing on the continuous multi-level health grade label output, and obtain a continuous health index .

4. The method for predicting sepsis risk based on health index according to claim 1, characterized in that: In step S30, for the continuous health index The modified ordered logistic regression method was used to obtain the probability distribution of sepsis grade, including: Step S301: pre-dividing sepsis risk into M levels, and setting M-1 risk level threshold parameters, where the risk level threshold parameters are used to define the conversion interval between the health index and each level of risk; Step S302: Continuous health index The improved ordered logistic regression method was used to obtain the probability distribution of sepsis grade. The probability in the probability distribution of sepsis grade was expressed as: ; ； ； in, ( ) is the Sigmoid function, is the probability of belonging to the first level of sepsis risk when the health index is H, When the health index is H, it belongs to The probability of grade 1 sepsis risk, When the health index is H, it belongs to The probability of grade 1 sepsis risk, is the preset first classification threshold, is the preset mth classification threshold, is the preset M-1th classification threshold.

5. The method for predicting sepsis risk based on health index according to claim 1, characterized in that: In step S40, a teacher-student model is introduced, and knowledge transfer supervised learning is performed in combination with the teacher-student model according to the sepsis grade probability distribution to obtain a first sepsis risk prediction result. The steps specifically include: introducing a teacher-student model, the teacher-student model includes a sub-teacher model and a sub-student model, using the sepsis grade probability distribution obtained in step S30 as a sub-teacher model to guide the sub-student model to learn the mapping from the standardized input feature vector to the sepsis grade probability distribution, and obtaining a trained sub-student model, and the trained sub-student model outputs the first sepsis risk prediction result.

6. A method for predicting sepsis risk based on health index according to claim 5, characterized in that: In step S40, during the training process, the loss function used is a joint loss function including the KL divergence between distributions, and the joint loss function is: ,in, is the joint loss function, is the weight balance coefficient between the cross entropy term and the KL divergence term in the loss function, is the number of categories of health grade labels, is the one-hot encoded indicator variable of the true label, Predict the probability label of the kth class for the student model, is the standardized input feature vector of the current sample, The probability label of the kth class output by the teacher model.

7. The method for predicting sepsis risk based on health index according to claim 1, characterized in that: In step S50, the first sepsis risk prediction result is interpreted and analyzed based on the SHAP method to obtain the second sepsis risk prediction result, specifically including: obtaining the first sepsis risk prediction result, introducing the SHAP method to perform an interpretable analysis on the prediction process of the model, including: quantifying the positive and negative effects of each input feature in the standardized input feature vector on the first sepsis risk prediction result, and obtaining a feature contribution result; correcting the parameters in the teacher-student model according to the feature contribution result, and outputting the second sepsis risk prediction result.

8. A sepsis risk prediction system based on health index, applied to a sepsis risk prediction method based on health index according to any one of claims 1 to 7, characterized in that: The sepsis risk prediction system based on health index includes: The health index collection and standardization module is used to collect the health index data of the subject within 24 hours and construct a standardized input feature vector based on the health index data; The health index scoring construction module is used to pre-build and train the cumulative health index scoring model, input the standardized input feature vector into the trained cumulative health index scoring model, and obtain the continuous health index ; Sepsis grade probabilistic modeling module for continuous health index The modified ordered logistic regression method was used to obtain the probability distribution of sepsis grade; The teacher-student modeling and risk prediction module is used to introduce the teacher-student model, perform knowledge transfer supervised learning based on the probability distribution of sepsis levels combined with the teacher-student model, and obtain the first sepsis risk prediction result; The risk interpretation and result correction module is used to interpret and analyze the first sepsis risk prediction result based on the SHAP method to obtain the second sepsis risk prediction result.

9. A sepsis risk prediction device based on health index, characterized in that: The sepsis risk prediction device based on the health index comprises: a memory, a processor, and a sepsis risk prediction program based on the health index stored in the memory and executable on the processor, wherein the sepsis risk prediction program based on the health index, when executed by the processor, implements a sepsis risk prediction method based on the health index according to any one of claims 1 to 7.

10. A computer program product, characterized in that The computer program product comprises a sepsis risk prediction program based on a health index, and when the sepsis risk prediction program based on a health index is executed by a processor, a sepsis risk prediction method based on a health index according to any one of claims 1 to 7 is implemented.

Images