A method for constructing a medical model, system and storage medium thereof

Abstract

A method for constructing a medical model and a medical model are disclosed. The medical model is configured to predict a diagnosis result based on information of target patients, and the method comprising: acquiring a first data set of a medical institution, wherein the first data set comprises a first structured data and a first unstructured data, and the first structured data and the first unstructured data are both in text form; preprocessing the first data set into a second data set, the second data set comprises a second structured data and a second unstructured data; obtaining a first type of data based on the second structured data, and processing the second unstructured data by a bidirectional long short-term memory network to obtain a second type of data, and extracting language features from the second unstructured data to obtain a third type of data; fusing the first type of data, the second type of data, and the third type of data, and processing the fused result using a softmax to obtain the diagnosis result. The medical model can predict patients' outcomes to identify potential critical patients.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more particularly to a method for constructing a medical model and the medical model itself. Prior Technology

[0002] In some scenarios, when a patient enters a hospital, a nurse-led triage process is first conducted to make an initial assessment of the patient's condition before transferring them to a doctor for further examination. Due to limited medical resources, if a patient is marked as a non-critical patient by a nurse, they may not receive a timely diagnosis from a doctor; conversely, patients marked as critical patients by nurses are more likely to receive priority for diagnosis from a doctor.

[0003] However, nurse-led triage can sometimes fail to accurately identify potential critical patients. For example, when a potentially critical patient is marked as non-critical by a nurse, their condition may worsen or even lead to death due to prolonged waiting. Therefore, there is an urgent need for effective technologies to identify potential critical patients who are marked as non-critical during nurse-led triage, thereby improving the patient experience.

[0004] Emergency departments (EDs) worldwide are frequently overcrowded, potentially delaying essential medical care. Given the critical sensitivity of critical patient outcomes to timely intervention, early identification of patients with deteriorating conditions is crucial. Identifying patients requiring hospitalization and appropriate medical care remains a challenge for ED physicians. Clinical judgment, relying on laboratory results and prior physician experience, can be influenced by individual factors, increasing morbidity and mortality. Among outpatients, 2–5% will return to the ED within 72 hours. Of these, 19%–30% will be hospitalized, and one-third will experience adverse events. Unnecessary testing and treatment during a second ED visit increase costs and prolong hospital stays.

[0005] Therefore, establishing a clinical decision support system that incorporates medical models to predict the treatment of emergency patients can help reduce human error. This system can also reduce length of stay (LOS), improve the quality of healthcare, and optimize the allocation of medical resources. Summary of the Invention

[0006] In view of this, the present invention discloses a method for constructing a medical model, a medical system including the medical model, and a storage medium storing the medical model. The medical model is used to predict the patient's condition or diagnosis based on the patient's information, thereby identifying potential critically ill patients and reducing the probability that the condition of potential critically ill patients will deteriorate due to long waiting times.

[0007] In a first aspect, the present invention discloses a method for constructing a medical model, the medical model being configured to predict diagnostic results based on information about a target patient, the method comprising: (A) acquiring a first dataset from a medical institution, wherein the first dataset includes a first structured data and a first unstructured data; (B) preprocessing the first dataset into a second dataset, wherein the second dataset includes a second structured data and a second unstructured data; (C) obtaining a first type of data based on the second structured data, processing the second unstructured data using a bidirectional long short-term memory network to obtain a second type of data, and extracting linguistic features from the second unstructured data to obtain a third type of data; and (D) fusing the first type of data, the second type of data, and the third type of data into a fusion result, and processing the fusion result using a normalized exponential function to obtain the diagnostic result.

[0008] In some specific embodiments, step (B) further includes: (B1) grouping the first structured data based on a preset index and performing a normalization process on the first structured data; (B2) optimizing the first unstructured data so that the first unstructured data conforms to the rules of the medical model, wherein the optimization includes at least one step of data cleaning, text translation, or text correction, wherein the medical model uses a natural language processing algorithm to process the first unstructured data, and the natural language processing algorithm is configured to process messages in a single language.

[0009] In some specific embodiments, the reference values ​​for grouping include age, mean arterial pressure, or body temperature.

[0010] In some specific embodiments, step (B1) further includes performing one-hot encoding on the grouped first structured data and normalizing the respiratory rate, pulse rate and blood oxygen saturation in the first structured data using a min-max normalization algorithm.

[0011] In some embodiments, data cleaning further includes: converting abbreviated text in the first unstructured data into terms in their full expanded form. Data cleaning further includes: retaining a preset punctuation mark in the first unstructured data and deleting stop words and punctuation marks other than the preset punctuation mark, which is a commonly used punctuation mark in the medical field. Data cleaning further includes: retaining negative words in the first unstructured data. Text translation further includes: translating non-English text in the first unstructured data into English text. Text correction further includes: correcting spelling errors in the first unstructured data.

[0012] In some specific embodiments, the preset punctuation marks include: slash, plus sign, minus sign, and space character.

[0013] In some specific embodiments, the text translation further includes: the text translation includes translating the non-English text in the first unstructured data into English text.

[0014] In some specific embodiments, the text correction further includes: the text correction includes correcting spelling errors in the first unstructured data.

[0015] In some specific embodiments, step (A) includes filtering the first dataset to exclude data of non-target patients present in the first dataset, wherein the data of non-target patients includes data of minors, data of people who have suffered out-of-hospital cardiac arrest, data of people who have abandoned medical advice, data of people who have left without notice, data of people who have withdrawn their registration, data of people who have been transferred to other hospitals, and data of people with incomplete data.

[0016] In some specific embodiments, the diagnostic result includes a critical result or a non-critical result, wherein the critical result includes in-hospital cardiac arrest or admission to the intensive care unit, and the non-critical result includes discharge or admission to a ward.

[0017] In some specific embodiments, the second structured data includes a vital sign and a demographic data point; the second unstructured data includes a chief complaint, a present medical history, and a past medical history; the second type of data includes a first sub-data and a second sub-data; and the third type of data includes a third sub-data, a fourth sub-data, and a fifth sub-data. The first sub-data is obtained from the chief complaint through a bidirectional long short-term memory network; the second sub-data is obtained from the present medical history through a bidirectional long short-term memory network; the third sub-data is obtained by extracting language features from the chief complaint; the fourth sub-data is obtained by extracting language features from the past medical history; and the fifth sub-data is obtained by extracting language features from the present medical history.

[0018] In some specific embodiments, the first sub-data is related to the life characteristics and the demographic data.

[0019] In some specific embodiments, the dimension of the data related to the chief complaint and the past medical history in the second type of data and the third type of data is 100 to 500.

[0020] In some specific embodiments, both the first sub-data and the second sub-data have a dimension of 300.

[0021] In some specific embodiments, the dimensions of both the third sub-data and the fifth sub-data are 200.

[0022] In some specific embodiments, the dimension of the fourth sub-data is 30.

[0023] In some specific embodiments, the dimension of the first type of data is 15.

[0024] In some specific embodiments, the fusion result has 64 dimensions.

[0025] In some specific embodiments, the first type of data, the second type of data, and the third type of data are fused into the fusion result, and the fusion result is processed using the normalized exponential function, with the calculation formula as follows:

[0026] Further, F represents the result of the fusion, DF represents a feature in the first type of data related to the demographic data, VF represents a feature in the first type of data related to the life representation, LF represents the third type of data, and LSF represents the second type of data. Here, is an approximate probability value, M is a parameter matrix of the connection layer, and b is the bias.

[0027] In some specific embodiments, the step "and extracting language features from the second unstructured data to obtain the third type of data" further includes: when the second unstructured data contains keywords, extracting keywords from the second unstructured data, and obtaining the third type of data based on each keyword and its corresponding weight; and when the second unstructured data does not contain keywords, calculating an average value of all single-word vectors in the second unstructured data, calculating the cosine similarity between the average value and all preset keyword vectors, and selecting the five nearest preset keyword vectors as the third type of data.

[0028] In some specific embodiments, the weights are calculated using the following formula: Wherein, HR represents high-risk patients, ¬HR represents non-high-risk patients, w represents the keyword for high risk, k represents the number of clinical narratives of high-risk cases containing w, l represents the number of clinical narratives of cases containing w but not high-risk cases, m represents the number of clinical narratives of high-risk cases without w, and n represents the number of clinical narratives of non-high-risk cases without w.

[0029] In some specific embodiments, the medical model is constructed using the following parameters: 10-fold cross-validation as the validation method, a random inactivation parameter of 0.35, 40 training cycles, a threshold of 0.5, an early stop strategy parameter of 3, and a maximum length of 350 for each text; the method also includes: using synthetic minority class oversampling technology to balance the data distribution so that the proportion of data in each class is the same.

[0030] Secondly, this disclosure provides a medical system comprising a medical model constructed using the method described above, wherein the medical model is used to predict a diagnosis based on information from a subject.

[0031] Secondly, this disclosure provides a non-transitory computer-readable recording medium that stores a computer-executable instruction that, when executed by one or more processors, causes the one or more processors to perform an operation including the aforementioned method for constructing a medical model to predict a diagnostic outcome based on information of a subject.

[0032] In summary, the method and medical model disclosed herein for constructing a medical model have at least the following beneficial effects: (1) The medical model disclosed in this invention can predict possible diagnostic results based on the patient’s medical history and vital signs routinely collected during triage. The predicted diagnostic results can assist doctors in making diagnoses and reduce the probability of misdiagnosis. (2) Compared with the existing technology that uses digital data for calculation, the medical model construction process uses both structured data and unstructured data. The data input into the medical model is in text form, and this data, together with the language processing module in the medical model, can make the predicted diagnosis results more accurate. (3) The area under the receiver operating characteristic curve (AUROC) of the medical model disclosed herein is 0.874, and the area under the precision-recall curve (AUPRC) is 0.207, which is better than other related medical models in the prior art; and it is also better than the doctor's processing level in terms of sensitivity and accuracy. Simple Explanation of the Diagram

[0033] The accompanying drawings illustrate one or more specific embodiments by way of example and not limitation. Unless otherwise disclosed, the drawings are not to scale. Those skilled in the art will understand this disclosure in light of the following detailed description of preferred embodiments and with reference to the accompanying drawings.

[0034] The features, technical characteristics, advantages, and implementation methods of this disclosure will be further described below in a clear and easy-to-understand manner with reference to the illustrations of preferred embodiments. A brief introduction to the accompanying drawings used in this disclosure is provided below:

[0035] Figure 1 shows a flowchart of a method for constructing a medical model provided by the present invention.

[0036] Figure 2 shows a schematic diagram of a medical model provided by the present invention.

[0037] Figure 3 illustrates a method for processing keyword-free clinical narratives provided by the present invention.

[0038] The figure shows a schematic diagram of a medical model provided by the present invention. Implementation

[0039] To more clearly illustrate the technical solutions of the specific embodiments described in this patent application, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. The accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without creative effort. Adjustments and improvements made without departing from the concept of the present invention are all within the protection scope of this disclosure.

[0040] To keep the drawings simple, only the parts related to the corresponding embodiments are shown schematically in each figure, and they do not represent the actual structure of the product. In addition, to make the drawings simple and easy to understand, some parts with the same structure or function are only shown schematically in some figures, and there may actually be more or fewer parts with the same structure or function.

[0041] It should be understood that when words such as "comprising," "including," or "comprise" are used in the specification and claims of this disclosure, they indicate the presence of the said features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or sets.

[0042] To ensure clarity, the accompanying drawings only schematically illustrate the parts relevant to this disclosure and do not represent the actual structure of the product. Furthermore, for simplicity and ease of understanding, in some figures, components with the same structure or function are schematically shown, or only one of them is indicated. Here, "a" can mean not only "only one" but also "more than one".

[0043] In this invention, unless otherwise explicitly specified and limited, ordinal numbers, such as "first," "second," etc., are used only to distinguish and describe related objects, and should not be construed as indicating or implying the relative importance or order between related objects; furthermore, they do not represent the quantity of related objects. "Multiple" includes two or more, and other quantifiers are similar. " / " is used to describe the relationship between related objects, indicating an "or" relationship between them. "And / or" is used to describe the relationship between related objects, including any combination relationship between them, such as "a and / or b," which includes: "a alone," "b alone," or "a and b." "One or more" or "at least one" of a plurality of objects means any one or any combination of a plurality of objects, such as "one or more of a1, a2, a3" or "at least one of a1, a2, a3" including: "a1 alone", "a2 alone", "a3 alone", "a1 and a2", "a1 and a3", "a2 and a3", or "a1, a2 and a3".

[0044] Due to increasing demand for medical care and a shortage of healthcare personnel, more and more hospitals are facing a shortage of medical resources. This shortage manifests in various aspects of hospitals, such as overloaded emergency rooms, insufficient intensive care unit (ICU) beds, long waiting times for surgery, and shortages of medications and equipment. Among these, emergency department (ED) overcrowding is a common phenomenon; and ED overcrowding can lead to delays in ICU care. Because the treatment outcomes for many emergency patients are time-sensitive, early identification of potentially critically ill patients is crucial for reducing morbidity and mortality. Furthermore, ED overcrowding affects the quality of care. Long ED hospital stays not only negatively impact patient satisfaction but also increase healthcare costs. Therefore, early identification of critically ill patients will help reduce hospital stays and costs, and improve the quality of care.

[0045] Nurse-led triage is the current international standard triage model. For example, in Taiwan, the Chinese version of the Canadian Emergency Department Triage and Acuity Scale (CTAS) is used in daily practice. Triage nurses first categorize patients into five urgency levels based on their vital signs and chief complaint. Then, an emergency physician examines the patient, records their medical history, performs a physical examination, and arranges serological and imaging tests. The five urgency levels of the CTAS are ranked in order of importance: resuscitation required, urgent, urgent, less urgent, and non-urgent. The corresponding treatment methods differ depending on the urgency level.

[0046] Due to limited medical resources, if a patient is marked as non-critical by a nurse, they may not receive a timely diagnosis from a doctor; conversely, patients marked as critical by nurses are more likely to receive priority for diagnosis. For example, a patient whose condition is deemed semi-urgent or non-urgent by a triage nurse will be marked as non-critical; conversely, a patient whose condition is deemed to require immediate treatment, highly urgent, or urgent by a triage nurse will be marked as critical. It should be noted that different hospitals may use different standards for these classifications; the above is for reference only.

[0047] However, nurse-led triage can sometimes fail to accurately identify potentially critically ill patients. For example, when a potentially critically ill patient is marked as non-critical by a nurse, their condition may worsen or even lead to death due to prolonged waiting. However, if the initial triage is performed directly by a professional physician, it can lead to a shortage of medical resources. Therefore, the core concept of this invention is to design a medical model using advanced technology, while maintaining existing nurse-led triage, to predict diagnostic results based on information about the target patient. This model's predicted diagnostic results can assist nurses or doctors in identifying potentially critically ill patients, thereby reducing the probability of misdiagnosis and, consequently, reducing the probability of potentially critically ill patients experiencing worsening conditions due to prolonged waiting, thus improving the patient's overall experience.

[0048] The following description is in conjunction with the accompanying drawings.

[0049] Figure 1 shows a flowchart of a method for constructing a medical model according to one specific embodiment of the present invention. Referring to Figure 1, the medical model is configured to predict diagnostic results based on information about a target patient. The method includes the following steps: Step S100: Obtain the first dataset from the medical institution, wherein the first dataset includes first structured data and first unstructured data, and both the first structured data and the first unstructured data are presented in text form. In other words, the first structured data and the first unstructured data are pre-converted into text form or directly presented in text form before being collected. Step S200: Preprocess the first dataset into a second dataset, wherein the second dataset includes second structured data and second unstructured data. Step S300: Based on the second structured data, a first type of data is obtained. The second unstructured data is then processed using a bidirectional long short-term memory network to obtain a second type of data. Language features are then extracted from the second unstructured data to obtain a third type of data. Furthermore, this language feature extraction is performed using natural language processing (NLP). Step S400: The first type of data, the second type of data, and the third type of data are fused into a fusion result, and the fusion result is processed using a normalized exponential function to obtain the diagnostic result.

[0050] The method for constructing a medical model disclosed in this application can be broadly divided into several steps: acquiring data, preprocessing the data, classifying the preprocessed data and performing different operations on each, fusing the data, and obtaining a predicted diagnostic result based on the fused result.

[0051] To improve the authenticity of the data and the reliability of the medical model, the dataset used in this invention comes from medical institutions. For example, a medical institution could be the Department of Health and Welfare at Taipei Medical University-Shuang Ho Hospital, and the dataset could be clinical data from the emergency department of that institution. Since this institution has approximately 95,000 emergency department data entries annually, the dataset obtained from it has significant reference value. In other words, this invention places high demands on the source of the dataset when constructing the medical model: all data used in constructing the medical model consists of real patient medical data.

[0052] For ease of distinction, the dataset from medical institutions is referred to as the first dataset; this first dataset includes a first structured dataset and a first unstructured dataset. The first structured dataset may contain: demographic data, such as age and gender; mode of arrival (at the hospital), such as ambulance, self-medication, assisted by a third party, or transfer from another medical institution; and vital signs, such as body temperature, pulse, respiratory rate, blood oxygen saturation, and blood pressure. The first unstructured dataset may contain: chief complaint, present illness, and medical history. Both the first structured and first unstructured datasets are presented in text format. The first unstructured dataset can also be called the patient's clinical narrative; this type of data may originally be in the form of audio recordings, which can be converted into text using devices with text-to-speech capabilities. For the first structured data, its original form may be digital; for example, for a patient with a body temperature of 39 degrees Celsius, the prior art can use the value 1 (V=1) to represent that the patient is in a fever state, while for a patient with a body temperature of 37 degrees Celsius, the prior art can use the value 0 (V=0) to represent that the patient is not in a fever state. Unlike the processing methods of the prior art, in this invention, for the patient with a body temperature of 39 degrees Celsius, this information will be directly expressed in the text form, such as "patient with fever". That is to say, regardless of the original form of the first structured data and the first unstructured data, before using these data to build a medical model, their form will first be converted into text form (if their original form is not text form). The reason for converting the above data into text form is that these data can cooperate with the language processing module in the medical model. In the subsequent data processing, the diagnosis result of the target patient will be predicted based on the similarity between each word in the text and preset keywords. Therefore, converting the data belonging to the medical model into text form in advance can make the predicted diagnosis result more accurate.

[0053] During model building, the first dataset undergoes preprocessing to correct errors or defects in the first structured and unstructured data. Preprocessing includes grouping, standardization, data cleaning, text translation, and text correction. For clarity, this preprocessed first dataset is referred to as the second dataset, which includes second structured data and second unstructured data, with the second structured data corresponding to the first structured data and the second unstructured data corresponding to the first unstructured data. Preprocessing the input data in a medical model can improve convergence speed. For example, standardization and normalization distribute feature values ​​within a specific range, reducing the magnitude differences between features and helping the model find the optimal solution faster, thus accelerating convergence. Furthermore, considering that some data in the first dataset consists of doctors' quick records during consultations, and that different doctors use different recording methods, as well as issues such as abbreviations and spelling errors in these quick records, further preprocessing is necessary. Therefore, preprocessing can be used to enable the model to learn on higher-quality data, thereby achieving better performance and prediction results.

[0054] After preprocessing, the second structured data and the second unstructured data in the second dataset are processed separately. For the second unstructured data, a bidirectional long short-term memory (BiLSTM) network and language feature extraction are used to obtain two different types of data, namely, the second type and the third type. Based on the second structured data, the first type of data is obtained. The first, second, and third types of data are then fused. The fusion result is a representation of the fusion distributed characteristics. Finally, the softmax function is used to process the fusion result to obtain the predicted diagnostic result. The calculation formula is as follows:

[0055]

[0056]

[0057] Among them, DF (demographic features) refers to demographic features, VF (vital sign features) refers to vital sign features, LF (linguistic features) refers to linguistic features, and LSF (latent semantic features) refers to latent semantic features. Here, is the approximate probability value, M is the parameter matrix of the connection layer (in the medical model), F is the representation of the fused distribution features, and b is the bias. Demographic and vital sign features are derived from the first type of data mentioned above, latent semantic features from the second type of data mentioned above, and linguistic features from the third type of data mentioned above. The above formula yields two approximate probability values; the larger of these is taken as the final result, which is the diagnostic result predicted by the medical model.

[0058] During the fusion process, the fusion result is represented in a 64-dimensional form. In this invention, experiments were conducted by increasing the dimension from 32 (2 to the power of 5) to 128 (2 to the power of 7) in an exponential manner. It was finally determined that the best effect can be achieved when the dimension is 64: that is, simplifying the vector representation while retaining key information.

[0059] Furthermore, considering convergence speed, the ReLU function can be used as the initiation function for nonlinear operations, and dropout can be used as a correction mechanism to prevent overfitting. The Adam optimizer is used to optimize the network's loss function, which is an effective and efficient backpropagation algorithm. This invention uses the cross-entropy function as the loss function, which can reduce the risk of gradient vanishing during stochastic gradient descent. This means it is better than classification error rate or mean squared error.

[0060] The model's loss rate can be calculated as follows:

[0061]

[0062] Where N is the number of training samples, and y is the label of the sample. It is the output of the model.

[0063] In some embodiments of the present invention, the diagnostic outcome includes a critical outcome or a non-critical outcome, wherein the critical outcome includes in-hospital cardiac arrest (IHCA) or admission to the intensive care unit (ICU), and the non-critical outcome includes discharge or admission to a ward.

[0064] The medical model constructed using the above methods can predict possible diagnoses based on routinely collected patient medical history and vital signs during triage. These predicted diagnoses can assist doctors in making diagnoses and reduce the probability of misdiagnosis. Compared to existing technologies that utilize digital data for computation, this medical model utilizes both structured and unstructured data. The data input into the model is in text format, and this text, combined with the language processing module within the model, can lead to more accurate diagnoses.

[0065] In one specific embodiment of the present invention, the aforementioned medical model is constructed to assess whether a patient visiting the emergency room is truly in a critical condition. After the medical model is constructed, its performance is tested. As shown in Table 1, the area under the receiver operating characteristic (AUROC) curve of the method of the present invention is 0.874, and the area under the precision-recall curve (AUPRC) is 0.207, which is superior to other related model methods in the prior art, and also superior to the physician's processing level in both sensitivity and accuracy (as shown in Table 2). Since the medical model of the present invention is based on the existing triage system (i.e., nurse-led triage as mentioned above), it can identify potentially critically ill patients. Therefore, it may also be applied to the training of emergency room physicians to cultivate their ability to make decisions under incomplete information and time pressure.

[0066] Table 1: Performance Comparison of Various Natural Language Processing Models and Methods Method AUROC P value AUPRC Sensitivity Specificity PPV NPV NB 0.785 < 0.001 0.143 0.556 0.868 0.098 0.987 RF 0.789 < 0.001 0.134 0.115 0.990 0.224 0.977 XGB 0.783 < 0.001 0.084 0.193 0.961 0.113 0.979 LR 0.828 < 0.001 0.133 0.685 0.779 0.074 0.990 MLP 0.765 < 0.001 0.100 0.210 0.965 0.135 0.979 TextCNN 0.786 < 0.001 0.090 0.285 0.931 0.096 0.981 BERT 0.810 < 0.001 0.180 0.069 0.998 0.490 0.977 BiLSTM 0.807 < 0.001 0.133 0.168 0.981 0.188 0.979 BiLSTM + TR 0.844 < 0.001 0.157 0.452 0.937 0.156 0.985 Present Model 0.847 - 0.207 0.499 0.933 0.161 0.986

[0067] Table 2: Comparison of the predictive efficacy of emergency physicians with the efficacy of this medical model Sensitivity Specificity PPV NPV Accuracy Set 1 Physicians 0.47 0.82 0.55 0.78 0.72 Present Model 1 0.8 0.9 1 0.93 Set 2 Physicians 0.23 0.64 0.26 0.6 0.5 Present Model 0.9 0.75 0.9 0.75 0.86 Overall Physicians 0.41 0.78 0.47 0.74 0.67 Present Model 0.95 0.77 0.9 0.87 0.9

[0068] In some specific embodiments of the present invention, step (A) of obtaining a first dataset from a medical institution further includes: filtering the first dataset to exclude data of non-target patients, wherein the data of non-target patients includes data of non-adults, data of people who have experienced out-of-hospital cardiac arrest (OHCA), data of people who have refused medical advice (AMA), data of people who have left without notice, data of people who have cancelled their registration, data of people who have transferred, and data of people with incomplete data.

[0069] Filtering and excluding data from non-target patients in the first dataset helps improve data quality and research validity. Removing data from non-target patients (such as minors, out-of-hospital cardiac arrest patients, etc.) ensures a more unified dataset, reducing bias caused by significant differences in patient types. Minors and adults may differ greatly in pathological characteristics and treatment methods; removing minor data helps make research on the adult population more precise. Some patients (such as those who abandoned medical advice, left without notice, withdrew their registration, or have incomplete data) may not have completed the treatment process, resulting in incomplete or non-standard data. Excluding this data avoids the impact of incomplete or abnormal data on research results, improving the scientific rigor and reliability of conclusions. The treatment process of transferred patients and out-of-hospital cardiac arrest patients may be affected by external interventions; removing this type of patient data ensures the representativeness of the sample, making the medical conditions of patients in the dataset consistent with the research objectives, facilitating subsequent analysis. Excluding ineligible patient data reduces the complexity of data processing, minimizes interference with model training and analysis, and thus improves analytical efficiency. Models and analytical algorithms can then focus more on compliant data, improving the stability of data processing and results. Furthermore, patients with incomplete data or abnormal treatment processes can introduce biases, leading to incorrect associations or inaccurate predictions in the model. Excluding this data helps reduce the negative impact of noise and outliers on the analysis, ensuring greater interpretability of the model and analysis results.

[0070] In some embodiments of the present invention, step (B) preprocesses the first dataset into a second dataset, comprising: (B1) grouping the first structured data based on a preset index and performing a normalization process on the first structured data; (B2) optimizing the first unstructured data to conform to the rules of the medical model, wherein the optimization comprises at least one step of data cleaning, text translation, or text correction, wherein the medical model uses a natural language processing (NLP) algorithm to process the first unstructured data, and the NLP algorithm is configured to process monolingual information. For example, the reference values ​​for grouping include age, mean arterial pressure (MAP), or temperature; step (B1) further comprises: performing one-hot encoding on the grouped first structured data; and normalizing the respiratory rate, pulse rate, and blood oxygen saturation in the first structured data using a min-max normalization algorithm.

[0071] For example, patients were divided into three groups based on age: young adults (18 to 40 years), middle-aged adults (41 to 64 years), and elderly adults (65 years or older). Mean arterial pressure was calculated to distinguish patients in shock from other patients, and patients were further divided into febrile and non-febrile groups using a cutoff of 38 degrees Celsius. Each category used unique thermal coding, and the numerical data of respiratory rate, pulse, and oxygen saturation were standardized using a min-maximum normalization algorithm.

[0072] By categorizing patients into youth, middle-aged, and elderly groups based on age, the physiological differences among patients of different age groups can be analyzed more precisely, thereby capturing the characteristic differences of each age group. For example, the mean arterial pressure and cardiovascular status often differ between the elderly and young people; such grouping helps identify risk factors in different age groups. Mean arterial pressure directly reflects the normality of blood circulation and can be used to distinguish shock patients from other patients, helping to quickly identify and label potentially at-risk patient groups. Through this distinction, the model can more accurately focus on the characteristics of shock patients, thereby improving the predictive ability and accuracy of judgment for this group. Using 38 degrees Celsius as a cutoff value to divide patients into fever and non-fever groups allows the model to identify the impact of body temperature on the condition. Fever may be related to infection or inflammation; therefore, this grouping method helps to highlight the impact of fever on the patient's condition, supporting more targeted analysis.

[0073] Using one-hot encoding for age groups, shock states, and fever states can convert categorical features into numerical features, making them easier for the model to process. One-hot encoding avoids the model misinterpreting the order of categorical data and ensures that the influence of each category feature on the model is independent, thereby improving model performance. Using min-max normalization to scale the numerical data of respiratory rate, pulse, and blood oxygen saturation to the same range can reduce the impact of dimensional differences, improve the model's convergence speed, and reduce the weight influence caused by a large feature value on model training. In addition, normalization can reduce the impact of outliers, making the model more stable and robust.

[0074] In some embodiments of the present invention, data cleaning further includes: converting abbreviated text in the first unstructured data into terms in their full expanded form; data cleaning further includes: retaining a preset punctuation mark in the first unstructured data and deleting stop words and punctuation marks other than the preset punctuation mark in the first unstructured data, wherein the preset punctuation mark is a commonly used punctuation mark in the medical field. Data cleaning further includes: retaining negative words in the first unstructured data. Text translation further includes: translating non-English text in the first unstructured data into English text. Text correction further includes: correcting spelling errors in the first unstructured data. The preset punctuation marks include: forward slash (" / "), plus sign ("+"), minus sign ("-"), and space (" ").

[0075] Because the first set of unstructured data is free text, it is difficult to analyze, and different doctors have different writing styles, such as varying usage of abbreviations, punctuation, and uppercase preferences. Furthermore, free text may contain incomplete and ungrammatical sentences, as well as Chinese characters; this is because Chinese is the official working language of Taiwan, and Mandarin is widely used (in this invention, the default data source is the Department of Health and Welfare, Shuang Ho Hospital, Taipei Medical University, mentioned above). These factors increase the difficulty of text analysis because most existing natural language processing algorithms are designed for processing monolingual information. Designing an additional natural language processing algorithm adapted to multiple languages ​​would significantly increase development costs. Moreover, multilingual text can lead to poor predictive performance of the model because the model can only follow the grammatical principles of one language. Therefore, preprocessing of the first dataset is necessary to address these issues. The following examples are excerpts from two original cases in the dataset. Case 1: "Present Illness:, VISIED ED 3 DAYS AGO, NO FEVER, NO ¬¬¬CONTUSION, NO ABDOMINAL PAIN, TODAY, LEFT INGUINAL CARBUNCLE SINCE LAST WEEK WITH DISCHARGE". Case 2: "Chief Complaint:, ABD op hx: niL, deined CHRONIC dz, nkda, NO COLD SWEATING, NO RADIATION TO BACK, NO diarrhea, CT in China: ureter stone, rlq pain on and off, feel pain when press it, then ruq pain, chest XR, In China, feel chest is stuck after a meal, similar episode? gerd location, no radiation pain to back, persistent, acute onset, chest pain the date before yesterday".

[0076] Due to diverse language conditions, the following text refinement can be carried out: text translation, abbreviation expansion, spelling error correction, capital letter adjustment to lowercase letters, stop word and punctuation deletion. Exemplarily, the Google Cloud Platform Translate application programming interface can be used to translate Chinese characters into English words. Then, abbreviations or inappropriate sentences are converted into common terms, such as converting "ABD" to "abdomen", "OP" to "operation", "Hx" to "history", "NKDA" to "no known drug allergy", "No 拉" to "no diarrhea", "RLQ" to "right lower quadrant", etc. Spelling errors are corrected (such as converting "Parkenson" to "Parkinson's Disease") and some capital letters are adjusted to lowercase letters (such as converting "NO RADIATION TO BACK" to "no radiation to back"). In addition, a common clinical terminology dictionary compiled by emergency department doctors can be used to map abbreviations in narrative text to fully expanded terms. For example, the NLTK (natural language toolkit) package is used to delete stop words and punctuation, except for the commonly used slash (" / "), plus sign (" + "), minus sign (" - "), and space (" ") characters in medical abbreviations. For example, laceration is abbreviated as L / W, abrasion is abbreviated as A / W. Some doctors prefer to use " + " or " - " to indicate whether a patient has a specific symptom. For example, dizziness (+) indicates that the patient is dizzy. To delete stop words, negative words in the text are retained because the negative results are important for clinical judgment. For example, in the original text, if "不" is deleted, "不發燒,不抽泣" will be converted to "發燒,抽泣", which turns into a completely different meaning.

[0077] In some specific embodiments of the present invention, the second structured data includes a vital sign and a demographic data, and the second unstructured data includes a chief complaint, a present medical history, and a past medical history; wherein the first sub-data is related to the vital sign and the demographic data, and the dimension of the first type of data is 15, and the dimension of the data related to the chief complaint and the past medical history in the second type of data and the third type of data is 100~500. The second type of data includes a first sub-data and a second sub-data, both of which have a dimension of 300. The third type of data includes a third sub-data, a fourth sub-data, and a fifth sub-data, both of which have a dimension of 200, and the fourth sub-data has a dimension of 30. The first sub-data is obtained by processing the chief complaint through the bidirectional long short-term memory network, the second sub-data is obtained by processing the present medical history through the bidirectional long short-term memory network, the third sub-data is obtained by extracting language features from the chief complaint, the fourth sub-data is obtained by extracting language features from the past medical history, and the fifth sub-data is obtained by extracting language features from the present medical history. The third type of data is obtained by extracting language features from the second unstructured data, including: when the second unstructured data contains keywords, extracting keywords from the second unstructured data, and obtaining the third type of data based on each keyword and its corresponding weight; and when the second unstructured data does not contain keywords, calculating the average value of all single-character vectors in the second unstructured data, calculating the cosine similarity between the average value and all preset keyword vectors, and selecting the five preset keyword vectors that are closest to each other as the third type of data.

[0078] Figure 2 shows a schematic diagram of a medical model provided by an embodiment of the present invention. Figure 2 contains several abbreviations, the meanings of which are as follows: CNTR - clinical narratives text representation, BiLSTM - bidirectional long short-term memory network, kw - keyword, and dim - dimension.

[0079] In this invention, different processing methods are used for different input data. For the chief complaint and present illness, the dimensions of these two types of data can be selected between 100 and 500. Experiments were conducted using 100, 200, 300, 400, and 500 as target dimensions, yielding the following conclusions: For the chief complaint (corresponding to the first sub-data) and present illness (corresponding to the second sub-data) obtained through bidirectional long short-term memory network processing, their dimensions are set to 300; for the chief complaint (corresponding to the third sub-data) and present illness (corresponding to the fifth sub-data) obtained through language feature extraction, their dimensions are set to 200 (with positive and negative vectors each having a dimension of 100). The dimension of past medical history (corresponding to the fourth sub-data) is set to 30, and the dimensions of the processed vital signs and demographic data (corresponding to the first type of data) are set to 15. This optimizes the model's performance. The processing methods for the first type of data include the one-hot encoding and min-max normalization algorithm disclosed above.

[0080] Clinical narrative perception deep neural networks predict key outcomes in the pre-diagnosis stage by processing text-intensive and standardized numerical data. Referring to Figure 2, the medical model disclosed in this invention consists of five main parts: preprocessing, language feature extraction, bidirectional long-term short-term memory network, multi-feature fusion mechanism, and key outcome prediction.

[0081] Since text classification performance benefits from efficient text representation, a vector-based representation of clinical narrative text is proposed. This invention uses this method to establish the association between key outcomes and clinical narratives (i.e., medical history, chief complaint, and present illness). CNTR is inspired by observations of how intensive care physicians identify ICU cases through important words or semantic content, thereby rapidly narrowing down the range of possible candidates. For example, when an expression contains keywords such as "ICH (intracerebral hemorrhage)," "SAH (subarachnoid hemorrhage)," and "aortic dissection," it can be concluded that the expression is likely to contain key information. This hypothesis can be used to explain how intensive care physicians can quickly browse clinical narratives to capture key information in ICU cases. The CNTR method first filters keywords related to critically ill patients (IHCA patients / ICU patients) from all categories of clinical text and combines them into a keyword dictionary. Based on a keyword dictionary, it is determined whether the clinical text contains keywords. If it does, the text is represented by a log-likelihood ratio (LLR) weight. If the clinical text does not contain any keywords, it is represented by a weighted combination of several keywords with the highest similarity to the entire clinical text, where the weight is determined by the similarity score between the keywords and the text.

[0082] To address this, this invention employs a category-based keyword extraction method, which calculates term weights based on the association between terms and categories. Thus, high-weight terms indicate a strong association with critical illness outcomes (ED). This association is determined by calculating the log-likelihood ratio of each word in the text. Patients at high risk of ICU or IHCA admission are defined as positive cases, while those at low risk are defined as negative cases. The calculation formula is as follows:

[0083]

[0084] Wherein, HR represents high-risk patients (high risk of ICU admission or IHCA), ¬HR represents non-high-risk patients, w represents a high-risk keyword, k represents the number of high-risk case clinical narratives containing w, l represents the number of clinical narratives containing w but not high-risk cases, m represents the number of high-risk case clinical narratives without w, and n represents the number of non-high-risk case clinical narratives without w. Words with higher LLR values ​​are more strongly associated with a certain ED disposition. Maximum likelihood estimation is performed to obtain the following probability values: p(w), p(w|HR), and p(w|¬HR). This formula measures the difference between having a high-risk (HR) keyword w and not having one, and this method is applied to clinical texts to extract key information from them.

[0085] Subsequently, this invention sorts the LLR values ​​of all words in the training data and retains the top 200 candidate words as a keyword list for the chief complaint and present illness. Since past medical history is listed as a case description consisting of several key phrases, the top 30 are selected as the keyword list for past medical history. Next, this invention trains word vectors using Gensim and employs the continuous bag-of-words (CBOW) method for text representation. In this invention, the word vectors are trained using clinical narratives from 171,275 ED visits, containing 11,060,526 tags, of which 299,591 unique words are encoded as 300-dimensional distributed vectors. The clinical narrative is jointly represented by the keyword vectors. For example, the clinical narrative text Tk is represented as the weighted average of the keyword vectors, and the weight λi of the keyword Ki is determined by its LLR value. Figure 3 illustrates a method for processing keyword-free clinical narratives provided by an embodiment of this invention. Referring to Figure 3, for a clinical description without any keywords, calculate the average of all word vectors in the description and the cosine similarity of all keyword vectors to find the keywords that best represent the clinical description. For example, five keywords are selected. Figure 3 visually demonstrates the similarity between the text and each keyword vector; the closer the vectors are, the higher the similarity; conversely, the farther apart they are, the lower the similarity.

[0086] In some specific embodiments of this invention, 10-fold cross-validation can be used for validation. For example, the dataset is divided into 10 equal parts, and the model is trained using 9 parts and validated using 1 part. This helps to evaluate the generalization ability of the CNTR model on clinical narrative text data and ensures the model's performance is stable on unseen data. When processing medical texts, the data may be small or uneven, and cross-validation can effectively reduce the impact of the randomness of data splitting on model evaluation. During validation, the random dropout parameter is set to 0.35. Dropout is a technique to prevent model overfitting by randomly dropping a portion of neurons (35% in this invention) during training, reducing the model's dependence on specific neurons. In CNTR, text data is usually high-dimensional and sparse; appropriate dropout can make the model more stable when processing clinical texts (such as doctor's notes and medical records). The number of training epochs is set to 40. The number of training epochs determines the length of time the model learns from the data. 40 epochs indicate that the model will undergo 40 complete training rounds on the clinical narrative text. If the number of training epochs is too low, the model may not be able to capture enough semantic information; if it is too high, it may lead to overfitting. This invention demonstrates through experiments that 40 training epochs are sufficient for effectively learning data features. The threshold is set to 0.5. Thresholds are commonly used in classification tasks, especially in binary or multi-class classification problems. 0.5 means that a predicted value greater than or equal to 0.5 is considered positive, and vice versa. In CNTR, the model may need to classify patient states or conditions based on the input text content (such as descriptions related to "shock"). The threshold setting directly affects the judgment criteria for classification results. Early stopping can be used to prevent overfitting; if the validation set performance does not improve within three consecutive epochs, training is stopped. In CNTR, overfitting causes the model to over-memorize training data, reducing its generalization ability to new texts. Early stopping helps optimize training efficiency and prevent performance degradation. The maximum length of each text is set to 350. When processing text, a maximum length of 350 means that each clinical text contains a maximum of 350 units; any excess is truncated, and any insufficient units are padded. Medical texts can vary considerably in length; by standardizing the length, the model can process text data more efficiently while avoiding the computational overhead of excessively long texts. Furthermore, data processing can be performed using the synthetic minority oversampling technique (SMOTE). CNTR may face the problem of class imbalance when processing clinical texts, such as fewer data points for certain disease descriptions while more data points for other categories.SMOTE balances the data distribution by synthesizing minority class samples. The resulting 1:1 balancing ratio ensures the model gives equal importance to each class, preventing bias towards multi-class data and improving predictive performance on minority classes (such as descriptions of rare diseases).

[0087] In addition to using CNTR for linguistic feature extraction, the texts of the chief complaint and present illness are also input into a deep neural network. First, an embedding layer is used to convert the input tokens of the sentences into vectors of dimension 300. Furthermore, this invention sets the maximum document length to 350, where longer documents are truncated and shorter documents are padded with zeros. Next, the latent semantic features behind the texts of the chief complaint and present illness are learned separately by two (i.e., dual-channel) BiLSTMs. The BiLSTMs consider past and future information by connecting two LSTM networks with opposite time signatures in the same output. Deep learning can discover latent features autonomously, relying on linguistic input rather than feature engineering. However, to improve performance, multiple feature vectors can be combined with the neural network. This invention integrates the generated clinical narrative CNTR into the neural network by connecting positive and negative vectors, specifically connecting a 30-dimensional past medical history vector and two 200-dimensional chief complaint and present illness vectors respectively. Therefore, incorporating these features into the deep learning model is discriminative.

[0088] In another specific embodiment, the medical model of the present invention is constructed to assess whether a patient visiting the emergency room is a patient who will actually be hospitalized. More specifically, the medical model of the present invention can assist in determining (1) whether a patient will be hospitalized and (2) whether a patient will return to the emergency room within 72 hours of discharge. This study used electronic health records from the Emergency Department of Taipei Medical University-Shuang Ho Hospital, Ministry of Health and Welfare, for model development, which included a total of 245,721 emergency room visits.

[0089] The model building method used in this specific embodiment is similar to that described above, so only the differences will be explained. Referring to Figure 4, in the preprocessing stage, this medical model building method separately preprocesses the first structured data and the first unstructured data in the first dataset into second structured data and second unstructured data. More importantly, in the preprocessing stage, text and numbers are preprocessed separately, and the text is further translated into corresponding English (if the first structured data and the first unstructured data are presented in a non-English text format). Therefore, the above content corresponds to steps S100 and S200.

[0090] After the medical model was constructed, its performance was tested. As shown in Table 3, the AUROC and AUPRC for patients who are hospitalized are 0.9014 and 0.7739, respectively, which are better than other related models in the prior art. Furthermore, the AUROC and AUPRC for out-of-call follow-up visits are 0.6475 and 0.0539, respectively, also better than other related models in the prior art. In addition, this medical model also shows that methods incorporating translated text and textualizing categorized digital data have the highest AUROC and AUPRC values ​​(as shown in Table 4).

[0091] Table 3: Performance Comparison of Various Natural Language Processing Models and Methods Experiment Disposition AUROC AUPRC Sensitivity Specificity PPV NPV CNB 0.8159 0.6284 0.3464 0.9639 0.7495 0.8254 KNN 0.7721 0.5935 0.3857 0.3991 0.6641 0.8305 RF 0.8494 0.6743 0.3084 0.9778 0.8125 0.8193 XGB 0.8521 0.6905 0.4280 0.9590 0.7651 0.8431 LGBM 0.8569 0.6970 0.4412 0.9582 0.7671 0.8461 BioClinicalBERT 0.8797 0.7345 0.6211 0.9201 0.7085 0.8862 BlueBERT 0.8798 0.7349 0.6238 0.9208 0.7106 0.8870 Present Model 0.9014 0.7739 0.8037 0.8608 0.6760 0.9184 Experiment Unscheduled Return Visits AUROC AUPRC Sensitivity Specificity PPV NPV CNB 0.5895 0.0516 0.0856 0.9804 0.0909 0.9791 KNN 0.5763 0.0518 0.1929 0.9018 0.0432 0.9799 RF 0.6091 0.0490 0.0876 0.9798 0.0906 0.9791 XGB 0.5923 0.0440 0.1178 0.9570 0.0592 0.9793 LGBM 0.6104 0.0516 0.1122 0.9642 0.0670 0.9793 BioClinicalBERT 0.6262 0.0498 0.1905 0.9181 0.0523 0.9802 BlueBERT 0.6334 0.0510 0.2172 0.9085 0.0522 0.9806 Present Model 0.6475 0.0539 0.2435 0.8991 0.0532 0.9811

[0092] Table 4: Performance Comparison of Different Data Processing Methods Experiment Disposition AUROC AUPRC Sensitivity Specificity PPV NPV Baseline 0.8798 0.7349 0.6238 0.9208 0.7106 0.8870 + TT 0.8891 0.7513 0.7948 0.8561 0.6508 0.9201 + TT + ND 0.7922 0.7303 0.7987 0.8574 0.6679 0.9164 + TT + CND 0.7873 0.7283 0.7966 0.8574 0.6536 0.9210 + TT + CNDT 0.9014 0.7739 0.8037 0.8608 0.6760 0.9184 Experiment Unscheduled Return Visits AUROC AUPRC Sensitivity Specificity PPV NPV Baseline 0.6334 0.0510 0.2172 0.9085 0.0522 0.9806 + TT 0.6455 0.0561 0.2092 0.9192 0.0567 0.9807 + TT + ND 0.5696 0.1554 0.2421 0.8971 0.0518 0.9810 + TT + CND 0.5668 0.1488 0.2262 0.9074 0.0540 0.9808 + TT + CNDT 0.6475 0.0539 0.2435 0.8991 0.0532 0.9811 Note: Baseline=Non-Translated Text, TT=Translated Text, ND=Nmneric Data, CND=Categorical Numeric Data, CNDT = Textualize Categorical Numeric Data

[0093] Based on a similar technical concept, this invention discloses a medical model constructed using the methods disclosed in the above embodiments, configured to predict diagnostic results based on information of the target patient.

[0094] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions in other embodiments. Furthermore, the above embodiments can be freely combined as needed.

[0095] S100, S200, S300, S400: Steps

Claims

1. A method for constructing a medical model, wherein the medical model is used to predict a diagnostic result based on information of a subject, comprising: (A) acquiring a first dataset from a medical institution, wherein the first dataset includes a first structured data and a first unstructured data; (B) preprocessing the first dataset into a second dataset, wherein the second dataset includes a second structured data and a second unstructured data; (C) obtaining a first type of data based on the second structured data, processing the second unstructured data using a bidirectional long short-term memory network to obtain a second type of data, and extracting linguistic features from the second unstructured data to obtain a third type of data; and (D) fusing the first type of data, the second type of data, and the third type of data into a fusion result, and processing the fusion result using a normalized exponential function to obtain the diagnostic result.

2. The method as described in claim 1, wherein step (B) further comprises: (B1) grouping the first structured data based on a preset index, and performing a normalization process on the first structured data; (B2) optimizing the first unstructured data to conform to the rules of the medical model, wherein the optimization comprises at least one step of data cleaning, text translation, or text correction, wherein, The medical model uses a natural language processing algorithm to process the first unstructured data, and the natural language processing algorithm is configured to process messages in a single language.

3. The method as described in claim 2, wherein the reference values ​​for the grouping include age, mean arterial pressure, or body temperature.

4. The method as described in claim 2, wherein step (B1) further comprises: performing one-hot encoding on the grouped first structured data; and normalizing the respiratory rate, pulse rate and blood oxygen saturation in the first structured data using a min-max normalization algorithm.

5. The method as described in claim 2, wherein the data cleaning further comprises: converting abbreviated text in the first unstructured data into terms in their full expanded form.

6. The method as described in claim 2, wherein the data cleaning further comprises: retaining a preset punctuation mark in the first unstructured data, and deleting stop words and punctuation marks other than the preset punctuation mark in the first unstructured data, wherein the preset punctuation mark is a commonly used punctuation mark in the medical field.

7. The method as described in claim 6, wherein the data cleaning further comprises: retaining negative words in the first unstructured data.

8. The method as described in claim 6, wherein the preset punctuation marks include: Forward slash, plus sign, minus sign, and space character.

9. The method as described in claim 2, wherein the text translation further comprises: the text translation includes translating non-English text in the first unstructured data into English text.

10. The method of claim 2, wherein the text correction further comprises: the text correction includes correcting spelling errors in the first unstructured data.

11. The method as described in claim 1, wherein step (A) further comprises: filtering the first dataset to exclude data of non-target patients present in the first dataset, wherein the data of non-target patients includes data of minors, data of people who have suffered out-of-hospital cardiac arrest, data of people who have waived medical advice, data of people who have left without notice, data of people who have withdrawn their registration, data of people who have been transferred to other hospitals, and data of people with incomplete data.

12. The method of claim 1, wherein the diagnostic result includes a critical result or a non-critical result, wherein the critical result includes in-hospital cardiac arrest or admission to the intensive care unit, and the non-critical result includes discharge or admission to a ward.

13. The method as described in claim 1, wherein the second structured data includes a vital sign and a demographic data point, the second unstructured data includes a chief complaint, a present medical history, and a past medical history, the second type of data includes a first sub-data and a second sub-data, the third type of data includes a third sub-data, a fourth sub-data, and a fifth sub-data, wherein the first sub-data is obtained by processing the chief complaint through the bidirectional long short-term memory network, the second sub-data is obtained by processing the present medical history through the bidirectional long short-term memory network, the third sub-data is obtained by extracting language features from the chief complaint, the fourth sub-data is obtained by extracting language features from the past medical history, and the fifth sub-data is obtained by extracting language features from the present medical history.

14. The method as described in claim 13, wherein the first sub-data relates to the life characteristics and the demographic data.

15. The method as described in claim 13, wherein the dimension of the data relating to the chief complaint and the past medical history in the second and third types of data is 100 to 500.

16. The method as described in request item 13, wherein both the first sub-data and the second sub-data have a dimension of 300.

17. The method as described in request item 13, wherein both the third sub-data and the fifth sub-data have a dimension of 200.

18. The method as described in request item 13, wherein the dimension of the fourth sub-data is 30.

19. The method as described in request item 1, wherein the dimension of the first type of data is 15.

20. The method as described in request item 1, wherein the dimension of the fusion result is 64.

21. The method as described in claim 13, wherein the first type of data, the second type of data, and the third type of data are fused into the fusion result, and the fusion result is processed using the normalization exponential function, calculated as follows: Wherein, F represents the result of the fusion, DF represents a feature in the first type of data that is related to the demographic data, VF represents a feature in the first type of data that is related to the life representation, LF represents the third type of data, LSF represents the second type of data, is an approximate probability value, M is a parameter matrix of the connection layer, and b represents the bias.

22. The method as described in claim 1, wherein the step "and extracting language features from the second unstructured data to obtain the third type of data" further comprises: when the second unstructured data contains keywords, extracting keywords from the second unstructured data, and obtaining the third type of data based on each keyword and its corresponding weight; and when the second unstructured data does not contain keywords, calculating an average value of all single-word vectors in the second unstructured data, calculating the cosine similarity between the average value and all preset keyword vectors, and selecting the five nearest preset keyword vectors as the third type of data.

23. The method as described in request item 22, wherein the weight is calculated using the following formula: Wherein, HR represents high-risk patients, ¬HR represents non-high-risk patients, w represents the keyword for high-risk, k represents the number of clinical narratives of high-risk cases containing w, l represents the number of clinical narratives of cases containing w but not high-risk cases, m represents the number of clinical narratives of high-risk cases without w, and n represents the number of clinical narratives of non-high-risk cases without w.

24. The method according to any one of claims 1 to 22, wherein the medical model is constructed using the following parameters: 10-fold cross-validation as the validation method, a random inactivation parameter of 0.35, 40 training epochs, a threshold of 0.5, an early termination strategy parameter of 3, and a maximum length of 350 characters per text; the method further includes: Synthetic minority class oversampling techniques are used to balance the data distribution so that the proportion of data in each class is the same.

25. A medical system comprising a medical model constructed by any one of claims 1 to 24, wherein the medical model is used to predict a diagnostic outcome based on information from a subject.

26. A non-transitory computer-readable recording medium storing computer-executable instructions that, when executed by one or more processors, cause the one or more processors to perform an operation comprising any one of claims 1 to 24, the method for constructing a medical model to predict a diagnostic outcome based on information of a subject.

Images