Doctor-patient communication risk assessment and early warning system and method based on Internet hospital

Abstract

The invention discloses a doctor-patient communication risk assessment and early warning system and method based on an internet hospital, and the system comprises a basic feature calculation module which is used for calculating the basic feature score of the early warning according to the collected online inquiry data of a doctor and a patient, and calculating the weight ratio; the dynamic feature analysis module is used for taking the dynamic feature factors as input of the constructed patient end prediction model and the constructed doctor end prediction model and respectively calculating to obtain patient dynamic prediction scores and doctor dynamic prediction scores; the risk fusion prediction module is used for performing feature fusion and final risk prediction by taking the patient and doctor dynamic prediction scores, the number of times of previous complaints of the patient, the number of times of previous complaints of the doctor and the basic feature scores as input of a constructed fusion prediction model, and performing calculation to obtain a doctor-patient communication risk score; and the early warning triggering module is used for triggering an intelligent early warning mechanism when the doctor-patient communication risk score exceeds a set threshold value. According to the invention, efficient and accurate doctor-patient communication risk assessment and intelligent early warning can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical fields of Internet medical treatment and artificial intelligence, and in particular relates to a doctor-patient communication risk assessment and early warning system and method based on an Internet hospital. Background Art

[0002] An Internet hospital is a new and increasingly mature medical treatment mode under the integration of current information technology and medicine. With the help of an Internet hospital, patients can complete a series of processes such as online registration, consultation, medical treatment, and drug delivery, and can complete an online consultation and follow-up visit without leaving home.

[0003] As a new medical treatment service mode, the Internet hospital provides an online communication channel (such as text-based medical consultation, online consultation, etc.) for patients and doctors, greatly improving the convenience and accessibility of medical services. However, due to the non-face-to-face nature of online communication and the lack of non-verbal information such as body language and tone of voice, communication misunderstandings may occur, thus increasing the risk of doctor-patient disputes. In addition, factors such as the doctor's response time, the patient's emotional fluctuations, and communication behavior patterns may all become potential dispute incentives. The existing methods for resolving doctor-patient disputes have the following disadvantages: 1. Most only consider a single subject (such as patient emotions or doctor behavior); 2. Often rely on static data (such as historical complaint records, satisfaction scores, etc.), and cannot timely capture the emotional fluctuations and risk accumulation trends in real-time communication; 3. Most are mainly for after-the-fact mediation, lacking an active risk identification and early warning mechanism, and it is difficult to timely intervene in potential doctor-patient conflicts; 4. Since it involves patient medical record data, doctor communication content, and sensitive medical information, how to protect the privacy of doctors and patients while accurately assessing risks is an important challenge. Summary of the Invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a doctor-patient communication risk assessment and early warning system and method based on an Internet hospital, which is a doctor-patient dual-perspective dynamic quantification method based on natural language processing (NLP), emotion computing, and multi-model fusion, and is used for doctor-patient communication risk assessment and early warning in an Internet hospital, helping to avoid doctor-patient disputes; by performing real-time analysis on the dual data streams of the patient side and the doctor side, combining emotion computing and historical interaction data, realizing efficient and accurate doctor-patient communication risk assessment and intelligent early warning, and at the same time ensuring the privacy security of both doctors and patients.

[0005] To achieve the above invention purpose, the present invention adopts the following technical solutions:

[0006] A doctor-patient communication risk assessment and early warning system based on an Internet hospital, which includes a basic feature calculation module, a dynamic feature analysis module, a risk fusion prediction module, and an early warning trigger module; wherein,

[0007] The basic feature calculation module is used to calculate the basic feature score S of this warning according to the collected online consultation data of doctors and patients and obtain the patient consultation data and doctor consultation data included in the consultation data base , and calculate the weight ratio NormFWR; the basic feature score S of this warning base is calculated according to the fixed feature values of the patient's disease sensitivity C severity , the integrity C of the patient's medical record completeness , the doctor's experience level D experience , and the doctor's online score D score ;

[0008] The patient's disease sensitivity C severity is calculated according to the disease type Base_Score filled in by the patient online, the disease duration bonus Duration_Modifier, and the length of the symptom description Symptom_Detail

[0009] C severity = w1*Base_Score + w2*Duration_Modifier + w3*Symptom_Detail

[0010] where w1, w2, and w3 respectively represent the weights of the disease type Base_Score, the disease duration bonus Duration_Modifier, and the length of the symptom description Symptom_Detail

[0011] The calculation formula for the basic feature score S of this warning base is as follows:

[0012] S base = C severity + w4*C completeness + w5*D experience + w6*D score

[0013] = w1*Base_Score + w2*Duration_Modifier + w3*Symptom_Detail + + w4*C completeness + w5*D experience + w6*D score

[0014] where w1 to w6 respectively represent the disease type Base_Score, the disease duration bonus Duration_Modifier, the length of the symptom description Symptom_Detail, the integrity C of the patient's medical record completeness , the doctor's experience level D experience, the online doctor score D score The weight of the feature is determined by logistic regression;

[0015] The dynamic feature analysis module is used to dynamically analyze the real-time communication content between doctors and patients by adopting a continuous data collection mechanism, and construct a patient-side prediction model and a doctor-side prediction model. Taking the dynamic feature factors as the inputs of the patient-side prediction model and the doctor-side prediction model, the patient dynamic prediction score S dynamic-patient and the doctor dynamic prediction score S dynamic-doctor are calculated respectively;

[0016] The risk fusion prediction module is used to construct a fusion prediction model. Taking the patient dynamic prediction score S dynamic-patient , the doctor dynamic prediction score S dynamic-doctor , the number of previous patient complaints, the number of previous doctor complaints, and the basic feature score S base as the inputs of the fusion prediction model, XGBoost is used for feature fusion and final risk prediction, and the doctor-patient communication risk score R final of the current online doctor-patient communication process is calculated;

[0017] The early warning trigger module is used to trigger the intelligent early warning mechanism when the doctor-patient communication risk score R final obtained by the risk fusion prediction module exceeds the set threshold.

[0018] Furthermore, the above-mentioned patient medical record integrity C completeness depends on the integrity and availability of the medical record materials. There are several items in the content filled in by the patient online, and each item filled in completely gets a predetermined score;

[0019]

[0020] Furthermore, the determination of the above-mentioned doctor experience level D experience is based on the practice years and score weights.

[0021] Furthermore, the above-mentioned online doctor score D score is the score evaluated by the patients who have completed the online diagnosis and treatment for the doctor; the calculation method of the online doctor score D score is as follows:

[0022] Adopting a linear-based normalization method, assuming x min and x max are respectively the minimum and maximum values of the online doctor score in the original data, and the normalization upper limit D score-max and the lower limit D score-min of this feature are respectively defined as: D score-max = 30, D score-min = 0;

[0023] For the original online score x of the doctor, the normalized online score D score The calculation formula is:

[0024]

[0025] The calculated D score The value range is 0 to 30.

[0026] Furthermore, the above-mentioned weight ratio NormFWR is calculated in the following steps:

[0027] According to the result of the weight confirmation, calculate the doctor feature weight W doctor and the patient feature weight W patient The comparison NormFWR, where NormFWR is the calculated normalized feature weight ratio;

[0028] W doctor = w5 + w6

[0029] W patient = w1 + w2 + w3 + w4

[0030]

[0031] The value of NormFWR is 0 to 1. The closer NormFWR is to 1, it indicates that the doctor feature weight W doctor has a larger value. If it is closer to 0, it indicates that the patient feature weight W patient is larger.

[0032] Furthermore, in the above-mentioned dynamic feature analysis module, a continuous data collection mechanism is adopted to dynamically analyze the real-time communication content between doctors and patients, including the collection of dynamic feature factors at the patient end and the collection of dynamic feature factors at the doctor end.

[0033] Even further, the above-mentioned dynamic feature factors at the patient end include: patient emotion privacy feature, patient sensitive word feature, communication behavior feature;

[0034] 1.1. The patient emotion privacy feature includes: emotional intensity S emotion , emotional fluctuation intensity ΔEmotion, and negative emotion ratio R neg ; where

[0035] (1). The emotional intensity S emotion : Calculate the emotion score of the chat content, including 3 emotions: positive, neutral, and negative. The calculation is as follows:

[0036] Use the Bert model to perform sentiment classification on the text of the chat record. The input is the chat record on the patient side, and the output is the sentiment intensity on the patient side. The output form is a 3D vector:

[0037] S emotion =[P pos ,P neu ,P neg

[0038] Among them, P pos , P neu , P neg respectively represent the probabilities of the patient's positive, neutral, and negative emotions;

[0039] (2), Emotional fluctuation intensity ΔEmotion: The amplitude of the patient's emotional change after receiving a new chat record. The calculation formula is:

[0040] ΔEmotion = |S emotion (t) - S emotion (t - 1)|

[0041] Among them, t represents the current time; t - 1 represents the previous time; S emotion (t) represents the emotional score of the patient at the t-th moment; S emotion (t - 1) represents the emotional score of the patient at the (t - 1)-th moment; "| |" represents the absolute value;

[0042] (3), Proportion of negative emotions R neg : The proportion of negative emotions in the total emotions within a certain time window. The calculation formula is:

[0043]

[0044] Among them, N neg represents the total number of times negative emotions appear; N total represents the total number of times all emotions appear;

[0045] 1.2. The patient sensitive word feature includes the number of times sensitive words appear and the density of sensitive words. Among them,

[0046] (1), Number of times sensitive words appear S sensitive : The number of sensitive words in the chat record; By collecting historical data, constructing a sensitive word library related to medical disputes, and retrieving the online communication text content sent by the patient, count the number of sensitive words n sensitie_i appearing in each text content, and record 1 point for each appearance;

[0047]

[0048] Among them, n represents the total number of sensitive words;​

[0049] (2) Sensitive word density D sensitive : The ratio of the total number of sensitive words to the total number of words in the message. The calculation formula is:

[0050]

[0051] 1.3. The communication behavior characteristics: By analyzing the message sending frequency λ msg and the message sending rhythm σ interval and the high-density message sending D high , identify the behavior pattern of the patient during the communication process. The calculation method is:

[0052] (1) Message sending frequency λ msg : The sending frequency of messages within a unit of time. The calculation formula is:

[0053]

[0054] where N messages represents the number of messages sent within the time window; T window represents the time window;

[0055] (2) Message sending rhythm σ interval : Measure the regularity of the time interval between message sendings by the patient or doctor during the conversation. The calculation formula is:

[0056]

[0057] where σ interval represents the standard deviation of the message sending rhythm; N represents the total number of messages that have been sent so far; t i -t i-1 represents the time interval between two adjacent messages; is the average value of the message intervals;

[0058]

[0059] (3) High-density message sending D high : Refers to the situation where the patient sends multiple messages continuously within a very short period of time. The calculation formula is:

[0060]

[0061] where T high is the set time window; N high represents the number of messages sent continuously within the T high time period.

[0062] Further, the above dynamic characteristic factors of the doctor side Including: emotion-related features, reply behavior features, message sending behavior features;

[0063] 2.1. Emotion-related features include: the proportion of doctors' negative emotions S doctor_neg , the emotional fluctuation of doctors S doctor_fluctuation , the proportion of doctors' negative emotions R doc_neg ; among them,

[0064] (1). The calculation method of the proportion of doctors' negative emotions S doctor_neg is as follows:

[0065] Use the Bert model to perform sentiment classification on the text of the chat record. The input is the chat record on the doctor's side, and the output is the emotional intensity on the doctor's side. The output form is a 3D vector:

[0066] S doctor_emotion = [P′ pos , P′ neu , P′ neg

[0067] Among them, P′ pos , P′ neu , P′ neg respectively represent the probabilities of doctors' positive, neutral, and negative emotions;

[0068] (2). The calculation formula of the emotional fluctuation of doctors S doctor_fluctuation is:

[0069] S doctor_fluctuation = |S′ emotion (t) - S′ emotion (t - 1)|

[0070] Among them, S′ emotion (t) represents the emotional score of the doctor at the t-th moment; S′ emotion (t - 1) represents the emotional score of the doctor's previous reply; "| |" represents the absolute value;

[0071] (3). The calculation formula of the proportion of doctors' negative emotions R doc_neg is

[0072]

[0073] Among them, N′ neg represents the number of times of doctors' negative emotion replies; N′ total represents the total number of times of doctors' emotional replies;

[0074] 2.2. Reply behavior features include: reply speed T response , the proportion of timeout replies R timeout , the average number of replies L reply ​, short reply ratio R short ;

[0075] (1) Reply speed T response The calculation formula is:

[0076] T response = t reply - t query

[0077] Among them, t reply represents the timestamp when the doctor replies to the message; t query represents the timestamp when the patient sends the message;

[0078] (2) Overtime reply ratio R timeout The calculation formula is:

[0079]

[0080] Among them, N timeout represents the number of overtime replies; N total represents the total number of replies from the doctor;

[0081] (3) Average number of replies L reply The calculation formula is:

[0082]

[0083] Among them, M i represents the i-th reply content of the doctor; len(M i ) represents the number of words in the reply of M i ;

[0084] (4) Short reply ratio R short The calculation formula is:

[0085]

[0086] Among them, N short represents the number of replies with a word count lower than the threshold;

[0087] 2.3. Message sending behavior characteristics include the doctor's sending frequency F doc and the doctor's message sending rhythm V doc ;

[0088] (1) Doctor's sending frequency F doc The calculation formula is:

[0089]

[0090] Among them, N msg represents the total number of messages sent by the doctor within the unit time T;

[0091] (2) Doctor message sending rhythm V doc The calculation formula is:

[0092] V doc = std({t i - t i-1 )

[0093] Among them, t i represents the timestamp when the doctor sends the i-th message; std represents the standard deviation operation.

[0094] Furthermore, in the above dynamic feature analysis module, the patient-side risk prediction model uses BERT and BiLSTM to model the patient-side conversation emotions. The operations are as follows:

[0095] (1) BERT is used for semantic feature extraction: using the BERT pre-trained model to obtain the deep semantic features of the patient's text;

[0096] (2) BiLSTM is used to capture temporal information: learning the context relationship in the conversation sequence;

[0097] The patient-side risk prediction model includes the dynamic monitoring features on the patient side, which are: emotional intensity S emotion , emotional fluctuation intensity ΔEmotion, negative emotion ratio R neg , number of sensitive word occurrences S sensitive , sensitive word density D sensitive , message sending frequency λ msg , message sending rhythm σ interval , high-density message sending D high ; The model output is the patient dynamic prediction score S dynamic-patient ;

[0098] S dynamic-patient = f(S emotion , ΔEmotion, R neg , S sensitive , D sensitive , λ msg , σ interval , D high ).

[0099] Furthermore, in the above dynamic feature analysis module, the doctor-side risk prediction model uses the GNN and Transformer structures to model the doctor's communication behavior. The operations are as follows:

[0100] GNN is used to model the doctor's communication pattern. In the GNN structure, the communication relationship between the doctor and the patient is modeled as a graph structure G(D, P, E):

[0101] Nodes: Doctor Di Each doctor is a node; patient P j Patients interacting with doctors are regarded as another type of node;

[0102] Edge: E(D i， P j ) represents the communication relationship between doctors and patients. The weight of the edge is based on the emotional changes of doctors, that is, the proportion S of negative emotions of doctors doctor_neg , the emotional fluctuation S of doctors doctor_fluctuation , the proportion R of negative emotions of doctors doc_neg , the sending frequency F of doctors doc ;

[0103] Through the aggregation calculation of GNN, the system learns the communication behavior of doctors and judges whether they have a high-risk communication mode;

[0104] The Transformer structure is used to encode the text features of doctors' replies;

[0105] The doctor-side risk prediction model includes the dynamic monitoring features on the doctor side, which are respectively: the proportion S of negative emotions of doctors doctor_neg , the emotional fluctuation S of doctors doctor_fluctuation , the proportion R of negative emotions of doctors doc_neg , the reply speed T response , the proportion R of overtime replies timeout , the average number of replies L reply , the proportion R of short replies short , the sending frequency F of doctors doc , the message sending rhythm V of doctors doc , and the output of the model is the doctor's dynamic prediction score S dynamic-doctor ;

[0106] S dynamic-doctor = XGBoost(GNN(S doctor_neg ,S doctor_fluctuation ,R doc_neg ,F doc ),Transformer(T response ,R timeout ,L reply ,R short ))。

[0107] Furthermore, in the above risk fusion prediction module, XGBoost learns through the cascading of multiple decision trees, and each tree is optimized based on the wrong predictions of the previous round. The process is as follows:

[0108] (1). Initialize the base model: Set the initial prediction value, usually the mean of the target variable in the dataset, and calculate the initial loss function to measure the deviation between the prediction value and the true value;

[0109] (2) Calculate the residuals: Calculate the prediction error for each sample, i.e., the residuals:

[0110]

[0111] where R i represents the residuals; y i represents the true value; represents the predicted value of the current model;

[0112] (3) Train a new decision tree: The new decision tree learns the residuals R i from the previous round, and the current prediction result is expressed as:

[0113]

[0114] where represents the predicted value of the t-th tree; represents the predicted values of the previous t - 1 prediction trees; f t (X) represents the prediction result of the current decision tree; α is the learning rate;

[0115] (4) Iterative update: Continuously add new decision trees, and optimize the errors of the previous round each time. Eventually, the model converges to an optimal prediction result;

[0116] (5) Calculate the final risk score: After multiple rounds of iteration, obtain the final doctor - patient communication risk score R final ;

[0117] R final = f(S dynamic-patient , S dynamic-doctor , H complaint-patient , H complaint-doctor , S base ).

[0118] Furthermore, during the above - mentioned model training process, normalize R final so that the value range of R final is between 0 and 100;

[0119]

[0120] where R raw represents the original predicted value of XGBoost; R min and R max represent the minimum and maximum predicted values in the training data respectively.

[0121] A doctor - patient communication risk assessment and early warning method based on the Internet hospital includes the following steps:

[0122] S1. Collect the online consultation data of doctors and patients, and obtain the patient consultation data and doctor consultation data included in the consultation data, and calculate the basic feature score S of this warning base , and calculate the weight ratio NormFWR; the basic feature score S of this warning base According to the patient's disease sensitivity C severity , the integrity C of the patient's medical record completeness , the doctor's experience level D experience , the online score D of the doctor score Calculate based on the fixed eigenvalue;

[0123] S2. Adopt a continuous data collection mechanism to dynamically analyze the real-time communication content of doctors and patients, collect the dynamic characteristic factors of the patient side and the doctor side, and construct a patient-side prediction model and a doctor-side prediction model. Take the dynamic characteristic factors of the patient side and the doctor side as the inputs of the patient-side prediction model and the doctor-side prediction model respectively, and calculate the patient dynamic prediction score S dynamic-patient and the doctor dynamic prediction score S dynamic-doctor ;

[0124] S3. Take the patient dynamic prediction score S dynamic-patient , the doctor dynamic prediction score S dynamic-doctor , the number of previous complaints of the patient, the number of previous complaints of the doctor, and the basic feature score S base As the input of the constructed fusion prediction model, use XGBoost for feature fusion and final risk prediction, and calculate the final doctor-patient communication risk score;

[0125] S4. When the final doctor-patient communication risk score exceeds the set threshold, the system triggers an intelligent warning mechanism.

[0126] Due to the adoption of the above-mentioned technical solution, the present invention has the following advantages:

[0127] The doctor-patient communication risk assessment and early warning system and method based on the Internet hospital integrates multi-source data such as chat records, medical record data, doctor behavior characteristics, and patient emotion analysis to construct a real-time risk assessment and early warning model, which can monitor, quantitatively analyze, and intelligently warn of doctor-patient communication risks in real time, so as to intervene before doctor-patient disputes occur and improve the medical quality and doctor-patient relationship management level of the Internet hospital. Respectively construct a patient-side risk prediction model and a doctor-side risk prediction model to dynamically capture key characteristics such as patient emotions, communication behaviors, sensitive word usage, and doctors' reply patterns and communication strategies, and comprehensively evaluate the risk level through a fusion model to provide a more accurate communication risk warning. By continuously collecting and analyzing doctor-patient dialogue content, using sentiment analysis, text mining, and behavior pattern recognition, dynamically quantify the negative emotions of patients and the reply characteristics of doctors, and combine historical data to achieve accurate prediction of communication risks. Through a risk quantification model and a fusion early warning mechanism, when a high-risk dialogue is detected, the system will automatically trigger an early warning notification and provide intervention suggestions to relevant medical management personnel or doctors to reduce the further escalation of doctor-patient conflicts. Adopt data anonymization, differential privacy, and access control technologies to ensure that patient and doctor privacy information is not leaked during data processing, and at the same time only provide necessary risk warning information to authorized personnel to achieve a safe and efficient risk assessment system. BRIEF DESCRIPTION OF THE DRAWINGS

[0128] Figure 1 is a structural diagram of the doctor-patient communication risk assessment and early warning system based on the Internet hospital of the present invention;

[0129] Figure 2 is a flowchart of the doctor-patient communication risk assessment and early warning method based on the Internet hospital of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0130] The technical solutions of the embodiments of the present invention will be further described in detail below with reference to the drawings and embodiments. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0131] As Figure 1 , 2 shown, a doctor-patient communication risk assessment and early warning system based on the Internet hospital includes a basic feature calculation module, a dynamic feature analysis module, a risk fusion prediction module, and an early warning trigger module; among them,

[0132] The basic feature calculation module is used to calculate the basic feature score S of this early warning according to the collected online consultation data of doctors and patients and obtain the patient consultation data and doctor consultation data included in the consultation database , first set S base with an initial value of 0, where the higher the score, the higher the warning level;

[0133] The basic feature score S of this warning base is calculated based on the disease sensitivity C of the patient severity , the completeness C of the patient's medical record completeness , the experience level D of the doctor experience , and the online score D of the doctor score from their fixed feature values;

[0134] 1. The disease sensitivity C of the patient severity depends on the severity and type of the disease and is calculated based on the medical record information filled in by the patient online; patients with severe or chronic diseases have higher communication sensitivity and are more likely to cause disputes; for inquiries involving privacy or sensitive topics (such as fertility, mental health), the risk of disputes is greater; the disease sensitivity C of the patient severity is calculated based on the disease type Base_Score filled in by the patient online, the duration bonus Duration_Modifier, and the length of the symptom description Symptom_Detail;

[0135] C severity = w1*Base_Score + w2*Duration_Modifier + w3*Symptom_Detail

[0136] where w1, w2, and w3 represent the weights of the disease type Base_Score, the duration bonus Duration_Modifier, and the length of the symptom description Symptom_Detail respectively;

[0137] (1). The disease type Base_Score: Compare the disease type filled in or selected by the patient online with the preset centralized disease types, and assign different weights to different disease types. The weight values can be dynamically adjusted according to the hospital's own situation. Preferably, the weight results are as shown in Table 1;

[0138] Table 1

[0139] Disease type Details Early warning risk level Early warning score weight Severe illness Cancer, heart disease High 50 Chronic disease Diabetes, hypertension Medium 30 Involving privacy or sensitive topics Related to fertility and mental health Special level 40 Others Other chronic diseases General 10

[0140] (2) Duration_Modifier of illness duration: Additional scores are added according to the illness duration. The longer the illness duration, the more complex the disease management is to a great extent, and the more anxiety and expectations the patient has about the treatment effect to a great extent, thus increasing the risk of disputes. The weight values can be dynamically adjusted according to the hospital's own situation. Preferably, the weight results determined according to the Duration_Modifier of illness duration are shown in Table 2;

[0141] Table 2

[0142] Duration of illness Early warning risk level Early warning score weight Less than 1 year No bonus 10 1 - 3 years Low bonus 20 More than 3 years High bonus 30

[0143] (3) Symptom_Detail of the symptom description: The degree of detail of the symptom description usually reflects the patient's attention to their own condition and expectations for medical results. A detailed description indicates that the patient has higher attention and expectations for the treatment. The weight values can be dynamically adjusted according to the hospital's own situation. Preferably, the weight results determined according to the Symptom_Detail of the symptom description are shown in Table 3;

[0144] Table 3

[0145] Length of symptom description Early warning risk level Early warning score weight Description less than 50 words Brief description 10 Description between 50 - 100 words Medium description 20 Description more than 100 words Detailed description 30

[0146] 2. The completeness C of the patient's medical record completeness Depends on the integrity and availability of the medical record materials. The integrity of the patient's medical record materials directly affects the doctor's understanding of the patient's condition and the accuracy of the diagnosis. There are 15 items filled in by the patient online, including but not limited to the illness duration, symptom description, and health record. The symptom description includes relevant examination forms, CT, medical records, affected parts, and medications. The health record includes marital status, occupation, fertility status, disease history, surgical trauma history, infectious disease history, family medical history, marriage and childbirth, drug allergy history, blood transfusion history, smoking history, alcohol consumption history, and drug exposure history. Each item filled in represents a higher degree of completeness, and each item filled in completely gets 2 points, with a full score of 30 points;

[0147]

[0148] 3. The doctor's experience level D experience Is an important factor in evaluating the risk of doctor-patient communication, especially in the online communication environment of Internet hospitals. Experienced doctors can more effectively manage the patient's expectations, accurately diagnose the condition, and reduce misunderstandings and potential disputes through communication. The weight values can be dynamically adjusted according to the hospital's own situation. Preferably, according to the doctor's experience level D experience The determined weight results are shown in Table 4;

[0149] Table 4

[0150] Doctor's experience level Years of practice Early warning score weight Less than 5 years Low experience 10 5 - 15 years High experience 20 More than 15 years Detailed description 30

[0151] 4. The online doctor rating D score is the score given by patients who have completed online diagnosis and treatment to the doctor. A lower rating for the doctor indicates that the doctor's service or communication is not satisfactory, thus increasing the likelihood of patients filing disputes; conversely, a high rating for the doctor usually means that patients are relatively satisfied with the doctor's service and the risk of disputes is lower. The online doctor rating D score is calculated as follows:

[0152] Using a linear-based normalization method, assuming x min and x max are respectively the minimum and maximum values of the online doctor rating in the original data. To maintain the balance with other basic features, the normalization upper limit D score-max and lower limit D score-mi of this feature are respectively defined as: D score-max = 30, D score-min = 0;

[0153] For the original online doctor rating x, its normalized online rating D score is calculated by the formula:

[0154]

[0155] After being calculated by the above method, the value range of D score is 0 to 30, which can ensure the balance between different basic features;

[0156] By collecting and calculating the above basic features, the basic feature score S base of this early warning is calculated;

[0157] S base = C severity + w4 * C completeness + w5 * D experience + w6 * D score

[0158] = w1 * Base_Score + w2 * Duration_Modifier + w3 * Symptom_Detail++ w4 * C completeness + w5 * D experience + w6 * D score

[0159] where w1 to w6 respectively represent the disease type Base_Score, the duration of illness bonus Duration_Modifier, the length of symptom description Symptom_Detail, and the completeness of the patient's medical record C completeness, Doctor experience level D experience , Doctor online rating D score weights; the magnitude of the weights represents the importance of different features and directly affects the final S base calculation. Since there are differences in Internet hospitals of different institutions, for the weights of features, logistic regression is used to confirm them;

[0160] According to the result of the confirmation of the weights, calculate the doctor feature weight W doctor and the patient feature weight W patient of the comparison NormFWR. NormFWR is the calculated normalized feature weight ratio;

[0161] W doctor = w5 + w6

[0162] W patient = w1 + w2 + w3 + w4

[0163]

[0164] The value of NormFWR is 0 to 1. If NormFWR is 0.5, it indicates that W doctor and W patient are the same. If it is closer to 1, it indicates that the doctor feature weight W doctor has a larger value. If it is closer to 0, it indicates that the patient feature weight W patient is larger;

[0165] The dynamic feature analysis module is used to adopt a continuous data collection mechanism to dynamically analyze the real-time communication content between doctors and patients, collect patient-side dynamic feature factors and doctor-side dynamic feature factors, and construct a patient-side prediction model and a doctor-side prediction model. The patient-side dynamic feature factors and doctor-side dynamic feature factors are respectively used as the inputs of the patient-side prediction model and the doctor-side prediction model, and the patient dynamic prediction score S dynamic-patient and the doctor dynamic prediction score S dynamic-doctor are calculated respectively;

[0166] (1). Dynamic feature data collection

[0167] 1. Patient-side dynamic feature factors Include: patient emotion privacy feature, patient sensitive word feature, communication behavior feature;

[0168] 1.1. The patient emotion privacy feature includes: emotional intensity S emotion , emotion fluctuation intensity ΔEmotion and negative emotion ratio R neg ; where

[0169] (1). Emotional intensity S emotion: Calculate the emotion scores of chat content, including three emotions: positive, neutral, and negative. The specific calculation is as follows:

[0170] Use the Bert model to perform sentiment classification on the text of the chat record. The input is the chat record on the patient side, and the output is the sentiment intensity on the patient side. The output form is a 3D vector:

[0171] S emotion =[P pos ,P neu ,P neg

[0172] Among them, P pos , P neu , P neg respectively represent the probabilities of the patient's positive, neutral, and negative emotions;

[0173] (2) Emotion fluctuation intensity ΔEmotion: After receiving a new chat record, it represents the amplitude of the patient's emotional change. The specific calculation formula is:

[0174] ΔEmotion=|S emotion (t)-S emotion (t - 1)|

[0175] Among them, t represents the current time; t - 1 represents the previous time; S emotion (t) represents the emotional score of the patient at the t-th moment; S emotion (t - 1) represents the emotional score of the patient at the (t - 1)-th moment; "| |" represents the absolute value;

[0176] (3) Negative emotion ratio R neg : It represents the proportion of negative emotions in the total emotions within a certain time window. The specific calculation formula is:

[0177]

[0178] Among them, N neg represents the total number of times negative emotions appear; N total represents the total number of times all emotions appear;

[0179] 1.2. The patient sensitive word feature includes the number of sensitive word occurrences and the sensitive word density. Among them,

[0180] (1) The number of sensitive word occurrences S sensitive : The number of sensitive words in the chat record. By collecting historical data, a sensitive word library related to medical disputes is constructed, including but not limited to "misdiagnosis", "irresponsible", "compensation", "complaint". Search the online communication text content sent by the patient and count the number of sensitive words n sensitie_i ​, each occurrence is scored 1 point;

[0181]

[0182] Among them, n represents the total number of sensitive words;

[0183] (2) Sensitive word density D sensitive : The ratio of the total number of sensitive words to the total number of words in the message. The specific calculation formula is:

[0184]

[0185] 1.3. The communication behavior characteristics are important indicators for measuring the patient's emotional state and potential dispute risks. By analyzing the message sending frequency λ msg , the message sending rhythm σ interval , and the high-density message sending D high , identify the patient's behavior pattern during the communication process, and then evaluate their emotional changes and communication risks. The specific calculation method is:

[0186] (1) Message sending frequency λ msg : The message sending frequency per unit time. The specific calculation formula is:

[0187]

[0188] Among them, N messages represents the number of messages sent within the time window; T window represents the time window, such as 1 hour or 3 hours;

[0189] (2) Message sending rhythm σ interval : Measure the regularity of the message sending time interval of the patient or doctor during the conversation. The standard deviation is used to calculate the fluctuation degree of the message interval. The larger the value, the more uneven the message sending, indicating greater emotional fluctuations or abnormal communication;

[0190]

[0191] Among them, σ interval represents the standard deviation of the message sending rhythm; N represents the total number of messages that have been sent so far, used to calculate the message sending rhythm; t i - t i-1 represents the time interval between two adjacent messages; is the average value of the message intervals;

[0192]

[0193] If σ interval is close tointerval Higher, indicating that the fluctuation of the message sending time interval is relatively large, and there are emotional fluctuations or abnormal communications of the patient or doctor;

[0194] (3), High-density message sending D high : Refers to the situation where the patient sends multiple messages continuously within a very short period of time; High-density message sending may indicate that the patient is dissatisfied with the doctor's reply or has high expectations for medical services, which is likely to lead to disputes; The calculation formula is:

[0195]

[0196] Among them, T high is the set time window; N high represents the number of messages sent continuously within the T high time period;

[0197] 2. Dynamic characteristic factors on the doctor side Include: emotion-related characteristics, reply behavior characteristics, message sending behavior characteristics;

[0198] 2.1. Emotion-related characteristics include: the proportion of the doctor's negative emotions S doctor_neg , the doctor's emotional fluctuation S doctor_fluctuation , the proportion of the doctor's negative emotions R doc_neg ; Among them,

[0199] (1), The proportion of the doctor's negative emotions S doctor_neg : Calculate the emotion score of the chat content sent by the doctor, including 3 kinds of emotions: positive, neutral and negative. The specific calculation method is:

[0200] Use the Bert model to perform sentiment classification on the text of the chat record. The input is the chat record on the doctor side, and the output is the emotional intensity on the doctor side. The output form is a 3D vector:

[0201] S doctor_emotion =[P′ pos , P′ neu , P′ neg

[0202] Among them, P′ pos , P′ neu , P′ neg respectively represent the probabilities of the doctor's positive, neutral and negative emotions;

[0203] (2), The doctor's emotional fluctuation S doctor_fluctuation : After receiving a new chat record, the change range of the doctor's emotion. The specific calculation formula is:

[0204] S doctor_fluctuation =|S′ emotion (t)-S′​emotion (t - 1)|

[0205] Among them, S′ emotion (t) represents the emotional score of the doctor at the t-th moment; S′ emotion (t - 1) represents the emotional score of the doctor's previous reply; "| |" represents the absolute value;

[0206] If the emotional fluctuation is too large, it means that the doctor is affected by the patient's emotion or there is pressure in the communication process;

[0207] (3) Doctor's negative emotion ratio R doc_neg : The proportion of negative emotions in the total emotions within a certain time window. The specific calculation formula is

[0208]

[0209] Among them, N′ neg represents the number of times of the doctor's negative emotion replies; N′ total represents the total number of times of the doctor's all emotion replies;

[0210] 2.2. Reply behavior characteristics include: reply speed T response 、overtime reply ratio R timeout 、average number of replies L reply 、short reply ratio R short , and the above characteristics directly affect the patient experience and the doctor-patient relationship;

[0211] (1) Reply speed T response : Measure the timeliness of the doctor's reply to the patient. Calculate the average time from when the doctor receives the patient's message to the reply. The calculation formula is:

[0212] T response = t reply - t query

[0213] Among them, t reply represents the timestamp of the doctor's reply message; t query represents the timestamp of the patient's sent message;

[0214] Overtime reply ratio R timeout : Measure the proportion of the number of times the doctor's reply time exceeds the threshold. The threshold is preset by each hospital according to experience, such as 10 minutes or 20 minutes. The calculation formula is:

[0215]

[0216] Among them, N timeout represents the number of overtime replies (exceeding the preset threshold); N total represents the total number of the doctor's replies;

[0217] (3) Average number of replies L reply : Calculate the average number of words in the doctor's replies to judge the detail level of the doctor's replies. The calculation formula is:

[0218]

[0219] where M i represents the i-th reply content of the doctor; len(M i ) represents the number of words in M i 's reply; N total represents the total number of replies of the doctor;

[0220] (4) Short reply ratio R short : Calculate the proportion of messages with the doctor's reply word count lower than the threshold, and measure whether the doctor has a high frequency of short replies. The threshold is preset by each hospital according to experience, such as 15 words, 20 words or 30 words. The calculation formula is:

[0221]

[0222] where N short represents the number of replies with a word count lower than the threshold;

[0223] 2.3 Message sending behavior characteristics: The doctor's message sending behavior pattern reflects their communication strategy and working status. Abnormal message patterns affect the patient's emotions and even lead to doctor-patient disputes;

[0224] Message sending behavior characteristics mainly include the doctor's sending frequency F doc and the doctor's message sending rhythm V doc ;

[0225] (1) Doctor's sending frequency F doc : Calculate the number of messages sent by the doctor within the unit time T to measure the communication activity of the doctor. Preferably, the unit time T is 30 minutes or 1 hour, and the hospital presets it according to different situations. The calculation formula is:

[0226]

[0227] where N msg represents the total number of messages sent by the doctor within the unit time T;

[0228] (2) Doctor's message sending rhythm V doc : Calculate the standard deviation of the doctor's message sending time intervals to measure whether the doctor's message sending is regular. The calculation formula is:

[0229] V doc =std({t i -ti-1 [[ID=1}}

[0230] Among them, t i represents the timestamp when the doctor sends the i-th message; std represents the standard deviation;

[0231] (2). Constructing a risk prediction model

[0232] 1. Patient-side risk prediction model

[0233] Use BERT (Bidirectional Encoder Representations from Transformers) and BiLSTM (Bidirectional Long Short-Term Memory) to model the patient's dialogue emotions. Specifically:

[0234] (1). BERT is used for semantic feature extraction: Use the BERT pre-trained model to obtain the deep semantic features of the patient's text to ensure efficient understanding of medical terms, implicit emotions, and situational associations;

[0235] (2). BiLSTM is used to capture temporal information: Since the patient's emotions fluctuate over time, BiLSTM can learn the context relationships in the dialogue sequence and effectively detect sudden changes in emotions;

[0236] The patient-side risk prediction model includes the dynamic monitoring features of the patient side, which are: emotional intensity S emotion , emotional fluctuation intensity ΔEmotion, negative emotion ratio R neg , the number of occurrences of sensitive words S sensitive , sensitive word density D sensitive , message sending frequency λ msg , message sending rhythm σ interval , high-density message sending D high ; The model output is the patient's dynamic prediction score S dynamic-patient , the higher the value, the higher the communication risk of the patient in the current dialogue;

[0237] S dynamic-patient = f(S emotion , ΔEmotion, R neg , S sensitive , D sensitive , λ msg , σ interval , D high ) k

[0238] 2. Doctor-side risk prediction model

[0239] The communication behavior of doctors is modeled using GNN (Graph Neural Network) and Transformer structures to more accurately assess the risk levels of doctors;

[0240] GNN is used to model the communication patterns of doctors: The communication styles of doctors are affected by various factors, such as response speed, emotional fluctuations, patient emotional feedback, and historical complaint records; GNN is suitable for modeling complex features and can capture the stability of communication patterns among doctors for different patients, and is used to determine whether doctors have high-risk communication behaviors;

[0241] In the GNN structure, the communication relationship between doctors and patients is modeled as a graph structure G(D, P, E):

[0242] Nodes: Doctor D i It means that each doctor is a node; Patient P j It means that the patients interacting with the doctor are regarded as another type of node;

[0243] Edges: E(D i , P j ) represents the communication relationship between doctors and patients, and the weight of the edge is based on the emotional changes of doctors, that is, the proportion of negative emotions of doctors S doctor_neg , the emotional fluctuation of doctors S doctor_fluctuation , the proportion of negative emotions of doctors R doc_neg , the sending frequency of doctors F doc ;

[0244] Through the aggregation calculation of GNN, the system can learn the communication behavior of doctors and determine whether they have high-risk communication patterns;

[0245] The Transformer structure is used to encode the text features of doctors' responses: Transformer has powerful text representation capabilities and can efficiently analyze the semantic integrity, emotional expression, and communication strategies of doctors' responses, ensuring that the model can accurately evaluate the communication quality of doctors; Through the self-attention mechanism, Transformer can analyze the language quality of doctors' responses and extract the communication style features of doctors;

[0246] The risk prediction model for doctors includes dynamic monitoring features on the doctor side, which are respectively: the proportion of negative emotions of doctors S doctor_neg , the emotional fluctuation of doctors S doctor_fluctuation , the proportion of negative emotions of doctors R doc_neg , response speed T response , the proportion of overtime responses R timeout , average number of responses L reply , the proportion of short responses R short , the sending frequency of doctors F doc, Doctor Message Sending Rhythm V doc , the model output is the doctor dynamic prediction score S dynamic-doctor , the higher the value, the higher the dispute risk of the patient in the current conversation;

[0247] S dynamic-doctor = XGBoost(GNN(S doctor_neg ,S doctor_fluctuation ,R doc_neg ,F doc ),Trabsformer(T response ,R timeout ,L reply ,R short ))

[0248] The risk fusion prediction module is used to construct a fusion prediction model, and the patient dynamic prediction score S dynamic-patient , doctor dynamic prediction score S dynamic-doctor , the number of previous complaints of the patient, the number of previous complaints of the doctor, and the basic feature score S base are used as the inputs of the fusion prediction model. XGBoost (Extreme Gradient Boosting) is used for feature fusion and final risk prediction, and the doctor-patient communication risk score R final is calculated;

[0249] XGBoost learns through the cascade of multiple decision trees, and each tree is optimized based on the mispredictions of the previous round. The specific process is as follows:

[0250] (1). Initialize the base model (the first decision tree): Set the initial prediction value, usually the mean of the target variable in the dataset, and calculate the initial loss function (such as mean squared error) to measure the deviation between the prediction value and the true value;

[0251] (2). Calculate the residuals: Calculate the prediction error of each sample, that is, the residuals:

[0252]

[0253] where, R i represents the residuals; y i represents the true value; represents the prediction value of the current model;

[0254] (3). Train a new decision tree: The new decision tree learns the residuals R i from the previous round and tries to reduce the error. The current prediction result is expressed as:

[0255]

[0256] where, Represents the predicted value of the t-th tree; Represents the predicted values of the first t - 1 prediction trees; f t (X) represents the prediction result of the current decision tree; α is the learning rate, controlling the impact of the new tree on the overall model;

[0257] (4), Iterative update: Continuously add new decision trees, optimizing the errors of the previous round each time, and finally the model converges to an optimal prediction result;

[0258] (5), Calculate the final risk score: After multiple rounds of iteration, obtain the final doctor-patient communication risk score R final ;

[0259] R final = f(S dynamic-patient , S dynamic-doctor , H complaint-patient , H complaint-doctor , S base )

[0260] During the process of model training, to more quickly achieve early warning level determination and triggering, normalize the doctor-patient communication risk score R final to ensure that the value range of R final is between 0 and 100;

[0261]

[0262] Among them, R raw represents the original predicted value of XGBoost; R min and R max represent the minimum and maximum predicted values in the training data respectively;

[0263] The early warning trigger module is used to trigger the intelligent early warning mechanism of the system when the doctor-patient communication risk score R final obtained by the risk fusion prediction module exceeds the set threshold, so as to take corresponding intervention measures to reduce potential doctor-patient contradictions; The main goal of early warning triggering is to identify high-risk communication scenarios in advance and provide reasonable intervention strategies, thereby optimizing doctor-patient interaction and improving the service quality of Internet hospitals.

[0264] According to the doctor-patient communication risk score R final threshold, formulate different treatment measures, as shown in Table 5.

[0265]

[0266] As Figure 1 , 2 shown, a doctor-patient communication risk assessment and early warning method based on an Internet hospital includes the following steps:

[0267] S1. Collect the online consultation data of doctors and patients, obtain the patient consultation data and doctor consultation data included in the consultation data, and calculate the basic feature score S of this warning base , and calculate the weight ratio NormFWR; the basic feature score S of this warning base According to the patient's disease sensitivity C severity , the completeness C of the patient's medical record completeness , the doctor's experience level D experience , the online score D of the doctor score Calculate based on the fixed eigenvalue;

[0268] S2. Adopt a continuous data collection mechanism to dynamically analyze the real-time communication content of doctors and patients, collect the dynamic characteristic factors of the patient side and the doctor side, and construct the prediction models of the patient side and the doctor side. Use the dynamic characteristic factors of the patient side and the doctor side as the inputs of the prediction models of the patient side and the doctor side respectively, and calculate the patient dynamic prediction score S dynamic-patient and the doctor dynamic prediction score S dynamic-doctor ;

[0269] S3. Use the patient dynamic prediction score S dynamic-patient , the doctor dynamic prediction score S dynamic-doctor , the number of previous complaints of the patient, the number of previous complaints of the doctor, and the basic feature score S base As the input of the constructed fusion prediction model, use XGBoost for feature fusion and final risk prediction, and calculate the final doctor-patient communication risk score;

[0270] S4. When the final doctor-patient communication risk score exceeds the set threshold, the system triggers the intelligent warning mechanism.

[0271] On the other hand, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned doctor-patient communication risk assessment and warning method based on the Internet hospital are realized.

[0272] On yet another aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned doctor-patient communication risk assessment and warning method based on the Internet hospital are realized.

[0273] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A doctor-patient communication risk assessment and early warning system based on the Internet hospital, characterized in that: It includes a basic feature calculation module, a dynamic feature analysis module, a risk fusion prediction module, and a warning trigger module; among them, The basic feature calculation module is used to calculate the basic feature score S of this warning according to the online consultation data of doctors and patients collected and obtain the patient consultation data and doctor consultation data included in the consultation data base , and calculate the weight ratio NormFWR; the basic feature score S of this warning base According to the patient's disease sensitivity C severity , the integrity of the patient's medical record C completeness , the doctor's experience level D experience , the online score D of the doctor score is calculated from the fixed feature values; The disease sensitivity C of the patient severity It is calculated based on the disease type Base_Score filled in by the patient online, the disease duration bonus Duration_Modifier, and the length of the symptom description Symptom_Detail; C severity = w1 * Base_Score + w2 * Duration_Modifier + w3 * Symptom_Detail wherein, w1, w2, and w3 respectively represent the weights of the disease type Base_Score, the disease duration bonus Duration_Modifier, and the length of the symptom description Symptom_Detail; The basic feature score S of this early warning base The calculation formula is as follows: S base = C severity + w4 * C completeness + w5 * D experience + w6 * D score = w1 * Base_Score + w2 * Duration_Modifier + w3 * Symptom_Detail++w4 * C completeness + w5 * D experience + w6 * D score Among them, w1 to w6 respectively represent the disease type Base_Score, the duration of illness bonus Duration_Modifier, the length of the symptom description Symptom_Detail, the completeness C of the patient's medical record completeness , the doctor's experience level D experience , the doctor's online score D score ; the weights of the features are confirmed by means of logistic regression; The dynamic feature analysis module is used to dynamically analyze the real-time communication content between doctors and patients by adopting a continuous data collection mechanism, and construct a patient-side prediction model and a doctor-side prediction model. Taking the dynamic feature factors as the inputs of the patient-side prediction model and the doctor-side prediction model, the patient dynamic prediction score S dynamic-patient and the doctor dynamic prediction score S dynamic-doctor are calculated respectively; The risk fusion prediction module is used to build a fusion prediction model, which takes the patient's dynamic prediction score S dynamic-patient , the doctor's dynamic prediction score S dynamic-doctor , the patient's previous complaint times, the doctor's previous complaint times, and the basic feature score S base as the input of the fusion prediction model, uses XGBoost for feature fusion and final risk prediction, and calculates the doctor-patient communication risk score R final ; The early warning trigger module is used to trigger the intelligent early warning mechanism of the system when the doctor-patient communication risk score R obtained by the risk fusion prediction module final exceeds the set threshold.

2. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 1, wherein: it also includes any one or more of the following features: (1), the completeness C of the patient's medical record completeness Depends on the integrity and availability of the medical record data. There are several items in the content filled in by the patient online, and each item filled in completely gets a predetermined score; (2) The doctor's experience level D experience is determined based on the years of practice and the score weight; (3) The doctor's online score D score is the score evaluated by patients who have completed online diagnosis and treatment for the doctor; the doctor's online score D score The calculation method is as follows: Using a linear-based normalization method, assume x min and x max are the minimum and maximum values of the doctor's online score in the original data respectively. Define the upper limit D score-max and the lower limit D score-min of this feature as follows: D score-max = 30, D score-min = 0; For the original online doctor rating \(x\), its normalized online rating \(D\) score is calculated by the following formula: Calculated D score The value range is 0 to 30.

3. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 1, characterized in that: In the basic feature calculation module, the weight ratio NormFWR is calculated, and the steps are as follows: Calculate the doctor feature weight W according to the result confirmed by the said weight doctor and the patient feature weight W patient of the comparison NormFWR, where NormFWR is the calculated normalized feature weight ratio; W doctor = w5 + w6 W patient = w1 + w2 + w3 + w4 The value of NormFWR ranges from 0 to 1. The closer NormFWR is to 1, the greater the weight W of the doctor's characteristics. doctor If it is closer to 0, it indicates that the weight W of the patient's characteristics patient is greater.

4. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 1, wherein: In the dynamic feature analysis module, a continuous data collection mechanism is adopted to dynamically analyze the real-time communication content between doctors and patients, including the collection of dynamic feature factors at the patient end and the collection of dynamic feature factors at the doctor end.

5. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 4, characterized in that: It also includes any one or more of the following features: (1), the dynamic characteristic factors of the patient side including: patient emotional privacy characteristics, patient sensitive word characteristics, communication behavior characteristics; 1.

1. The emotional privacy characteristics of the patient include: emotional intensity S emotion , emotional fluctuation intensity ΔEmotion, and negative emotion ratio R neg ; where 1.

11. Emotional intensity S emotion : Calculate the emotional score of the chat content, including three emotions: positive, neutral, and negative. The calculation is as follows: The text of the chat record is sentiment classified using the Bert model. The input is the chat record at the patient end, and the output is the sentiment intensity at the patient end. The output form is a 3D vector: S emotion = [P pos , P neu , P neg ​ Among them, P pos , P neu , P neg respectively represent the probabilities of the patient's positive, neutral, and negative emotions; 1.

12. Emotional fluctuation intensity ΔEmotion: After receiving a new chat record, the degree of change in the patient's emotions. The calculation formula is: ΔEmotion = |S emotion (t) - S emotion (t - 1)| Among them, t represents the current time; t - 1 represents the previous time; S emotion (t) represents the emotional score of the patient at the t-th moment; S emotion (t - 1) represents the emotional score of the patient at the (t - 1)-th moment; "| |" represents the absolute value; 1.

13. Negative emotion ratio R neg : The proportion of negative emotions in the total emotions within a certain time window. The calculation formula is as follows: Among them, N neg represents the total number of occurrences of negative emotions; N total represents the total number of occurrences of all emotions; 1.

2. The patient sensitive word feature includes the number of occurrences of sensitive words and the sensitive word density, where 1.

21. Number of occurrences of sensitive words S sensitive : The number of sensitive words in the chat record; by collecting historical data, constructing a sensitive word library related to medical disputes, retrieving the online communication text content sent by patients, and counting the number of sensitive words n sensitie_i that appear in each text content, and each occurrence is recorded as 1 point; where n represents the total number of sensitive words; 1.

22. Sensitive word density D sensitive : The ratio of the total number of sensitive words to the total number of words in the message. The calculation formula is: 1.

3. The communication behavior characteristics: By analyzing the message sending frequency λ msg , the message sending rhythm σ interval , the high-density message sending D high , identify the behavior pattern of the patient during the communication process, and the calculation method is as follows: 1.

31. Message sending frequency λ msg : The message sending frequency per unit time, and the calculation formula is as follows: Among them, N messages represents the number of messages sent within the time window; T window represents the time window; 1.

32. Message sending rhythm σ interval : It measures the regularity of the time intervals between message sendings by patients or doctors during a conversation. The calculation formula is as follows: Among them, σ interval represents the standard deviation of the message sending rhythm; N represents the total number of messages that have been sent so far; t i -t i-1 represents the time interval between two adjacent messages; is the average value of the message intervals; 1.

33. High-density message sending D high : It refers to the situation where a patient sends multiple messages continuously within an extremely short period of time. The calculation formula is as follows: Among them, T high is the set time window; N high represents the number of messages continuously sent within the time period T high ; (2), the dynamic characteristic factors of the doctor side including: emotion-related characteristics, reply behavior characteristics, message sending behavior characteristics; 2.

1. Emotional-related features include: the proportion of doctors' negative emotions S doctor_neg , the emotional fluctuations of doctors S doctor_fluctuation , the proportion of doctors' negative emotions R doc_neg ; among which, 2.11 Proportion S of Doctors' Negative Emotions doctor_neg The calculation method is as follows: The text of the chat record is sentiment classified using the Bert model. The input is the chat record at the doctor end, and the output is the sentiment intensity at the doctor end. The output form is a 3D vector: S doctor_emotion = [P' pos , P' neu , P' neg ​ Among them, P′ pos , P′ neu , P′ neg respectively represent the probabilities of the doctor's positive, neutral, and negative emotions; 2.

12. Doctor's Emotional Fluctuation S doctor_fluctuation The calculation formula is as follows: S doctor_fluctuation = |S′ emotion (t) - S′ emotion (t - 1)| Among them, S' emotion (t) represents the emotional score of the doctor at the t-th moment; S' emotion (t - 1) represents the emotional score of the doctor's previous reply; "| |" represents the absolute value; 2.

13. Proportion R of negative emotions of doctors doc_neg The calculation formula is Among them, N' neg represents the number of negative emotional responses from doctors; N' total represents the total number of emotional responses from doctors; 2.

2. Reply behavior characteristics include: reply speed T response , timeout reply ratio R timeout , average number of replies L reply , short reply ratio R short ; 2.

21. Response speed T response The calculation formula is as follows: T response = t reply - t query where t reply represents the timestamp of the doctor's reply message; t query represents the timestamp of the patient's sent message; 2.

22. Timeout reply ratio R timeout The calculation formula is as follows: Among them, N timeout represents the number of timeout replies; N total represents the total number of replies from the doctor; 2.

23. Average number of replies L reply The calculation formula is as follows: Among them, M i represents the i-th reply content of the doctor; len(M i ) represents the number of words in the reply of M i ; N total represents the total number of replies of the doctor; 2.

24. Short reply ratio R short The calculation formula is as follows: Among them, N short represents the number of responses with a word count lower than the threshold; 2.

3. The message sending behavior characteristics include the doctor's sending frequency F doc , the doctor's message sending rhythm V doc ; 2.

31. Doctor sending frequency F doc The calculation formula is as follows: Among them, N msg represents the total number of messages sent by the doctor within the unit time T; 2.

32. Doctor message sending rhythm V doc The calculation formula is as follows: V doc = std({t i - t i-1 ) where t i represents the timestamp when the doctor sends the i-th message; std represents the standard deviation.

6. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 1, characterized in that: In the dynamic feature analysis module, the patient end risk prediction model uses BERT and BiLSTM to model the patient end dialogue emotions. The operation is as follows: (1). BERT is used for semantic feature extraction: Use the BERT pre-trained model to obtain the deep semantic features of the patient's text; (2). BiLSTM is used to capture temporal information: Learn the context relationship in the dialogue sequence; The patient - side risk prediction model includes the dynamic monitoring features on the patient side, which are respectively: emotional intensity S emotion , intensity of emotional fluctuation ΔEmotion, proportion of negative emotions R neg , number of occurrences of sensitive words S sensitive , density of sensitive words D sensitive , message sending frequency λ msg , message sending rhythm σ interval , high - density message sending D high ; The model output is the patient's dynamic prediction score S dynamic-patient ; S dynamic-patient ＝ f(S emotion , ΔEmotion, R neg , S sensitive , D sensitive , λ msg , σ interval , D high )。 7. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 1, wherein: In the dynamic feature analysis module, the doctor end risk prediction model uses the GNN and Transformer structures to model the doctor's communication behavior. The operation is as follows: GNN is used to model the doctor's communication pattern. In the GNN structure, the communication relationship between the doctor and the patient is modeled as a graph structure G(D, P, E): Node: Doctor D i Each doctor is a node; Patient P j Indicates that the patients interacting with the doctor are another type of node; Edge: E(D i， P j ) represents the communication relationship between doctors and patients. The weight of the edge is based on the emotional changes of doctors, that is, the proportion S of doctors' negative emotions doctor_neg , the emotional fluctuation S of doctors doctor_fluctuation , the proportion R of doctors' negative emotions doc_neg , the sending frequency F of doctors doc ; Through the aggregation calculation of GNN, the system learns the doctor's communication behavior and judges whether it has a high-risk communication pattern; The Transformer structure is used to encode the text features of the doctor's reply; The doctor - side risk prediction model includes the dynamic monitoring features on the doctor side, which are respectively: the proportion of doctor's negative emotions S doctor_neg , the emotional fluctuation of the doctor S doctor_fluctuation , the proportion of doctor's negative emotions R doc_neg , the response speed T response , the proportion of overtime responses R timeout , the average number of responses L reply , the proportion of short responses R short , the sending frequency of the doctor F doc , the message - sending rhythm of the doctor V doc , and the model output is the doctor's dynamic prediction score S dynamic-doctor ; S dynamic-doctor = XGBoost(GNN(S doctor_neg , S doctor_fluctuation , R doc_neg , F doc ), Transformer(T response ,R timeout ,L reply ,R short ))。 8. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 1, characterized in that: In the risk fusion prediction module, XGBoost learns through the cascading of multiple decision trees. Each tree is optimized based on the wrong predictions of the previous round. The process is as follows: (1). Initialize the base model: Set the initial prediction value, usually the mean of the target variable in the dataset, and calculate the initial loss function to measure the deviation between the prediction value and the true value; (2). Calculate the residuals: Calculate the prediction error of each sample, that is, the residuals; Among them, R i represents the residual; y i represents the true value; represents the predicted value of the current model; (3) Training a new decision tree: The new decision tree learns the residual R from the previous round i , and the current prediction result is expressed as: Among them, represents the predicted value of the t-th tree; represents the predicted values of the first t - 1 predicted trees; f t (X) represents the prediction result of the current decision tree; α is the learning rate; (4). Iterative update: Continuously add new decision trees, and optimize the mistakes of the previous round each time. Finally, the model converges to an optimal prediction result; (5) Calculate the final risk score: After multiple rounds of iteration, the final doctor-patient communication risk score R is obtained. final ; R final ＝ f(S dynamic-patient ,S dynamic-doctor ,H complaint-patient ,H complaint-doctor ,S base )。 9. The doctor-patient communication risk assessment and early warning system based on the Internet hospital according to claim 8, wherein: During the process of model training, the doctor-patient communication risk score R final is normalized so that the value range of R final is between 0 and 100; Among them, R raw represents the original predicted value of XGBoost; R min and R max represent the minimum and maximum predicted values in the training data, respectively.

10. A method for evaluating and warning doctor-patient communication risks based on an Internet hospital, which includes the following steps: S1. Collect the online consultation data of doctors and patients, and obtain the patient consultation data and doctor consultation data included in the consultation data, and calculate the basic feature score S of this warning base , and calculate the weight ratio NormFWR; the basic feature score S of this warning base According to the patient's disease sensitivity C severity , the integrity C of the patient's medical record completeness , the doctor's experience level D experience , the online score D of the doctor score Calculate based on the fixed feature values of S2. Adopt a continuous data collection mechanism to dynamically analyze the real-time communication content between doctors and patients, collect dynamic characteristic factors on the patient side and the doctor side, and construct a prediction model on the patient side and a prediction model on the doctor side. Use the dynamic characteristic factors on the patient side and the doctor side as the inputs of the prediction model on the patient side and the prediction model on the doctor side respectively, and calculate the patient dynamic prediction score S dynamic-patient and the doctor dynamic prediction score S dynamic-doctor ; S3. Use the patient's dynamic prediction score S dynamic-patient , the doctor's dynamic prediction score S dynamic-doctor , the patient's previous complaint times, the doctor's previous complaint times, and the basic feature score S base as the input of the constructed fusion prediction model, use XGBoost for feature fusion and final risk prediction, and calculate the final doctor-patient communication risk score; S4. When the final doctor-patient communication risk score exceeds the set threshold, the system triggers the intelligent warning mechanism.

Images