Deep learning-based burn patient fluid infusion prediction method and system, and storage medium

Abstract

The invention relates to the field of medical artificial intelligence, in particular to a burn patient fluid infusion prediction method and system based on deep learning and a storage medium, and the method comprises the steps that a burn patient fluid infusion prediction model based on deep learning is obtained, and the prediction model takes baseline data and time sequence data as input and takes fluid infusion data as output; the baseline data comprises physique data and burn condition data of the patient, and the time sequence data comprises vital sign data and diagnosis and treatment inspection data of the patient; and acquiring baseline data and time sequence data of a target patient, inputting the baseline data and the time sequence data into the burn patient fluid infusion prediction model based on deep learning, and outputting predicted fluid infusion data of the target patient. The system can comprehensively consider multi-dimensional data of a patient, automatically generate prediction suggestions of liquid types and fluid infusion flow rates, and dynamically adjust a fluid infusion scheme according to real-time data changes, so that accurate and effective fluid infusion treatment is provided for the burn patient, and the optimal treatment effect is achieved.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence, and particularly to a method, system and storage medium for predicting fluid replacement for burn patients based on deep learning. Background Art

[0002] Burn is one of the common types of injuries in wartime. Especially in emergencies such as fires, explosions, and chemical leaks, the incidence of burns is more significant. According to statistical data, a fire occurs in China every 40 seconds on average, and group severe burn incidents are not uncommon. In modern warfare, more than 70% of the wounded will have varying degrees of burns. The mortality rate of severe burns can be as high as 30% - 85%, and the disability rate is as high as 80% - 100%. Its harmfulness is huge and has a profound impact on society. Therefore, how to effectively improve the treatment effect of burn patients, especially early shock resuscitation, has become an urgent problem to be solved in the medical field.

[0003] Burn patients often experience severe shock in the initial stage, which is one of the main causes of death. Timely treatment of shock is crucial for the survival rate and subsequent rehabilitation of burn patients. However, traditional burn fluid replacement plans rely on experienced doctors to manually adjust according to the patient's clinical situation. Traditional fluid replacement plans usually rely on doctor experience and are difficult to achieve personalization and real-time adjustment. Especially in a complex clinical environment, doctors need to consider multiple factors (such as age, weight, burn area, vital signs, etc.) to determine the type of fluid replacement and the fluid replacement rate. Most existing fluid replacement advice systems are based on fixed formulas or rules and cannot dynamically adapt to the special needs of each patient. Therefore, there is an urgent need for an intelligent and highly real-time fluid replacement decision support method to help doctors accurately and quickly formulate fluid replacement plans and improve the accuracy and timeliness of treatment. Summary of the Invention

[0004] The purpose of the present invention is to provide a method, system and storage medium for predicting fluid replacement for burn patients based on deep learning, which can comprehensively consider the multi-dimensional data of patients, automatically generate prediction suggestions for the type of fluid and the fluid replacement flow rate, and can dynamically adjust the fluid replacement plan according to real-time data changes, so as to provide accurate and effective fluid replacement treatment for burn patients to achieve the best treatment effect.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] In the first aspect, the present invention provides a method for predicting fluid replacement for burn patients based on deep learning, which includes:

[0007] Obtain a burn patient fluid replacement prediction model based on deep learning. This prediction model takes baseline data and time-series data as inputs and fluid replacement data as outputs. The baseline data includes the patient's physical data and burn situation data, and the time-series data includes the patient's vital sign data and diagnosis and treatment test data;

[0008] Collect and obtain the baseline data and time-series data of the target patient, and input them into the burn patient fluid replacement prediction model based on deep learning to output the predicted fluid replacement data of the target patient.

[0009] Furthermore, the patient's physical data includes gender, age, height, weight, and BMI index;

[0010] The patient's burn situation data includes the injury time, admission time, burn area, and inhalation injury;

[0011] The patient's vital sign data includes body temperature, heart rate, respiration, blood pressure, central venous pressure, peripheral oxygen saturation, and urine output;

[0012] The patient's diagnosis and treatment test data includes blood routine test data, blood biochemical test data, blood gas analysis test data, liver and kidney function test data, C-reactive protein test data, procalcitonin test data, and coagulation function test data;

[0013] The fluid replacement data includes the type of fluid and the fluid replacement flow rate.

[0014] Furthermore, obtaining the burn patient fluid replacement prediction model based on deep learning specifically includes:

[0015] Obtain multiple sample data sets. A single sample data set includes baseline data, time-series data, and fluid replacement data;

[0016] Divide the multiple sample data sets into a training set, a test set, and a validation set;

[0017] Construct a basic neural network model based on deep learning, substitute the data of the training set into the basic neural network model for training, and use the test set and the validation set to test and validate the model to obtain a burn patient fluid replacement prediction model based on deep learning that meets the preset requirements.

[0018] Furthermore, the basic neural network model includes a first encoder, a second encoder, a decoder, and an output module;

[0019] Extract and fuse the features of the baseline data through the first encoder, and output the baseline feature vector;

[0020] Extract and fuse the features of the time-series data through the second encoder, and output the time-series feature vector;

[0021] The feature representation obtained by concatenating the baseline feature vector and the time series feature vector is input into a decoder for decoding to obtain an output feature vector;

[0022] The output feature vector is processed by an output module to obtain predicted fluid replacement data.

[0023] Furthermore, the feature extraction and fusion of the baseline data by the first encoder to output the baseline feature vector specifically includes:

[0024] Input the baseline data into the first encoder, and after self-attention calculation, obtain a weighted output and perform normalization processing to obtain an intermediate result;

[0025] Use the intermediate result as the new input of the first encoder and perform self-attention calculation again to obtain the baseline feature vector.

[0026] Furthermore, the feature extraction and fusion of the time series data by the second encoder to output the time series feature vector specifically includes:

[0027] Divide each feature vector in the time series data according to a preset time step, and the divided time series data is expressed as:

[0028] where N i is the time series data corresponding to the i-th preset time step, and t is the total number of divisions;

[0029] Use the MLP layer to map to where T is the time step, N is the number of original data features, is the input variable after mapping, H is the latent feature dimension, and perform independent positional encoding on each single feature variable;

[0030] Input the input variable after mapping into the second encoder, and perform feature extraction and fusion on the input variable after mapping through multiple stacked Transformer modules in the second encoder to output the time series feature vector.

[0031] Furthermore, the output feature vector is processed by the output module to obtain the predicted fluid replacement data specifically including:

[0032] Determination of liquid types: The multi-layer perceptron model is used to reduce the dimension of the output feature vector, obtaining several intermediate vectors corresponding to different liquid types. Each liquid type includes several sub-types. The dimension of the intermediate vector corresponding to a certain liquid type is equal to the total number of sub-types corresponding to that liquid type. Several elements in the intermediate vector respectively correspond to several sub-types of the liquid type. Then, the intermediate vector is input into the Softmax network for probability distribution prediction to obtain the probabilities corresponding to each element in each intermediate vector. The element with the highest probability is used as the output item, and the sub-type corresponding to the element with the highest probability is selected as the fluid replacement.

[0033] Determination of fluid replacement flow rate: The multi-layer perceptron model is used to process the output feature vector to obtain the predicted values of the fluid replacement flow rates corresponding to different liquid types. The Sigmoid activation function is used to standardize the predicted values of the fluid replacement flow rates output to [0, 1]. Then, the standardized predicted values of the fluid replacement flow rates are multiplied by the preset flow rate to obtain the output values of the fluid replacement flow rates corresponding to each liquid type.

[0034] Furthermore, the time series data further includes the actual fluid replacement data of the patient in the previous diagnosis and treatment cycle.

[0035] In a second aspect, the present invention provides a fluid replacement prediction system for burn patients based on deep learning, which includes:

[0036] The first acquisition unit is used to acquire a fluid replacement prediction model for burn patients based on deep learning. This prediction model takes the baseline data and time series data as inputs and the fluid replacement data as outputs. The baseline data includes the physical data and burn condition data of the patient, and the time series data includes the vital sign data and diagnosis and treatment test data of the patient.

[0037] The second acquisition unit is used to collect and acquire the baseline data and time series data of the target patient.

[0038] The prediction unit is used to input the collected baseline data and time series data of the target patient into the fluid replacement prediction model for burn patients based on deep learning, and output the predicted fluid replacement data of the target patient.

[0039] In a third aspect, the present invention provides a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to cause a computer to execute the above-mentioned fluid replacement prediction method for burn patients based on deep learning.

[0040] The present invention has the following unexpected beneficial effects.

[0041] 1. The burn patient fluid replacement prediction model based on deep learning according to the present invention integrates various data sources such as the patient's physical data, burn condition data, vital sign data, and diagnosis and treatment test data. Compared with the traditional single-data input model, this prediction model can provide a more comprehensive and accurate decision-making basis, significantly improving the accuracy and personalization of fluid replacement prediction and avoiding decision-making errors caused by insufficient or biased data. Moreover, traditional decision-making models usually rely on static or outdated patient data, while the burn patient fluid replacement prediction model based on deep learning according to the present invention can dynamically adjust the fluid replacement decision by obtaining and updating the patient's vital signs and diagnosis and treatment test data in real time. The ability to adjust in real time ensures that the model can accurately respond to changes in the patient's condition and provide timely and effective interventions. Through this dynamic adjustment mechanism, the response speed to changes in the condition of burn patients and the timeliness of decision-making are significantly improved, ensuring that the treatment plan always matches the patient's real-time state.

[0042] 2. The burn patient fluid replacement prediction method based on deep learning according to the present invention has strong compatibility and can be deployed on various platforms and systems and can quickly adapt to different treatment environments and hardware devices. Whether in the emergency room, intensive care unit or mobile medical device, the model can achieve seamless docking, ensuring popularization and application in a variety of medical environments. This high compatibility enables this model to be not only applied in specific medical institutions but also widely promoted in various treatment environments, improving the accuracy and standardization of the treatment of burn patients. The burn patient fluid replacement prediction method based on deep learning can not only provide an efficient and accurate fluid replacement plan for medical institutions at all levels but also provide a convenient auxiliary decision-making tool for doctors and medical staff of different specialties, helping to improve the overall medical service level. Especially in complex scenarios such as emergency treatment and handling multiple patients, it plays an important role. Compared with traditional methods, the prediction method according to the present invention can quickly provide accurate and standardized fluid replacement plans for medical staff with different professional backgrounds and different treatment levels in medical institutions at all levels and can adjust the fluid replacement strategy in real time, significantly improving the accuracy, standardization and traceability of burn shock fluid replacement.

[0043] 3. The introduction of the burn patient fluid replacement prediction method based on deep learning according to the present invention greatly reduces the workload of rescue personnel. Especially in high-intensity emergency scenarios, through automated and intelligent decision support, the burden on medical staff in complex decision-making is reduced. Medical staff can concentrate on dealing with the core treatment links, avoiding decision-making mistakes caused by lack of experience or fatigue, thereby improving the overall treatment quality. Especially in the environment with limited primary medical conditions, the application of the burn patient fluid replacement prediction method based on deep learning according to the present invention can effectively improve the decision-making level of doctors, improve the treatment efficiency, and ensure that burn patients receive timely and accurate treatment. Description of the Drawings

[0044] Figure 1 It shows a schematic flowchart of the burn patient fluid replacement prediction method based on deep learning according to an embodiment of the present invention.

[0045] Figure 2 It shows a schematic flowchart of the construction process of the burn patient fluid replacement prediction model based on deep learning according to an embodiment of the present invention.

[0046] Figure 3 It shows a schematic structural diagram of the burn patient fluid replacement prediction system based on deep learning according to an embodiment of the present invention. Detailed implementation manners

[0047] The following will describe the implementation manners of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for explaining the present invention and not for limiting the protection scope of the present invention.

[0048] In one embodiment, as shown in Figure 1 the present invention provides a burn patient fluid replacement prediction method based on deep learning, which includes:

[0049] Obtain a burn patient fluid replacement prediction model based on deep learning. This prediction model takes baseline data and time-series data as inputs and fluid replacement data as outputs. The baseline data includes the physical data and burn condition data of the patient, and the time-series data includes the vital sign data and diagnosis and treatment test data of the patient.

[0050] Collect and obtain the baseline data and time-series data of the target patient, and input them into the burn patient fluid replacement prediction model based on deep learning to output the predicted fluid replacement data of the target patient.

[0051] The burn patient fluid replacement prediction method based on deep learning described in the present invention has strong compatibility and can be deployed on various platforms and systems, and can quickly adapt to different treatment environments and hardware devices. Whether in the emergency room, intensive care unit or mobile medical equipment, the model can achieve seamless docking to ensure popularization and application in various medical environments. This high compatibility enables this model to be not only applied in specific medical institutions, but also widely promoted in various treatment environments, improving the accuracy and standardization of burn patient treatment. The burn patient fluid replacement prediction method based on deep learning can not only provide efficient and accurate fluid replacement plans for medical institutions at all levels, but also provide a convenient auxiliary decision-making tool for doctors and medical staff of different specialties, helping to improve the overall medical service level. Especially in complex scenarios such as emergency treatment and multi-patient handling, it plays an important role. Compared with traditional methods, the prediction method described in the present invention can quickly provide accurate and standardized fluid replacement plans for medical staff with different professional backgrounds and treatment levels in medical institutions at all levels, and adjust the fluid replacement strategy in real time, significantly improving the accuracy, standardization and traceability of burn shock fluid replacement.

[0052] The introduction of the burn patient fluid replacement prediction method based on deep learning described in the present invention greatly reduces the workload of rescue personnel. Especially in high-intensity first-aid scenarios, through automated and intelligent decision support, the burden on medical staff in complex decision-making is reduced. Medical staff can concentrate on dealing with the core treatment links, avoiding decision-making mistakes caused by lack of experience or fatigue, thereby improving the overall treatment quality. Especially in environments with limited primary medical conditions, the application of the burn patient fluid replacement prediction method based on deep learning described in the present invention can effectively improve the decision-making level of doctors, improve the treatment efficiency, and ensure that burn patients receive timely and accurate treatment.

[0053] As a preferred implementation manner of the present invention, the physical data of the patient includes gender, age, height, weight and BMI index.

[0054] Specifically, there are differences in physiological structures and functions between men and women. For example, men usually have a higher muscle content than women, while women have a relatively higher fat content. Such differences will affect the distribution and metabolism of drugs in the body, and also affect the physiological responses and fluid replacement needs after burns. With different genders, the hormone levels in the body are also different, and hormones play an important role in processes such as inflammatory responses and metabolic regulation after burns, thereby affecting the fluid replacement needs and strategies.

[0055] Patients of different age groups have different organ functions and reserve capacities. The organ functions of children are not yet fully developed, while those of the elderly gradually decline. This makes their tolerance to burns and recovery ability different from that of adults, and the organ's tolerance and regulatory ability to fluids need to be considered during fluid replacement. Moreover, age is closely related to the basal metabolic rate. Generally speaking, the younger the age, the higher the basal metabolic rate, and the degree of energy consumption and metabolic disorder after burns may be more severe. It is necessary to adjust the nutritional components and fluid volume in the fluid replacement according to age.

[0056] Height and weight are important parameters for calculating the body surface area, and the fluid replacement volume for burn patients is usually related to the body surface area. For example, in commonly used burn fluid replacement formulas, the calculation of the fluid replacement volume often takes the body surface area per square meter as the basis. Height and weight also affect the dosage of drugs. Various drugs, such as antibiotics and electrolytes, may be added in the fluid replacement, and it is necessary to accurately calculate the drug dosage according to the patient's weight to ensure the safety and effectiveness of the treatment.

[0057] The BMI index can reflect the nutritional status of the patient. A too low BMI may indicate that the patient is malnourished and has poor resistance and recovery ability to burns, and it is necessary to appropriately increase the supplement of nutrients in the fluid replacement; a too high BMI may suggest that the patient is obese. Obese patients have more adipose tissue, and the risk of complications such as fat liquefaction and infection after burns is relatively high, and the distribution of drugs in the body may also be affected. These factors need to be considered during fluid replacement to adjust the type and speed of the fluid replacement. The BMI index is also related to the patient's metabolic status. Obese patients may have metabolic problems such as insulin resistance, and the metabolic disorder after burns may be more complex. It is necessary to adjust the use of drugs such as insulin in the fluid replacement and the supply of nutrients such as glucose according to the BMI index.

[0058] The burn situation data of the patient include the injury time, admission time, burn area, and inhalation injury.

[0059] Specifically, obtaining the injury time can help determine which stage the patient is in after burns. At different time periods after burns, the pathophysiological changes of the patient are different. For example, in the early stage of burns (within 24 - 48 hours after injury), a large amount of fluid leaks from the blood vessels into the tissue space, which is a high-incidence period of shock. The main requirement for fluid replacement is to quickly replenish the blood volume and correct shock; in the late stage of burns, more attention may be paid to aspects such as infection prevention and treatment and nutritional support in fluid replacement. And combined with other factors, it can assist in evaluating the severity of the damage caused by burns to the body. Patients with a long injury time and without timely treatment may have more serious tissue damage, infections and other complications, and the fluid replacement plan needs to be more complex. In addition to replenishing fluids, it is also necessary to consider anti-infection and correcting electrolyte disorders, etc.

[0060] Obtaining the admission time can clarify the time interval from the burn to the start of formal treatment. If the admission time is short after the injury, the patient has a relatively greater chance of recovery after active fluid replacement and other treatments. If the admission time is late, serious complications may have occurred, such as shock and organ dysfunction. At this time, when replenishing fluids, not only the lost fluids need to be replaced, but also organ function protection needs to be considered. The speed and type of fluid replacement should be finely adjusted according to the specific situation. Starting from the admission time, considering factors such as the injury time and burn area, the fluid replacement volume at different stages after the burn can be calculated. For example, when calculating the fluid replacement volume in the first 24 hours after the burn, it needs to be calculated from the injury time. However, in actual clinical practice, it is more important to use the admission time as a reference to determine the subsequent fluid replacement plan, and the fluid replacement volume and speed should be adjusted according to the remaining time.

[0061] The burn area is a key factor in predicting the fluid replacement volume. Generally speaking, the larger the burn area, the more fluid is lost and the greater the fluid replacement volume. Patients with extensive burns need a faster fluid replacement speed in the early stage of the burn to correct shock and maintain effective circulating blood volume. For example, patients with a burn area exceeding 50% may need to rapidly infuse a large amount of fluid in a short time to ensure blood perfusion of important organs.

[0062] Inhalation injury can cause congestion, edema of the respiratory mucosa, and even pulmonary edema, affecting gas exchange and respiratory function. In addition to the fluid replacement needs for the body surface burn, for such patients, the fluid exudation and inflammatory response caused by the respiratory tract injury also need to be considered. When replenishing fluids, more caution is required to avoid aggravating pulmonary edema due to excessive fluid replacement. Since inhalation injury may be accompanied by risks of complications such as pulmonary infection, antibiotics and other drugs may need to be used earlier during fluid replacement, and the speed and type of fluid replacement should be adjusted according to the pulmonary condition, such as appropriately increasing the proportion of colloid solution to reduce the risk of pulmonary edema. At the same time, the respiratory tract should be kept unobstructed, and mechanical ventilation and other supportive treatments may be required when necessary, all of which interact with the fluid replacement strategy.

[0063] The vital sign data of the patient include body temperature, heart rate, respiration, blood pressure, central venous pressure, peripheral oxygen saturation, and urine output.

[0064] Specifically, body temperature can reflect the condition and metabolism: an increase in body temperature may be caused by the inflammatory response or infection after the burn, which will increase the body's metabolic rate and water evaporation, resulting in an increased need for fluid replacement. For every 1°C increase in body temperature, the insensible water loss of the body will increase by about 13%. If the body temperature is too low, it may indicate shock or poor circulation in the patient. At this time, attention should be paid to keeping warm and improving circulation during fluid replacement. During the fluid replacement process, the gradual return of body temperature to normal usually indicates that the fluid replacement and treatment measures are effective, and the internal environment and metabolism of the body are gradually stabilizing; if the body temperature remains abnormal, the fluid replacement plan may need to be adjusted, such as increasing the fluid replacement volume or adjusting the fluid replacement components.

[0065] Heart rate is an important indicator reflecting the circulatory status. After burns, a decrease in blood volume can lead to compensatory tachycardia. Generally, a heart rate exceeding 120 beats per minute often indicates insufficient blood volume, and it is necessary to increase the fluid infusion rate to correct shock and maintain effective perfusion of the heart. If the heart rate gradually decreases and tends to return to the normal range after fluid infusion, it indicates that the amount and rate of fluid infusion are appropriate and the circulation has improved; if the heart rate continues to increase or arrhythmia occurs, it may indicate insufficient or excessive fluid infusion, and further adjustment of the fluid infusion strategy is required.

[0066] Respiration can be used to judge ventilation and oxygenation: changes in respiratory rate, depth, and rhythm can reflect the patient's ventilation and oxygenation functions. When burn patients are complicated with inhalation injury or pulmonary edema, respiration becomes rapid and difficult. At this time, fluid infusion should be cautious to avoid aggravating pulmonary edema, and at the same time, it may be necessary to increase oxygen supply or take measures such as mechanical ventilation. Abnormal respiration may also be related to acid-base balance disorders. For example, in metabolic acidosis, patients may present with deep and rapid breathing. During fluid infusion, it is necessary to adjust the composition of the infused fluid according to the results of blood gas analysis to correct acid-base balance disorders.

[0067] Blood pressure is an important indicator reflecting circulatory blood volume and cardiac function. After burns, a decrease in blood pressure, especially when the systolic blood pressure is lower than 90 mmHg, often indicates severe insufficient blood volume, and rapid and large amounts of fluid infusion are required to increase blood pressure and ensure blood perfusion of vital organs. During fluid infusion, if the blood pressure gradually rises and stabilizes within the normal range, it indicates that the fluid infusion is effective; if the blood pressure remains low or fluctuates greatly, it may be necessary to continue increasing the amount of fluid infusion or adjust the type of infused fluid, such as appropriately increasing the use of colloidal fluid or vasoactive drugs.

[0068] Central venous pressure can reflect the pressure in the right atrium and the blood volume situation. The normal range is 5 - 12 cmH2O. A too low central venous pressure indicates insufficient blood volume and requires active fluid infusion; a too high central venous pressure may indicate heart failure or excessive fluid infusion, and it is necessary to control the rate and amount of fluid infusion and further evaluate cardiac function. According to the relationship between central venous pressure and blood pressure, the fluid infusion plan can be adjusted more accurately. For example, a low central venous pressure and a low blood pressure indicate severe insufficient blood volume, and rapid and large amounts of fluid infusion should be given; a high central venous pressure and a low blood pressure indicate heart failure, and fluid infusion should be restricted and cardiac stimulants should be given.

[0069] Peripheral oxygen saturation reflects the oxygen content in the blood and the oxygen supply to tissues. A decrease in peripheral oxygen saturation in burn patients may be due to oxygenation disorders or insufficient tissue perfusion caused by inhalation injury, pulmonary edema, shock, etc. At this time, while actively improving ventilation, it is necessary to adjust the fluid infusion plan to ensure blood perfusion and oxygen supply to tissues. During fluid infusion and treatment, if the peripheral oxygen saturation gradually increases and maintains within the normal range, it indicates that the fluid infusion and treatment measures are effective and the tissue oxygen supply has improved; if the peripheral oxygen saturation does not rise or continues to decline, it may be necessary to further optimize the fluid infusion plan and strengthen treatment measures such as respiratory support.

[0070] Urine output is the most intuitive and sensitive indicator reflecting renal perfusion and the effect of fluid replacement. For burn patients, the urine output should be maintained above 30 - 50 ml per hour. If the urine output decreases, it often indicates insufficient blood volume and poor renal perfusion, and additional fluid replacement is needed. By monitoring urine output, the fluid replacement rate can be adjusted in a timely manner. If the urine output is excessive, it may indicate over - fluid replacement or abnormal renal function, and the fluid replacement rate and volume need to be appropriately controlled; if the urine output is too low, the fluid replacement rate should be increased to maintain normal renal function and internal environment stability.

[0071] The patient's diagnosis and treatment test data include blood routine test data, blood biochemical test data, blood gas analysis test data, liver and kidney function test data, C - reactive protein test data, procalcitonin test data, and coagulation function test data.

[0072] Specifically, the blood routine test data include red blood cell count, hemoglobin and hematocrit, white blood cell count and classification, and platelet count. Red blood cell count, hemoglobin and hematocrit: After burns, due to factors such as blood concentration or blood loss, these indicators may change. If the values increase, it indicates blood concentration, and additional fluid replacement may be needed to dilute the blood and improve circulation; if the values decrease, there may be blood loss or anemia, and in addition to fluid replacement, blood transfusion and other treatments may also need to be considered. White blood cell count and classification: An elevated white blood cell count often indicates infection or inflammatory response. Infection can lead to increased body metabolism and increased water demand, and this factor needs to be considered during fluid replacement, and antibiotics may also be needed. An increased proportion of neutrophils may indicate bacterial infection, and abnormal lymphocyte proportion may be related to viral infection or immune function. Different situations have different impacts on fluid replacement and treatment plans. Platelet count: Thrombocytopenia may be related to abnormal coagulation function, blood dilution or bone marrow suppression after burns. During the fluid replacement process, pay attention to observing platelet changes. If the platelets are too low, it may increase the bleeding risk, and the fluid replacement plan needs to be adjusted, and platelet transfusion may be necessary when necessary.

[0073] Blood biochemical test data: The levels of electrolytes such as potassium, sodium, and chloride are crucial for fluid replacement. Burn patients may experience electrolyte disorders due to massive exudation, vomiting, diarrhea, etc. For example, hyperkalemia may be related to tissue damage, renal insufficiency, etc. During fluid replacement, large amounts of potassium - containing fluids should be avoided, and measures should be taken to lower blood potassium; hyponatremia may require appropriate supplementation of hypertonic saline for correction. The stress response after burns can lead to elevated blood sugar. During fluid replacement, large amounts of glucose solutions should be avoided, and insulin should be used to control blood sugar according to blood sugar levels. If the blood sugar is too low, it may indicate poor nutritional status or other metabolic problems of the patient, and appropriate glucose supplementation is needed in the fluid replacement.

[0074] The blood gas analysis test data specifically include:

[0075] pH: It reflects the body's acid-base balance. Burn patients often suffer from metabolic acidosis due to shock, infection, etc., and the pH value decreases. At this time, it is necessary to consider using alkaline drugs such as sodium bicarbonate to correct acidosis during fluid replacement, while adjusting the composition and speed of fluid replacement to improve tissue perfusion to correct acid-base imbalance.

[0076] Oxygen partial pressure (PaO2) and carbon dioxide partial pressure (PaCO2): A decrease in PaO2 indicates oxygenation dysfunction, which may be related to inhalation injury, pulmonary edema, etc. It is necessary to strengthen respiratory support and increase oxygen supply while rehydrating. An increase in PaCO2 may indicate ventilation dysfunction, and attention should be paid to adjusting respiratory parameters. Mechanical ventilation should be performed when necessary, and rehydration should be done to avoid aggravating respiratory insufficiency.

[0077] Excess alkali (BE) and bicarbonate ions (HCO3 - ): BE and HCO3 - It can reflect the body's alkaline reserve. - If it is low, it indicates metabolic acidosis, and the amount of alkaline drugs to be supplemented needs to be calculated based on the specific values ​​to correct the acid-base imbalance.

[0078] Lactate: The normal range is 0.5-2.2mmol / L. Increased lactate (>2mmol / L) indicates tissue hypoxia or enhanced anaerobic metabolism, which is common in the early stages of burn shock (hypovolemia leading to hypoperfusion), septic shock (microcirculatory disorders) or sepsis. Hyperlactatemia (>4mmol / L) is significantly associated with multiple organ failure (MOF) and increased mortality. During fluid rehydration in burns, dynamic monitoring of lactate clearance (a decrease of >10% per hour indicates effective resuscitation) is used to guide the adjustment of fluid rehydration rate and fluid type (such as adding colloid fluid or vasoactive drugs). It should be noted that burn patients may have interference with test results due to wound exudation, red blood cell destruction or drugs (such as epinephrine), and clinical dynamic analysis is required.

[0079] Liver and kidney function indicators: Increased liver function indicators such as alanine aminotransferase and aspartate aminotransferase may indicate liver damage. When rehydrating, pay attention to the effects of drugs on the liver and avoid using hepatotoxic drugs. Increased kidney function indicators such as creatinine and urea nitrogen indicate that kidney function may be impaired. The amount and speed of rehydration need to be adjusted according to creatinine clearance to avoid increasing the burden on the kidneys.

[0080] Liver and kidney function test data include:

[0081] Liver function indicators: In addition to the liver enzyme indicators mentioned above, indicators such as albumin and globulin are also very important. A decrease in albumin often indicates a decline in liver synthesis function or excessive protein loss. Burn patients may develop hypoproteinemia due to factors such as wound exudation. In this case, colloidal solutions such as albumin can be appropriately supplemented in the fluid replacement to increase the colloid osmotic pressure and reduce tissue edema.

[0082] Kidney function indicators: In addition to creatinine and urea nitrogen, attention should also be paid to urine specific gravity, urine osmolality, etc. An increase in urine specific gravity and urine osmolality indicates enhanced renal concentrating function, which may be due to insufficient blood volume and requires an increase in the fluid replacement volume; if the urine specific gravity and urine osmolality decrease, it may indicate impaired kidney function. When performing fluid replacement, attention should be paid to avoiding aggravating the kidney burden and adjusting the type and speed of fluid replacement according to the kidney function.

[0083] C-reactive protein is a non-specific inflammatory indicator. An increase in C-reactive protein after burns indicates a strong inflammatory response. Inflammation can lead to an accelerated body metabolism, increased consumption of water and nutrients. When performing fluid replacement, appropriate increases in energy and nutrient supplementation should be considered. At the same time, according to the degree of inflammation, enhanced anti-infection treatment may be required. During the fluid replacement and treatment process, a gradual decrease in the C-reactive protein level indicates that the inflammation is under control and the fluid replacement and treatment plan is effective; if the C-reactive protein continues to increase or does not decrease, the treatment plan may need to be adjusted, such as changing antibiotics, increasing the fluid replacement volume, or adjusting the fluid replacement components.

[0084] Procalcitonin significantly increases during bacterial infections and generally does not increase or only slightly increases during non-infectious inflammations. If procalcitonin increases in burn patients, it indicates that there may be a bacterial infection. When performing fluid replacement, appropriate antibiotics should be selected according to the infection situation, and the fluid replacement volume should be appropriately increased to maintain circulatory stability. At the same time, attention should be paid to correcting electrolyte disorders, etc.

[0085] The specific coagulation function test data includes:

[0086] Prothrombin time (PT), activated partial thromboplastin time (APTT): Prolongation of PT and APTT indicates abnormal coagulation function, which may be related to vascular endothelial damage, consumption or dilution of coagulation factors after burns. During the fluid replacement process, excessive fluid replacement should be avoided to prevent excessive blood dilution and aggravation of coagulation dysfunction.

[0087] Fibrinogen: A decrease in fibrinogen may affect coagulation function and increase the risk of bleeding. If the fibrinogen is too low in burn patients, fibrinogen or fresh frozen plasma may need to be transfused. At the same time, during fluid replacement, attention should be paid to maintaining an appropriate coagulation state to avoid bleeding or thrombosis.

[0088] D-dimer: An elevated D-dimer indicates hyperfibrinolysis or thrombosis in the body. For burn patients with elevated D-dimer, it may be necessary to appropriately use anticoagulant drugs or antifibrinolytic drugs in fluid replacement, and at the same time, closely observe the patient's coagulation status and changes in the condition, and adjust the fluid replacement and treatment plans.

[0089] The fluid replacement data includes the type of fluid and the fluid replacement flow rate. Exemplarily, the types of fluids include: crystalloids (lactated Ringer's injection, 0.9% sodium chloride injection, 5% sodium bicarbonate injection, 10% potassium chloride injection), colloids (plasma, human albumin), and water (5% glucose injection).

[0090] The deep learning-based fluid replacement prediction model for burn patients according to the present invention integrates various data sources such as the patient's physical constitution data, burn condition data, vital sign data, and diagnosis and treatment test data. Compared with traditional single-data input models, this prediction model can provide a more comprehensive and accurate decision-making basis, significantly improving the accuracy and personalization of fluid replacement prediction, and avoiding decision-making errors caused by insufficient or biased data. And traditional decision-making models usually rely on static or outdated patient data, while the deep learning-based fluid replacement prediction model for burn patients according to the present invention can dynamically adjust fluid replacement decisions by real-time obtaining and updating the patient's vital sign data and diagnosis and treatment test data. The ability of real-time adjustment ensures that the model can accurately respond to changes in the patient's condition and provide timely and effective interventions. Through this dynamic adjustment mechanism, the response speed to changes in the condition of burn patients and the timeliness of decision-making are significantly improved, ensuring that the treatment plan always matches the patient's real-time state.

[0091] It should be noted that when performing fluid replacement prediction, the weight coefficients of the vital sign data in the time series data and the various indicators in the diagnosis and treatment test data are preset. In fluid replacement prediction, various data indicators are involved, such as heart rate, blood pressure, blood routine, blood biochemical indicators, etc. The importance of different indicators for judging the patient's fluid replacement needs is different. The weight coefficient can quantify this importance. For example, the heart rate may be relatively important in reflecting the patient's circulatory status and then judging the fluid replacement needs, and a relatively high weight coefficient can be assigned; while some relatively less important indicators are assigned lower weight coefficients, so as to clearly define the status of each indicator in fluid replacement prediction. By reasonably setting the weight coefficients, the prediction model can more accurately reflect the actual situation. For example, when a patient shows shock symptoms, indicators such as blood pressure and central venous pressure are crucial for judging the fluid replacement volume. By giving these indicators higher weights, the model can more accurately predict the appropriate fluid replacement volume based on the changes in these key indicators, avoiding prediction biases caused by insufficient consideration of important indicators or excessive attention to less important indicators, thereby improving the overall accuracy of fluid replacement prediction.

[0092] Fluid replacement prediction requires comprehensive consideration of multi-source data such as vital sign data and diagnostic and treatment test data. The introduction of weight coefficients helps to integrate these different types and dimensions of data, enabling the model to evaluate the patient's condition as a whole. For example, data such as heart rate and blood pressure in vital signs are combined with data such as hemoglobin and electrolytes in blood tests according to their respective weight coefficients to obtain a more comprehensive prediction result that can better reflect the patient's true fluid replacement needs, achieving the effective integration and utilization of multi-source data.

[0093] In clinical practice, doctors can quickly determine which indicators need to be focused on in the current fluid replacement decision based on the weight coefficients. When abnormal values occur for high-weight indicators, doctors need to consider the fluid replacement plan more carefully. Moreover, the weight coefficients can provide a quantitative reference standard for doctors, helping them make more scientific and reasonable fluid replacement decisions in the face of complex patient conditions and a large amount of data, reducing the influence of subjective factors, and improving the scientificity and standardization of clinical decisions. Also, when constructing and optimizing the fluid replacement prediction model, the weight coefficients are an important basis for evaluating the model performance and conducting optimization. By adjusting the weight coefficients and observing the changes in the model prediction results, the optimal weight combination can be found to make the model achieve the best prediction effect. At the same time, the changes in the weight coefficients can also reflect the model's dependence on different indicators, helping researchers discover potential problems and deficiencies in the model, and then improving and perfecting the model accordingly.

[0094] As a preferred embodiment of the present invention, refer to Figure 2 As shown, the steps for obtaining a fluid replacement prediction model for burn patients based on deep learning are specifically as follows:

[0095] S11, obtain multiple sample data sets, and a single sample data set includes baseline data, time series data, and fluid replacement data.

[0096] Specifically, the sample data sets can be obtained from multiple channels, such as the burn department medical record system of large general hospitals, professional burn specialty hospital databases, etc. These data are usually collected by medical staff through various examination and monitoring means during the patient's visit.

[0097] Baseline data: includes the patient's physical data, such as gender, age, height, weight, BMI index, etc., which reflect the patient's basic physical condition; and burn situation data, such as injury time, admission time, burn area, inhalation injury, etc., which are crucial for judging the severity and progression stage of the burn.

[0098] Time series data: It covers the vital sign data of patients, such as body temperature, heart rate, respiration, blood pressure, central venous pressure, peripheral oxygen saturation, and urine output, etc., which can reflect the physiological state of patients in real time; and diagnosis and treatment test data, including blood routine test data, blood biochemical test data, blood gas analysis test data, liver and kidney function test data, C-reactive protein test data, procalcitonin test data, and coagulation function test data, etc. These test data help to deeply understand the internal environment of the patient's body and pathophysiological changes.

[0099] Fluid replacement data: It includes the types of fluids, such as crystalloid fluids, colloid fluids, and water; and the fluid replacement flow rate, that is, the amount of fluid input into the patient's body per unit time.

[0100] S12. Divide multiple sample data sets into a training set, a test set, and a validation set.

[0101] Training set: It is used for the training of the model, enabling the model to learn the patterns and rules in the data, and adjusting the parameters of the model so that the model can make reasonable predictions of the fluid replacement data for the input baseline data and time series data.

[0102] Test set: After the model training is completed, the test set is used to evaluate the generalization ability of the model, that is, the prediction ability of the model for unseen data. Through the test set, it can be initially judged whether there are problems of overfitting or underfitting in the model.

[0103] Validation set: During the model training process, it is used to verify the performance of the model and adjust the hyperparameters of the model (such as the learning rate, the number of layers of the neural network, etc.) to optimize the performance of the model.

[0104] The common division ratios of the training set, the test set, and the validation set are divided according to the ratio of 7:2:1 or 6:2:2. The random division method can be adopted to shuffle the order of all sample data sets, and then divide them into the training set, the test set, and the validation set according to the above ratios. To ensure the consistency of data distribution, the stratified sampling method can also be adopted. For example, stratify according to factors such as the burn area and the degree of burns, and then conduct random sampling and division in each layer.

[0105] S13. Construct a basic neural network model based on deep learning.

[0106] S14. Substitute the data of the training set into the basic neural network model for training, and use the test set and the validation set to test and verify the model to obtain a burn patient fluid replacement prediction model based on deep learning that meets the preset requirements.

[0107] Use the data of the test set to test the trained model, calculate the evaluation indicators of the model on the test set, and reflect the prediction accuracy of the model through the evaluation indicators.

[0108] Validate the model using the data from the validation set. After multiple adjustments and validations, until the performance of the model on the validation set meets the preset requirements, such as the error being within an acceptable range, the resulting model is the deep learning-based burn patient fluid replacement prediction model that meets the preset requirements.

[0109] The deep learning-based burn patient fluid replacement prediction method described in the present invention can comprehensively consider the patient's baseline data and time series data to perform personalized fluid replacement prediction and improve the accuracy of prediction. And it can respond in a timely manner to the dynamic changes in the patient's condition and adjust the fluid replacement plan. It reduces the workload of medical staff in manually calculating and adjusting the fluid replacement plan and improves the treatment efficiency.

[0110] Further, the basic neural network model includes a first encoder, a second encoder, a decoder, and an output module;

[0111] The first encoder is used to extract and fuse features from the baseline data and output a baseline feature vector; the second encoder is used to extract and fuse features from the time series data and output a time series feature vector; the feature representation obtained by concatenating the baseline feature vector and the time series feature vector is input into the decoder for decoding to obtain an output feature vector; the output feature vector is processed by the output module to obtain the predicted fluid replacement data.

[0112] This basic neural network model structure can fully utilize the multi-dimensional information of the patient by separately processing the baseline data and the time series data and performing effective feature fusion and decoding, thereby improving the accuracy and reliability of burn patient fluid replacement prediction.

[0113] As a preferred embodiment of the present invention, the process of using the first encoder to extract and fuse features from the baseline data and output a baseline feature vector specifically includes:

[0114] S21, input the baseline data into the first encoder. After self-attention calculation, a weighted output is obtained and normalized to obtain an intermediate result. Specifically, the self-attention mechanism calculates the similarity scores between each element in the input sequence and other elements, and then performs weighted summation on the elements according to these scores to obtain a weighted output. Specifically, for the input baseline data, it will be linearly transformed into three matrices: query, key, and value. Then, the dot product of the query and the key is calculated to obtain the similarity scores, and then normalized through the softmax function to obtain the attention weights. Finally, the value matrix is weighted and summed using the attention weights to obtain the weighted output. Normalizing the weighted output, for example, using Layer Normalization, can make the data have a similar distribution in different feature dimensions, which is helpful for the training and convergence of the model.

[0115] S22. Use the intermediate result as the new input of the first encoder, and perform self-attention calculation again to further explore the features and dependencies in the data, obtaining the baseline feature vector.

[0116] The self-attention mechanism allows the model to focus on the relationships between elements at different positions within the input sequence when processing it, thereby better capturing the features and dependencies in the data. In this preferred embodiment, through two self-attention calculations and normalization processes, the information in the baseline data can be explored more deeply, generating a more representative baseline feature vector.

[0117] In this way, the first encoder can effectively perform feature extraction and fusion on the baseline data, extract richer feature information, generate a more representative baseline feature vector, and provide more valuable information for the subsequent fluid replacement prediction model.

[0118] Exemplarily, the preprocessed baseline data N baseline Obtain the baseline feature vector X through two stacked first encoders based on the self-attention mechanism baseline . First, the baseline data N baseline is input into the self-attention module of the first encoder. After self-attention calculation, a weighted output is obtained and normalized by LayerNorm to obtain the intermediate result Then, use the intermediate result as the new input and perform self-attention calculation again to obtain the final baseline feature vector X baseline . Each self-attention calculation captures the dependencies between different positions in the data through linear transformations and weighted averages of the query, key, and value. Through two stacked self-attention mechanisms, the model can extract richer feature information and provide an effective input representation for subsequent tasks. The specific process is as follows:

[0119]

[0120] where Q is the query matrix, K is the key matrix, V is the value matrix, d k is the feature dimension, T is the matrix transpose, are the learnable feature transformation matrices corresponding to the query matrix, key matrix, and value matrix respectively during the first self-attention calculation, are the learnable feature transformation matrices corresponding to the query matrix, key matrix, and value matrix respectively during the second self-attention calculation.

[0121] Q1 = K1 = V1 = N baseline , Obtain the baseline feature vector X through two stacked above processes baseline。

[0122] As a preferred embodiment of the present invention, the second encoder is used to extract and fuse features from the time-series data, and the output time-series feature vector specifically includes:

[0123] Each feature vector in the time-series data is divided according to a preset time step, and the divided time-series data is expressed as:

[0124] where N i is the time-series data corresponding to the i-th preset time step, and t is the total number of divisions;

[0125] Use the MLP layer to map to where T is the time step, N is the number of original data features, is the input variable after mapping, H is the latent feature dimension, and for each single feature variable, independent position encoding is performed;

[0126] The input variable after mapping is input into the second encoder, and the second encoder uses multiple stacked Transformer modules to extract and fuse features from the input variable after mapping, and the output time-series feature vector is obtained.

[0127] Exemplarily, assume that the timestamps of each time-series data are continuous, and historical data for the past 12 hours is retrieved from the current time t. Set the time step per hour to 1, then the time-series data for the most recent 12 hours is expressed as:

[0128]

[0129] where N t represents the time-series data at time t;

[0130] First, obtain the sequence encoding of a single feature variable. Use a simple MLP layer to map to where T is the time step, N is the number of original data features, is the input variable after mapping, H is the latent feature dimension, and for each single feature variable, independent position encoding is performed. Then, the input variable after mapping is input into the second encoder, and the second encoder uses multiple stacked Transformer modules to extract and fuse features from the input variable after mapping, and the output time-series feature vector is obtained.

[0131] When performing feature extraction and fusion, Multivariate-Attention is used to analyze the correlation between different variable tokens and update the vector representation. After each Multivariate-Attention, a residual connection and layer normalization are connected. Standardization is performed separately for each variable. Finally, a feed-forward neural network is used to further model the temporal features within each variable to obtain an efficient temporal representation.

[0132] As a preferred embodiment of the present invention, the feature representation obtained by concatenating the baseline feature vector and the temporal feature vector is input into the decoder for decoding to obtain an output feature vector, which specifically includes:

[0133] Concatenate the baseline feature vector X baseline and the flattened temporal feature vector X temporal to obtain the feature representation X total , that is:

[0134]

[0135] Flatten the feature representation into a one-dimensional vector and process it through the fully connected layer of each layer of the decoder. The output of the fully connected layer of each layer undergoes a non-linear transformation through the LeakyReLU activation function to enhance the expression ability of the network. To accelerate training and improve the stability of the network, Dropout regularization and residual connection are performed after each fully connected layer, and batch normalization operation is carried out to process complex input features and generate the output feature vector Z dim .

[0136] As a preferred embodiment of the present invention, the prediction of the replenishment fluid data obtained after the output feature vector is processed by the output module specifically includes:

[0137] Determination of the type of liquid: Use a multi-layer perceptron model to perform dimensionality reduction processing on the output feature vector to obtain several intermediate vectors corresponding to different liquid types. Each liquid type includes several subclasses. The dimension of the intermediate vector corresponding to a certain liquid type is equal to the total number of subclasses corresponding to that liquid type. Several elements in the intermediate vector respectively correspond to several subclasses of the liquid type; then input the intermediate vector into the Softmax network for probability distribution prediction to obtain the probabilities corresponding to each element in each intermediate vector, and use the element with the highest probability as the output item, and select the subclass corresponding to the element with the highest probability as the replenishment fluid.

[0138] Exemplarily, there are three types of liquids, namely crystalloid solutions, colloid solutions, and water. Specifically, the subclasses of crystalloid solutions include sodium lactate Ringer's injection, 0.9% sodium chloride injection, 5% sodium bicarbonate injection, 10% potassium chloride injection, three-liter bags, and unselected crystalloid solutions; the subclasses of colloid solutions include plasma, human albumin, red blood cell suspension, and unselected colloid solutions; the subclasses of water include 5% glucose injection and unselected water.

[0139] It should be noted that in the present invention, by setting a subclass of "not selecting this type of liquid" in the subclasses of each type of liquid, after subsequent probability distribution prediction, if the probability of the subclass of "not selecting this type of liquid" is the largest, it means that this type of liquid is not used in the fluid replacement prediction result of this period. That is, a new dimension is added to the conventional subclass items to better adapt to the prediction and selection of burn fluid replacement.

[0140] The three-liter bag is a commonly used parenteral nutrition preparation in clinical practice. The seven major nutritional elements required by the body, such as carbohydrates, amino acids, fat emulsion, vitamins, trace elements, electrolytes, and water, are proportionally formulated in a 3L nutrition bag in a strictly sterile environment as required, and then it is infused into the body through the peripheral vein or central vein to participate in the blood circulation.

[0141] Determination of the fluid replacement flow rate: The output feature vector is processed by a multi-layer perceptron model to obtain the predicted fluid replacement flow rate values corresponding to different types of liquids, and the Sigmoid activation function is used to standardize the predicted fluid replacement flow rate values output to [0,1]. The formula of the Sigmoid function is which can map any real number to the interval [0,1]. Then, the standardized predicted fluid replacement flow rate values are multiplied by a preset flow rate to obtain the fluid replacement flow rate output values corresponding to each type of liquid.

[0142] Specifically, after obtaining the predicted fluid replacement data of the target patient, according to the probability distribution of each element in each intermediate vector output, the position with the highest probability output in each intermediate vector is selected as 1, and the others are 0, to represent the selection of the corresponding subclass and fluid replacement flow rate. To ensure the rationality of the liquid selection, the final selection result needs to be optimized or adjusted for the liquid category according to clinical requirements or other constraints.

[0143] Moreover, outliers in the model output are identified and adjusted, such as situations where the flow rate is extremely high or low, or the selection of the liquid type is unreasonable. To ensure the smooth change of the predicted flow rate over time, a moving average or other smoothing method is used to smooth the predicted flow rate to ensure continuity and stability.

[0144] As a preferred embodiment of the present invention, the time series data further includes the actual fluid replacement data of the patient in the previous treatment cycle.

[0145] Incorporating the actual fluid replacement data of patients in the previous treatment cycle into the time-series data enables the model to more comprehensively understand the fluid replacement situation and its effects of patients at different treatment stages, helps capture the responses of patients to different fluid replacement regimens, enables the model to adjust the current prediction based on historical fluid replacement situations, and thus improves the accuracy and adaptability of fluid replacement prediction. For example, if there are specific changes in the vital signs and test indicators of a patient after adopting a fluid replacement regimen of a certain fluid type and flow rate in the previous cycle, the model can learn this association and then optimize the current prediction.

[0146] For the data on the fluid categories and fluid replacement flow rates, at the beginning, since no treatment has been carried out, it is initialized by filling with 0.

[0147] In one embodiment, the present invention provides a fluid replacement prediction system for burn patients based on deep learning. Refer to Figure 3 As shown, the prediction system 10 includes a first acquisition unit 20, a second acquisition unit 30, and a prediction unit 40.

[0148] The first acquisition unit 20 is used to acquire a fluid replacement prediction model for burn patients based on deep learning. This prediction model takes baseline data and time-series data as inputs and fluid replacement data as outputs. The baseline data includes the physical constitution data and burn condition data of the patient, and the time-series data includes the vital sign data and diagnostic test data of the patient.

[0149] The second acquisition unit 30 is used to collect and acquire the baseline data and time-series data of the target patient.

[0150] The prediction unit 40 is used to input the collected baseline data and time-series data of the target patient into the fluid replacement prediction model for burn patients based on deep learning, and output the predicted fluid replacement data of the target patient.

[0151] In one embodiment, the present invention provides a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to cause a computer to execute the fluid replacement prediction method for burn patients based on deep learning as described in any of the above embodiments.

[0152] The computer-readable storage medium in this embodiment and the fluid replacement prediction method for burn patients based on deep learning in any of the above embodiments belong to the same concept. The specific implementation process is described in detail in the corresponding method embodiment, and the technical features in the method embodiment are equally applicable in this device embodiment, which will not be elaborated here.

[0153] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not mean the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0154] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For parts not detailed or recorded in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0155] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A professional technician can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of this application.

[0156] In the embodiments provided in this application, it should be understood that the disclosed method, controller, and device for estimating the coolant flow rate of a motor controller can be implemented in other ways. For example, the embodiments of the motor control system described above are merely illustrative.

[0157] All or part of the processes of the methods in the above embodiments of this application can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or interface switching device, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc., that can carry the computer program code.

[0158] The above has described the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the knowledge scope of those of ordinary skill in the art to which the present invention pertains, various changes can be made without departing from the purpose of the present invention. In addition, the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

Claims

1. A method for predicting fluid rehydration for burn patients based on deep learning, characterized in that: include: Obtain a deep learning-based prediction model for fluid repletion in burn patients, the prediction model taking baseline data and time series data as input and taking fluid repletion data as output, the baseline data including the patient's physical data and burn condition data, the time series data including the patient's vital signs data and diagnosis and treatment test data; The baseline data and time series data of the target patients are collected and input into the deep learning-based fluid rehydration prediction model for burn patients, and the predicted fluid rehydration data of the target patients is output.

2. The method for predicting fluid replacement for burn patients based on deep learning according to claim 1, characterized in that: The patient's physical data included sex, age, height, weight, and BMI index; Patients’ burn data included time of injury, time of hospital admission, burn area, and inhalation injury; The patients’ vital sign data included temperature, heart rate, respiration, blood pressure, central venous pressure, peripheral oxygen saturation, and urine output; The patient's diagnosis and treatment test data include blood routine test data, blood biochemistry test data, blood gas analysis test data, liver and kidney function test data, C-reactive protein test data, procalcitonin test data and coagulation function test data; The fluid infusion data includes the type of fluid and the fluid infusion flow rate.

3. The method for predicting fluid rehydration for burn patients based on deep learning according to claim 1, characterized in that: Obtaining a deep learning-based prediction model for fluid rehydration in burn patients specifically includes: Obtain multiple sample data sets, where a single sample data set includes baseline data and time series data; Divide multiple sample data sets into training sets, test sets, and validation sets; A basic neural network model based on deep learning was constructed, the data of the training set was substituted into the basic neural network model for training, and the model was tested and verified using the test set and validation set to obtain a deep learning-based fluid rehydration prediction model for burn patients that met the preset requirements.

4. The method for predicting fluid rehydration for burn patients based on deep learning according to claim 3, characterized in that: The basic neural network model includes a first encoder, a second encoder, a decoder and an output module; Performing feature extraction and fusion on the baseline data by the first encoder, and outputting a baseline feature vector; The second encoder extracts and fuses features of the time series data, and outputs a time series feature vector; The feature representation obtained by concatenating the baseline feature vector and the time series feature vector is input into a decoder for decoding to obtain an output feature vector; The output feature vector is processed by the output module to obtain predicted fluid infusion data.

5. The method for predicting fluid rehydration for burn patients based on deep learning according to claim 3, characterized in that: The first encoder extracts and fuses the baseline data to obtain a baseline feature vector, which specifically includes: The baseline data is input into the first encoder, and after self-attention calculation, the weighted output is obtained, and normalized to obtain the intermediate result; Use the intermediate result as the new input of the first encoder and perform self-attention calculation again to obtain the baseline feature vector.

6. The method for predicting fluid rehydration for burn patients based on deep learning according to claim 3, characterized in that: The second encoder extracts and fuses the time series data to obtain a time series feature vector, which specifically includes: Each feature vector in the time series data is divided according to the preset time step, and the divided time series data is expressed as: Among them, N i is the time series data corresponding to the i-th preset time step, and t is the total number of partitions; Using the MLP layer Mapping Where T is the time step, N is the number of original data features, is the input variable after mapping, H is the potential feature dimension, and Each single-feature variable is independently positionally encoded; The mapped input variables are input into the second encoder, and features are extracted and fused on the mapped input variables through a plurality of stacked Transformer modules in the second encoder, and a time series feature vector is output.

7. The method for predicting fluid rehydration for burn patients based on deep learning according to claim 3, characterized in that: The output feature vector is processed by the output module to obtain the predicted fluid infusion data, which specifically includes: Determination of liquid types: A multi-layer perceptron model is used to reduce the dimension of the output feature vector to obtain several intermediate vectors corresponding to different liquid types. Each liquid type includes several subclasses. The dimension of the intermediate vector corresponding to a certain liquid type is equal to the total number of subclasses corresponding to the liquid type. Several elements in the intermediate vector correspond to several subclasses of the liquid type. The intermediate vector is then input into the Softmax network for probability distribution prediction to obtain the probability corresponding to each element in each intermediate vector. The element with the highest probability is used as the output item, and the subclass corresponding to the element with the highest probability is selected as the rehydration fluid. Determination of fluid infusion flow rate: The multi-layer perceptron model is used to process the output feature vector to obtain the predicted values ​​of fluid infusion flow rate corresponding to different types of liquids, and the Sigmoid activation function is used to standardize the output predicted values ​​of fluid infusion flow rate to [0,1]. The standardized predicted values ​​of fluid infusion flow rate are then multiplied by the preset flow rate to obtain the output values ​​of fluid infusion flow rate corresponding to each type of liquid.

8. The method for predicting fluid rehydration for burn patients based on deep learning according to claim 1, characterized in that: The time series data also includes the actual fluid replacement data of the patient in the previous treatment cycle.

9. A deep learning-based fluid replacement prediction system for burn patients, characterized in that: include: A first acquisition unit is used to acquire a deep learning-based prediction model for fluid repletion of burn patients, the prediction model taking baseline data and time series data as input and taking fluid repletion data as output, the baseline data including the patient's physical data and burn condition data, the time series data including the patient's vital signs data and diagnosis and treatment test data; The second acquisition unit is used to collect the baseline data and time series data of the target patient; The prediction unit is used to input the acquired baseline data and time series data of the target patient into the deep learning-based burn patient fluid repletion prediction model, and output the predicted fluid repletion data of the target patient.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the deep learning-based fluid rehydration prediction method for burn patients as described in any one of claims 1 to 8.

Images