Hematopoietic stem cell transplant cinv risk assessment and antiemetic umbilical management system

Abstract

The application discloses a hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system and relates to the technical field of medical health. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system comprises a risk stratification unit, a plaster preparation unit, a time sequence decision unit and a sterile operation guiding unit. The risk stratification unit obtains a Dranitsaris basic score, age, sleep duration and previous vomiting history of a patient, calculates a risk factor, inputs the risk factor into a logistic regression model to generate a comprehensive risk probability value, and outputs an activation signal when the risk is high. The plaster preparation unit prepares medicinal paste in proportion and sterilizes the medicinal paste. The time sequence decision unit determines a plastering time window according to the high-risk signal and sleep data and issues an instruction. The sterile operation guiding unit completes the plastering operation. A closed-loop adaptive unit records a nausea and vomiting grading score after plastering and updates the model weight, thereby forming a closed-loop regulation. The application realizes the dynamic optimization of the delayed nausea and vomiting risk stratification prompt and the plastering management scheme through risk quantification, time window dynamic decision, standardized plastering and closed-loop adaptive updating.

Description

Technical Field

[0001] This invention relates to the field of medical and health technology, specifically to a hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system. Background Technology

[0002] Hematopoietic stem cell transplantation (HSCT) is a crucial treatment for hematologic malignancies. Pre-transplantation treatment typically involves high-dose myeloablative chemotherapy, such as cyclophosphamide and busulfan, which are highly emetogenic. Delayed intravascular coagulation (CINV) usually occurs 24 hours to 7 days after chemotherapy, and its pathological mechanism involves delayed activation of the substance P / NK-1 pathway. While triple antiemetic regimens are recommended by guidelines, clinical adherence is low in HSCT patients, and dexamethasone use can interfere with immune reconstitution and increase the risk of infection. Delayed CINV occurs in 60% to 80% of HSCT patients, severely impacting their nutritional status, quality of life, and treatment adherence.

[0003] The limitations of existing technologies include at least the following issues: antiemetic regimens are limited to the single indicator of the emetic grade of chemotherapy drugs, and have not yet established a dynamic correlation between individualized high-risk factors such as patient age, sleep status, and previous vomiting history and the pathological mechanisms of delayed CINV, such as the delayed activation characteristics of the substance P / NK-1 pathway, resulting in a lack of closed-loop linkage between risk assessment and intervention measures; conventional oral or injectable antiemetics are significantly less effective in controlling delayed nausea and vomiting than acute vomiting, and are prone to causing adverse reactions such as constipation, headache, and central drowsiness; in particular, the use of dexamethasone in HSCT patients can interfere with immune reconstitution and increase the risk of infection; although traditional Chinese medicine acupoint application has the advantage of transdermal drug delivery, it lacks Chinese medicine compositions with clearly defined ingredients and reproducible ratios, as well as standard operating procedures for Class 100 laminar flow sterile environments, the preparation process is crude, there are potential infection risks in clinical application, and it is difficult to balance stable efficacy and safety in special scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system. This system solves the problems of lack of closed-loop linkage in risk assessment and intervention measures for delayed chemotherapy-related nausea and vomiting in hematopoietic stem cell transplantation patients, limited control of delayed symptoms by conventional antiemetic drugs and prominent side effects, and lack of standardized formulas and sterile chamber adaptation schemes for traditional Chinese medicine umbilical therapy.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system, comprising: a risk stratification unit, used to acquire the Drantsaris baseline score, age, sleep duration, and number of vomiting episodes; sleep duration below a preset threshold, age below a preset threshold, and number of vomiting episodes greater than or equal to a preset threshold are used as risk factors; the baseline score and risk factors are input into a logistic regression model trained with historical HSCT data; this model sets independent weight coefficients for the delayed activation characteristics of the NK-1 pathway; and outputs a calibrated comprehensive risk probability value; when the equivalent score is higher than a preset risk score threshold, a high-risk activation signal is generated; an umbilical therapy ointment preparation unit, used to weigh Pinellia ternata, clove, persimmon calyx, dried ginger, and Codonopsis pilosula according to a preset ratio, pulverize and sieve them, mix them with a petroleum jelly matrix to form an ointment, and then sterilize it by irradiation after packaging; and a timing decision unit, used to receive high-risk activation signals. After the live signal is received and chemotherapy ends, the sleep duration of the previous night is read when the preset duration is reached. If the sleep duration is greater than or equal to the preset threshold, the application time window is set to Chenshi (7-9 AM); otherwise, it is set to an offset period after Chenshi. When the current time falls into the time window, the application execution command is output. The aseptic operation guidance unit is used to respond to the application execution command, guide hand disinfection, wearing sterile gloves, disinfecting the Shenque acupoint twice with iodine swabs and letting it dry, take out the preset dose of ointment, apply it to the center of sterile gauze, apply it to the Shenque acupoint and fix it with sterile breathable tape, continue the application for the preset duration, remove the gauze and clean the navel with sterile saline swabs. The closed-loop adaptive unit is used to record the nausea and vomiting grade score within the preset time window after each application, compare the score with the comprehensive risk probability value before this application to calculate the prediction error, update the NK-1 pathway weight coefficient in the logistic regression model according to the prediction error backpropagation, and send the updated model to the risk stratification unit.

[0006] Furthermore, the specific steps for risk calculation performed by the risk stratification unit are as follows: The acquired age, sleep duration, and number of vomiting episodes are segmented and encoded. Sleep duration greater than or equal to a high threshold is encoded as a low-value code, less than a high threshold but greater than or equal to a low threshold is encoded as a medium-value code, and less than a low threshold is encoded as a high-value code. Age and number of vomiting episodes are encoded as binary features according to preset thresholds. The encoded features are normalized in the same direction as the Drantsaris baseline score. The normalized data is input into a logistic regression model, and independent weighting coefficients are applied to the delayed activation features of the NK-1 pathway. The calibrated comprehensive risk probability value is output based on the weighting result.

[0007] Furthermore, the specific steps for the timing decision unit to perform time window matching are as follows: Perform validity verification on the previous night's sleep duration data; calculate the sleep offset index based on the verified sleep duration data; when the sleep offset index is greater than zero and not greater than the mild offset threshold, shift the application time window backward by a mild offset amount; when the sleep offset index is greater than the mild offset threshold, shift the application time window backward by a severe offset amount, where the severe offset amount is greater than the mild offset amount; perform a matching comparison between the current system time and the time window obtained after the offset, and output the application execution command after successful matching.

[0008] Furthermore, the specific steps for error calculation and weight update in the closed-loop adaptive unit are as follows: Collect nausea and vomiting grading scores within a preset time window after each application; perform numerical mapping between the grading scores and the comprehensive risk probability value before the current application to calculate the prediction error; use a weighted moving average method to perform a weighted summation of the prediction errors from the three most recent applications, with the decreasing weights of the three errors being the recent weight, medium-term weight, and long-term weight, respectively, with the recent weight being greater than the medium-term weight, which is greater than the long-term weight; use the weighted summation result as the basis for backpropagation update, and perform NK-1 pathway weight coefficient update; after applying regularization constraints to the updated weight coefficients, distribute them to the risk stratification unit.

[0009] Furthermore, the specific steps of the risk stratification unit in performing the graded identification processing are as follows: Determine the relationship between the equivalent score corresponding to the calibrated comprehensive risk probability value and the preset risk score threshold; when the equivalent score is greater than the preset risk score threshold but less than the next higher risk score threshold, generate a high-risk activation signal and attach a basic application duration identifier; when the equivalent score is greater than or equal to the next higher risk score threshold, generate a high-risk activation signal and attach an extended application duration identifier; send the high-risk activation signal to the timing decision unit, and send the basic application duration identifier or the extended application duration identifier to the aseptic operation guidance unit.

[0010] Furthermore, the specific steps of the aseptic operation guidance unit in performing the patch application operation are as follows: After responding to the patch application instruction, the operator is guided to perform hand cleaning and aseptic disinfection; the operator is guided to wear double-layer sterile gloves, with an inner layer of polyethylene film gloves and an outer layer of nitrile rubber gloves, and a glove airtightness check is performed; if an extended application time indicator is received, the preset application time is increased by a preset percentage; the Shenque acupoint area is disinfected twice in a circular manner with iodine swabs and allowed to dry, then the iodine is removed once with an alcohol swab and allowed to dry; the preset dose of ointment is evenly spread in the center of sterile gauze, applied to the Shenque acupoint, and fixed with sterile breathable tape.

[0011] Furthermore, the specific steps for refining the ointment in the umbilical therapy ointment preparation unit are as follows: Pinellia ternata, cloves, persimmon calyx, dried ginger, and Codonopsis pilosula are weighed according to a preset ratio and then coarsely pulverized; the coarsely pulverized medicinal materials are finely pulverized using a vibrating ultra-micro pulverizer until the particle size is smaller than a preset particle size threshold, and then sieved; the sieved powder is thoroughly mixed with petrolatum matrix to form an ointment; the ointment is then sealed and sterilized by irradiation using cobalt-60 rays within a preset irradiation dose range.

[0012] Furthermore, the specific steps for the closed-loop adaptive unit to perform incremental learning are as follows: when the nausea and vomiting grade scores after two consecutive applications are both greater than the preset score threshold, a model retraining trigger signal is generated; after responding to the trigger signal, the training samples of the most recent preset number of times are extracted from the historical database, and incremental learning is performed; during the incremental learning process, only the NK-1 pathway weight coefficients in the logistic regression model are updated; the updated model parameters are then sent to the risk stratification unit.

[0013] Furthermore, the specific steps of the timing decision unit in performing fault-tolerant processing of sleep data are as follows: Discrete-point smoothing filtering is applied to the read sleep duration from the previous night; the sleep duration data is checked for missing or invalid data. If missing or invalid, the application time window is set to Chenshi (7-9 AM) by default, and a sleep monitoring missing alarm signal is generated; if the sleep duration data is valid, the filtered sleep duration is precisely compared with a preset threshold to lock a unique application time window; conflict verification is performed on the locked application time window, and if there is no conflict, an application execution command is output; an alarm signal is sent to the aseptic operation guidance unit, with an additional prompt for the operator to inquire about the patient's subjective sleep quality.

[0014] Furthermore, the specific steps for feature correlation screening in the risk stratification unit are as follows: perform feature correlation calculations between the base score and each risk factor; eliminate redundant input features based on the correlation calculation results, and retain three effective features: age, sleep duration, and number of vomiting episodes; transform the sleep duration feature using a segmented encoding method; perform a linear weighted summation of the three transformed features with the Drantsaris base score, with the initial weights of the three features being the baseline weight value, auxiliary weight value, and adjustment weight value, respectively; dynamically adjust the above three weights and the NK-1 pathway weight coefficient during initial model training or offline training.

[0015] The present invention has the following beneficial effects:

[0016] (1) The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system obtains the Drantsaris baseline score, patient age, sleep duration, and number of vomiting episodes through risk stratification units. These factors are input into a logistic regression model trained with historical HSCT data. The model sets independent weight coefficients for the delayed activation characteristics of the NK-1 pathway and outputs a calibrated comprehensive risk probability value. Traditional antiemetic regimens only rely on the emetic grade of chemotherapy drugs and ignore individual patient differences. This system dynamically links the patient's physiological characteristics with the disease pathological mechanism, identifies high-risk groups that truly need intervention, and improves the accuracy of risk stratification through segmented coding, feature correlation screening, and graded labeling. When the equivalent score exceeds the preset risk score threshold, a high-risk activation signal is generated, providing a clear start point for umbilical therapy intervention and solving the problem of lack of closed-loop linkage between risk assessment and intervention measures.

[0017] (2) The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system receives a high-risk activation signal through the timing decision unit and starts timing from the end of chemotherapy. When the preset start time is reached, the system reads the patient's sleep duration the night before. Based on whether the sleep duration reaches the preset threshold, the application time window is set to the Chenshi (7-9 AM) or a period after Chenshi. The calculation of the sleep offset index allows the application timing to be dynamically adjusted according to the patient's actual sleep state. The more insufficient the sleep, the more the application time window is shifted backward. This design adapts the TCM time medicine theory to the patient's individual rhythm. The closed-loop adaptive unit records the nausea and vomiting grade score after each application. The score is compared with the comprehensive risk probability value before the current application to calculate the prediction error. The NK-1 pathway weight coefficient in the logistic regression model is updated through backpropagation. After the update, the model is re-issued to the risk stratification unit, forming a closed-loop system of assessment, intervention, feedback, and reassessment. The model is continuously optimized as the treatment progresses.

[0018] (3) The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system weighs Pinellia ternata, clove, persimmon calyx, dried ginger and Codonopsis pilosula according to the preset ratio through the umbilical therapy ointment preparation unit, crushes and sieves them and mixes them with Vaseline matrix to form an ointment, and then seals and sterilizes them by irradiation to ensure that the composition of the Chinese medicine composition is clear, the ratio is repeatable and sterile. The vibration ultra-micro pulverization makes the particle size of the medicinal materials smaller than the preset threshold, which is conducive to transdermal absorption. The cobalt-60 irradiation sterilization is adapted to the Class 100 laminar flow sterile chamber environment, avoiding the infection risk caused by the rough preparation process of traditional Chinese medicine patch. The sterile operation guidance unit provides a standardized patching process, including hand disinfection, wearing double sterile gloves, iodine ring disinfection, ethanol deiodination, ointment application, sterile fixation and cleaning of the umbilicus, etc. When the extended patching time indicator is received, the preset patching time is automatically increased.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] Figure 1 This is a block diagram of the hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system of the present invention.

[0021] Figure 2 This is a flowchart illustrating the specific steps involved in risk calculation within the risk stratification unit of the hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system of the present invention. Detailed Implementation

[0022] Please see Figure 1 This invention provides a technical solution: a hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system, comprising: a risk stratification unit, used to acquire the Dranthasaris baseline score, age, sleep duration, and number of vomiting episodes; sleep duration below a preset threshold, age below a preset threshold, and number of vomiting episodes greater than or equal to a preset threshold are used as risk factors; the Dranthasaris baseline score and risk factors are input into a logistic regression model trained with historical HSCT data; this model sets independent weight coefficients for the delayed activation characteristics of the NK-1 pathway; and outputs a calibrated comprehensive risk probability value; a high-risk activation signal is generated when the equivalent score is higher than a preset risk score threshold; an umbilical therapy ointment preparation unit, used to weigh Pinellia ternata, clove, persimmon calyx, dried ginger, and Codonopsis pilosula according to a preset ratio, pulverize and sieve them, mix them with Vaseline matrix to form an ointment, and then sterilize it by irradiation after packaging; and a timing decision unit, used to receive the high-risk activation signal and then start chemotherapy. The system is designed to: 1. Read the previous night's sleep duration when the preset duration is reached. If the sleep duration is greater than or equal to a preset threshold, the application time window is set to Chenshi (7-9 AM); otherwise, it is set to an offset period after Chenshi. When the current time falls within the time window, an application execution command is output. 2. A sterile operation guidance unit responds to the application execution command, guiding hand disinfection, wearing sterile gloves, circumferential disinfection of the Shenque acupoint with iodine swabs, allowing it to dry, applying the preset dose of ointment to the center of sterile gauze, applying it to the Shenque acupoint, and securing it with sterile breathable tape. After a preset application time, the gauze is removed, and the navel is cleaned with sterile saline swabs. 3. A closed-loop adaptive unit records the nausea and vomiting grade score within a preset time window after each application. The score is compared with the comprehensive risk probability value before the current application to calculate the prediction error. The NK-1 pathway weight coefficient in the logistic regression model is updated based on the prediction error through backpropagation. The update magnitude is equal to the absolute value of the prediction error multiplied by the preset learning rate. The updated model is then sent to the risk stratification unit.

[0023] The Drantaris baseline score is calculated by collecting basic information such as the patient’s emetic level, chemotherapy dose, and previous chemotherapy history, and then according to the Drantaris scoring scale standard.

[0024] Historical data from HSCT include the age, sleep duration, number of vomiting episodes, occurrence of delayed CINV, NK-1 pathway-related indicators, and clinical intervention efficacy data of at least 100 hematopoietic stem cell transplant patients.

[0025] The preset risk scoring threshold corresponds to the judgment criteria for high risk of delayed CINV in the Drantsaris scoring system. The equivalent score is obtained by converting the comprehensive risk probability value and the scoring threshold.

[0026] The initial preset duration corresponds to 24 hours after the end of chemotherapy;

[0027] The preset ratio (6:3:3:5:3) of Pinellia ternata, clove, persimmon calyx, dried ginger, and Codonopsis pilosula in the Chinese herbal composition is determined based on the theory of antiemetics in traditional Chinese medicine and modern pharmacological research. It can achieve targeted intervention for delayed CINV while avoiding the adverse reactions caused by Western antiemetics.

[0028] Applying patches to the Shenque acupoint utilizes the thin skin and rapid absorption of the Shenque acupoint to achieve transdermal drug delivery, avoiding the first-pass effect of the liver, and is suitable for the special physical condition of hematopoietic stem cell transplant patients.

[0029] Irradiation sterilization is compatible with the environmental requirements of a Class 100 laminar flow sterile chamber, effectively eliminating microbial contamination during the ointment and packaging process, and reducing the risk of infection for patients.

[0030] Specifically, such as Figure 2 As shown, the specific steps for risk calculation performed by the risk stratification unit are as follows:

[0031] The acquired data on age, sleep duration, and number of vomiting episodes were segmented and coded. Sleep duration greater than or equal to a high threshold was coded as a low coded value, sleep duration less than a high threshold but greater than or equal to a low threshold was coded as a medium coded value, and sleep duration less than a low threshold was coded as a high coded value. Specifically:

[0032] First, the collected sleep duration data is denoised to remove obvious outliers (such as sleep data less than 2 hours or more than 12 hours), and then high and low thresholds for sleep duration are set.

[0033] In one implementation, the high threshold for sleep duration is set to 7 hours, and the low threshold is set to 5 hours.

[0034] When sleep duration is ≥7 hours, the code is a low coding value of 1;

[0035] When sleep duration is 5 hours or less and less than 7 hours, the code is coded as medium code value 2;

[0036] When sleep duration is less than 5 hours, the code is a high coding value of 3;

[0037] Segmented encoding converts continuous sleep duration data into discrete features, which facilitates feature calculation in logistic regression models.

[0038] Age and the number of times a person has a history of vomiting are encoded as binary features according to preset thresholds, specifically as follows:

[0039] Set a preset age threshold;

[0040] In one implementation, the age preset threshold is set to 60 years old. When the patient's age is <60 years old, it is coded as 1 (indicating the presence of the risk factor), and when the patient's age is ≥60 years old, it is coded as 0 (indicating the absence of the risk factor).

[0041] Set a preset threshold for the number of vomiting episodes;

[0042] In one implementation, the preset threshold for the number of vomiting history is set to 1 time;

[0043] When a patient has a history of vomiting during chemotherapy ≥1 time, the code is 1 (indicating the presence of this risk factor).

[0044] When a patient has a history of vomiting during chemotherapy less than once, the code is 0 (indicating that the risk factor does not exist).

[0045] The encoded features are then normalized in the same direction as the Drantisaris base score, specifically as follows:

[0046] The min-max normalization method was used to uniformly map the encoded age, sleep duration, number of vomiting history features and Drantsaris baseline score to the interval [0, 1].

[0047] The normalized data is input into the logistic regression model, and independent weighting operations are performed on the delayed activation features of the NK-1 pathway, specifically as follows:

[0048] The inputs to the logistic regression model are the normalized Drantsaris baseline score, age encoding value, sleep duration encoding value, vomiting history frequency encoding value, and NK-1 pathway delayed activation feature value. The model pre-sets initial weights for each input feature; among them, an independent weight coefficient is set for the NK-1 pathway delayed activation feature. This weight coefficient is used to amplify the impact of the NK-1 pathway delayed activation pathological mechanism on the risk of delayed CINV. The weighting calculation process is as follows:

[0049] Each input feature is multiplied by its corresponding weight coefficient, summed, and then the sum is mapped to a comprehensive risk probability value between 0 and 1 using the sigmoid function.

[0050] ;

[0051] in, For model bias terms, For Drantisaris's base rating, This is an age-coded value. This is a value encoded for sleep duration. This is the coded value for the number of times a person has a history of vomiting. The delayed activation characteristic value of the NK-1 pathway; , , , They are respectively , , , The corresponding weights These are the independent weighting coefficients for the delayed activation features of the NK-1 pathway;

[0052] In one implementation, The initial value is greater than other weight values, highlighting the impact of pathological mechanisms on risk assessment;

[0053] The calibrated overall risk probability value is output based on the weighted calculation result, specifically as follows:

[0054] The comprehensive risk probability value between 0 and 1 output by the logistic regression model is matched with the preset risk probability-score mapping relationship and converted into an equivalent score. This equivalent score is consistent with the score range of the Drantsaris rating scale (0-30 points).

[0055] In one implementation, a comprehensive risk probability value of 0.8 corresponds to an equivalent score of 24 points, a comprehensive risk probability value of 0.5 corresponds to an equivalent score of 15 points, and a comprehensive risk probability value of 0.3 corresponds to an equivalent score of 9 points.

[0056] In this implementation plan, segmented encoding and binary encoding are performed on age, sleep duration, and number of vomiting episodes to convert continuous data into discrete features; unidirectional normalization eliminates dimensional differences and avoids a single feature from overly dominating the assessment results; the logistic regression model sets independent weight coefficients for the delayed activation features of the NK-1 pathway, and outputs a comprehensive risk probability value after weighted calculation and maps it to an equivalent score, enabling medical staff to intuitively determine the patient's risk level.

[0057] Specifically, the time window matching steps performed by the time-series decision unit are as follows:

[0058] Perform a validity check on the retrieved sleep duration data from the previous night, specifically as follows:

[0059] First, determine if there are any missing sleep duration data (e.g., no data was collected, or the data is empty). If there are any missing data, the data is considered invalid.

[0060] Secondly, determine whether the sleep duration data is within a reasonable range (2-12 hours). If it exceeds this range, it is considered invalid data.

[0061] Next, it is determined whether there are abnormal fluctuations in the data (such as a difference of more than 3 hours from the patient's previous sleep duration). If there are abnormal fluctuations, they are further verified. If the verification is correct, the data is determined to be valid. If the verification is unsuccessful, the data is determined to be invalid. Validity verification can ensure the accuracy of sleep duration data and avoid invalid data causing time window matching errors.

[0062] The sleep deviation index is calculated based on the verified sleep duration data. The sleep deviation index is equal to the preset sleep threshold minus the measured sleep duration, divided by the preset sleep threshold. Specifically:

[0063] The preset sleep threshold is the standard for appropriate sleep duration for hematopoietic stem cell transplant patients;

[0064] In one implementation, the preset sleep threshold is set to 7 hours, and the sleep deviation index is calculated using the following formula:

[0065] ;

[0066] Where λ is the sleep deviation index, T0 is the preset sleep threshold, and T1 is the measured sleep duration;

[0067] When the measured sleep duration is 6 hours, the sleep deviation index is (7-6) / 7≈0.14;

[0068] When the measured sleep duration is 4 hours, the sleep deviation index is (7-4) / 7≈0.43. The larger the sleep deviation index, the more severe the patient's sleep deprivation, and the higher the corresponding risk of delayed CINV. It is necessary to improve the intervention effect by adjusting the application time window.

[0069] When the sleep offset index is greater than zero and not greater than the mild offset threshold, the application time window is shifted backward by a mild offset amount, specifically:

[0070] Set a slight offset threshold;

[0071] In one implementation, the mild offset threshold is set to 0.2. When the sleep offset index 0 < λ ≤ 0.2, it is determined to be mild sleep deprivation. At this time, the application time window (Chen Shi, i.e., 7-9 o'clock) is shifted backward by a mild offset amount, which is set to 30 minutes. That is, the adjusted application time window is 7:30-9:30.

[0072] When the sleep shift index is greater than the mild shift threshold, the application time window is shifted backward by the severe shift amount, which is greater than the mild shift amount. Specifically:

[0073] When the sleep offset index λ > 0.2, it is determined to be severe sleep deprivation. At this time, the application time window is shifted backward by the severe offset amount.

[0074] In one implementation, the severe offset is set to 60 minutes, which means the adjusted application time window is 8:00-10:00.

[0075] The current system time is compared with the time window obtained after offset. If the match is successful, the application execution command is output, which is as follows:

[0076] The timing decision unit synchronizes the current system time in real time (accurate to the minute), compares the current system time with the adjusted application time window, and if the current system time falls within the application time window (e.g., if the adjusted time window is 7:30-9:30 and the current time is 8:15, then the match is successful), it immediately outputs the application execution command, which includes key information such as the application time window and the preset application duration.

[0077] If the current system time does not fall within the application time window, the system time will be continuously monitored until a successful time match is achieved. Once this is achieved, the application execution command will be output, and the time matching result will be recorded to provide data support for the subsequent model update of the closed-loop adaptive unit.

[0078] In this implementation plan, the validity of the previous night's sleep duration is verified, and missing, out-of-range, or abnormally fluctuating data are removed to ensure the input quality of time window matching. The sleep offset index quantifies the degree of sleep deficiency. For mild sleep deficiency, the application time window is shifted backward by 30 minutes, and for severe sleep deficiency, it is shifted backward by 60 minutes, so that the application timing is adapted to the patient's individual rhythm. After the current system time is successfully matched with the shifted time window, the application execution command is output to avoid the application operation from interfering with the patient's rest.

[0079] Specifically, the steps for error calculation and weight update performed by the closed-loop adaptive unit are as follows:

[0080] The nausea and vomiting grading data were collected within a preset time window after each patch application, specifically as follows:

[0081] The preset time window is 24-72 hours after application, which coincides with the peak period for delayed CINV. The nausea and vomiting grading uses the clinically commonly used nausea and vomiting grading standard, as shown in Table 1.

[0082] Table 1 Specific examples of classification

[0083] Classification score Detailed description Level 0 0 points No nausea or vomiting Level 1 1 point Mild nausea, no vomiting, does not affect eating. Level 2 2 points Moderate nausea, with occasional vomiting (1-2 times / day), affecting food intake. Level 3 3 points Severe nausea, frequent vomiting (3-5 times / day), inability to eat. Level 4 4 points Severe nausea and vomiting (more than 5 times per day) require emergency treatment.

[0084] During the data collection, medical staff will assign a score based on the patient's actual performance within a preset time window, in accordance with the above-mentioned grading criteria, to ensure the accuracy and objectivity of the scoring data.

[0085] The grading and scoring data are numerically mapped to the comprehensive risk probability value before this application, and the prediction error is calculated as follows:

[0086] First, the nausea and vomiting grading score (0-4 points) is mapped to the 0-1 range. The mapping method is to divide the score by 4, that is, 0 points are mapped to 0, 1 point is mapped to 0.25, 2 points are mapped to 0.5, 3 points are mapped to 0.75, and 4 points are mapped to 1. This mapping result is consistent with the range of the comprehensive risk probability value (0-1).

[0087] The formula for calculating prediction error is:

[0088] ;

[0089] in, For prediction error, The mapped nausea and vomiting grading score, This is the overall risk probability value before this application;

[0090] In one implementation, if the mapped score is 0.25 and the overall risk probability value is 0.3, then the prediction error is |0.25-0.3|=0.05. The smaller the prediction error, the better the risk assessment result matches the actual intervention effect.

[0091] The weighted moving average method is used to perform a weighted summation of the prediction errors of the three most recent applications. The decreasing weights of the three errors are, in descending order, the recent weight, the medium-term weight, and the distant-term weight, with the recent weight being greater than the medium-term weight, which is greater than the distant-term weight. Specifically:

[0092] Set near-term weights, medium-term weights, and long-term weights;

[0093] In one implementation, the near-term weight is set to 0.5, the medium-term weight is set to 0.3, and the far-term weight is reset to 0.2, with 0.5 + 0.3 + 0.2 = 1;

[0094] The formula for weighted summation is:

[0095] ;

[0096] in, This is the total error after weighted summation. This represents the prediction error for the most recent application (recent). This represents the prediction error (intermediate) of the previous application. This is the prediction error from the previous application (long-term). Recent weighting For medium-term weighting, For long-term options weight;

[0097] The weighted summation result is used as the basis for backpropagation updates, and the NK-1 pathway weight coefficients are updated as follows:

[0098] Using the weighted summation of the total error E as the input to the backpropagation algorithm, the independent weight coefficients of the NK-1 pathway in the logistic regression model are adjusted using gradient descent. The update formula is as follows:

[0099] ;

[0100] in, The updated NK-1 pathway weighting coefficients. The weighting coefficients before the update. To preset the learning rate, Total error right The partial derivatives;

[0101] In one implementation, a preset learning rate is used. Setting it to 0.01 ensures that the weight coefficients are updated smoothly, avoiding excessive update amplitude that could lead to abnormal model fitting.

[0102] The updated weight coefficients are then processed with regularization constraints and distributed to the risk stratification unit, specifically as follows:

[0103] The updated NK-1 pathway weight coefficients are constrained using the L2 regularization method. The regularization calculation formula is as follows:

[0104] ;

[0105] in, The loss function after regularization. The regularization coefficient (in one implementation) Set to 0.001). The weight coefficients for each feature of the model are defined. Regularization constraints can prevent the model from overfitting. After the constraint processing is completed, the updated weight coefficients and complete model parameters are sent to the risk stratification unit to replace the original model parameters and achieve adaptive optimization of the model.

[0106] In this implementation plan, the preset time window is set to 24 to 72 hours after application, which coincides with the peak period of delayed CINV. The nausea and vomiting grade score is mapped to the 0 to 1 interval and compared with the comprehensive risk probability value to calculate the prediction error and quantify the model prediction bias. The weighted moving average method is used to sum the prediction errors of the three most recent predictions, with a weight of 0.5 for the recent period, 0.3 for the medium period, and 0.2 for the long period. This highlights the impact of recent errors while taking into account medium and long-term data, and avoids excessive fluctuations in weight updates due to a single error. The gradient descent method updates the NK-1 pathway weight coefficients with the weighted summation result, and L2 regularization is used to constrain the updated coefficients to prevent model overfitting.

[0107] Specifically, the steps for the risk stratification unit to perform hierarchical identification processing are as follows:

[0108] The relationship between the equivalent score corresponding to the calibrated comprehensive risk probability value and the preset risk score threshold is determined as follows:

[0109] The preset risk scoring threshold is the threshold for determining high risk of delayed CINV, and the next higher risk scoring threshold is the threshold for determining extremely high risk of delayed CINV.

[0110] In one implementation, the preset risk scoring threshold is set to 16 points, and the next higher risk scoring threshold is set to 22 points.

[0111] First, obtain the equivalent score corresponding to the comprehensive risk probability value. Then, compare the equivalent score with the preset risk score threshold and the next higher risk score threshold to determine the patient's risk level. The comparison results are divided into three categories:

[0112] Equivalent score ≤ 16 points (not high risk);

[0113] 16 points < equivalent score < 22 points (high risk);

[0114] Equivalent score ≥ 22 points (extremely high risk);

[0115] Of these, only the latter two cases generate high-risk activation signals;

[0116] When the equivalent score is greater than the preset risk score threshold but less than the next higher risk score threshold, a high-risk activation signal is generated and a basic patch duration identifier is attached, specifically:

[0117] When 16 points < equivalent score < 22 points, the patient is identified as a high-risk group for delayed CINV. At this time, a high-risk activation signal is generated, and a basic patching duration indicator is added. The basic patching duration indicator corresponds to the preset regular patching duration.

[0118] In one implementation, the basic application time is set to 8 hours. This time is determined in combination with the drug release rate of the traditional Chinese medicine umbilical therapy ointment and the absorption characteristics of human skin, to ensure that the effective ingredients of the drug are fully absorbed and to avoid skin irritation caused by excessive application time.

[0119] When the equivalent score is greater than or equal to the threshold of the next higher risk score, a high-risk activation signal is generated and an extended application time indicator is attached, specifically as follows:

[0120] When the equivalent score is ≥22 points, the patient is determined to be a very high-risk group for delayed CINV. At this time, a high-risk activation signal is generated, and an extended application time indicator is added. The extended application time indicator corresponds to adding a certain time on the basis of the basic application time.

[0121] In one implementation, the application time is extended to 12 hours, which is 50% longer than the basic application time. By extending the application time, the absorption of the effective components of traditional Chinese medicine can be increased, thereby improving the intervention effect on extremely high-risk patients.

[0122] A high-risk activation signal is sent to the timing decision unit, and a basic application duration indicator or an extended application duration indicator is sent to the aseptic operation guidance unit. Specifically:

[0123] The risk stratification unit sends the generated high-risk activation signal to the timing decision unit in real time, triggering the timing decision unit to start the timing and time window matching process;

[0124] At the same time, based on the patient's risk level, the corresponding application duration identifier is sent to the aseptic operation guidance unit. After receiving the identifier, the aseptic operation guidance unit determines the specific application duration. If the received identifier is the basic application duration identifier, the operation is performed according to the basic application duration.

[0125] If the received instruction is for extended application time, the operation will be performed according to the extended application time to achieve a precise match between risk level and application time.

[0126] In this implementation plan, the equivalent score is compared with the preset risk score threshold of 16 points and the next higher threshold of 22 points in a triple comparison to classify patients into three categories: non-high-risk, high-risk, and very high-risk. When the equivalent score is between 16 and 22 points, a high-risk activation signal is generated and a basic application duration indicator is attached, corresponding to an application duration of 8 hours. When the equivalent score is greater than or equal to 22 points, an extended application duration indicator is attached, corresponding to an application duration of 12 hours. The high-risk activation signal is sent to the timing decision unit to start timing and time window matching, and the application duration indicator is sent to the aseptic operation guidance unit to determine the application duration, so as to achieve precise matching between risk level and intervention intensity.

[0127] Specifically, the aseptic operation guidance unit performs the following steps for the application procedure:

[0128] Upon receiving the patch application instruction, the operator is guided to perform hand cleaning and aseptic disinfection, specifically as follows:

[0129] The operator of the patch application should be a medical staff or a trained nursing staff. The operator should first rinse their hands with running sterile water for at least 2 minutes, focusing on cleaning areas such as between the fingers and under the nails where dirt can easily accumulate. After rinsing, dry the hands with a sterile towel.

[0130] Then, 75% medical alcohol was used to disinfect the hands, covering the palms, backs of hands, fingers, and 5cm above the wrists. After applying the alcohol, it was allowed to dry naturally for at least 30 seconds to ensure that the hands were sterile and to avoid introducing microorganisms during the operation.

[0131] Instruct the operator to wear double-layered sterile gloves, with an inner layer of polyethylene film gloves and an outer layer of nitrile rubber gloves, and perform a glove airtightness check, specifically as follows:

[0132] Instruct operators to first put on inner polyethylene film gloves, ensuring that the gloves are undamaged and wrinkle-free, and that the glove opening covers 3cm above the wrist.

[0133] Wear nitrile rubber gloves on the outside. Nitrile rubber gloves have good sterility, abrasion resistance and fit, which can further improve the sterile protection effect.

[0134] After the gloves are put on, guide the operator to perform an airtightness check. The check method is as follows:

[0135] Make fists with both hands and inflate the gloves. Observe whether the gloves are leaking or damaged. If there is any leakage or damage, replace the gloves immediately until the gloves are airtight enough to avoid the risk of infection due to glove damage.

[0136] If an extended application time notification is received, the preset application time will be increased by a preset percentage, specifically:

[0137] The aseptic operation guidance unit receives the application duration indicator sent by the risk stratification unit. If the received indicator is for extended application duration, the application duration is increased by a preset percentage.

[0138] In one implementation, the preset percentage is set to 50%. If the basic application time is 8 hours, the extended application time is 8 × (1 + 50%) = 12 hours.

[0139] If no extended application time indicator is received (i.e., the basic application time indicator is received), the preset application time remains unchanged to ensure that the application time matches the patient's risk level.

[0140] Apply iodine-soaked cotton swabs to the Shenque acupoint area twice in a circular motion and allow it to dry. Then, apply ethanol-soaked cotton swabs to remove the iodine once and allow it to dry. The specific steps are as follows:

[0141] First, locate the patient's Shenque acupoint (i.e., the center of the navel). Use sterile forceps to pick up a sterile cotton swab soaked in iodine. With the Shenque acupoint as the center, perform circular disinfection. The disinfection area is a 5cm area around the Shenque acupoint. Rotate the cotton swab during disinfection to ensure that no area is missed. After the first disinfection is completed, replace it with a new iodine cotton swab and perform the second disinfection in the same way.

[0142] After disinfection, wait for the iodine solution to dry naturally for at least 1 minute to avoid the ointment deteriorating or affecting skin absorption if the iodine solution is not dry.

[0143] Then, using a sterile cotton swab soaked in 75% medical ethanol, perform iodine removal treatment on the Shenque acupoint area in the same circular pattern. After iodine removal, allow it to dry naturally for at least 30 seconds. Iodine removal can avoid iodine residue irritating the skin and reduce the risk of skin allergies or inflammation.

[0144] The pre-prepared ointment is evenly applied to the center of a sterile gauze pad, which is then applied to the Shenque acupoint and secured with sterile breathable adhesive tape. Specifically:

[0145] Use sterile tweezers to remove the pre-set dose of traditional Chinese medicine umbilical therapy ointment. The pre-set dose is determined based on the patient's body type and risk level.

[0146] In one embodiment, the preset dosage is 2g. The ointment is evenly applied to the center of a sterile gauze, with a coating diameter of about 2cm, to ensure that the ointment thickness is uniform (about 0.5cm) and to avoid the ointment being too thick or too thin, which would affect the absorption effect.

[0147] Then, align the center of the sterile gauze with the ointment with the patient's Shenque acupoint, and gently press it to ensure that the gauze is in close contact with the skin, without air bubbles or wrinkles;

[0148] Finally, use sterile breathable tape to secure the gauze around the edge, covering an area 1cm outside the edge of the gauze. This ensures the gauze is firmly fixed and not easily falls off. At the same time, the sterile breathable tape ensures skin breathability, reduces skin irritation, and is suitable for the sensitive skin condition of hematopoietic stem cell transplant patients.

[0149] In this implementation plan, hand cleaning and 75% alcohol disinfection are adapted to the operation requirements of a Class 100 laminar flow sterile chamber; the inner layer of polyethylene film gloves and the outer layer of nitrile rubber gloves constitute double-layer sterile protection, and the airtightness of the gloves is checked to ensure that there is no damage; when receiving the indication of extended application time, the application time is increased by 50% to match the application time with the patient's risk level; after two circular disinfections with povidone-iodine, allow it to dry, and after one removal of iodine with an alcohol swab, allow it to dry to avoid irritation of the sensitive skin of hematopoietic stem cell transplant patients by povidone-iodine residue; the preset dose of 2 grams of ointment is evenly applied to the center of sterile gauze, applied to the Shenque acupoint, and fixed with sterile breathable tape.

[0150] Specifically, the steps for refining the ointment in the umbilical therapy ointment preparation unit are as follows:

[0151] After weighing out the ingredients (Pinellia ternata, cloves, persimmon calyx, dried ginger, and Codonopsis pilosula) according to the preset proportions, coarse pulverization is performed, specifically as follows:

[0152] The dried Pinellia ternata, cloves, persimmon calyx, dried ginger, and Codonopsis pilosula were weighed using an electronic balance according to a preset ratio, with a weighing accuracy of 0.1g.

[0153] After weighing, mix the five medicinal materials evenly and put them into a universal grinder for coarse grinding. Set the speed of coarse grinding to 1000 r / min and the grinding time to 5 minutes to grind the medicinal materials to a particle size of 100-150 mesh. The coarsely ground medicinal materials have uniform particles, which can facilitate subsequent fine grinding and retain the effective components of the medicinal materials.

[0154] A vibrating ultrafine pulverizer is used to further pulverize the coarsely pulverized medicinal material particles until the particle size is smaller than a preset particle size threshold, and then sieve separation is completed. Specifically:

[0155] The coarsely pulverized medicinal material granules were placed into a vibrating ultra-micro pulverizer, with the vibration frequency set to 2000 r / min and the pulverization time set to 20 minutes for fine pulverization.

[0156] The preset particle size threshold is set to 200 mesh. After fine grinding, the medicinal powder is put into a sterile sieve machine and sieved through a 200-mesh sterile sieve to remove coarse particles that do not meet the particle size requirements. The coarse particles are then put back into the ultrafine pulverizer for further grinding until all powder particles are smaller than 200 mesh.

[0157] The sieving process is carried out in a Class 100 sterile environment to avoid microbial contamination of the powder. The finely pulverized medicinal powder has a larger specific surface area, which can improve the dissolution rate of the active ingredients and enhance the efficacy of the umbilical therapy ointment.

[0158] The sieved powder is thoroughly mixed with the petroleum jelly base to form a paste, specifically as follows:

[0159] Take the sieved Chinese medicine powder and petroleum jelly base according to the preset ratio. The petroleum jelly base is medical sterile petroleum jelly.

[0160] In one embodiment, the mass ratio of the traditional Chinese medicine powder to the petrolatum matrix is ​​1:4;

[0161] Place the petroleum jelly base in a sterile mixing bowl and heat it to 50°C to melt it. Stir at 500 rpm. After the petroleum jelly is completely melted, slowly add the Chinese herbal powder while stirring. Stir for 30 minutes to ensure that the Chinese herbal powder and petroleum jelly base are fully mixed without clumping or layering.

[0162] After mixing, stop heating and allow it to cool naturally to room temperature to form a uniform and smooth paste. The cooling process is carried out in a sterile environment to avoid contamination of the paste.

[0163] After the ointment is sealed in a sealed container, it is sterilized by irradiation using cobalt-60 rays within a preset irradiation dose range, specifically as follows:

[0164] The cooled ointment is placed in a sterile sealed bag, and each sealed bag is filled with a preset dose of ointment (2g / bag in one embodiment). The bag is sealed using a heat sealer. During the sealing process, it is ensured that the sealed bag is undamaged and leak-proof to prevent the ointment from coming into contact with air and deteriorating.

[0165] After packaging, the sealed ointment is placed in a cobalt-60 irradiation sterilization device. The preset irradiation dose range is 25-30 kGy, and the irradiation time is 30 minutes. The cobalt-60 rays kill bacteria, fungi and other microorganisms in the ointment and sealed bag, and the sterilization qualification rate reaches more than 99.9%.

[0166] After irradiation sterilization is completed, the ointment is sampled and tested. Once the test is qualified, it is stored in a sterile storage environment to meet the requirements of Class 100 laminar flow sterile chamber and reduce the risk of infection for patients.

[0167] In this implementation plan, the coarse pulverizer grinds the medicinal materials to 100 to 150 mesh, which facilitates subsequent fine pulverization while preserving the effective components of the medicinal materials; the vibrating ultrafine pulverizer performs fine pulverization to a particle size of less than 200 mesh, and the sieving and separation are completed in a Class 100 sterile environment to avoid microbial contamination of the powder; the Chinese medicine powder is mixed with medical sterile petroleum jelly at a mass ratio of 1:4, heated to 50 degrees Celsius and stirred for 30 minutes, and then cooled to form a uniform and delicate paste; sterilization is performed by cobalt-60 irradiation at a dose of 25 to 30 kGy, with a sterilization qualification rate of over 99.9%, and after sealing and packaging, it is adapted to the requirements of Class 100 laminar flow sterile chamber, eliminating the risk of infection caused by the rough preparation process of traditional Chinese medicine plasters.

[0168] Specifically, the steps for the closed-loop adaptive unit to perform incremental learning are as follows:

[0169] When the nausea and vomiting severity score is greater than the preset threshold after two consecutive applications of the patch, a model retraining trigger signal is generated, which is as follows:

[0170] The preset scoring threshold is the critical value for determining that the intervention effect is not good;

[0171] In one implementation, the preset scoring threshold is set to 2 points (i.e., level 2 nausea and vomiting). The closed-loop adaptive unit records the nausea and vomiting grade score after each application in real time. When the score after two consecutive applications is greater than 2 points, it indicates that the current model's risk assessment accuracy is insufficient and the intervention effect is poor. At this time, a model retraining trigger signal is generated. The trigger signal contains the triggering reason, the score data of the two consecutive applications, and the corresponding comprehensive risk probability value, providing a basis for incremental learning.

[0172] If no two consecutive scores > 2 occur, no trigger signal is generated, and the model retains its current parameters.

[0173] Upon responding to the trigger signal, the training samples with the most recent preset number of iterations are extracted from the historical database, and incremental learning is performed, specifically as follows:

[0174] The preset number of applications is set to 20, meaning that the relevant data from the most recent 20 applications are extracted from the historical database as training samples. Each training sample includes:

[0175] Data on Drantsaris baseline score, age, sleep duration, and number of vomiting episodes before patch application; comprehensive risk probability value before patch application; and nausea and vomiting grade score after patch application.

[0176] After extraction, the training samples are divided into a training set and a validation set in a 7:3 ratio. The training set is used for model training, and the validation set is used for model performance verification.

[0177] Incremental learning uses gradient descent to train the model with the goal of minimizing the prediction error of the training set, avoiding retraining on all historical data and improving model update efficiency.

[0178] During incremental learning, only the NK-1 pathway weight coefficients in the logistic regression model are updated, while other model parameters are frozen and remain unchanged. Specifically:

[0179] During incremental learning, the weight coefficients corresponding to the Drantsaris baseline score, age, sleep duration, and number of vomiting episodes in the logistic regression model are locked and their values ​​are kept unchanged. Only the independent weight coefficients of the NK-1 pathway are updated.

[0180] The updated model parameters are then distributed to the risk stratification unit, specifically as follows:

[0181] After incremental learning is completed, the updated model is validated using a validation set. The validation metric is the average prediction error. If the average prediction error of the validation set is less than the average prediction error before the update, it indicates that the model optimization is effective. The updated NK-1 pathway weight coefficients and complete model parameters are then sent to the risk stratification unit to replace the original model parameters.

[0182] If the average prediction error of the validation set is greater than or equal to the average prediction error before the update, it indicates that the model optimization is ineffective. In this case, the update should be abandoned, the original model parameters should be kept unchanged, and the validation results should be recorded.

[0183] In this implementation plan, when the nausea and vomiting grade score is greater than 2 after two consecutive applications, a model retraining trigger signal is generated to determine that the current model's risk assessment accuracy is insufficient. The most recent 20 application training samples are extracted from the historical database and divided into training and validation sets in a 7:3 ratio. Gradient descent is used to perform incremental learning. During the incremental learning process, only the NK-1 pathway weight coefficients in the logistic regression model are updated, while the weight coefficients corresponding to the Drantsaris baseline score, age, sleep duration, and number of vomiting episodes are frozen to ensure model stability. When the average prediction error of the validation set is less than the average before the update, the updated model parameters are sent to the risk stratification unit to replace the original parameters, realizing adaptive optimization of the model as the treatment progresses.

[0184] Specifically, the time-series decision unit performs fault-tolerant processing of sleep data as follows:

[0185] Discrete-point smoothing filtering is applied to the read sleep duration from the previous night, specifically as follows:

[0186] The moving average filtering method is used to smooth the sleep duration data. The filtering window is set to 3, that is, the average of the sleep duration data and the two adjacent data (if they exist) is taken as the filtered sleep duration data.

[0187] In one implementation, the sleep duration of the previous night was 5.5 hours, and the sleep durations of the two adjacent nights were 6 hours and 5.8 hours, respectively. Then the filtered sleep duration is (6+5.5+5.8) / 3≈5.77 hours.

[0188] If the sleep duration read is the first collection and there is no adjacent data, the read sleep duration is directly used as the filtered data;

[0189] Smoothing filters can eliminate random noise in sleep duration data and avoid time window matching errors caused by data fluctuations.

[0190] The system checks for missing or invalid sleep duration data. If data is missing or invalid, the application time window is set to Chenshi (7-9 AM) by default, and a sleep monitoring missing alarm signal is generated. Specifically:

[0191] The testing content includes:

[0192] If the sleep duration data is missing, exceeds the reasonable range (2-12 hours), or has obvious abnormalities (such as a difference of more than 3 hours from the patient's previous average sleep duration), the sleep duration data is determined to be missing or invalid.

[0193] At this time, the default application time window is set to Chenshi (7-9 am) to ensure that the application operation is carried out normally and to generate a sleep monitoring missing alarm signal. The alarm signal contains information such as the type of missing / invalid data and the current default application time window.

[0194] If the sleep duration data is valid, the filtered sleep duration will be precisely compared with a preset threshold to lock in a unique application time window, specifically as follows:

[0195] The preset sleep threshold is set to 7 hours. The filtered sleep duration is accurately compared with 7 hours, with a comparison accuracy of 0.1 hours. If the filtered sleep duration is ≥7 hours, the application time window is locked to Chen Shi (7-9 am).

[0196] If the filtered sleep duration is less than 7 hours, the application time window after offset is determined according to the sleep offset index, and a unique time window is locked to avoid multiple time windows causing operational confusion.

[0197] Perform conflict checking within the locked application time window. If no conflict is found, output the application execution command, which is as follows:

[0198] The content of the conflict check is:

[0199] Does the locked application time window conflict with the patient's other diagnostic and treatment procedures (such as chemotherapy, infusion, examination, etc.)? The timing decision unit obtains the patient's current treatment plan from the hospital's diagnostic and treatment system, compares the application time window with the operation time in the treatment plan, and if there is no time overlap, it is determined that there is no conflict and the application execution instruction is output.

[0200] If there is a time overlap, adjust the application time window (fine-tune within the offset time period to ensure that it does not conflict with other diagnostic and treatment operations), and perform conflict check again after adjustment until there is no conflict before outputting the application execution command;

[0201] An alarm signal is sent to the aseptic operation guidance unit, with an additional prompt for the operator to inquire about the patient's subjective sleep quality, specifically:

[0202] The timing decision unit sends the generated sleep monitoring missing alarm signal to the aseptic operation guidance unit in real time. The alarm signal is accompanied by a prompt message, which reads: "Sleep duration data is missing / invalid. The default application time window is set to Chenshi (7-9 AM). Please ask the patient about their subjective sleep quality the night before and record the feedback."

[0203] After receiving the prompt, the operator asks the patient about their sleep experience the previous night (such as whether they had difficulty falling asleep, whether they woke up easily, and whether they got enough sleep), and records the feedback in the system to provide supplementary data for the subsequent model update of the closed-loop adaptive unit.

[0204] In this implementation scheme, the moving average filtering method is used to smooth the sleep duration of the previous night, and the filtering window is set to 3 to eliminate random noise.

[0205] When sleep duration data is missing or invalid, the default application time window is set to 7 to 9 AM, and a sleep monitoring missing alarm signal is generated to ensure that the application operation is not interrupted due to abnormal data.

[0206] The filtered sleep duration is accurately compared with the 7-hour preset threshold to lock a unique application time window and avoid multiple time windows that could cause operational confusion.

[0207] The locked application time window is checked for conflicts with the patient's chemotherapy, infusion and other operation time in the hospital diagnosis and treatment system. If there is no conflict, the application execution instruction is output. If there is a conflict, it is fine-tuned within the offset time period. The alarm signal is sent to the aseptic operation guidance unit, and the operator is prompted to ask the patient about subjective sleep quality. The feedback result is recorded in the system as supplementary data for model update.

[0208] As another implementation method, the specific steps for the risk stratification unit to perform feature correlation screening are as follows:

[0209] The correlation between the baseline score and each risk factor is calculated using the following characteristics:

[0210] The association between the Drantaris baseline score and three risk factors—age, sleep duration, and number of vomiting episodes—was calculated using the Pearson correlation coefficient method. The Pearson correlation coefficient r ranged from [-1, 1], and the closer the absolute value of r was to 1, the stronger the association between the two features.

[0211] The closer the absolute value of r is to 0, the weaker the correlation between the two features.

[0212] After the calculation is completed, the correlation coefficients between the base score and each risk factor are obtained;

[0213] In one implementation, the correlation coefficient between baseline score and sleep duration is -0.65 (negative correlation, the shorter the sleep duration, the higher the baseline score), the correlation coefficient between baseline score and age is -0.42, and the correlation coefficient between baseline score and the number of vomiting episodes is 0.71 (positive correlation, the more vomiting episodes, the higher the baseline score).

[0214] Based on the correlation calculation results, redundant input features were removed, and three effective features—age, sleep duration, and number of vomiting episodes—were retained. Specifically:

[0215] Set a correlation threshold;

[0216] In one implementation, the correlation threshold is set to 0.3, and features with an absolute correlation coefficient less than 0.3 are identified as redundant features and removed.

[0217] Features with an absolute correlation coefficient ≥ 0.3 are considered valid features and will be retained.

[0218] In the calculation results, the absolute values ​​of the correlation coefficients between age, sleep duration, and number of vomiting history and the baseline score were all ≥0.3, indicating no redundant features. Therefore, these three effective features were retained.

[0219] If other risk factors (such as weight, gender, etc.) exist, their correlation coefficient absolute value is <0.3, and they are removed. Feature screening can reduce the number of input features to the model, reduce the computational complexity of the model, and avoid the impact of redundant features on the model's prediction accuracy.

[0220] The sleep duration feature is transformed using a segmented encoding method, specifically as follows:

[0221] The segmented coding method assigns different weights to different sleep duration intervals based on the strength of the correlation between sleep duration and delayed CINV risk. The stronger the correlation, the greater the coding weight.

[0222] In one implementation, sleep duration ≥ 7 hours is encoded as a low encoding value 1;

[0223] Sleep duration of 5 hours ≤ sleep duration < 7 hours is coded as medium code value 2;

[0224] Sleep duration < 5 hours, coded as a high coding value 3;

[0225] During the encoding conversion, the encoded value corresponding to the sleep duration is multiplied by the corresponding weight to obtain the converted sleep duration feature value. For example, if the sleep duration is 6 hours (with an encoded value of 2), the converted feature value is 2 × 0.3 = 0.6.

[0226] Segmented coding can highlight the impact of insufficient sleep duration on risk assessment;

[0227] The transformed three features are then linearly weighted and summed with the Drantisaris baseline score. The initial weights for the three features are the baseline weight, auxiliary weight, and adjustment weight, respectively, as follows:

[0228] Set baseline weight values ​​(corresponding to age characteristics), auxiliary weight values ​​(corresponding to sleep duration characteristics), and adjustment weight values ​​(corresponding to the number of vomiting history characteristics);

[0229] In one implementation, the baseline weight is set to 0.2, the auxiliary weight is set to 0.3, the adjustment weight is set to 0.3, and the weight of the Drantsaris baseline score is set to 0.2.

[0230] The formula for calculating the linear weighted sum is:

[0231] ;

[0232] in, For the weighted summation result, These are the converted age feature values. These are the converted sleep duration feature values. The converted vomiting history frequency characteristic value, The weighted sum of the Drantsaris base score is used as one of the inputs to the subsequent logistic regression model, which can enhance the impact of each effective feature on risk assessment.

[0233] The weights of the above three items and the NK-1 pathway weight coefficients are dynamically adjusted during initial model training or offline training, specifically as follows:

[0234] During the initial training or offline training of the model, with the goal of minimizing the prediction error, the weights of three features—age, sleep duration, and number of vomiting episodes—and the independent weight coefficients of the NK-1 pathway are dynamically adjusted using the gradient descent method. The Drantsaris baseline score weights are used as baseline score weights and participate in the model calculation separately, without being included in the dynamic adjustment range of the above three feature weights.

[0235] In one implementation, after training, the baseline weight value is adjusted to 0.25, the auxiliary weight value is adjusted to 0.35, the adjustment weight value is adjusted to 0.25, and the NK-1 pathway weight coefficient is adjusted to 0.15. Through dynamic adjustment, the model weights can be made to better fit the actual clinical data and improve the accuracy of risk assessment.

[0236] In this implementation scheme, the Pearson correlation coefficient method is used to calculate the association between the Dranthasaris baseline score and age, sleep duration, and the number of times of vomiting. The closer the absolute value is to 1, the stronger the association. The association threshold is set to 0.3. Features with an absolute value of correlation coefficient less than 0.3 are considered redundant features and are removed to reduce the computational complexity of the model. Sleep duration is segmented and coded into intervals with a high threshold of 7 hours and a low threshold of 5 hours. The higher the degree of sleep deprivation, the higher the risk code value, highlighting the impact of sleep deprivation on risk assessment. The three transformed features are linearly weighted and summed with the baseline score. The initial weights are 0.2, 0.3, 0.3, and 0.2, respectively. The gradient descent method is used to dynamically adjust the initial weights of the three features and the NK-1 pathway weight coefficient.

[0237] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0238] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system, characterized in that, include: The risk stratification unit is used to obtain the Drantsaris baseline score, age, sleep duration, and number of vomiting episodes. Sleep below a preset threshold, age below a preset threshold, and number of vomiting episodes greater than or equal to a preset number are used as risk factors. The baseline score and risk factors are input into a logistic regression model trained with historical HSCT data. This model sets independent weight coefficients for the delayed activation features of the NK-1 pathway and outputs a calibrated comprehensive risk probability value. When the equivalent score is higher than the preset risk score threshold, a high-risk activation signal is generated. The umbilical therapy ointment preparation unit is used to weigh out Pinellia ternata, cloves, persimmon calyx, dried ginger, and Codonopsis pilosula according to a preset ratio, pulverize and sieve them, mix them with petroleum jelly base to form an ointment, and then package and sterilize by irradiation. The timing decision unit is used to receive a high-risk activation signal and start timing from the end of chemotherapy. When the preset start time is reached, it reads the sleep duration of the previous night. If the sleep duration is greater than or equal to the preset threshold, the application time window is set to Chenshi (7-9 AM). Otherwise, it is set to the offset time period after Chenshi. When the current time falls into the time window, the application execution command is output. The aseptic operation guidance unit is used to respond to the application execution command, guide hand disinfection, put on sterile gloves, disinfect the Shenque acupoint twice with iodine swabs and let it dry, take out the preset dose of ointment and apply it to the center of sterile gauze, apply it to the Shenque acupoint and fix it with sterile breathable tape, continue to apply for the preset time, remove the gauze and clean the navel with sterile saline swabs. The closed-loop adaptive unit records the nausea and vomiting grade score within a preset time window after each application. It compares the score with the comprehensive risk probability value before the current application to calculate the prediction error. Based on the prediction error, it backpropagates to update the NK-1 pathway weight coefficient in the logistic regression model and sends the updated model to the risk stratification unit.

2. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for risk calculation performed by the risk stratification unit are as follows: The acquired data on age, sleep duration, and number of vomiting episodes were segmented and encoded. Sleep duration greater than or equal to the high threshold was encoded as a low value, sleep duration less than the high threshold but greater than or equal to the low threshold was encoded as a medium value, and sleep duration less than the low threshold was encoded as a high value. Age and the number of times a person has a history of vomiting are encoded as binary features according to preset thresholds; Perform co-directional normalization between the encoded features and the Drantisaris base score; The normalized data is input into the logistic regression model, and independent weighting calculations are performed on the delayed activation features of the NK-1 pathway. Output the calibrated overall risk probability value based on the weighted calculation results.

3. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for time window matching performed by the time-series decision unit are as follows: Perform a validity check on the read sleep duration data from the previous night; Calculate the sleep deviation index based on the verified sleep duration data; When the sleep offset index is greater than zero and not greater than the mild offset threshold, the application time window will be shifted backward by the mild offset amount. When the sleep offset index is greater than the mild offset threshold, the application time window will be shifted backward by the severe offset amount, which is greater than the mild offset amount. The current system time is matched and compared with the time window obtained after offset. If the match is successful, the application execution command is output.

4. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for error calculation and weight update of the closed-loop adaptive unit are as follows: Collect nausea and vomiting grading data within a preset time window after each patch application; The graded scoring data is numerically mapped to the comprehensive risk probability value before this application, and the prediction error is calculated. The weighted moving average method is used to perform a weighted summation of the prediction errors of the three most recent applications. The decreasing weights of the three errors are, in order, the recent weight, the medium-term weight, and the long-term weight, with the recent weight being greater than the medium-term weight, which is greater than the long-term weight. The weighted summation result is used as the basis for backpropagation updates to perform NK-1 pathway weight coefficient updates; The updated weight coefficients are then processed with regularization constraints and distributed to the risk stratification unit.

5. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for performing hierarchical identification processing in the risk stratification unit are as follows: Determine the relationship between the equivalent score corresponding to the calibrated comprehensive risk probability value and the preset risk score threshold; When the equivalent score is greater than the preset risk score threshold but less than the higher-level risk score threshold, a high-risk activation signal is generated and a basic application duration indicator is attached. When the equivalent score is greater than or equal to the threshold of the next higher risk score, a high-risk activation signal is generated and an extended application time indicator is attached. A high-risk activation signal is sent to the timing decision unit, and a basic application duration indicator or an extended application duration indicator is sent to the aseptic operation guidance unit.

6. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for the aseptic procedure guidance unit to perform the dressing procedure are as follows: Upon receiving the application instruction, guide the operator to perform hand cleaning and aseptic disinfection; Guide the operator to wear double sterile gloves, with an inner layer of polyethylene film gloves and an outer layer of nitrile rubber gloves, and perform a glove airtightness check. If an extended application time notification is received, the preset application time will be increased by a preset percentage. Apply iodine-soaked cotton swabs to the Shenque acupoint area twice in a circular motion and allow it to dry. Then apply ethanol-soaked cotton swabs to remove the iodine once and allow it to dry. Apply the pre-set dose of ointment evenly to the center of a sterile gauze, place it on the Shenque acupoint, and secure it with sterile breathable tape.

7. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for refining the ointment in the umbilical therapy ointment preparation unit are as follows: Weigh out the following ingredients according to the preset ratio: Pinellia ternata, cloves, persimmon calyx, dried ginger, and Codonopsis pilosula, and then perform coarse grinding. A vibrating ultrafine pulverizer is used to finely pulverize the coarsely pulverized medicinal material particles until the particle size is smaller than the preset particle size threshold, and then sieve separation is completed; The sieved powder is thoroughly mixed with the petroleum jelly base to form a paste. After the ointment is sealed in a sealed container, it is sterilized by irradiation using cobalt-60 rays within a preset irradiation dose range.

8. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for the closed-loop adaptive unit to perform incremental learning are as follows: When the nausea and vomiting scores after two consecutive applications are both greater than the preset score threshold, a model retraining trigger signal is generated. Upon responding to the trigger signal, the training samples with the most recent preset number of times are extracted from the historical database, and incremental learning is performed. During incremental learning, only the NK-1 pathway weight coefficients in the logistic regression model are updated; The updated model parameters are then distributed to the risk stratification unit.

9. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for the time-series decision unit to perform fault-tolerant processing of sleep data are as follows: Perform discrete-point smoothing filtering on the read sleep duration from the previous night; It detects whether sleep duration data is missing or invalid. If it is missing or invalid, it sets the application time window to Chenshi (7-9 AM) by default and generates a sleep monitoring missing alarm signal. If the sleep duration data is valid, the filtered sleep duration will be precisely compared with the preset threshold to lock in a unique application time window; Perform conflict checking on the locked application time window; if there is no conflict, output the application execution command. The alarm signal is sent to the aseptic operation guidance unit, with an additional prompt for the operator to inquire about the patient's subjective sleep quality.

10. The hematopoietic stem cell transplantation CINV risk assessment and antiemetic umbilical therapy management system according to claim 1, characterized in that, The specific steps for performing feature correlation screening in the risk stratification unit are as follows: Calculate the correlation between the baseline score and each risk factor; Redundant input features were removed based on the correlation calculation results, and three effective features were retained: age, sleep duration, and number of vomiting episodes. The sleep duration features are transformed using a segmented encoding method; The transformed three features are linearly weighted and summed with the Drantsaris baseline score. The initial weights for the three features are the baseline weight, auxiliary weight, and adjustment weight, respectively. The weights of the above three items and the weight coefficient of the NK-1 pathway are dynamically adjusted during the initial training of the model or during offline training.

Images