Medical aid decision-making system, computer equipment and storage medium

Abstract

The invention relates to the technical field of medical data processing, and discloses a medical aid decision making system, computer equipment and a storage medium, and the system comprises an original decision obtaining module which is used for determining original decision making data of a target object based on original multi-source data of the target object; wherein the original decision data is determined by using an original multi-task model; the original multi-task model comprises a bottom coding layer, a scheme recommendation branch and a state prediction branch; the updated data acquisition module is used for acquiring the updated multi-source data of the target object in the current intervention period; the target decision determination module is used for determining target decision data of the target object by utilizing the updated multi-source data; wherein the target decision data is determined by using a target multi-task model; and the follow-up intervention auxiliary module is used for guiding the follow-up intervention process of the target object based on the target decision data. The whole system forms a complete management closed loop, and the accuracy and dynamic adaptability of health management decisions are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of medical data processing technology, and in particular to a medical decision support system, computer device and storage medium. Background Technology

[0002] Children's normal growth and development is a core concern for families and healthcare professionals. However, existing assistive methods or systems suffer from limitations such as single data sources and fragmented analysis processes, resulting in insufficient personalized accuracy and dynamic adaptability. Therefore, there is an urgent need for a medical decision support system capable of personalized and dynamic adjustments. Summary of the Invention

[0003] This application aims to at least partially solve one of the technical problems in related technologies. To this end, this application proposes a medical auxiliary decision-making system, a computer device, and a storage medium. The main technical solutions adopted in this application include: In a first aspect, this application provides a medical decision support system, the system comprising: The original decision acquisition module is used to determine the original decision data of the target object based on the original multi-source data of the target object. The original decision data is determined using the original multi-task model. The original multi-task model includes a bottom coding layer, a scheme recommendation branch and a state prediction branch, and the scheme recommendation branch and the state prediction branch are respectively connected to the bottom coding layer. The updated data acquisition module is used to acquire updated multi-source data of the target object within the current intervention period; The target decision determination module is used to determine the target decision data of the target object using updated multi-source data; wherein, the target decision data is determined using a target multi-task model; The follow-up intervention support module is used to guide the follow-up intervention process for the target subjects based on the target decision data.

[0004] The system first generates initial decisions based on raw, multi-source data by using a shared underlying code and a collaborative multi-task branch model. This solves the problem of fragmented analysis and improves the comprehensiveness and relevance of the decisions. By collecting and updating data within the current intervention period, the system can dynamically track changes in health status and use this data to optimize the model and generate more personalized target decisions, ultimately achieving effective integration between decision-making and execution. Finally, the entire system forms a complete management loop of "assessment, recommendation, monitoring, and optimization," significantly improving the accuracy and dynamic adaptability of health management decisions.

[0005] Optionally, the original decision data includes the state prediction results of the target object and the original recommendation scheme; the original decision acquisition module also includes a model execution unit, used for: extracting features from the underlying encoding layer based on the original multi-source data to obtain shared feature data; wherein, the shared feature data refers to general feature data that can simultaneously describe the syndrome characteristics and growth characteristics of the target object; performing regression prediction based on the shared feature data using the state prediction branch to obtain the state prediction results; and performing probabilistic recommendation based on the shared feature data and the preset prescription library using the scheme recommendation branch to obtain the original recommendation scheme.

[0006] Feature extraction from raw multi-source data is performed at the underlying encoding layer to obtain shared feature data that simultaneously covers syndrome and growth characteristics, improving the comprehensiveness of feature representation. Regression prediction is performed through the state prediction branch to obtain state prediction results including increments, confidence intervals, and prediction curves, improving the accuracy and intuitiveness of growth prediction. Through collaborative work with the pre-set prescription library and knowledge graph, a structured original recommendation scheme is generated by the scheme recommendation branch, enhancing the personalization and practicality of intervention suggestions. Parallel operation of the two main task branches enables synchronous output of decision data, improving the system's decision-making efficiency.

[0007] Optionally, the model execution unit includes a decision subunit, used for: using the scheme recommendation branch to determine candidate recommendation schemes based on shared feature data and a pre-set prescription library; using the candidate recommendation schemes and shared feature data to determine candidate intervention suggestions using a pre-set knowledge graph; wherein, the candidate intervention suggestions include prescription-related suggestions and prescription-irrelevant suggestions; and performing structured integration processing based on the candidate recommendation schemes and candidate intervention suggestions to obtain the original recommendation scheme.

[0008] First, candidate recommendations are identified using shared feature data, visually demonstrating the suitability of each option. Then, a pre-defined knowledge graph is used to retrieve personalized intervention suggestions that are relevant to or irrelevant to the prescription, improving the comprehensiveness of the intervention guidance. Finally, all data is structured and integrated to generate original recommendation plans that are both accurate and actionable, providing high-quality auxiliary references for clinical intervention.

[0009] Optionally, the target decision determination module further includes a decision update unit, used for: optimizing the original multi-task model based on the updated multi-source data to obtain a target multi-task model; wherein the optimization is implemented using an incremental learning algorithm; and determining target decision data based on the updated multi-source data and the target multi-task model.

[0010] In the above implementation, the original multi-task model is first optimized using an incremental learning algorithm, which realizes personalized adaptation of model parameters and improves the model's fit to the individual response patterns of the target object. Then, the updated multi-source data is processed by the target multi-task model to generate targeted target decision data, which improves the dynamic adaptability of the decision scheme. The whole process constructs a closed-loop adaptive decision mechanism through data-driven model optimization and decision update, providing more accurate auxiliary references for subsequent intervention.

[0011] Optionally, the decision update unit further includes a trigger management subunit, used to: optimize the original multi-task model based on the updated multi-source data when the updated multi-source data meets the preset update conditions, to obtain the target multi-task model; wherein the preset update conditions include time conditions and / or data conditions.

[0012] The above methods achieve two goals: firstly, by utilizing time constraints to enable regular optimization within a normal cycle, ensuring the model's continuous adaptation to the long-term growth and development process; and secondly, by utilizing data constraints to enable timely triggering of abnormal situations, ensuring the model responds quickly to deviations in growth trajectory or changes in symptoms.

[0013] Optionally, the system also includes a raw data acquisition module, used to: acquire digital vital sign data and quantitative growth data of the target object; wherein, the digital vital sign data includes waveform feature data and image feature data; the quantitative growth data includes current feature data and historical feature data; and perform fusion processing based on the digital vital sign data and quantitative growth data to obtain raw multi-source data.

[0014] By combining TCM vector data with growth vector data, the growth status and syndrome status of the target object can be comprehensively characterized, laying a solid data foundation for subsequent accurate decision-making.

[0015] Optionally, the raw data acquisition module also includes a multimodal fusion unit, used for: performing vectorization transformation processing on digitized vital sign data to obtain TCM vector data; performing normalization processing on quantified growth data to obtain growth vector data; and performing splicing processing on TCM vector data and growth vector data to obtain raw multi-source data.

[0016] Digital vital sign data is transformed into TCM vector data, quantifying subjective TCM vital signs with objective indicators and improving the computability of the data. Furthermore, normalizing the quantitative growth data, which is already a quantitative indicator, eliminates the dimensional differences between different growth indicators, making the data comparable.

[0017] Secondly, this application also provides a computer device that includes the medical decision support system of any of the above.

[0018] Thirdly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the functions of any of the above-mentioned medical auxiliary decision-making systems.

[0019] Fourthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the functions of any of the above-mentioned medical auxiliary decision-making systems. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a structural block diagram of a medical auxiliary decision-making system according to an embodiment of this application; Figure 2 This is a structural block diagram of a medical auxiliary decision-making system according to yet another embodiment of this application; Figure 3 This is a structural block diagram of a medical auxiliary decision-making system provided according to another embodiment of this application; Figure 4 This is a structural block diagram of a medical decision support system according to another embodiment of this application; Figure 5 This is an internal structural diagram of a computer device provided according to an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Specifically, methods for managing children's growth and development in related technologies can be broadly divided into two directions: one is intelligent auxiliary systems for traditional Chinese medicine diagnosis and treatment, and the other is children's growth and development assessment and prediction tools based on modern medicine.

[0024] In the area of ​​intelligent TCM, related technologies mostly use artificial intelligence to analyze single modal information such as tongue and pulse to assist in constitution identification or recommend fixed prescriptions. Although this reflects the personalized diagnosis and treatment ideas of TCM, it lacks continuous tracking and integrated analysis of key quantitative indicators of children's growth such as bone age and body composition. As a result, intervention suggestions cannot be closely integrated with measurable growth trajectories, making it difficult to objectively assess the substantial impact on parameters such as height growth rate.

[0025] In modern medical assessment, related technologies mainly rely on quantitative data such as bone age assessment and growth curve comparison to predict adult height or screen for growth and development disorders through statistical models. However, they completely ignore the individual spleen and stomach, kidney qi and other TCM syndrome states of children. The prediction results and suggestions given are mostly general and cannot provide personalized conditioning plans based on TCM theory. They have limitations in "preventive treatment" and optimizing growth potential.

[0026] In recent years, although there have been attempts to integrate traditional Chinese medicine and Western medicine, these have mostly involved simply displaying information side-by-side or piecing together simple logic. They have not yet achieved a deep integration of the characteristics of TCM qualitative signs and modern quantitative indicators at the data level, nor have they constructed a closed-loop intelligent system that can self-optimize and dynamically adjust based on intervention feedback data. This leaves related solutions facing three core problems: First, at the data level, it is difficult to effectively extract, standardize, and deeply correlate unstructured TCM qualitative signs with structured modern quantitative data, making it impossible to comprehensively digitally represent children's growth and development status. Second, at the decision-making level, most suggestions are one-off, lacking a dynamic mechanism for automatically optimizing prescriptions and predictive models based on feedback data within the intervention period. Third, at the system level, a complete closed loop of "status assessment - intelligent recommendation - dynamic monitoring - feedback optimization" has not been formed, making it difficult to continuously adapt and adjust along with children's growth. Therefore, there is an urgent need for a medical auxiliary decision-making solution for children's growth and development that can solve the above problems and achieve deep integration of multi-source data, dynamic decision-making, and closed-loop optimization.

[0027] This embodiment provides a medical decision support system 100, such as Figure 1 As shown, the system includes an original decision acquisition module 110, an updated data acquisition module 120, a target decision determination module 130, and a subsequent intervention assistance module 140.

[0028] The original decision acquisition module 110 can refer to a module with multi-source data analysis and multi-task decision-making functions, which can be used to determine the original decision data of the target object based on the original multi-source data of the target object.

[0029] The target group can refer to individual children receiving growth and development health management. The original multi-source data can refer to a structured dataset that comprehensively characterizes the physiological state and growth characteristics of the target group. For example, it can simultaneously include health-related information of the target group in different dimensions, such as traditional Chinese medicine signs and modern quantitative data, including pulse waveform characteristics, tongue image characteristics, bone age, height, and weight. Specifically, intelligent instruments can be used to collect the aforementioned multi-dimensional data of the target group, and feature extraction and fusion processing can be performed on these data to obtain the original multi-source data.

[0030] Original decision data can refer to a comprehensive dataset used to assist a target subject in making initial health interventions, and may include the target subject's state prediction results and original recommended plans. For example, the original recommended plan may be a personalized plan that includes traditional Chinese medicine prescriptions and lifestyle adjustment suggestions, while the state prediction results may refer to a dataset showing the target subject's growth status over a future period of time, such as the increase in a child's height and weight over the next 6 months and their confidence intervals. In addition, it may also include prediction curves composed of the predicted height and predicted weight values ​​for a child over the next 6 months.

[0031] It should be noted that the original decision data can be determined using an original multi-task model. This original multi-task model can refer to a pre-trained deep learning model that can collaboratively process multiple parallel tasks. Based on the original multi-source data of the target object, it can simultaneously perform two major tasks: scheme recommendation and state prediction, so as to synchronously output the state prediction results and the original recommended scheme of the target object.

[0032] Specifically, the original multi-task model may include a bottom-level encoding layer, a scheme recommendation branch, and a state prediction branch, with the scheme recommendation branch and state prediction branch each connected to the bottom-level encoding layer. The bottom-level encoding layer can be a network layer capable of extracting high-order shared features from the original multi-source data. The scheme recommendation branch can be a classification network capable of calculating prescription probabilities based on shared features, and the state prediction branch can be a regression network capable of predicting future growth states based on shared features.

[0033] Furthermore, the working process of the original multi-task model can be summarized as follows: First, the original multi-source data is input into the bottom coding layer. After feature extraction, general feature data that comprehensively reflects the characteristics of children's syndromes and growth patterns can be obtained. Subsequently, this general feature data is simultaneously fed into the scheme recommendation branch and the state prediction branch. The scheme recommendation branch outputs the probability distribution and related suggestions for the preset prescription library as the original recommended scheme. The state prediction branch outputs the future growth situation (which can be an increment, confidence interval, and predicted sequence curve) as the state prediction result. Finally, the outputs of these two branches are integrated to form the original decision data.

[0034] For example, assuming the target is an 8-year-old child with shorter than average height, the features reflected in their original multi-source data may include feature A: pale tongue with white coating; feature B: weak pulse; feature C: bone age 7.5 years; feature D: height data from the past year; and feature E: weight data from the past year. The model first extracts the feature "spleen deficiency" by combining features A and B through the bottom coding layer, and then extracts the feature "growth retardation" by combining vectors C, D, and E. After integration, it generates general feature data that reflects the characteristics of "spleen deficiency syndrome: pale tongue with white coating and weak pulse; growth retardation: height, weight, and composition analysis data from the past year".

[0035] Subsequently, the recommendation branch, based on this feature, calculates the recommendation probability of all commonly used prescriptions in the preset prescription library, and finally outputs candidate recommendation schemes according to the probability values. These candidate recommendation schemes can take the following form: the preferred prescription is Liu Jun Zi Tang (recommendation weight 0.85), the alternative prescription 1 is Si Jun Zi Tang (recommendation weight 0.51)... and the alternative prescription N is Shen Ling Bai Zhu San (recommendation weight 0.12), where N is the total number of commonly used prescriptions in the preset prescription library. Then, combining the modification rules corresponding to "spleen deficiency" in the preset knowledge graph, candidate intervention suggestions are output, including prescription-related suggestions (such as adding 6g of hawthorn and 9g of malt, and a dosage suggestion of one dose per day, decocted in 150ml of water and taken twice warm) and prescription-unrelated suggestions (such as lifestyle suggestions like 15 minutes of exercise per day). Finally, the candidate recommendation schemes and candidate intervention suggestions are integrated and output as the original recommendation scheme.

[0036] Meanwhile, the state prediction branch uses the height, weight, and composition analysis data from the past year in the general feature data to predict the height increase and its confidence interval (e.g., an increase of 3.5±1cm) and the weight increase and its confidence interval (e.g., an increase of 2.5±1kg) in the next 6 months using a regression analysis algorithm. It also plots the height change curve (e.g., initially 120cm, 120.5cm in the first month, etc.) and the weight change curve (e.g., initially 26kg, 26.2kg in the first month, etc.) and integrates them as the state prediction result output.

[0037] Finally, the original recommended schemes and state prediction results are integrated and processed according to the preset output format to obtain comprehensive original decision data.

[0038] The updated data acquisition module 120 can refer to a module with continuous data acquisition and preprocessing functions, which can be used to acquire updated multi-source data of the target object within the current intervention period.

[0039] It's important to note that because child growth and development is a continuous and dynamic process, and the effects of TCM interventions require observation and adjustment, the concept of an intervention cycle is introduced. Within the current intervention cycle, decision-makers (physicians or parents) can refer to the original decision data to manage the child's health and development, and continuously collect the latest data on the target individual during implementation to optimize decisions. This implementation period is the current intervention cycle, used to observe the intervention's effects and collect dynamic change data. Updated multi-source data refers to a structured dataset reflecting the latest health status of the target individual, which is re-collected and integrated within the current intervention cycle.

[0040] Specifically, similar to the acquisition method of original multi-source data, the latest digital vital signs data and the latest quantitative growth data of the target subject are collected during the current intervention period. For example, assuming the current intervention period is set to 8 weeks, parents can collect digital vital signs data such as the child's new tongue appearance and new pulse appearance daily or weekly through smart sensing devices, and collect quantitative growth data such as the child's new height and new weight using a height chart and a weight scale.

[0041] In addition to the original multi-source data, the updated multi-source data can also include symptom assessment data reflecting recent changes, compared to the original multi-source data. This symptom assessment data refers to scoring data that quantifies the subjective changes in symptoms of the target population. It can be collected through standardized numerical scales or structured questionnaires and can describe parents' subjective observations of changes in their children's symptoms. For example, this subjective symptom change data can include appetite and eating habits (changes in loss of appetite, picky eating, food intake, and eating speed), digestion and bowel movements (changes in bloating, abdominal pain, stool consistency, frequency, and regularity of bowel movements), sleep quality (changes in difficulty falling asleep, sleep stability, nighttime awakenings, teeth grinding, etc.), mental state and physical strength (changes in daily fatigue, willingness to engage in activities, play stamina, and concentration), and specific related symptoms (changes in night sweats, hot palms and soles, bad breath, and recurrent colds, etc.). Each item can be evaluated using a 0-4 point Likert five-point scale or a five-point option of "significant improvement / improvement / no change / worsening / significantly worsening". After collecting the above scoring data, the system can automatically encode it into a structured vector S_symptom=[s1,s2,s3,s4,s5], where each component corresponds to a quantitative score of a symptom item.

[0042] Subsequently, through a vectorization and standardization fusion process similar to that used to acquire the original multi-source data, updated multi-source data is obtained. That is, the system integrates these new digitized vital signs data, new quantitative growth data, and scoring data to generate an updated multi-source data vector that can characterize the child's new state after intervention.

[0043] The target decision determination module 130 can refer to a module with model optimization, data operation and decision update functions, which can be used to determine the target decision data of the target object using updated multi-source data.

[0044] Among them, target decision data can refer to decision reference information generated after optimization and adjustment based on updated multi-source data, which is used to guide the next stage of health management intervention process. It can also include updated prediction results of growth and development and optimized recommendations for prescription intervention.

[0045] Specifically, the target decision data can be determined using a target multi-task model. This target multi-task model can be an adaptive model that is based on the original multi-task model and optimized by updating multi-source data to better reflect the real-time individual state of the target object.

[0046] Understandably, the target multi-task model shares the same architecture as the original multi-task model, both including a low-level encoding layer, a scheme recommendation branch, and a state prediction branch. Only the model parameters are adjusted and optimized based on the updated multi-source data of the target object. For example, assuming the target object is an 8-year-old child with shorter stature, the original multi-task model could be a general model trained on large-scale population data, outputting an original recommended scheme ("Liu Jun Zi Tang") and candidate intervention suggestions (adding 6g of hawthorn and 9g of malt, one dose daily, decocted in 150ml of water and taken twice warm, and monthly massage), as well as state prediction results including a height increase of 3.5±1cm and a weight increase of 2.5±1kg. After obtaining updated multi-source data of the child during the 8 weeks of health management intervention (e.g., a height increase of 1.2cm, no significant weight gain but a significantly improved appetite), the system can use an incremental learning algorithm to fine-tune the parameters of the original multi-task model using this new data. For example, the main adjustment might be to the neuron connection weights related to "strengthening the spleen" in the scheme recommendation branch, generating a target multi-task model that better understands the child's response patterns. Subsequently, the updated multi-source data is input into this target multi-task model, which will output the adjusted optimized recommendation scheme (such as suggesting to maintain the main prescription of Liu Jun Zi Tang, but increase the dosage of hawthorn to 9 grams) and updated prediction results (such as adjusting the increment of predicted height and weight for the next 6 months and redrawing the height and weight change curves). These outputs can be integrated and used as target decision data.

[0047] The follow-up intervention support module 140 can refer to a module with decision information presentation and intervention guidance generation functions, capable of guiding the follow-up intervention process for the target subject based on the target decision data. The follow-up intervention process can refer to the next health management stage immediately following the current intervention cycle; that is, the process of further implementing child health management with reference to the target decision data.

[0048] Specifically, the target decision data, as the core reference for formulating and implementing the next stage of personalized intervention measures, can first be displayed using the system's visualization interface, so that the decision-makers can determine the intervention plan that actually needs to be implemented based on rich clinical experience and actual conditions.

[0049] For example, assuming the preset output format is a report, the final standardized and integrated target decision data can be displayed on the visualization interface as follows: After the previous cycle of treatment, appetite has significantly improved, but growth rate is still slow. The recommended optimized plan for the next cycle (next 8 weeks) is: continue taking Liu Jun Zi Tang (Six Gentlemen Decoction), with the hawthorn dosage adjusted to 9 grams; and increase daily jumping exercise by 15 minutes. The updated prediction for the next cycle is: the expected height increase at the end of the next stage is 2.0±1.2 cm, and the weight increase is 1.9±1 kg. Height and weight change curves are also attached. This report allows doctors to review it and, in conjunction with clinical examination, ultimately determine the specific prescription dosage, dietary adjustments, and exercise intervention plan for the next cycle.

[0050] Furthermore, in the next intervention cycle, the above steps can be repeated, that is, to collect subsequent multi-source data of the target object again, so as to optimize the parameters of the target multi-task model again, so that it can better fit the real-time individual state of the target object.

[0051] It is important to note that all raw decision data and target decision data generated by this system are auxiliary reference information and cannot be directly used as medical implementation plans. The final diagnosis and prescription adjustment must be made by a licensed physician based on a comprehensive assessment of the target patient's specific situation and clinical experience. In other words, this implementation claims protection for a purely computer-based data processing system. It processes objective data (such as image pixels, digital signals, and standardized scores), and outputs another set of data (such as prescription probability distributions, predicted values, and structured suggestion text). This result itself does not constitute a clinically meaningful diagnosis or treatment, but is merely reference information calculated using complex algorithms. Whether and how to adopt this information depends entirely on the physician's independent professional judgment.

[0052] In the above implementation, the system first generates initial decisions based on original multi-source data through a shared underlying code and a collaborative multi-task branch model. This solves the problem of fragmented analysis and improves the comprehensiveness and relevance of the decisions. By collecting and updating data within the current intervention cycle, the system can dynamically track changes in health status and use this data to optimize the model and generate more personalized target decisions, ultimately achieving effective integration between decision-making and execution. Finally, the entire system forms a complete management closed loop of "assessment, recommendation, monitoring, and optimization," significantly improving the accuracy and dynamic adaptability of health management decisions.

[0053] In some implementation methods, please refer to the appendix. Figure 2 The system also includes a raw data acquisition module 210, which is used to: acquire digital vital sign data and quantitative growth data of the target object; and perform fusion processing based on the digital vital sign data and quantitative growth data to obtain raw multi-source data.

[0054] Digital vital sign data refers to a set of objective data that characterizes the physical signs of a target subject in Traditional Chinese Medicine (TCM). Specifically, digital vital sign data can include waveform feature data and image feature data. Waveform feature data refers to the characteristic values ​​that quantitatively describe dynamic vital signs such as pulse, i.e., feature data extracted from the waveform of human physiological signals captured by sensing devices. For example, the radial artery pulse wave acquired using a digital pulse diagnostic instrument, after filtering, denoising, and feature point recognition, can extract waveform feature data such as pulse rate, pulse strength, rhythm, and specific frequency domain energy. Image feature data refers to numerical parameters that quantitatively describe the morphological and color characteristics of the tongue, i.e., feature data obtained by analyzing images of external physical signs of the human body captured by image acquisition devices. For example, the tongue image captured by a tongue image acquisition device, after color space conversion, region segmentation, and texture analysis, can extract image feature data such as the RGB mean of tongue color, tongue coating coverage, crack index, and ecchymosis density.

[0055] Quantitative growth data refers to a set of numerical data that accurately reflects the growth and development status of a target individual. Specifically, it can include current characteristic data and historical characteristic data. Current characteristic data refers to quantitative data reflecting the current growth status, i.e., the various growth indicators measured at the most recent assessment point. This may include currently measured bone age, height, and weight, and may also include less frequently measured body composition data (body fat percentage, muscle mass), as appropriate.

[0056] Historical characteristic data can refer to quantitative data reflecting the long-term growth trajectory and trends of an individual. This includes growth and development indicators and their derived data collected and recorded periodically over a predetermined time period. For example, this could include time-series measurements of height, weight, and bone age measured and archived every three months or six months over the past 1 to 3 years. Body composition data can also be included as appropriate. Furthermore, it can include key growth rate derived data automatically calculated by the system based on historical height and weight sequences, such as the height growth rate (e.g., 5 cm / year) and weight growth rate (e.g., 2 kg / year) over the past 6 months.

[0057] For example, the system can be pre-connected remotely to a dedicated data acquisition device. After data collection on the target subject using an intelligent pulse diagnostic instrument and a tongue image acquisition device, the system can directly obtain the original pulse waveform signal and tongue color image through a dedicated data interface as digital vital sign data. For quantitative growth data, the system can query and retrieve the target subject's current bone age X-ray interpretation results, the most recent body composition analysis report, and the latest height and weight measurements from electronic health records from a medical information system or dedicated growth and development assessment equipment (height scale or electronic scale, etc.) as current feature data. Simultaneously, the system retrieves the child's historical archived height, weight, and bone age measurements sorted by time from the database, and uses the recent growth rate and other derived data calculated from these time-series data to constitute historical feature data together with the original time-series measurement records.

[0058] Subsequently, the digital vital signs data and quantitative growth data can be fused together to obtain the original multi-source data.

[0059] Specifically, please continue to refer to Figure 2 The raw data acquisition module 210 also includes a multimodal fusion unit 211, which is used to: perform vectorization conversion processing on the digitized vital sign data to obtain TCM vector data; perform normalization processing on the quantified growth data to obtain growth vector data; and perform splicing processing on the TCM vector data and growth vector data to obtain raw multi-source data.

[0060] In this context, TCM vector data can refer to a high-dimensional numerical array that integrates waveform and image features. This involves converting waveform and image feature data into standardized numerical vectors and then concatenating them to obtain a comprehensive vector. For example, assuming the waveform feature data is a pulse rate of 80 beats / min and a pulse strength of 0.7, mapping it yields a numerical vector of the form [80, 0.7]. Similarly, assuming the image feature data is a tongue color quantification value of 0.6 and a tongue coating coverage of 0.3, mapping it yields a numerical vector of the form [0.6, 0.3]. Then, concatenating these two types of vectors sequentially results in TCM vector data of the form [80, 0.7, 0.6, 0.3].

[0061] Growth vector data can refer to a high-dimensional numerical array that integrates the current state and historical trends; that is, a comprehensive vector formed after normalizing current and historical feature data. For example, normalizing quantified growth data is essentially a process of standardizing growth data with different dimensions and ranges. For instance, after obtaining various growth data, the Min-Max normalization method can be used. For example, assuming the normal height range for a 7-year-old child is 100-150cm, if the detected height of the target object is 120cm, then it can be standardized using the following formula: (120-100) / (150-100) = 0.4, to obtain the standard value of this growth data. Subsequently, the same normalization process is performed on all growth data, and all standard values ​​are integrated. These are then concatenated according to time order or feature category to form a one-dimensional numerical vector, which serves as the growth vector data.

[0062] The TCM vector data and growth vector data can then be directly concatenated in a preset order to obtain the original multi-source data. For example, assuming the TCM vector data is [80, 0.7, 0.6, 0.3] (including pulse rate, pulse strength, tongue color, and tongue coating coverage ratio), and the growth vector data is [0.4, 0.4, 0.5, 0.3] (including normalized bone age, height, weight, and height growth rate), the two vectors can be directly concatenated in order to obtain a high-dimensional vector of [80, 0.7, 0.6, 0.3, 0.4, 0.4, 0.5, 0.3], which is the original multi-source data.

[0063] By transforming digitized vital sign data into TCM vector data, subjective TCM vital signs are quantified using objective indicators, improving the data's computability. Normalizing the already quantified growth data eliminates dimensional differences between various growth indicators, making the data comparable. Finally, concatenating the TCM vector data with the growth vector data provides a comprehensive characterization of the target object's growth and syndrome states, laying a solid data foundation for subsequent precise decision-making.

[0064] In some implementations, the raw decision data includes the state prediction results of the target object and the original recommended solutions; please refer to the appendix. Figure 3 The original decision acquisition module 110 also includes a model execution unit 310, which is used to: extract features based on the original multi-source data using the underlying coding layer to obtain shared feature data; perform regression prediction using the state prediction branch based on the shared feature data to obtain state prediction results; and perform probabilistic recommendation using the scheme recommendation branch based on the shared feature data and the preset prescription library to obtain the original recommendation scheme.

[0065] Shared feature data refers to general feature data that can simultaneously describe the syndrome characteristics and growth characteristics of the target object, which is abstract information that is valuable for both subsequent prescription recommendation and growth prediction tasks.

[0066] For example, the bottom encoding layer can be a deep neural network module composed of multiple fully connected layers stacked sequentially. The specific number of fully connected layers can be adjusted according to actual needs and computational costs, preferably 2 to 4. Taking a 3-layer fully connected layer structure as an example, the first fully connected layer (FC_Shared1) can contain 512 neurons, also using the ReLU activation function, followed by a Dropout layer with a dropout rate of 0.3 to prevent overfitting; the second fully connected layer (FC_Shared2) can contain 256 neurons, also using the ReLU activation function, followed by a Dropout layer with a dropout rate of 0.2; the third fully connected layer (FC_Shared3) can contain 128 neurons, also using the ReLU activation function. In specific processing, taking a high-dimensional vector of dimension [batch_size, 1280] as the original multi-source data as an example, it first undergoes preliminary nonlinear transformation and feature selection in the first fully connected layer of the encoding layer. The processed result is then input into the second and third fully connected layers to gradually extract higher-order and more abstract common features. After these processing layers, the final output is a vector with fixed dimensions (e.g., [batch_size, 128]). That is, the shared feature data no longer directly corresponds to a specific tongue color or height value, but is encoded to describe a comprehensive state such as "spleen deficiency and growth retardation". For example, if the target is an 8-year-old child who is shorter than average, the features reflected in their original multi-source data might include: Feature A: pale tongue with white coating; Feature B: weak pulse; Feature C: bone age 7.5 years; Feature D: height data from the past year; Feature E: weight data from the past year. The bottom coding layer will further extract these scattered features. For example, based on features A and B, the feature "spleen deficiency" is extracted, and then combined with features C, D, and E, the feature "growth retardation" is extracted. After integration, shared feature data that reflects the characteristics of "spleen deficiency syndrome: pale tongue with white coating and weak pulse; growth retardation: height, weight, and composition analysis data from the past year" is generated. It should be noted that the actual input and output of the bottom coding layer are standardized numerical vectors; the meanings are described in words here for ease of understanding.

[0067] Subsequently, the state prediction branch can be used to perform regression prediction based on shared feature data to obtain the state prediction result.

[0068] For example, the state prediction branch can be a regressive neural network with a structure consisting of two fully connected layers and a linear output layer. The first fully connected layer (FC_Growth1) can contain 64 neurons using the ReLU activation function; the second fully connected layer (FC_Growth2) can contain 32 neurons using the ReLU activation function. The number of neurons in the output layer (Output_Growth) can be determined based on the number of future time points to be predicted. For example, if the goal is to predict height and weight for the next four quarters (four time points), the output layer would be set to eight neurons (each time point corresponding to two values ​​for height and weight), and would directly output the predicted physical values ​​using a Linear activation function.

[0069] In practice, shared feature data is passed sequentially through each layer of the state prediction branch. After two fully connected layers extract growth-related specific features, the final output layer shows predicted height and weight values ​​for each future time point, i.e., a vector sequence, such as [H1,W1,H2,W2,H3,W3,H4,W4], representing predicted height and weight for the first to fourth quarters of the future, respectively. Subsequently, based on the prediction error distribution (e.g., 95%) statistically analyzed during model training by the state prediction branch, a confidence interval is calculated for each predicted value to obtain the confidence range for height and weight increments (e.g., height increase of 3.5±1cm and weight increase of 2.5±1kg in the next 6 months). Finally, the predicted height values ​​(e.g., initial 120cm, first month 120.5cm, second month 121cm... sixth month 123.5cm) and predicted weight values ​​(e.g., initial 26kg, first month 26.2kg, second month 26.5kg... sixth month 28.5kg) at each time point are arranged in chronological order to plot the height prediction curve and weight prediction curve. The three together constitute the state prediction result.

[0070] Furthermore, please continue to refer to Figure 3 The model execution unit 310 may further include a decision subunit 311, which is used to: determine candidate recommendation schemes based on shared feature data and a preset prescription library using the scheme recommendation branch; determine candidate intervention suggestions based on the candidate recommendation schemes and shared feature data using a preset knowledge graph; and perform structured integration processing based on the candidate recommendation schemes and candidate intervention suggestions to obtain the original recommendation scheme.

[0071] The preset formula library can refer to a dataset containing several classic Chinese medicine formulas commonly used for regulating children's growth and development, including information such as the basic compatibility and efficacy indications of each formula. The candidate recommendation scheme can be a list containing all formulas and their recommendation weights; that is, after performing probability calculations on all formulas in the preset formula library, a list containing each formula and its corresponding recommendation weight is obtained, reflecting the suitability of different formulas for the target population.

[0072] A pre-defined knowledge graph can refer to a digital knowledge base embedded with TCM diagnostic logic and intervention rules. It is a structured TCM knowledge database that graphically links entities such as prescriptions, syndrome types, symptoms, medicinal ingredients, treatment methods, and lifestyle care suggestions, as well as the logical relationships between them.

[0073] Candidate intervention suggestions refer to targeted intervention guidance information generated based on the specific state of the target individual. These suggestions, generated based on a pre-defined knowledge graph, supplement and refine candidate recommended solutions. Specifically, candidate intervention suggestions can include prescription-related suggestions and prescription-independent suggestions. Prescription-related suggestions may involve adjustments to the prescription itself, such as suggestions on adding herbs, adjusting dosages, and methods of administration for the top-ranked prescription with the highest weight in the candidate recommended solution. Prescription-independent suggestions may refer to non-pharmacological conditioning suggestions that do not involve prescriptions, such as suggestions on diet, exercise, and rest.

[0074] For example, the scheme recommendation branch can be a classification network whose structure may include two fully connected layers and an output layer. The first fully connected layer (FC_Prescription1) may contain 64 neurons and use the ReLU activation function; the second fully connected layer (FC_Prescription2) may contain 32 neurons and use the ReLU activation function. The number of neurons in the output layer (Output_Prescription) may be equal to the total number of prescriptions K in the preset prescription library (e.g., K=50), and use the Softmax activation function.

[0075] In practice, the shared feature data is first input into the scheme recommendation branch. After extracting syndrome-related specific features through two fully connected layers, the output layer outputs a K-dimensional probability distribution vector. Each element in the vector corresponds to the recommendation probability of a prescription in the preset prescription library. That is, it can be in the form of W=[w1,w2,...,w... K ], where w i Representing the i The recommendation probability of each prescription was determined. Subsequently, all prescriptions were sorted from highest to lowest recommendation probability to form candidate recommendation schemes. For example, the first choice prescription was Liu Jun Zi Tang (recommendation weight 0.85), the first choice prescription was Si Jun Zi Tang (recommendation weight 0.71), the second choice prescription was Bu Zhong Yi Qi Tang (recommendation weight 0.62)... and the fourth choice prescription was Shen Ling Bai Zhu San (recommendation weight 0.12).

[0076] Next, based on the preferred formula (Liu Jun Zi Tang) in the candidate recommendation schemes and the key symptoms (such as loss of appetite) extracted from the shared feature data, a pre-defined knowledge graph is queried. This pre-defined knowledge graph stores modification rules such as "Liu Jun Zi Tang - Spleen Deficiency - Loss of Appetite," as well as dosage rules such as "Children's dosage needs to be converted according to weight." Through the query, targeted formula-related suggestions can be retrieved. For example, when the core formula is Liu Jun Zi Tang and accompanied by loss of appetite, hawthorn and malt are often added; and based on the target subject's age (8 years old) and weight (25 kg), the dosage of "one dose per day, decocted in water to 150 ml, taken warm in two divided doses" is calculated. These are integrated into formula-related suggestions.

[0077] At the same time, it will also match non-drug intervention rules related to "spleen deficiency syndrome" from the preset knowledge graph, such as increasing the intake of yam and millet in the diet or massaging the abdomen for 5 minutes every day, and then organize them to obtain prescription-irrelevant suggestions.

[0078] Finally, the candidate recommendation schemes, formula-related suggestions, and formula-irrelevant suggestions are structured and integrated. The above data can be merged according to a preset format (e.g., organized in the order of core formula recommendations (including weights), formula-related suggestions, and formula-irrelevant suggestions) to form a complete original recommendation scheme.

[0079] First, candidate recommendations are identified using shared feature data, visually demonstrating the suitability of each option. Then, a pre-defined knowledge graph is used to retrieve personalized intervention suggestions that are relevant to or irrelevant to the prescription, improving the comprehensiveness of the intervention guidance. Finally, all data is structured and integrated to generate original recommendation plans that are both accurate and actionable, providing high-quality auxiliary references for clinical intervention.

[0080] Furthermore, the original multi-task model can be trained using a supervised learning paradigm. By constructing a large-scale labeled dataset, setting a multi-task loss function, configuring training parameters, and performing iterative training, the model can learn an accurate mapping from the original multi-source data to the original decision data.

[0081] For example, the training process for the model is as follows: A training dataset is constructed. This dataset contains no fewer than 10,000 complete electronic medical records of children's growth and development in Traditional Chinese Medicine (TCM), all of which are double-blindly labeled to ensure quality. Specifically, each data point includes the original multi-source data, a prescription label determined by a senior TCM physician (Y1), and a height and weight sequence label obtained from actual follow-up (Y2). This dataset can be divided into a training set (approximately 7,000 records), a validation set (approximately 2,000 records), and a test set (approximately 1,000 records) in a 7:2:1 ratio.

[0082] Next, the training parameters are configured. This can be achieved using the TensorFlow framework and the AdamW optimizer. The initial learning rate is set to 3e-4, the weight decay coefficient is 1×10⁻², and the momentum parameters β1=0.9 and β2=0.999. The batch size is set to 32, the maximum training epochs are 150, and a cosine annealing decay strategy is used to adjust the learning rate. A Dropout layer can be added after the fully connected layer of each branch to prevent overfitting. An early stopping strategy is also implemented: if the validation set loss does not decrease for 15 consecutive epochs, training is terminated and the optimal parameters are rolled back.

[0083] Next, the loss function is set. The total loss function can be defined as L_total = α × L_prescription + β × L_growth, where L_prescription is the classification cross-entropy loss, measuring the difference between prescription recommendations and expert labels; L_growth is the smoothing L1 loss, also known as Huber loss, measuring the difference between growth predictions and the true values; α and β are the task weights, initially set to α = 1.0 and β = 1.0. However, in actual training, this can be finely adjusted using grid search combined with validation set performance. To achieve the optimal balance between the Top-3 accuracy of prescription recommendations and the mean absolute percentage error (MAPE) of growth prediction, α = 0.7 and β = 0.3 are preferred to balance the weights of the two tasks. It is understood that the training objective is to minimize the loss function, enabling the model to accurately predict both prescriptions and growth curves simultaneously.

[0084] Ultimately, the criteria for determining whether the model training is complete are as follows: On the independent test set, the model must simultaneously achieve the following performance thresholds: First, prescription recommendation performance, meaning the consistency rate between the model's top-1 preferred prescription and the expert prescription label is ≥65%; the recall rate of the top three candidate prescriptions (Top-3) containing expert prescriptions is ≥90%. Second, growth prediction performance, meaning the mean absolute percentage error (MAPE) between the predicted and actual height and weight for the next 6 months is ≤5%. Finally, clinical rationality, meaning that ≥85% of the prescriptions output from 100 randomized test sets are deemed clinically reasonable and usable by blind review by TCM experts who were not involved in the training.

[0085] In the above implementation, feature extraction is performed on the original multi-source data through the underlying encoding layer to obtain shared feature data that simultaneously covers syndrome and growth characteristics, thereby improving the comprehensiveness of feature representation. Regression prediction is performed through the state prediction branch to obtain state prediction results containing increments, confidence intervals, and prediction curves, improving the accuracy and intuitiveness of growth prediction. Through the collaborative work of the scheme recommendation branch with the preset prescription library and knowledge graph, structured original recommendation schemes are generated, improving the personalization and practicality of intervention suggestions. The parallel operation of the two major task branches enables the synchronous output of decision data, improving the system's decision-making efficiency.

[0086] In some implementation methods, please refer to the appendix. Figure 4 The target decision determination module 130 also includes a decision update unit 410, which is used to: optimize the original multi-task model based on the updated multi-source data to obtain the target multi-task model; and determine the target decision data based on the updated multi-source data and the target multi-task model.

[0087] Alternatively, please refer to the appendix. Figure 4 The decision update unit 410 also includes a trigger management subunit 411, which is used to: optimize the original multi-task model based on the updated multi-source data when the updated multi-source data meets the preset update conditions, and obtain the target multi-task model.

[0088] The preset update conditions refer to a series of rules that determine whether the system needs to initiate the model optimization process. Specifically, the preset update conditions include time conditions and / or data conditions. The time conditions can be trigger conditions set based on a fixed cycle, meaning that the original multi-task model is optimized when a preset regular monitoring time point is reached. This time condition can be a regular cycle that conforms to the principles of traditional Chinese medicine conditioning and the rationality of growth monitoring, such as 4 to 12 weeks, typically 8 weeks.

[0089] Data conditions refer to triggering conditions set based on the deviation between the latest data of the target object and historical predictions. These conditions can be quantitative growth data deviating from the prediction confidence interval or the target object's symptom assessment data indicating a significant worsening of key symptoms. For example, if the target object originally had vivid dreams during sleep, the latest feedback data shows that the symptoms of difficulty falling asleep and night awakenings have significantly worsened (obtained through a scale).

[0090] For example, to determine whether the updated multi-source data meets the preset update conditions, the system can perform this by monitoring two types of signals. First, regarding the time condition, the system presets a regular monitoring period T for each target object, for example, a default setting of 8 weeks. When the system clock detects that 8 weeks have passed since the last assessment or intervention, it automatically determines that the time condition is met, prepares to collect the updated multi-source data, and triggers the optimization process.

[0091] Regarding data conditions, the system will continuously monitor newly collected data. For example, the system compares the child's latest height measurement with the predicted value at the corresponding time point in the status prediction results and its 95% confidence interval. If the latest measurement falls outside the confidence interval (e.g., the predicted height after three weeks is 100.0 ± 1.5 cm, but the actual measurement is 98.0 cm), it is recorded as a deviation. To eliminate single measurement errors, the system can set a continuous confirmation principle. After a deviation is recorded, the system will remind the decision-making end that a retest is required after a period of time (e.g., two weeks). That is, the system requires that the same indicator (e.g., height) deviates from the prediction interval twice consecutively (with an interval of no less than 2 weeks) before finally determining that the data conditions are met. In addition, if key symptoms (e.g., appetite, sleep) are marked as significantly aggravated from the structured symptom questionnaire, the data conditions can also be directly determined to be met.

[0092] It should be noted that the specific triggering conditions can be flexibly set according to requirements. For example, due to the requirement of model accuracy, it can be set that as long as either the time condition or the data condition is met, the updated multi-source data is determined to meet the preset update conditions, and the model optimization process is initiated. However, due to the requirement of computing resources, it can be set that the updated multi-source data can only be determined to meet the preset update conditions and initiate the model optimization process if both the time condition and the data condition are met simultaneously.

[0093] The above methods achieve two goals: firstly, by utilizing time constraints to enable regular optimization within a normal cycle, ensuring the model's continuous adaptation to the long-term growth and development process; and secondly, by utilizing data constraints to enable timely triggering of abnormal situations, ensuring the model responds quickly to deviations in growth trajectory or changes in symptoms.

[0094] Furthermore, when the updated multi-source data meets the preset update conditions, the original multi-task model can be optimized to obtain the target multi-task model. Subsequently, the target decision data can be determined based on the updated multi-source data and the target multi-task model.

[0095] The optimization process can be implemented using an incremental learning algorithm.

[0096] For example, the specific process of optimization is as follows: First, algorithm selection and data preparation are performed. An elastic weight consolidation algorithm can be used to calculate the importance of the original multi-task model parameters to historical training tasks. During fine-tuning, constraints are imposed on important parameters, aiming to absorb new knowledge while protecting existing model capabilities and preventing catastrophic forgetting. Simultaneously, the multi-source data of the target object after the current and previous updates, the corresponding original decision data, and the actual feedback results are combined into a personalized micro-time series dataset.

[0097] Then, a tiered fine-tuning strategy is implemented. In practice, the feedback data from each iteration of the target object (including triggered multi-source data) can be combined into a personalized micro-training set. Then, using the original multi-task model as the initial state, tiered fine-tuning is performed using this training set. Because the scheme recommendation branch and state prediction branch are directly responsible for the output and are more sensitive to individual differences, the parameters of these two branches can be primarily updated. Simultaneously, in the initial stage of fine-tuning, the parameters of the underlying encoding layer are temporarily frozen to protect the general syndrome and growth association patterns encoded in the model. Training uses a significantly low learning rate (e.g., 1 / 10 of the initial learning rate of 3e-4, i.e., 3e-5), and only 1-5 epochs of lightweight training are performed on this micro-training set to avoid overfitting. After accumulating 3 complete feedback cycles for the target object, the underlying encoding layer is then unfrozen and fine-tuned using an extremely small learning rate of 1 / 10 to 1 / 20 of the initial learning rate (i.e., 3e-5 to 1.5e-5).

[0098] It is worth noting that, based on the deviation between the updated multi-source data and the original recommendation scheme after triggering the optimization process, a preset optimization rule base can be invoked to enable the target multi-task model to output more accurate target decision data.

[0099] The pre-defined optimization rule base can be a rule mapping relationship library consisting of premises and conclusions, storing adjustment rules based on TCM syndrome differentiation logic and clinical experience. The premises are derived from the analysis of updated multi-source data, while the conclusions correspond to specific prescription adjustment strategies or priority suggestions.

[0100] For example, the analytical dimensions of the premise section include three core types of information: First, prediction bias analysis, which compares the actual growth indicators (height or weight) in the updated multi-source data with the corresponding state prediction results in the original decision data, and calculates the magnitude of the deviation (e.g., the difference between the actual height growth value and the predicted value). Second, objective indicator change analysis, which assesses the changing trends of digitized vital signs data in the updated multi-source data (e.g., the improvement or aggravation of pulse and tongue characteristics). Third, subjective symptom vector analysis, which analyzes the changes in structured symptom scores in the updated multi-source data (e.g., fluctuations in scores for core symptoms such as appetite, sleep, and digestion). These three types of analytical results together constitute the condition combination of the premise, ensuring the targeted nature of rule triggering.

[0101] The prescription adjustment strategies in the conclusion section closely align with the logic of TCM syndrome differentiation and treatment, providing tiered optimization suggestions for different scenarios corresponding to the preconditions. These may include retaining or replacing the primary prescription, adding or subtracting modified herbs, adjusting specific dosage ratios, and strengthening or adjusting non-pharmacological interventions. The conclusions of each rule have been verified by TCM experts to ensure clinical rationality. For example, when "growth indicators do not meet expectations but symptoms improve," the dosage of auxiliary herbs is adjusted first rather than replacing the primary prescription to avoid affecting the continuity of treatment due to excessive adjustments. When "growth indicators meet targets but new concurrent symptoms appear," a light dose of symptomatic herbs is added to the primary prescription to balance the original treatment direction with the improvement of new symptoms.

[0102] Understandably, while the system initiates incremental learning to optimize model parameters, it can simultaneously input updated multi-source data into the optimization rule base for rule matching. If a unique matching optimization rule is found, the adjustment suggestion in the conclusion is directly extracted as a new recommendation in the target decision data. If multiple related rules exist, the optimal suggestion is selected based on rule priority (e.g., core pathogenesis-related rules take precedence over symptom fine-tuning rules). Subsequently, the adjustment suggestion output by the rule base is compared and integrated with the model parameter adjustment direction obtained from incremental learning. If the two suggestions are consistent, they are directly integrated into the output logic of the target multi-task model. If there is a conflict (e.g., data-driven suggestions to change the primary prescription, while the rule base suggests maintaining the primary prescription and adjusting the dosage), the conflict points and the reasoning basis of both sides are presented to the decision-making end in a visual manner. The physician makes a final decision based on clinical experience, and this decision result is fed back to the system as high-quality labeled data for subsequent iterative optimization of the rule base.

[0103] For example, the process of determining the target decision data is similar to obtaining the original decision data, but uses an optimized model and the latest data. Specifically, the updated multi-source data corresponding to this optimization is input into the obtained target multi-task model. After personalized fine-tuning, the model's solution recommendation branch outputs new candidate recommendation solutions that better match the child's recent response (e.g., the weight probabilities of each alternative prescription may have been adjusted), and its state prediction branch outputs growth prediction data updated based on the latest trends. The system then performs structured processing on these outputs according to the same integration logic as generating the original recommendation solutions and state prediction results (e.g., combining knowledge graphs, calculating confidence intervals, etc.), ultimately generating a new set of decision reference information that reflects the latest evaluation and optimization results, i.e., the target decision data.

[0104] In the above implementation, the original multi-task model is first optimized using an incremental learning algorithm, which realizes personalized adaptation of model parameters and improves the model's fit to the individual response patterns of the target object. Then, the updated multi-source data is processed by the target multi-task model to generate targeted target decision data, which improves the dynamic adaptability of the decision scheme. The whole process constructs a closed-loop adaptive decision mechanism through data-driven model optimization and decision update, providing more accurate auxiliary references for subsequent intervention.

[0105] It should be understood that the medical decision support system in this embodiment is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the aforementioned functions. Each module in the system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0106] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 5 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations. Figure 5 Take a processor 10 as an example.

[0107] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0108] The memory 20 stores instructions executable by at least one processor 10 to enable the at least one processor 10 to perform the functions of the system shown in the above embodiments.

[0109] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0110] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0111] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0112] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touchscreen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some alternative embodiments, the display device may be a touchscreen.

[0113] This application also provides a computer-readable storage medium. The functions of the system shown in the embodiments of this application can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the functions described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that a computer, processor, microprocessor controller, or programmable hardware includes storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the functions of the system shown in any embodiment of this application are implemented.

[0114] This application provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the functions of the system shown in any embodiment of this application.

[0115] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

[0116] It is understood that before using the technical solutions disclosed in the embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant regulations, and user authorization should be obtained. For example, in response to receiving a user's active request, a prompt message can be sent to the user to clearly inform the user that the requested operation will require the acquisition and use of the user's personal information. This allows the user to choose whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media that perform the operations of the technical solutions of this disclosure, based on the prompt message. As an optional but non-limiting implementation, the way to send a prompt message to the user in response to receiving a user's active request can be, for example, a pop-up window, in which the prompt message can be presented in text form. Furthermore, the pop-up window can also include a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device. It is understood that the above notification and user authorization process is merely illustrative and does not constitute a limitation on the implementation of this disclosure; other methods that comply with relevant regulations can also be applied to the implementation of this disclosure. It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws and regulations. It is understood that in the specific embodiments of this application, user information, location information, navigation data and other related data are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data shall comply with the relevant laws and standards of relevant countries and regions.

[0117] The systems, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0118] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0119] Those skilled in the art will understand that embodiments of this application can be provided as a system or a computer program product. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This application is described with reference to flowchart illustrations and / or block diagrams of systems and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0122] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0123] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. Similar embodiments are described more simply; relevant parts can be referred to the descriptions of other embodiments.

[0124] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0125] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A medical decision support system, characterized in that, The system includes: The original decision acquisition module is used to determine the original decision data of the target object based on the original multi-source data of the target object; wherein, the original decision data is determined using an original multi-task model; the original multi-task model includes a bottom coding layer, a scheme recommendation branch and a state prediction branch, and the scheme recommendation branch and the state prediction branch are respectively connected to the bottom coding layer; The data acquisition module is updated to acquire updated multi-source data of the target object within the current intervention period. The target decision determination module is used to determine the target decision data of the target object using the updated multi-source data; wherein the target decision data is determined using a target multi-task model; The subsequent intervention support module is used to guide the subsequent intervention process of the target object based on the target decision data.

2. The system according to claim 1, characterized in that, The original decision data includes the state prediction results of the target object and the original recommendation scheme; the original decision acquisition module further includes a model execution unit, used for: Based on the original multi-source data, feature extraction is performed using the underlying coding layer to obtain shared feature data; wherein, the shared feature data refers to general feature data that can simultaneously describe the syndrome characteristics and growth characteristics of the target object; Based on the shared feature data, regression prediction is performed using the state prediction branch to obtain the state prediction result. Based on the shared feature data and the preset prescription library, the original recommendation scheme is obtained by using the scheme recommendation branch for probabilistic recommendation.

3. The system according to claim 2, characterized in that, The model execution unit includes a decision subunit, used for: Using the proposed branch, candidate recommendation schemes are determined based on the shared feature data and the preset prescription library; Based on the candidate recommendation scheme and the shared feature data, a preset knowledge graph is used to determine candidate intervention suggestions; wherein, the candidate intervention suggestions include prescription-related suggestions and prescription-irrelevant suggestions; The original recommendation scheme is obtained by performing structured integration processing based on the candidate recommendation scheme and candidate intervention suggestions.

4. The system according to claim 1, characterized in that, The target decision determination module further includes a decision update unit, used for: The original multi-task model is optimized based on the updated multi-source data to obtain the target multi-task model; wherein the optimization is implemented using an incremental learning algorithm. The target decision data is determined based on the updated multi-source data and the target multi-task model.

5. The system according to claim 4, characterized in that, The decision update unit further includes a trigger management subunit, used for: When the updated multi-source data meets the preset update conditions, the original multi-task model is optimized based on the updated multi-source data to obtain the target multi-task model; wherein, the preset update conditions include time conditions and / or data conditions.

6. The system according to claim 1, characterized in that, The system also includes a raw data acquisition module, used for: The target object's digital vital signs data and quantitative growth data are acquired; wherein, the digital vital signs data includes waveform feature data and image feature data; the quantitative growth data includes current feature data and historical feature data; The original multi-source data is obtained by fusing the digital vital signs data and the quantitative growth data.

7. The system according to claim 6, characterized in that, The raw data acquisition module further includes a multimodal fusion unit, used for: Based on the digitized vital sign data, vectorization conversion processing is performed to obtain traditional Chinese medicine vector data; Based on the quantized growth data, normalization processing is performed to obtain growth vector data; The original multi-source data is obtained by splicing the TCM vector data and the growth vector data.

8. A computer device, characterized in that, Including the medical decision support system as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, implements the functions of the medical decision support system as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the functions of the medical decision support system as described in any one of claims 1 to 7.

Images