Traditional Chinese medicine five-tone disease treatment personalized music generation system based on large language model and construction method thereof

Abstract

The invention discloses a traditional Chinese medicine five-tone treatment personalized music generation system based on a large language model and a construction method thereof. The invention designs a double-layer architecture technical route of firstly performing traditional Chinese medicine syndrome differentiation decision and then generating five-tone music. Meanwhile, according to the characteristics of traditional Chinese medicine syndrome differentiation treatment individuation and five-tone therapy accuracy, an intelligent syndrome differentiation engine based on a large language model and a five-tone music generation framework based on traditional music pre-training are provided, and the treatment thought of traditional Chinese medicine treatment based on syndrome differentiation and according to different persons is achieved through multi-dimensional health information fusion and a five-tone feature constraint mechanism. The technical problem of fusion of the traditional Chinese medicine theory and the modern AI technology is solved. The intelligent level and the individuation degree of the treatment scheme are remarkably improved while the theory integrity of the traditional Chinese medicine five-tone therapy is kept, and a complete technical solution is provided for modern application of traditional Chinese medicine music treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence medical application technology, in particular to a traditional Chinese medicine five-tone therapy personalized music therapy system based on a large language model and music generation technology and a construction method thereof. BACKGROUND

[0002] Traditional Chinese medicine music therapy has a long research history in China. According to the Chinese medical classic Huangdi Neijing (The Yellow Emperor's Classic of Internal Medicine) written more than 2,000 years ago, "The liver belongs to wood, the sound is a horn, and the emotion is anger; the heart belongs to fire, the sound is a symbol, and the emotion is joy; the spleen belongs to earth, the sound is a palace, and the emotion is thinking; the lungs belong to gold, the sound is a business, and the emotion is worry; the kidneys belong to water, the sound is a feather, and the emotion is fear." The five tones (1-Do, 2-Re, 3-Mi, 5-So, and 6-La) form the five basic scales of Chinese classical music, known as "five tones."

[0003] With the continuous development of technology, the technical solutions of traditional Chinese medicine music therapy are also developing. At present, there are various technical routes to explore intelligent music therapy systems, mainly including the following schemes:

[0004] I. Traditional five-tone therapy: Based on the "five-tone therapy" theory of Huangdi Neijing, music therapy is achieved through the corresponding relationship between the five tones (palace, business, horn, symbol, and feather) and the five internal organs (heart, liver, spleen, lungs, and kidneys). The existing five-tone therapy system mainly adopts a fixed music library mode, pre-recording music pieces corresponding to different syndromes. Traditional Chinese medicine doctors manually select and play music according to their clinical experience. Typical implementations include: a five-tone-organ correspondence database based on traditional Chinese medicine theory; a syndrome-music matching system using a simple mapping algorithm; and a music playback platform integrating traditional instrument timbres. This type of system maintains the theoretical integrity of traditional five-tone therapy and has a strong traditional Chinese medicine cultural background. However, it has four main technical defects: first, it relies on the personal experience of traditional Chinese medicine doctors for syndrome differentiation and music selection, has low standardization, and different doctors have different treatment ideas and five-tone prescriptions, making it difficult to form a unified treatment standard and meet the needs of large-scale clinical applications; second, it uses a fixed music library mode with limited pre-set music pieces, which cannot adapt to individual physical differences, symptom changes, and personal music preferences, lacking the ability to customize individualized music; third, it lacks a dynamic adjustment mechanism and cannot optimize the music therapy program in real time based on the improvement of patient symptoms during treatment, seasonal changes, and user feedback; fourth, it lacks an objective efficacy evaluation system and mainly relies on patient subjective feelings to judge treatment effectiveness, making it difficult to quantitatively evaluate the effectiveness of music therapy and optimize the program.

[0005] II. General music generation system based on deep learning: In recent years, music generation technology based on deep learning has made significant progress. Representative models include MusicGen, Stable Audio, Jukebox, and InspireMusic. These systems usually use an autoregressive Transformer architecture or diffusion models to obtain music creation capabilities through large-scale music data training. In terms of technical implementation, it mainly includes: audio quantization coding technology based on VQ-VAE; discrete audio representation method using multiple codebook mechanism; and conditional generation text-to-music conversion algorithm. Among them, MusicGen uses EnCodec for audio compression, and improves sound quality through parallel processing of multiple codebooks; Stable Audio uses latent diffusion model to achieve high-fidelity audio generation; InspireMusic integrates super-resolution flow matching technology to achieve long-form music generation. These systems perform well in terms of music generation quality and diversity, but face three fundamental problems in the application of traditional Chinese medicine: First, there is a lack of guidance and constraint mechanism of traditional Chinese medicine theory. Existing models such as MusicGen and Stable Audio are mainly trained based on modern music theory and cannot understand the correspondence between five notes and five internal organs. The generated music lacks treatment specificity and is difficult to achieve the expected effect of regulating visceral function; Second, it cannot maintain the unique pitch characteristics of five notes and the charm of traditional music. These models mainly use popular music, classical music and other modern music data in training, and lack understanding of the pitch rules, mode characteristics and cultural connotations of Chinese traditional five-note music. The generated music is inclined to modern style, losing the core characteristics of traditional Chinese medicine music therapy; Third, there is a lack of deep association mechanism with patient health status. It cannot adjust the generation strategy according to the patient's specific symptoms, constitution type, seasonal and solar terms, and other traditional Chinese medicine diagnosis elements, and cannot realize the individualized treatment principle of "tailor-made for different people and different times".

[0006] III. Music therapy recommendation system based on emotion recognition: This music therapy recommendation system mainly uses emotion computing technology to recommend corresponding therapeutic music by recognizing the current emotional state of the user. The main technical route includes: emotion detection algorithm based on facial expression recognition; psychological state assessment method using voice emotion analysis; and emotion quantification technology based on physiological signals (heart rate, blood pressure, etc.). The system architecture usually includes an emotion data acquisition module, an emotion classification and recognition module, and a music recommendation and matching module. In terms of music selection strategy, the main recommendation algorithms used are based on collaborative filtering and content-based music matching technology. This type of system has certain effect in improving user emotions, but there are three key technical limitations in Chinese medicine applications: First, there is a lack of direct correspondence between emotional state and TCM syndrome type. Existing systems are mainly based on modern psychological theories of emotion, which cannot accurately reflect the functional state of Zangfu organs, the running state of Qi and blood, and the balance of Yin and Yang in TCM theory. The emotion recognition results are difficult to guide TCM diagnosis and treatment; Second, the recommended music lacks a theoretical basis in TCM. The recommendation algorithm mainly uses collaborative filtering or content matching, and the recommended music works may improve emotions, but do not match the treatment mechanism of five-tone therapy, and cannot achieve the treatment goal of regulating Zangfu organs and dredging meridians; Third, there is a lack of individualized TCM constitution identification and seasonal adaptation capability, which cannot reflect the treatment principles of TCM "three principles of treatment" (treatment according to individual constitution, local conditions, and time), and cannot develop personalized music therapy programs according to the patient's constitution characteristics and current solstice characteristics.

[0007] In addition to the above defects, for symptom recognition and TCM terminology standardization, existing technologies mainly use simple text matching or basic semantic similarity calculation methods, but in the specific context of TCM, there is a problem that patient's colloquial symptom description is difficult to accurately map to standardized TCM terminology. Existing symptom recognition methods cannot effectively handle the diverse and non-standardized symptom expressions used by patients, especially the lack of understanding of local disease descriptions and TCM professional terminology.

[0008] Regarding the five-tone feature preservation technology of music generation, existing music generation models lack a special coding and constraint mechanism for Chinese traditional five-tone music. Traditional audio coding technologies such as EnCodec, HuBERT, etc. are mainly optimized for Western music and lack effective representation and preservation capabilities for the unique pitch relationship, mode structure, and timbre characteristics of five-tone music, resulting in the generated music losing the pitch accuracy required for TCM treatment. SUMMARY

[0009] In view of the above technical problems, the present application provides a traditional Chinese medicine five-tone disease treatment personalized music generation system based on a large language model and a construction method thereof, aiming to solve the core technical challenges of deep integration of traditional Chinese medicine theory and AI technology, intelligent syndrome differentiation of multi-dimensional health information, accurate preservation of five-tone characteristics, dynamic optimization of personalized treatment plan, and construction of efficacy evaluation and feedback mechanism, so that the traditional Chinese medicine five-tone music treatment truly has modern clinical application value.

[0010] A traditional Chinese medicine five-tone disease treatment personalized music generation system based on a large language model, characterized in that it comprises:

[0011] A multi-modal health information acquisition module is used to collect and process the patient's symptom description, physical examination report, geographical location, and seasonal information, and convert the above information into standard structured traditional Chinese medicine symptom information.

[0012] A traditional Chinese medicine syndrome differentiation decision module comprises a zang-fu state evaluation expert module, a pathogenesis analysis expert module, a five-tone treatment expert module, a brain wave regulation expert module, and a personalized adjustment expert module. The zang-fu state evaluation expert module is used to analyze the functional status of the five zang organs and six fu organs according to the input traditional Chinese medicine symptom information, and calculate the deficiency and excess degree and mutual relationship of each zang and fu organ based on the symptom characteristics. The pathogenesis analysis expert module is used to analyze the pathological mechanism according to the input traditional Chinese medicine symptom information, and identify the root cause, pathogenesis and transmission law of the disease. The five-tone treatment expert module is used to formulate specific five-tone treatment parameters according to the differentiation results output by the zang-fu state evaluation expert module and the pathogenesis analysis expert module. The brain wave regulation expert module is used to set the brain wave regulation target and intensity parameters according to the differentiation results output by the zang-fu state evaluation expert module and the pathogenesis analysis expert module. The personalized adjustment expert module is used to optimize the scheme according to the patient's age, gender, constitution, and multi-dimensional factors such as season, geographical location. The traditional Chinese medicine syndrome differentiation decision module outputs the differentiation results, five-tone treatment scheme, brain wave regulation target, and personalized adjustment scheme through a gating network.

[0013] A five-tone music generation module is used to generate a personalized music treatment scheme according to the differentiation results, five-tone treatment scheme, brain wave regulation target, and personalized adjustment scheme output by the traditional Chinese medicine syndrome differentiation decision module.

[0014] Preferably, it further comprises a personalized optimization module, which realizes adaptive adjustment of treatment parameters and dynamic optimization of the scheme through a multi-dimensional evaluation system construction and a self-adaptive parameter adjustment mechanism.

[0015] The present application also discloses a construction method of the above-mentioned traditional Chinese medicine five-tone disease treatment personalized music generation system based on a large language model:

[0016] The construction of the multi-modal health information collection module includes the following steps:

[0017] S11: Construction of TCM symptom ontology, construct a symptom ontology covering TCM four diagnostic information, the symptom ontology adopts a hierarchical structure, and a multi-layer mapping relationship is established from symptom category, symptom name, symptom description to symptom degree;

[0018] S12 Multi-modal feature fusion training, using a multi-modal fusion architecture based on Transformer, integrating text, speech, image and other input modalities, realizing symptom text understanding, medical image recognition and test index analysis, and constructing a TCM attribute mapping table of medical test indexes, converting modern medical indexes into TCM constitution characteristics;

[0019] S13 Spatiotemporal information mapping optimization, construct a mapping model of geographical location-climate characteristics and seasonal solstice-TCM attributes, automatically obtain the local climate characteristics according to the geographical location of the patient, analyze the corresponding TCM environmental attributes, and establish the corresponding relationship between solstice and the strength of the viscera based on the theory of the twenty-four solar terms;

[0020] S14 Information quality evaluation: evaluate the completeness, accuracy and consistency of the collected information through multi-dimensional evaluation indicators.

[0021] The construction of the TCM syndrome differentiation decision module includes the following steps:

[0022] S21: Multi-modal data integration preprocessing, collect and integrate multi-source heterogeneous data required for TCM syndrome differentiation, standardize, align, desensitize and structure the data, establish a unified training corpus for five-tone therapy, and construct a relationship database of symptoms-viscera-five tones, deeply associate traditional TCM theory with five-tone therapy;

[0023] S22: Pre-training of TCM large model, pre-training based on large-scale pre-training model in TCM five-tone therapy field corpus, enhancing the model's deep understanding of TCM syndrome differentiation thinking and five-tone therapy, realizing self-recurrent language modeling, syndrome-organ association prediction, five-tone-viscera mapping task and syndrome differentiation reasoning chain reconstruction;

[0024] S23: Fine-tuning of hybrid expert architecture, based on the pre-trained TCM large model, construct viscera state evaluation expert model, pathogenesis analysis expert model, five-tone therapy expert model, brain wave regulation expert model, and individualized adjustment expert model, each expert is responsible for a specific syndrome differentiation decision task;

[0025] S24: Syndrome differentiation evaluation optimization, evaluate and iteratively optimize the model performance through multi-dimensional evaluation indicators.

[0026] Preferably, the zang-fu state evaluation expert model is used to analyze the functional state of the five zang organs and six fu organs, calculate the deficiency and excess degree of each zang and fu organ and their mutual relationship based on the symptom characteristics, and the training loss function is:

[0027] , wherein is the state label of the i-th zang-fu organ;

[0028] The pathogenesis analysis expert model is used for deep pathological mechanism analysis to identify the root cause, pathogenesis and transmission rule of the disease, and the training loss function is: , wherein is the cross-entropy loss, is the consistency loss.

[0029] The five-tone music generation module includes the following steps:

[0030] S31: Pre-training of traditional music model, collecting and integrating Chinese traditional music data, performing audio quality screening, format standardization, pitch feature labeling and cultural attribute classification processing on the data, establishing a unified training corpus for five-tone generation, and pre-training a traditional music model using an autoregressive transformer music generation model;

[0031] S32: Five-tone feature deep fine-tuning, using specially labeled five-tone music data to fine-tune the traditional music model, strengthening the model's understanding and generation ability of the core requirements of five-tone therapy, each music is labeled with corresponding zang-fu meridian, treatment efficacy, pitch matching and clinical verification result, the fine-tuning process uses a multi-task learning framework to realize five-tone recognition, zang-fu correspondence, frequency precise control and treatment effect prediction functions;

[0032] S33: Pitch constraint optimization, establishing a pitch constraint mechanism to make the generated music strictly follow the pitch requirements and theoretical specifications of traditional Chinese five-tone therapy;

[0033] S34: Rhythm preservation enhancement, using a super-resolution enhancer and a cultural feature preservation mechanism to maintain the rhythm and cultural characteristics of traditional Chinese music while ensuring sound quality, using a super-resolution model based on flow matching to convert low-resolution music tokens into high-fidelity audio output, and integrating a rhythm preservation constraint to ensure that traditional characteristics are not lost.

[0034] Preferably, the five-tone feature deep fine-tuning multi-task learning framework includes five-tone recognition, zang-fu correspondence, frequency precise control and treatment effect prediction.

[0035] Preferably, the five-tone recognition task uses a multi-label classification loss: , wherein ​​is the number of time steps, For the The time step Tone tags, Predict probabilities for the model to ensure that it accurately identifies the five-note components in the music.

[0036] Preferably, the organ-corresponding tasks are achieved through regression loss: ,middle For the The true weight of each organ, Predict weights for the model to ensure the accurate correspondence between the generated music and the target organs.

[0037] Preferably, the frequency precision control task adopts frequency domain loss: ,middle is the target frequency set, To generate audio at frequency The energy of The target energy value ensures precise control of key treatment frequencies.

[0038] Preferably, the treatment effect prediction task is: through effect regression modeling: ,in To predict treatment efficacy, For real clinical effects, is the regularization coefficient, which improves the model's ability to predict treatment effects.

[0039] The technical solution of this invention has the following beneficial effects: It innovatively proposes a two-layer, large-scale language model collaborative architecture specifically designed for the unique characteristics of TCM five-tone therapy, achieving deep decoupling and precise coordination between TCM theoretical reasoning and music content generation. This architecture comprises two specialized layers: a large-scale TCM syndrome differentiation and reasoning model and a large-scale five-tone music generation model. Through a carefully designed semantic interface and parameter transfer mechanism, it successfully transforms the traditional Chinese medicine principle of "differentiation and treatment based on individual needs" into a computable intelligent decision-making process. The upper-layer TCM model focuses on symptom analysis, constitution identification, seasonal adaptation, and syndrome differentiation and reasoning. Based on foundational models such as DeepSeek, it continues pre-training on a large-scale collection of TCM literature to gain a deep theoretical foundation. The lower-layer music generation model focuses on preserving the characteristics of the five-tones, precisely controlling the musical scale, and inheriting the rhythmic essence. Through traditional music pre-training and five-tone fine-tuning, it achieves high-quality music creation. This architecture not only ensures the integrity and professionalism of TCM theory, but also enhances the artistry and therapeutic relevance of the generated music through specialized division of labor, successfully addressing the technical challenges of integrating traditional TCM theory with modern AI technology. While maintaining the theoretical integrity of TCM five-tone therapy, the present invention significantly improves the intelligence and personalization of treatment plans, providing a complete technical solution for the modern application of TCM music therapy.

[0040] For a better understanding of the features and technical contents of the present application, please refer to the following detailed description and drawings of the present application. However, the drawings provided are only for reference and illustration, and are not intended to limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The system structure block diagram of the embodiment of the present application.

[0042] Figure 2 The construction flow chart of the multi-modal health information acquisition module of the embodiment of the present application.

[0043] Figure 3 The construction flow chart of the traditional Chinese medicine syndrome differentiation decision module of the embodiment of the present application.

[0044] Figure 4 The construction flow chart of the five-tone music generation module of the embodiment of the present application.

[0045] Figure 5 The construction flow chart of the personalized optimization module of the embodiment of the present application. DETAILED DESCRIPTION

[0046] The following is to illustrate the disclosed embodiments of the present application through specific specific embodiments, and those skilled in the art can understand the advantages and effects of the present application from the disclosed content of the present application. The present application can be implemented or applied through other different specific embodiments, and the details in the present application can be modified and changed based on different viewpoints and applications without departing from the concept of the present application. In addition, the drawings of the present application are only simple schematic illustrations, not the depiction of actual size, and the prior declaration is made. The following embodiments will further illustrate the related technical contents of the present application, but the disclosed content is not intended to limit the protection scope of the present application.

[0047] As shown in Figure 1 The present application discloses a traditional Chinese medicine five-tone disease treatment personalized music generation system based on a large language model, which mainly includes a multi-modal health information acquisition module, a traditional Chinese medicine syndrome differentiation decision module, a five-tone music generation module, and a personalized optimization module.

[0048] The multi-modal health information acquisition module is used to collect and process the symptom description, physical examination report, geographical location, seasonal information, and time information of the patient, and convert the above information into standard structured traditional Chinese medicine symptom information. The module mainly includes three core sub-modules of symptom natural language understanding, medical report analysis, and space-time information adaptation. The symptom natural language understanding sub-module. The module uses text understanding technology based on traditional Chinese medicine symptom ontology to extract symptom features, uses a multi-modal fusion network to process image and test data, and supports multiple information input methods such as text description, voice input, and picture upload.

[0049] The TCM syndrome differentiation decision module is the core of the system, which is constructed based on DeepSeek and other base large language models. Through the continued pre-training of TCM classical literature, modern clinical data, and five-tone therapy professional knowledge and multi-task fine-tuning, intelligent syndrome differentiation is achieved. This module includes four sub-modules: symptom analyzer, constitution recognizer, seasonal adapter, and syndrome differentiation reasoning engine, which generates accurate five-tone treatment plans through a hierarchical decision mechanism. It includes five sub-modules: zang-fu state evaluation expert module, pathogenesis analysis expert module, five-tone therapy expert module, brain wave regulation expert module, and individualized adjustment expert module.

[0050] The zang-fu state evaluation expert module is used to analyze the functional status of the five zang and six fu organs based on the input TCM symptom information, and to calculate the deficiency and excess degree of each zang and fu organ and their mutual relationship based on symptom characteristics. The pathogenesis analysis expert module is used to analyze the pathological mechanism based on the input TCM symptom information, and to identify the root cause, pathogenesis, and transmission law of the disease. The five-tone therapy expert module is used to formulate specific five-tone treatment parameters based on the syndrome differentiation results output by the zang-fu state evaluation expert module and the pathogenesis analysis expert module. The brain wave regulation expert module is used to set the brain wave regulation target and intensity parameters based on the syndrome differentiation results output by the zang-fu state evaluation expert module and the pathogenesis analysis expert module. The individualized adjustment expert module is used to optimize the plan based on the patient's age, gender, constitution, and multi-dimensional factors such as solar term and geographical location. The above syndrome differentiation results, five-tone treatment plan, brain wave regulation target, and individualized adjustment plan are output through a gating network.

[0051] The five-tone music generation module is used to generate individualized music therapy plans based on the syndrome differentiation results, five-tone treatment plan, brain wave regulation target, and individualized adjustment plan output by the TCM syndrome differentiation decision module. The five-tone music generation module uses a music generation architecture that has been pre-trained on Chinese traditional music and fine-tuned on five-tone features, and includes a five-tone feature encoder, an autoregressive generator, and a super-resolution enhancer. This module uses WavTokenizer for audio tokenization, and ensures that the generated music meets the requirements of five-tone therapy through pitch consistency constraints and TCM flavor preservation mechanisms

[0052] The individualized optimization module realizes adaptive adjustment of treatment parameters and dynamic optimization of plans through a multi-dimensional evaluation system and an adaptive parameter adjustment mechanism. The individualized optimization module realizes continuous optimization of treatment plans through user feedback collection, treatment record management, and reinforcement learning algorithms, including multi-dimensional evaluation system construction and adaptive parameter adjustment, to prevent plan solidification and improve individualization effect.

[0053] The system's workflow includes health information entry, intelligent symptom analysis, constitution type identification, seasonal solar term adaptation, TCM syndrome differentiation and reasoning, five-tone scheme generation, personalized music creation, user experience feedback, and treatment plan optimization. The TCM syndrome differentiation and reasoning phase draws on multi-dimensional information collected previously and the TCM theoretical framework to comprehensively analyze the patient's visceral function, qi and blood circulation, and yin-yang balance, determining the primary pathogenesis and treatment principles. This embodies the core TCM philosophy of "differentiation and treatment, a holistic approach." The five-tone scheme generation phase, based on the syndrome differentiation results, determines the dominant tempo, accompanying tempo, mode selection, and musical structure, providing precise therapeutic guidance for subsequent music creation.

[0054] The present invention also discloses a method for constructing the above-mentioned large language model-based personalized music generation system for TCM five-tone disease treatment.

[0055] like Figure 2 As shown, the construction of its multimodal health information collection module includes the following steps:

[0056] S11. Construction of TCM Symptom Ontology:

[0057] First, we constructed a symptom ontology encompassing information about the four diagnostic methods of Traditional Chinese Medicine. This symptom ontology employed a hierarchical structure, establishing a multi-layered mapping relationship between symptom category, symptom name, symptom description, and symptom severity. The symptom ontology employed a three-layered structure: an organ symptom layer, a four-diagnosis information layer, and a symptom quantification layer.

[0058] Organ symptom layer: A dictionary of organ-specific symptoms is created based on the five internal organs (liver, heart, spleen, lung, and kidney). For example, liver-related symptoms include "flank pain, irritability, dry eyes, and insomnia."

[0059] Four-diagnosis information layer: covers the standardized description system of the four diagnoses of inspection, auscultation, inquiry and palpation, including tongue characteristics, pulse classification, voice texture, facial complexion, etc.

[0060] Symptom quantification layer: Establish a five-level quantitative standard for symptom severity (mild, mild, moderate, severe, and extremely severe) to provide a basis for setting personalized treatment intensity.

[0061] S12, multimodal feature fusion training stage:

[0062] A Transformer-based multimodal fusion architecture is used to integrate multiple input modalities, including text, speech, and images. This phase includes three key training tasks: symptom text understanding, medical image recognition, and test indicator analysis. A TCM attribute mapping table for medical test indicators is constructed, converting modern medical indicators into TCM constitution characteristics. For example, high blood pressure is mapped to "hyperactivity of liver yang," and abnormal blood sugar corresponds to "malfunction of the spleen and stomach." The multimodal fusion loss function is defined as: ,in Text understanding loss, Image recognition loss, Numerical analysis loss, Modal alignment loss, 、 、 、 Weight coefficient.

[0063] S13, spatiotemporal information mapping optimization:

[0064] A mapping model of geographical location-climate characteristics and seasonal solstice-TCM attributes is constructed. According to the geographical location of the patient, the local climate characteristics (temperature, humidity, seasonal changes) are automatically obtained, and the corresponding TCM environmental attributes are analyzed. For example, it is easy to produce spleen dampness in a hot and humid area, and the function of the spleen and stomach needs to be focused on. Based on the theory of the 24 solar terms, the corresponding relationship between solar terms and the growth and decline of viscera is established. For example, in the spring, the liver is in the vigorous period, and the liver is focused on; in the winter, the kidney is in the closed period, and the kidney is focused on. Considering the interaction of geographical location and solar term, a dynamic environmental impact evaluation model is constructed. Graph neural network is used to model the influence of geographical and temporal factors on human health. The spatiotemporal mapping loss function is:

[0065]

[0066] where L geo is the geographical feature mapping loss, L season is the solstice attribute prediction loss, and L interaction is the spatiotemporal interaction effect loss. Solstice attribute prediction uses a multi-task learning framework to simultaneously predict five sound attributes, yin and yang characteristics, and climate impact:

[0067] S14, information quality evaluation:

[0068] The completeness, accuracy, and consistency of the collected information are evaluated through multi-dimensional evaluation indicators. The quality evaluation function is:

[0069]

[0070] where Completeness evaluates the coverage of the information, Accuracy is verified by expert annotation, Consistency detects the logical consistency of multi-modal information, and Reliability evaluates the credibility of the information source.

[0071] The input of the multi-modal health information collection module includes: patient text symptom description, voice symptom description, medical image data (tongue appearance, complexion, etc.), physical examination report, geographical location information, and current time information. The output of this module is standardized structured data:

[0072] Patient basic information: {age} {gender} {body type}

[0073] Main symptoms: {standardized symptom description}

[0074] Physical examination results: {key indicator abnormal items}

[0075] Environmental factors: {current solar term} {geographical location} {climate characteristics}.

[0076] Through the processing of the above four stages, this module realizes the intelligent conversion from multi-modal raw input to structured Chinese medicine information, providing a high-quality data basis for subsequent syndrome differentiation decision-making. The innovation of this method lies in the construction of a multi-modal fusion architecture oriented towards Chinese medicine applications, especially the symptom ontology mapping and spatiotemporal information adaptation technology, enabling the system to accurately understand and process the diverse health information expressions of patients.

[0077] As shown in Figure 3 , the construction flowchart of the Chinese medicine syndrome differentiation decision-making module, which is the core decision-making module of the system, is responsible for intelligent Chinese medicine syndrome differentiation based on multi-dimensional health information and generating accurate five-tone treatment plans. This module is based on large language models such as DeepSeek, and through the continued pre-training of five-tone therapy-specific corpus and the fine-tuning of hybrid expert architecture, a multi-modal intelligent syndrome differentiation system with Chinese medicine theory foundation and five-tone therapy professional ability is constructed. The syndrome differentiation engine construction method is divided into multi-modal data integration preprocessing phase, Chinese medicine large model pre-training phase, hybrid expert architecture fine-tuning phase, and syndrome differentiation evaluation optimization phase. The construction of this Chinese medicine syndrome differentiation decision-making module specifically includes the following steps.

[0078] S21, multi-modal data integration preprocessing:

[0079] First, collect and integrate multi-source heterogeneous data required for Chinese medicine syndrome differentiation, including classical Chinese medicine literature, modern clinical medical records, five-tone therapy professional literature, and multi-modal diagnosis data. After standardization, alignment, desensitization, and structuralization, a unified training corpus for five-tone therapy is established. The preprocessing process uses a fusion method based on multi-modal alignment, and a special visual feature extraction and text description generation mechanism is designed for tongue image and facial color data. In particular, a three-layer mapping relationship database of symptoms-viscera-five tones is constructed, which deeply associates traditional Chinese medicine theory with five-tone therapy. The multi-modal data alignment loss function is defined as:

[0080] where is the text-image alignment loss, is the text-value alignment loss, is the cross-modal consistency loss, , 、 is a balance weight.

[0081] S22, pre-training of the traditional Chinese medicine large model:

[0082] Based on large-scale pre-training models such as DeepSeek-7B or Qwen2.5-7B, continue pre-training in the field of traditional Chinese medicine five-tone therapy corpus to enhance the model's deep understanding of traditional Chinese medicine syndrome thinking and five-tone therapy. This stage adopts a hybrid training objective, including autoregressive language modeling, symptom-disease association prediction, five-tone-viscera mapping tasks, and syndrome reasoning chain reconstruction. The pre-training loss function is defined as:

[0083]

[0084] where is the standard language modeling loss, is the symptom association prediction loss, is the five-tone mapping task loss, is the syndrome reasoning chain reconstruction loss, 、 、 、 is the weight coefficient for balancing each task. Enhance the model's deep understanding of traditional Chinese medicine syndrome and five-tone theory.

[0085] a) The symptom-viscera association prediction task is achieved through multi-label classification:

[0086] ,

[0087] where is the number of samples, is the label of the sample corresponding to the viscera , is the model prediction probability.

[0088] b) The five-tone-viscera mapping task adopts a multi-label classification framework: where is the five-tone label of the viscera, is the corresponding probability predicted by the model.

[0089] c) The syndrome reasoning chain reconstruction loss is based on causal reasoning modeling:

[0090]

[0091] where is the reasoning step , is the symptom input, is the model parameter, To infer the chain length.

[0092] S23, fine-tuning of the hybrid expert architecture:

[0093] After completing the pre-training of the TCM large model, a hybrid expert architecture containing five specialized expert modules is constructed for fine-tuning. Each expert is responsible for a specific syndrome decision-making task. In this stage, the parameter-efficient fine-tuning (LoRA) technique is used, and the incremental weight parameter matrix is: where , , represents low-rank matrix decomposition.

[0094] The hybrid expert architecture includes five specialized modules: visceral state evaluation expert module, pathogenesis analysis expert module, five-tone therapy expert module, brain wave regulation expert module, and individualized adjustment expert module.

[0095] a) Visceral state evaluation expert module: responsible for analyzing the functional state of five Zang organs and six Fu organs, calculating the deficiency and excess degree of each Zang and Fu organ and their mutual relationship based on symptom characteristics. This expert uses a multi-classification loss function:

[0096] where is the th state label (deficiency, excess, and balance) of the th Zang or Fu organ, ensuring accurate identification of Zang and Fu organ function states.

[0097] b) Pathogenesis analysis expert module: focuses on deep pathological mechanism analysis, identifying the root cause, pathogenesis, and transmission rules of the disease. The training loss is: where is the cross-entropy loss, is the consistency loss, ensuring the logical coherence of pathogenesis analysis.

[0098] c) Five-tone therapy expert module: the core professional module, based on the syndrome differentiation results to develop specific five-tone therapy parameters, including main tone selection, five-tone ratio, frequency setting, and duration arrangement. This expert uses a multi-task learning framework to simultaneously optimize three sub-tasks: tone selection, ratio calculation, and parameter setting:

[0099]

[0100] where is the main tone selection loss, is the five-tone ratio loss, is the treatment parameter setting loss. The five-tone ratio is modeled using Dirichlet distribution: ensuring the rationality of the five-tone ratio and the consistency of TCM theory.

[0101] d) Brain wave regulation expert module: combined with modern neuroscience theory, set brain wave regulation target and intensity parameters, combine traditional Chinese medicine theory with modern scientific verification. Training loss is:

[0102]

[0103] wherein is the brain wave frequency band prediction loss, is the regulation intensity loss, is the efficacy prediction loss.

[0104] e) Personalized adjustment expert module: considering individual differences, environmental factors and seasonal characteristics of patients, individualized optimization of the scheme. Adaptive weight learning is adopted: . Wherein is the individualization factor set (age, gender, constitution, solar term, geographical location), is the dynamic weight.

[0105] The output calculation of the gating network of the mixed expert architecture is: , wherein the gating weight is calculated through the attention mechanism: .

[0106] S24, syndrome evaluation optimization:

[0107] Through multi-dimensional evaluation index, the model performance is comprehensively evaluated and iteratively optimized, and a three-fold quality assurance system including expert evaluation, clinical verification and effect feedback is established. The core evaluation function is:

[0108]

[0109] wherein each is a weight coefficient, which is adjusted according to the actual application scene. The accuracy of syndrome differentiation is calculated by comparison with expert annotation; the accuracy of scheme is evaluated by the quality of five-tone scheme; the symptom recall rate measures the identification ability of key symptoms; the theoretical coherence evaluates the TCM theoretical basis of the generated scheme; the clinical efficacy is verified by real world data.

[0110] Through the above four stages of processing, the traditional Chinese medicine syndrome differentiation decision module realizes the transformation from general large language model to professional traditional Chinese medicine five-tone treatment decision system, and provides accurate treatment guidance for subsequent personalized music generation. The innovation of this method lies in the design of specific pre-training tasks and mixed expert architecture for five-tone therapy, especially the professional design of five-tone therapy experts and the multi-task joint fine-tuning technology, which enables the model to deeply master the core capabilities of traditional Chinese medicine syndrome differentiation and treatment and five-tone therapy, and realizes the deep integration of traditional Chinese medicine theory and modern AI technology.

[0111] Input and output specifications of the TCM syndrome differentiation module:

[0112] Input data format: structured multi-modal patient information, including:

[0113] - Patient basic information: {age: 35, gender: "female", constitution type: "Qi depression constitution"}

[0114] - Main symptoms: {standardized symptom descriptions: ["hypochondrium pain", "irritability", "insomnia and dreaminess", "bitter taste", "dizziness"]}

[0115] - Physical examination results: {key indicator abnormalities: {"blood pressure": "140 / 90", "heart rate": "85", "liver function": "ALT slightly elevated"}}

[0116] - Environmental factors: {current solar term: "Vernal Equinox", geographical location: "South China", climate characteristics: "hot and humid"}

[0117] Output scheme format:

[0118] Syndrome differentiation results:

[0119] - Main symptoms: "liver depression and Qi stagnation",

[0120] - Pathogenesis analysis: "liver Qi stagnation, dysfunction of dispersion, and unsmooth Qi movement",

[0121] - Zang-fu state: {"liver": 0.3, "heart": 0.7, "spleen": 0.5, "lungs": 0.6, "kidneys": 0.8},

[0122] - Syndrome nature: "mainly excess",

[0123] - Treatment principle: "soothing liver and relieving depression, regulating Qi movement"

[0124] Five-tone treatment scheme:

[0125] - Dominant tone: determine the main treatment tone (one of palace, commercial, corner, zhi, and feather)

[0126] - Auxiliary tone: combination of tone

[0127] - Mode selection: five-tone mode or seven-tone mode

[0128] - Music structure: tempo, speed, and strength variation mode

[0129] - Frequency parameters: main frequency, resonance frequency

[0130] - Treatment duration: recommended listening time and frequency

[0131] Brainwave regulation target:

[0132] - Target frequency band: "alpha wave enhancement, beta wave balance",

[0133] - Regulation intensity: {"theta": 0.2, "alpha": 0.8, "beta": 0.6, "delta": 0.1},

[0134] - Expected effect: "Liver soothing and depression relieving, stable mood"

[0135] Personalized adjustment:

[0136] - Age factor: "young people",

[0137] - Cultural preference: "classical and elegant",

[0138] - Solar term characteristics: "liver soothing in spring"

[0139] As Figure 4 shown in the construction flowchart of the five-tone music generation module, the five-tone music generation module is the core creative engine of the system, responsible for generating high-quality treatment music that meets the requirements of traditional Chinese medicine theory based on the results of traditional Chinese medicine diagnosis and the five-tone treatment plan. This module uses a music generation architecture that has been pre-trained on Chinese traditional music and fine-tuned for five-tone features, and includes three core components: a five-tone feature encoder, an autoregressive generator, and a super-resolution enhancer. The module uses a specially optimized WavTokenizer for audio tokenization, and through rhythm consistency constraints and traditional Chinese medicine flavor preservation mechanisms, it ensures that the generated music meets the treatment requirements of five-tone therapy while maintaining the cultural characteristics and artistic qualities of traditional Chinese music. The music generation module construction method is divided into traditional music pre-training phase, five-tone feature deep fine-tuning phase, rhythm constraint optimization phase, and flavor preservation enhancement phase. The construction method of the five-tone music generation module includes the following steps.

[0140] S31, traditional music model pre-training:

[0141] First, collect and integrate large-scale Chinese traditional music data, including guqin music, national music, opera arias, and palace music, etc. Different types of traditional music works, build a pre-training data set containing 200,000 traditional music works. The data is processed through audio quality screening, format standardization, rhythm feature labeling and cultural attribute classification to establish a unified training corpus for five-tone generation. The model pre-training process uses an autoregressive transformer music generation model, and a special audio tokenization scheme is designed for the five-tone scale features of Chinese traditional music. WavTokenizer is trained with traditional Chinese music features to accurately identify and maintain the rhythm features and interval relationships of the five tones of do, re, mi, sol, and la. The pre-training model loss function is defined as:

[0142]

[0143] where is the audio reconstruction loss, is the pentatonic scale preservation loss, is the cultural feature preservation loss, , , is the balance weight. The pentatonic scale preservation loss ensures that the generated music meets the requirements of the pentatonic scale through spectral analysis: where is the spectral feature of the generated audio, is the spectral template of the target pentatonic, is the weight coefficient of each pitch.

[0144] S32, pentatonic feature deep fine-tuning

[0145] On the basis of traditional music model pre-training, use specially labeled pentatonic music data for deep fine-tuning, strengthen the model's understanding and generation ability of the core requirements of the pentatonic therapy. This stage constructs a high-quality labeled dataset containing 50,000 pieces of pentatonic therapy music, each piece of music is labeled with the corresponding meridian, therapeutic effect, pitch ratio and clinical verification result. The fine-tuning process uses a multi-task learning framework, including five core tasks of pentatonic recognition, visceral correspondence, frequency precision control and therapeutic effect prediction. Parameter efficient fine-tuning (LoRA) technology is used for training, and the incremental weight parameter matrix is: where , , , represents the low-rank matrix decomposition.

[0146] a) Pentatonic precision recognition task: use multi-label classification loss: where is the number of time steps, is the th pitch label at the th time step, is the model prediction probability, ensuring that the model accurately identifies the pentatonic components in the music.

[0147] b) Visceral correspondence learning task: achieved through regression loss: , where is the true weight of the th visceral, is the model predicted weight, ensuring that the generated music accurately corresponds to the target visceral.

[0148] c) Frequency precision control task: use frequency domain loss: , where For the target frequency set, For generating energy at the frequency , For the target energy value, ensure accurate control of key treatment frequencies.

[0149] d) Treatment effect prediction task: through effect regression modeling: , For predicting treatment effect, For real clinical effect, For regularization coefficient, improve the prediction ability of the model to the treatment effect.

[0150] S33, pitch constraint optimization:

[0151] Establish strict pitch constraint mechanism to ensure that the generated music strictly follows the pitch requirements and theoretical specifications of traditional Chinese five-tone therapy. This stage designs four constraint systems: five-tone matching constraint, frequency range constraint, mode structure constraint and rhythm pattern constraint. Pitch constraint optimization uses constraint optimization framework to integrate constraint conditions into the generation process. The total constraint loss function is defined as:

[0152] a) Five-tone matching constraint: ensure that the length ratio of each tone in the generated music meets the requirements of the treatment plan: , is the actual five-tone matching, is the target matching, is the entropy regularization term to prevent extreme matching.

[0153] b) Frequency range constraint: ensure that the frequency of each tone is within the range specified by traditional Chinese medicine theory: , is the indicator function, is the main frequency of the th tone, is the allowed range, is the frequency projection function.

[0154] c) Mode structure constraint: maintain the interval relationship of five-tone mode: , is the interval relationship set, is the actual interval, is the standard five-tone mode interval.

[0155] d) Rhythm pattern constraint: ensure that the rhythm meets the requirements of traditional Chinese music therapy: , is the rhythm pattern, is the template rhythm, is the rhythm change gradient, and the smoothing term ensures natural rhythm change.

[0156] S4, Rhythm Preservation Enhancement:

[0157] Through the super-resolution enhancer and cultural feature preservation mechanism, the traditional Chinese medicine music's rhythm and cultural characteristics are preserved while ensuring the sound quality. In this stage, a super-resolution model based on flow matching is used to convert low-resolution music tokens into high-fidelity audio output, while integrating rhythm preservation constraints to ensure that traditional characteristics are not lost. The rhythm preservation mechanism includes four dimensions: timbre quality preservation, performance technique simulation, cultural atmosphere expression, and emotional resonance enhancement. The super-resolution loss function is: where is the fidelity loss, is the perceptual quality loss, is the cultural style preservation loss.

[0158] a) Timbre quality preservation: achieved through spectral envelope matching: where is the spectral envelope extraction function, is the reference traditional music spectrum.

[0159] b) Performance technique simulation: using time-frequency feature contrast learning: where is the generated music feature, is the traditional performance technique feature, is the temperature parameter.

[0160] c) Cultural atmosphere expression: through semantic embedding alignment: where is the generated music's emotional semantic embedding, is the target cultural atmosphere embedding.

[0161] d) Emotional resonance enhancement: using emotional consistency loss: where and are the emotional distributions of the generated music and the target music, respectively.

[0162] Through the above four stages of processing, the five-tone music generation module realizes the intelligent conversion from structured treatment plans to high-quality personalized treatment music, providing professional music therapy services that meet both the requirements of traditional Chinese medicine theory and modern sound quality standards. The innovation of this method lies in the design of five-tone feature deep learning and rhythm preservation mechanism, especially the pitch constraint optimization and cultural feature preservation technology, which enables the system to generate high-quality music works with true Chinese medicine treatment effects and traditional cultural connotations.

[0163] The input data format of the five-tone music generation module: structured five-tone treatment plan from the traditional Chinese medicine diagnosis engine; the output music format is:

[0164] High-fidelity audio files (WAV / FLAC format)

[0165] Music feature annotation: pentatonic component analysis, frequency spectrum analysis, rhythm structure, mode characteristics

[0166] Quality assessment report: musical accuracy, matching conformity, cultural fidelity, and therapeutic suitability.

[0167] like Figure 5 The figure shows the construction flow chart of the personalized optimization module. The personalized optimization module is the adaptive learning engine of the system, which is responsible for the continuous optimization and personalized improvement of the treatment plan through user feedback collection, treatment file management and reinforcement learning algorithm. This module prevents the solidification of the treatment plan and continuously improves the personalized effect through the construction of a multi-dimensional evaluation system and an adaptive parameter adjustment mechanism, ensuring that each user can obtain the five-tone music treatment plan that best suits their individual characteristics and treatment response. Based on the deep reinforcement learning framework and personalized recommendation algorithm, this module constructs a complete personalized optimization system that includes user behavior analysis, treatment effect tracking, dynamic adjustment of the plan and long-term effect prediction. The construction method of the personalized optimization module is divided into a multi-dimensional feedback data collection stage, a personalized file construction and management stage, a reinforcement learning algorithm optimization stage and an adaptive parameter adjustment stage. The construction method of the personalized optimization module is as follows:

[0168] S41, Multi-dimensional feedback data collection stage:

[0169] First, a comprehensive user feedback collection system is established to collect users' treatment response and effect data through four dimensions: subjective evaluation, objective monitoring, behavioral data, and physiological indicators. Subjective evaluation includes the user's preference for music, treatment experience, self-assessment of symptom improvement, and assessment of changes in quality of life. Objective monitoring collects physiological indicators such as heart rate variability, blood pressure changes, sleep quality, and EEG activity through wearable devices or medical devices. Behavioral data records user usage patterns such as frequency of use, listening time, number of interruptions, and repeat playback preferences. Physiological indicators include quantitative symptom scores before and after treatment, changes in physical examination indicators, and tracking of rehabilitation progress. Data collection adopts a multimodal fusion architecture to ensure the integrity and accuracy of the data. The feedback data quality evaluation function is defined as:

[0170] ,

[0171] in Score data completeness, Score the reliability of the data. Score the data consistency, Score the timeliness of the data. 、 、 , is the weight coefficient.

[0172] a) Subjective feedback standardization processing: using semantic analysis and sentiment recognition: where is the sentiment analysis loss, is the consistency loss between the front and back feedback, ensuring accurate analysis of subjective evaluation.

[0173] b) Objective index anomaly detection: identifying abnormal data through statistical methods: where is the observation value, and is the mean and standard deviation, is the anomaly detection threshold, is the data weight.

[0174] c) Multi-dimensional data fusion: using attention mechanism weighted fusion: where is the feedback feature of the dimension, is the dynamic weight, is the learning parameter.

[0175] S42, personalized profile construction:

[0176] Based on the collected multi-dimensional feedback data, construct a personalized treatment profile for each user, recording the user's treatment history, reaction pattern, preference characteristics and effect trend. The personalized profile adopts a dynamic updating mechanism, and the user portrait is continuously enriched and optimized as the treatment process progresses. The archive management system includes four core modules: user feature modeling, treatment trajectory tracking, reaction pattern recognition and prediction model construction. User feature modeling constructs a user's personalized representation vector through multi-dimensional feature learning, including physiological characteristics, psychological characteristics, behavioral preferences and treatment reactions. The profile construction loss function is defined as:

[0177]

[0178] where is the user representation learning loss, is the treatment trajectory modeling loss, is the effect prediction loss, is the privacy protection loss.

[0179] d) User feature representation learning:

[0180] Using deep embedding method: where is the demographic feature, is the symptom feature, Behavioral features, Physiological features.

[0181] e) Treatment trajectory dynamic modeling:

[0182] With recurrent neural networks: , and The hidden state at the moment, The current symptom state, The treatment action.

[0183] f) Reaction pattern clustering identification: Discover user groups through unsupervised learning: , and The probability that the user belongs to the th group, The center of the th group.

[0184] g) Long-term effect prediction modeling: Use time series prediction: , where is the predicted effect at the future moment, is the historical symptom sequence, is the historical treatment sequence.

[0185] S43, reinforcement learning algorithm optimization:

[0186] Adopting a deep reinforcement learning framework, individualized treatment optimization is modeled as a Markov decision process, learning the optimal treatment strategy through interaction with the environment (user treatment response). This stage uses an improved deep deterministic policy gradient (DDPG) algorithm, combined with multi-agent reinforcement learning to handle multi-user parallel optimization problems. The reinforcement learning environment design includes four key components: state space definition, action space design, reward function construction, and policy network training. The state space includes the user's current symptom state, historical treatment effect, individualized features, and environmental factors. The action space is defined as continuous treatment parameter adjustment, including five sound matching, frequency parameter, duration setting, and intensity control. The reinforcement learning objective function is defined as: , where is the policy network, is the reward at the th step, is the discount factor, is the treatment cycle length.

[0187] a) Reward function design: Consider treatment effect and user satisfaction: , where is the degree of symptom improvement, is the user satisfaction, compliance indicators, treatment costs.

[0188] b) Policy network update: with deterministic policy gradient: where is the action value function, is the state distribution.

[0189] c) Experience replay optimization: with prioritized experience replay: where is the priority of experience , is the priority exponent.

[0190] d) Multi-agent coordination: handling multi-user optimization conflicts: where is the policy of user , is the user similarity.

[0191] S44, adaptive parameter adjustment:

[0192] Based on the optimization results of reinforcement learning and user feedback, the adaptive adjustment of treatment parameters and dynamic optimization of the scheme are realized. This stage establishes a four-fold protection system of parameter sensitivity analysis, adjustment amplitude control, stability guarantee and rollback mechanism to ensure the safety and effectiveness of parameter adjustment. The adaptive adjustment adopts a hybrid strategy combining gradient descent optimization and Bayesian optimization to ensure convergence while avoiding local optimum. Uncertainty quantification is introduced in the parameter adjustment process, and the adjustment risk is evaluated through confidence interval estimation. The adaptive optimization objective function is defined as:

[0193] , where is the treatment effect loss, is the stability loss, is the safety loss, is the individualization loss.

[0194] a) Parameter sensitivity analysis: over-gradient calculation evaluation: where is the sensitivity coefficient of parameter to the output .

[0195] b) Adjustment amplitude control: with adaptive step strategy: where is the adaptive learning rate, is the gradient clipping threshold.

[0196] c) Stability guarantee mechanism: over-Lyapunov function verification: wherein is a Lyapunov function, is a stability constant.

[0197] d) Intelligent rollback mechanism: at the effect threshold judgment: . In and is the effect of the new and old scheme, is the effect threshold, is the safety threshold.

[0198] Compared with the traditional Chinese medicine five-tone therapy application system, the application not only solves the fundamental limitations of the traditional system relying on artificial experience and fixed music library, but also innovatively proposes an intelligent syndrome differentiation decision-making technical path based on a large language model, so that the Chinese medicine five-tone therapy has modern clinical application capability. The specific advantages are as follows: 1) intelligent syndrome differentiation and standardization: through the double-layer large language model collaborative architecture, end-to-end intelligent decision-making from patient multi-modal information to precise treatment scheme is realized, while the traditional system completely relies on the personal experience of TCM doctors, and the standardization degree is extremely low, which is difficult to promote and apply on a large scale; 2) personalized music generation and dynamic optimization: the music generation technology with five-tone feature deep preservation can generate personalized treatment music in real time according to the patient's specific symptoms, constitution type, seasonal and solar terms, etc., while the traditional system can only select pre-recorded music from a limited music library, and cannot adapt to individual differences, scientific verification and continuous improvement capability; 3) integrated brain wave regulation target and reinforcement learning optimization mechanism, through objective physiological index monitoring and user feedback to realize continuous optimization of treatment scheme, while the traditional system lacks objective efficacy evaluation system and mainly relies on patient subjective feeling to judge the effect.

[0199] Compared with the general music generation system based on deep learning, the advantages of the application are as follows: 1) deep integration of TCM theory and treatment targeting: through the TCM intelligent syndrome differentiation engine of the mixed expert architecture, the traditional TCM concepts of "syndrome differentiation and treatment, and individualized treatment" are transformed into a computable decision-making process, and the generated music has clear treatment targets and TCM theoretical basis, while the general system lacks medical professionalism and the generated music has no therapeutic value; 2) accurate preservation of five-tone features and inheritance of traditional charm: the innovative four-stage music generation technology (traditional pre-training-five-tone fine-tuning-constraint optimization-charm enhancement) ensures that the generated music strictly preserves the pitch characteristics of the five tones of palace, commercial, corner, zhi, and feather, and the traditional cultural connotation, while the general system is mainly based on modern music training and cannot understand and preserve the special requirements of TCM five tones; 3) multi-dimensional constraint and quality guarantee mechanism: four constraint systems of five-tone matching constraint, frequency range constraint, mode structure constraint and rhythm pattern constraint are established, and the constraint optimization framework is used to ensure that the music meets the TCM treatment specifications, while the general system lacks professional constraints and the generated music is highly arbitrary.

[0200] Compared with the music therapy recommendation system based on emotion recognition: the advantages of the present application are: 1) the deep docking of TCM syndrome types and modern emotion theory: the precise mapping from patient symptoms to TCM syndrome types is established by the multi-modal health information fusion method, instead of simple emotion recognition, which can accurately reflect the function state of viscera and the running condition of qi and blood, while the emotion system cannot reflect the holistic concept of TCM theory; 2) individualized TCM constitution identification and seasonal adaptation: the individualized adjustment mechanism integrating multiple factors such as age, gender, constitution, solar term, geographical location, etc. is embodied, which embodies the treatment principle of TCM "three principles of treatment" (treatment according to individual, treatment according to location, treatment according to time), while the emotion system is mainly based on general psychology principles, lacking TCM characteristics; 3) the essential difference between active treatment and passive recommendation: the present application generates five-tone music with clear treatment mechanism based on TCM syndrome differentiation, realizing active treatment intervention, while the emotion system mainly uses recommendation algorithms such as collaborative filtering, belonging to passive music matching, lacking treatment specificity.

[0201] The above disclosed content is only the preferred feasible embodiment of the present application, and does not limit the patent application range of the present application, so any equivalent technical changes made by applying the content of the specification and drawings of the present application are included in the patent application range of the present application.

Claims

1. A personalized music generation system for TCM five-tone therapy based on a large language model, characterized by: It includes: A multimodal health information collection module, which is used to collect and process patients' symptom descriptions, physical examination reports, geographic locations, and seasonal information, and convert the above information into standard structured Traditional Chinese Medicine symptom information; A TCM syndrome differentiation and decision-making module, comprising an organ state evaluation expert module, a pathogenesis analysis expert module, a five-tone treatment expert module, a brainwave regulation expert module, and a personalized adjustment expert module; the organ state evaluation expert module is used to analyze the functional states of the five internal organs according to the input TCM symptom information, and calculate the degree of deficiency and excess of each organ and their mutual relationships based on the symptom characteristics; the pathogenesis analysis expert module is used to perform pathological mechanism analysis according to the input TCM symptom information, and identify the root cause, pathogenesis, and transmission law of the disease; the five-tone treatment expert module is used to formulate specific five-tone treatment parameters according to the syndrome differentiation results output by the organ state evaluation expert module and the pathogenesis analysis expert module; the brainwave regulation expert module is used to set brainwave regulation targets and intensity parameters according to the syndrome differentiation results output by the organ state evaluation expert module and the pathogenesis analysis expert module; the personalized adjustment expert module is used to optimize the plan according to the patient's age, gender, constitution, solar terms, and geographical location; the TCM syndrome differentiation and decision-making module outputs the dialectical results, the five-tone treatment plan, the brainwave regulation target, and the personalized adjustment plan; The five-tone music generation module is used to generate a personalized music therapy plan based on the dialectical results output by the traditional Chinese medicine syndrome differentiation decision module, the five-tone treatment plan, the brain wave regulation target, and the personalized adjustment plan.

2. The personalized music generation system for TCM five-tone therapy based on a large language model according to claim 1 is characterized in that: It also includes a personalized optimization module, which realizes adaptive adjustment of treatment parameters and dynamic optimization of plans through the construction of a multi-dimensional evaluation system and an adaptive parameter adjustment mechanism.

3. A method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model as described in claim 1, characterized in that: The construction of the multimodal health information collection module includes the following steps: S11: Construction of TCM symptom ontology, which covers the four diagnostic information of TCM. The symptom ontology adopts a hierarchical structure, establishing a multi-layer mapping relationship from symptom category, symptom name, symptom description to symptom degree; S12: Multimodal feature fusion training, using a Transformer-based multimodal fusion architecture to integrate multiple input modalities such as text, speech, and images, to achieve symptom text understanding, medical image recognition, and test indicator analysis. It also constructs a TCM attribute mapping table for medical test indicators and converts modern medical indicators into TCM constitution characteristics. S13: Optimize spatiotemporal information mapping, constructing a mapping model between geographic location and climate characteristics and between seasonal solar terms and TCM attributes. Based on the patient's geographic location, the local climate characteristics are automatically obtained, the corresponding TCM environmental attributes are analyzed, and based on the theory of the 24 solar terms, the corresponding relationship between solar terms and the prosperity and decline of internal organs is established; S14: Information quality assessment: Evaluate the completeness, accuracy and consistency of collected information through multi-dimensional evaluation indicators.

4. A method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model as described in claim 1, characterized in that: The construction of the TCM syndrome differentiation decision-making module includes the following steps: S21: Multimodal data integration and preprocessing: Collect and integrate multi-source heterogeneous data required for TCM syndrome differentiation, standardize, align, desensitize, and structure the data, establish a unified training corpus for Five-tone therapy, and construct a symptom-viscera-five-tone relationship database, deeply linking traditional TCM theory with Five-tone therapy. S22: Pre-training of a large-scale TCM model. Based on a large-scale pre-trained model, pre-training is performed on TCM Five-tone Therapy corpus. This enhances the model's deep understanding of TCM dialectical thinking and Five-tone Therapy, enabling autoregressive language modeling, symptom-pathogenesis association prediction, Five-tone-viscera mapping tasks, and reconstruction of dialectical reasoning chains. S23: Fine-tuning the hybrid expert architecture. Based on the pre-trained TCM model, we construct expert models for viscera status assessment, pathogenesis analysis, five-tone therapy, brainwave regulation, and personalized adjustment. Each expert is responsible for a specific dialectical decision-making task. S24: Dialectical evaluation and optimization: Evaluate and iteratively optimize model performance through multi-dimensional evaluation indicators.

5. The method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model according to claim 4 is characterized in that: The viscera status assessment expert model is used to analyze the functional status of the five internal organs and calculate the degree of deficiency and excess of each viscera and their mutual relationship based on symptom characteristics. Its training loss function is: ,in For the The first organ Status tags; The pathogenesis analysis expert model is used for deep pathological mechanism analysis to identify the root cause, pathogenesis and transmission pattern of the disease. Its training loss function is: ,in is the cross entropy loss, is the consistency loss.

6. A method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model as described in claim 1, characterized in that: The five-tone music generation module is constructed by the following steps: S31: Pre-training of traditional music models: Collect and integrate traditional Chinese music data, screen the data for audio quality, standardize the format, annotate the musical characteristics, and classify the cultural attributes. Establish a unified training corpus for pentatonic generation, and use an autoregressive transformer music generation model for pre-training to generate traditional music models. S32: Deep tuning of five-tone features. Based on the traditional music model, this approach uses specially labeled five-tone music data for deep tuning, strengthening the model's understanding and generation of the core requirements of five-tone therapy. Each piece of music is annotated with its corresponding zang-fu meridian, therapeutic efficacy, musical ratio, and clinical validation results. The fine-tuning process utilizes a multi-task learning framework to achieve five-tone recognition, zang-fu correspondence, precise frequency control, and therapeutic effect prediction. S33: Optimize the musical temperament constraint and establish a musical temperament constraint mechanism to ensure that the generated music strictly follows the musical temperament requirements and theoretical specifications of the five-tone therapy of traditional Chinese medicine; S34: Flavor Preservation Enhancement, which uses a super-resolution enhancer and a cultural feature preservation mechanism to preserve the flavor and cultural characteristics of traditional Chinese medicine music while ensuring sound quality. It uses a stream matching-based super-resolution model to convert low-resolution music tokens into high-fidelity audio output, while integrating flavor preservation constraints to ensure that traditional characteristics are not lost.

7. The method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model according to claim 6 is characterized in that: The multi-task learning framework for deep fine-tuning of five-tone features includes four tasks: five-tone recognition, organ correspondence, precise frequency control, and treatment effect prediction.

8. The method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model according to claim 7 is characterized in that: The five-tone recognition task uses multi-label classification loss: ,in is the number of time steps, For the The time step Tone tags, Predict probabilities for the model to ensure that it accurately identifies the five-note components in the music.

9. The method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model according to claim 7 is characterized in that: Frequency domain loss is used for precise frequency control tasks: ,middle is the target frequency set, To generate audio at frequency The energy of The target energy value ensures precise control of key treatment frequencies.

10. The method for constructing a personalized music generation system for TCM five-tone therapy based on a large language model according to claim 7 is characterized in that The treatment effect prediction task is modeled through effect regression: ,in To predict treatment efficacy, For real clinical effects, is the regularization coefficient, which improves the model's ability to predict treatment effects.

Images