LoRA fine-tuning medical decision Agent system based on AI and implementation method

Abstract

The invention belongs to the technical field of medical auxiliary diagnosis, and particularly relates to an AI-based LoRA fine-tuning medical decision Agent system and an implementation method. The multi-modal data acquisition module is used for acquiring vital sign data, medical image data, laboratory inspection data and electronic medical record text data of a patient in real time; the dynamic weight adjustment LoRA fine adjustment module is used for carrying out low-rank adaptive fine adjustment on a pre-trained medical decision model by introducing a time attenuation factor, a data importance weight and a modal emergency degree coefficient based on the multi-modal data so as to generate a personalized decision model for the current patient; the multi-modal risk assessment module is used for fusing feature vectors output by the fine-tuned personalized decision-making model, constructing a comprehensive assessment function in combination with an attention mechanism and a feature interaction coefficient, and realizing dynamic quantitative assessment of the illness state risk of the patient; according to the method, the timeliness, the importance and the scene emergency degree of the multi-modal data can be dynamically balanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical auxiliary diagnosis, and particularly relates to an AI-based LoRA fine-tuning medical decision agent system and an implementation method. BACKGROUND

[0002] In the field of medical artificial intelligence, fine-tuning technology based on pre-trained models has been widely used in scenarios such as auxiliary diagnosis and treatment plan recommendation. In the prior art, the low-rank adaptation (LoRA) fine-tuning method for medical decision models usually adopts a fixed weight distribution strategy, and lacks a dynamic adjustment mechanism for processing multi-modal medical data such as medical images, vital signs, and laboratory test results. Specifically, the traditional LoRA fine-tuning does not fully consider the timeliness characteristics of medical data, and gives equal weight to recent key data and long-term non-key data, making it difficult for the model to accurately capture the real-time changes in the patient's condition. At the same time, its data importance evaluation relies on static clinical guideline coefficients and does not dynamically correct them in combination with data freshness, so that outdated high-importance data still dominates model training, reducing the model's ability to adapt to the current state of individual patients. In addition, the modality weight is not adjusted according to the urgency of the diagnosis and treatment scene, so that real-time vital signs and other key data in emergency scenarios cannot preferentially affect model decision-making, affecting the response accuracy in emergency situations. The above problems result in limitations of existing medical decision models in terms of individualized adaptation, real-time disease response, and scene adaptability, making it difficult to meet the demand for precise and dynamic medical decision-making in clinical practice.

[0003] Based on the above problems, there is an urgent need for a model fine-tuning technology scheme that can dynamically balance the timeliness, importance, and scene urgency of multi-modal data. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art and to propose an AI-based LoRA fine-tuning medical decision agent system, which comprises: A multi-modal data acquisition module, which includes real-time acquisition of patients' vital sign data, medical image data, laboratory test data, and electronic medical record text data; A dynamically weighted LoRA fine-tuning module, which, based on the multi-modal data, fine-tunes a pre-trained medical decision model through the introduction of a time decay factor, a data importance weight, and a modality urgency coefficient, to generate an individualized decision model for the current patient; A multi-modal risk assessment module configured to fuse the feature vectors output by the fine-tuned individualized decision model, construct a comprehensive evaluation function in combination with an attention mechanism and a feature interaction coefficient, and realize dynamic quantitative assessment of the patient's disease risk; An adaptive decision generation module is configured to generate an adaptive treatment plan adjustment strategy according to the risk assessment result in combination with a treatment effect feedback factor, a patient tolerance parameter and a dynamic response coefficient; A human-computer collaborative interaction module is configured to fuse the treatment plan adjustment strategy with professional judgment of a clinician to form a final medical decision suggestion.

[0005] Preferably, the multi-modal data acquisition module comprises a wearable device interface configured to acquire heart rate, blood pressure and blood oxygen saturation vital sign data of the patient in real time. A medical image interface is configured to access CT, MRI and ultrasonic medical image devices and parse image data based on a DICOM standard protocol. A laboratory data interface is integrated with a hospital LIS system to acquire blood routine, biochemical index and pathological report laboratory test data. A natural language processing unit is configured to perform entity recognition, relationship extraction and semantic representation learning on electronic medical record text data to generate a structured medical record feature vector.

[0006] Further preferably, the LoRA fine-tuning module with dynamic weight adjustment comprises: A data time effectiveness analysis unit is configured to calculate a time decay coefficient of each modality data, wherein the decay coefficient of recent data is greater than that of long-term data. A data importance evaluation unit is configured to assign an importance weight to different modality data based on clinical guidelines and expert knowledge, wherein the weight of medical image data is higher than that of conventional vital sign data. A low-rank matrix decomposition unit is configured to perform singular value decomposition on a full connection layer weight matrix of a pre-trained model, extract a core low-rank subspace, and dynamically update the low-rank matrix through the time decay coefficient and the data importance weight. A model fusion unit is configured to superimpose the updated low-rank matrix and the original model parameters to generate a personalized medical decision model.

[0007] Further preferably, the multi-modal risk assessment module comprises: A modality alignment unit is configured to map feature vectors of different modalities to a unified semantic space through a cross-modal contrast learning method to eliminate semantic gaps between modalities. An attention mechanism unit is configured to calculate attention weights of features of each modality, wherein the weight of a disease-related feature is higher than that of a non-related feature. A risk quantification unit is configured to construct a comprehensive risk scoring system comprising disease severity, complication risk and treatment response prediction dimensions based on the attention weights and the semantically aligned feature vectors. A dynamic threshold adjustment unit adaptively adjusts the threshold range of the risk score according to the basic health status and disease progression stage of the patient.

[0008] Further preferably, in the dynamic weight adjustment LoRA fine-tuning module, the calculation formula of the time decay factor β, the data importance weight and the modality urgency coefficient is: ; Wherein, t is the time difference between the data acquisition time and the current time, is the time decay inflection point threshold, is the decay rate parameter, is the clinical benchmark importance coefficient of the i-th modality data, is the freshness index of the i-th modality data, n is the total number of modality data, is the urgency coefficient of the i-th modality, the vital signs =0.8-1.0 in emergency scenarios, image data =0.5-0.7, σ is the emergency degree coefficient of the current diagnosis and treatment stage, is the final modality dynamic weight.

[0009] Further preferably, in the multi-modality risk assessment module, the expression of the comprehensive evaluation function is: ; Wherein, is the attention weight of the j-th modality, is the independent influence coefficient of the k-th feature, is the nonlinear transformation function of the k-th feature in the j-th modality, m is the number of modalities, and p is the number of single modality features, is the interaction influence coefficient of the k-th and l-th features.

[0010] Further preferably, in the adaptive decision generation module, the calculation formula of the treatment scheme adjustment strategy is: ; Wherein, is the treatment scheme adjustment amount, γ is the treatment effect feedback factor, E is the current efficacy index, is the expected efficacy index, ξ is the effect fluctuation tolerance coefficient, δ is the patient tolerance parameter, T is the implemented treatment time, is the maximum tolerance time, is the interval time between two adjustments, is the adjustment interval reference value, is the interval sensitivity coefficient.

[0011] Further preferably, the human-computer collaborative interaction module comprises: A doctor decision fusion unit that weights and fuses the AI-generated treatment plan adjustment strategy and the doctor's decision suggestion, wherein the weight of the doctor's suggestion is higher than that of the AI strategy; A visual interactive interface that displays the risk assessment results, treatment plan adjustment basis and historical decision records in the form of interactive charts and structured reports; A privacy protection unit that uses homomorphic encryption and secure multi-party computation technology to ensure the security and privacy of patient data during cross-institutional transmission and sharing.

[0012] Further preferably, it further comprises a lightweight model deployment module that compresses the personalized decision model after LoRA fine-tuning to mobile devices through model pruning and knowledge distillation technology.

[0013] An implementation method applied to any one of the AI-based LoRA fine-tuning medical decision Agent systems described above, characterized in that it comprises: Real-time acquisition of multi-modal data of patients, fine-tuning of pre-trained models based on time decay factor, data importance weight and modality urgency coefficient through a dynamic weight adjustment LoRA fine-tuning module, generation of personalized decision models; Fusion of feature vectors through a multi-modal risk assessment module, and quantification of disease risk using a comprehensive evaluation function containing feature interaction terms; Based on the risk assessment results, the adaptive decision generation module generates a treatment plan in combination with treatment effect feedback, patient tolerance and adjustment interval parameters; Fusion of AI decisions and doctor suggestions to form the final medical decision; and realization of system edge running through a lightweight deployment module.

[0014] Technical effects: The present application introduces time decay factor, data importance weight and modality urgency coefficient in LoRA fine-tuning, solving the problem of fixed weight distribution of multi-modal medical data, ignoring time effectiveness and scene urgency differences in the prior art. Its creativity lies in dynamically balancing data time effectiveness stratification and clinical scene adaptability, combining multi-modal feature interaction and adaptive decision mechanism, so that the medical decision model can accurately capture real-time disease changes, improve the response accuracy to emergency scenes, enhance the individualized adaptation ability, and meet the demand for dynamic and accurate medical decision making in clinical practice. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The figure is a block diagram of the AI-based LoRA fine-tuning medical decision Agent system of the present application; Figure 2 The figure is a block diagram of the multi-modal data acquisition module of the present application; Figure 3 Block diagram of the LoRA fine-tuning module of the dynamic weight adjustment of the application; Figure 4 Block diagram of the multi-modal risk assessment module of the application; Figure 5 Block diagram of the human-computer collaborative interaction module of the application; Figure 6 Flow chart of the AI-based LoRA fine-tuning medical decision Agent implementation method of the application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the application clearer and more apparent, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0017] In the description of the application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the application and simplify the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In addition, in the description of the application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0018] The conventional technical solutions have the following technical problems: In the LoRA fine-tuning process of the existing medical decision model, the time decay characteristics, the differences in clinical importance and the urgency of the diagnosis and treatment scene of the multi-modal data are not fully considered, resulting in a lag in the model's response to real-time data of emergency patients; there is a lack of dynamic weight adjustment mechanism in the fusion of multi-modal data, making it difficult to balance the contribution of different types of data; risk assessment relies only on single modal features or simple weighting, and cannot capture the synergistic association between symptoms and test results; the generation of treatment plans does not combine the dynamic changes of patient real-time tolerance and efficacy feedback, and the adaptability is insufficient; the fusion mechanism of AI decision and doctor's judgment in human-computer interaction is imperfect, and decision conflicts are prone to occur.

[0019] Based on this, please refer to Figure 1The embodiment provides an AI-based LoRA fine-tuning medical decision agent system, which comprises a multi-modal data acquisition module configured to acquire real-time vital sign data, medical image data, laboratory test data and electronic medical record text data of a patient; a dynamic weight adjustment LoRA fine-tuning module configured to fine-tune a pre-trained medical decision model through low-rank adaptation based on the multi-modal data by introducing a time decay factor, a data importance weight and a modality urgency coefficient, so as to generate a personalized decision model for the current patient; a multi-modal risk assessment module configured to fuse a feature vector output by the fine-tuned personalized decision model, combine an attention mechanism and a feature interaction coefficient to construct a comprehensive evaluation function, and realize dynamic quantitative assessment of the patient's condition risk; an adaptive decision generation module configured to generate an adaptive treatment plan adjustment strategy according to the risk assessment result, in combination with a treatment effect feedback factor, a patient tolerance parameter and a dynamic response coefficient; and a man-machine collaborative interaction module configured to fuse the treatment plan adjustment strategy and the professional judgment of a clinician to form a final medical decision suggestion.

[0020] The scheme solves the deficiencies of the traditional model in data processing, fine-tuning adaptation, risk assessment, scheme generation and man-machine interaction through multi-module cooperation, and realizes the individualization, dynamization and precision of medical decision. The dynamic weight adjustment LoRA fine-tuning module optimizes the adaptability of the model through multi-dimensional parameters, the multi-modal risk assessment module captures clinical correlations through feature interaction, the adaptive decision generation module realizes real-time adjustment of the treatment plan, and the man-machine collaborative module guarantees the reliability of the decision.

[0021] It is worth mentioning that: the multi-modal data acquisition module covers physiological signals, images, tests and texts and other full-dimensional data, ensuring information integrity; the dynamic weight adjustment LoRA fine-tuning module adjusts the model through three adjustments of time decay, importance weight and urgency coefficient, so that the model is more suitable for the clinical actual scene; the multi-modal risk assessment module not only considers the influence of a single feature, but also focuses on the synergistic effect between features, such as the combined risk of high fever and white blood cell elevation; the adaptive decision generation module dynamically combines the efficacy feedback and the patient tolerance to avoid rigid schemes; and the man-machine collaborative interaction module takes the professional judgment of the doctor as the core, and the AI decision as the auxiliary, to form a complement.

[0022] The technical effects achieved by the above embodiments include: the dynamic weight adjustment mechanism makes the model more sensitive to real-time data of emergency patients, improving the decision-making speed in emergency scenarios; the multi-modal risk assessment is more in line with the clinical diagnosis logic through feature interaction analysis, reducing the risk of missed or misdiagnosed; the adaptive treatment plan generation can be optimized in real time according to patient feedback, improving treatment compliance; human-computer collaborative interaction balances AI efficiency and physician experience, enhancing the credibility and safety of decision-making. The system as a whole realizes the whole-process optimization from data collection to final decision-making, adapting to different diagnosis and treatment scenarios and individual patient differences.

[0023] The traditional technical solution has the following technical problems: existing medical data collection systems are mostly single modality independent collection, lacking standardized interfaces and unified processing mechanisms, resulting in heterogeneous data formats and integration difficulties of multiple sources such as vital signs, medical images, laboratory data, and electronic medical records; wearable device data lacks real-time performance, medical image analysis relies on manual format conversion, and the degree of structuring of laboratory data and electronic medical records is low, affecting subsequent model training and decision-making efficiency; the interface protocol is not optimized for medical scenarios during data collection, resulting in data transmission delays or losses, especially affecting the timeliness of decision-making in emergency diagnosis and treatment.

[0024] Based on this, please refer to Figure 2 , the multi-modal data collection module includes: a wearable device interface configured to acquire real-time patient vital signs such as heart rate, blood pressure, and blood oxygen saturation; a medical image interface configured to access CT, MRI, ultrasound, and other medical imaging devices and parse image data based on DICOM standard protocol; a laboratory data interface configured to integrate hospital LIS systems to obtain laboratory test data such as blood routine, biochemical indicators, and pathology reports; and a natural language processing unit configured to perform entity recognition, relationship extraction, and semantic representation learning on electronic medical record text data to generate structured medical record feature vectors.

[0025] This solution solves the heterogeneity and integration difficulties of multi-modal medical data collection through standardized interfaces and specialized processing units, achieving whole-process standardization from data collection to structured processing. The wearable device interface ensures the real-time performance of vital signs, the DICOM protocol parsing ensures the integrity of image data, the LIS system integration realizes efficient docking of laboratory data, and the natural language processing improves the degree of structuring of electronic medical records, providing a high-quality data foundation for subsequent model fine-tuning and risk assessment.

[0026] It is worth mentioning that: the wearable device interface adopts a mixed communication protocol of Bluetooth 5.0 and LoRa, which reduces power consumption while ensuring real-time performance, and is suitable for long-term monitoring scenarios; the medical image interface has a built-in DICOM analysis engine, which supports automatic conversion of image formats of multiple vendors, extraction of key parameters such as pixel value, layer thickness, window width, and window level; the laboratory data interface is connected to the LIS system through the HL7FHIR standard, realizing the structured extraction of test items, values, reference ranges, and other data; the natural language processing unit uses a pre-trained model in the medical field to perform entity recognition and relationship extraction on symptom descriptions, diagnostic terms, medication records, and other information in medical records, and generates feature vectors that meet clinical standards.

[0027] The technical effects achieved by the above embodiments include: standardized interfaces improve the efficiency of multi-modal data acquisition, reduce manual intervention in format conversion, and reduce data preprocessing costs; real-time data acquisition from wearable devices ensures the continuity of vital sign monitoring and provides real-time basis for dynamic disease assessment; DICOM protocol analysis of medical images ensures complete extraction of image features and improves the input quality of subsequent image analysis models; structured processing of laboratory data and electronic medical records enables models to directly use high-quality labeled data for training, improving fine-tuning efficiency and decision accuracy; multi-interface collaboration reduces data transmission delay, especially in emergency scenarios, providing data support for rapid decision-making, and overall improving the standardization, real-time performance, and usability of medical data acquisition.

[0028] The existing LoRA fine-tuning module in the medical model training has the following technical problems: the weight distribution of multi-modal data is fixed, and the time decay characteristics of the data are not considered, resulting in equal influence of long-term non-critical data and recent important data on the model, affecting the capture of real-time disease changes; data importance assessment only relies on static clinical guidelines, without dynamic adjustment based on data freshness, so that outdated data still occupies a high weight, reducing the model's ability to adapt to individual needs; the low-rank matrix decomposition and model fusion process are not optimized for medical scenarios, resulting in insufficient ability of the fine-tuned model in multi-modal data correlation analysis, especially in complex disease diagnosis.

[0029] Based on this, please refer to Figure 3, the dynamic weight adjustment LoRA fine-tuning module comprises: a data timeliness analysis unit configured to calculate a time decay coefficient of each modality data, wherein the decay coefficient of recent data is greater than that of long-term data; a data importance evaluation unit configured to assign an importance weight to different modality data based on clinical guidelines and expert knowledge, wherein the weight of medical image data is higher than that of conventional vital sign data; a low-rank matrix decomposition unit configured to perform singular value decomposition on the full connection layer weight matrix of the pre-trained model, extract a core low-rank subspace, and dynamically update the low-rank matrix through the time decay coefficient and the data importance weight; and a model fusion unit configured to superimpose the updated low-rank matrix and the original model parameters to generate a personalized medical decision-making model.

[0030] The scheme solves the problems of data weight static and insufficient model adaptability in traditional fine-tuning by optimizing the LoRA fine-tuning process in multiple units. The data timeliness analysis unit dynamically adjusts the time weight to ensure that recent data has a greater impact. The data importance evaluation unit assigns weights in combination with clinical priorities to highlight the value of key modalities. The low-rank matrix decomposition and fusion unit uses a dynamic updating mechanism to make the model more suitable for individual patient data characteristics and improve personalized decision-making capabilities.

[0031] It is worth mentioning that: the data timeliness analysis unit uses a segmented decay strategy, with a decay coefficient of 1.0 within 48 hours, 0.7 for 48-72 hours, and 0.3 for more than 72 hours, which adapts to the time-sensitive characteristics of medical data; the data importance evaluation unit establishes a modality weight table, with medical images (CT / MRI) having a weight of 0.35, laboratory test data having a weight of 0.25, vital signs having a weight of 0.2, and electronic medical records having a weight of 0.2, which can be dynamically fine-tuned according to the type of illness; the low-rank matrix decomposition unit extracts the low-rank subspace corresponding to the first 20% singular values for the Attention layer of the Transformer model to ensure that key features are retained; and the model fusion unit uses a weighted superposition mechanism, with the low-rank matrix weight being dynamically adjusted according to the data importance to avoid excessive coverage of the original model parameters.

[0032] The above embodiments achieve the following technical effects: the dynamic adjustment of the time decay coefficient makes the model pay more attention to recent data changes of the patient, improving the sensitivity to illness deterioration or improvement; the data importance weight assignment highlights the role of key data such as medical images, in line with the principle of "image priority" in clinical diagnosis; the dynamic updating and fusion of the low-rank matrix enhances the adaptability to individual patient data based on the retention of pre-trained model medical knowledge, reducing the risk of overfitting; the personalized decision-making model can better capture the unique illness characteristics of the patient, especially in chronic disease management and complex complication diagnosis, and overall improves the fine-tuning efficiency and decision-making accuracy of the model.

[0033] The traditional technical solution has the following technical problems: the existing multi-modal risk assessment module mainly adopts a single mode independent evaluation and simple weighting method, which does not solve the semantic gap problem of different modal data, resulting in insufficient cross-modal feature fusion; the risk assessment only depends on the independent influence of the features, ignoring the synergistic correlation between symptoms and test results, such as the combination risk of "chest pain + elevated troponin", which does not conform to the clinical diagnosis logic; the attention mechanism weight distribution is static, and the importance of each mode is not dynamically adjusted according to the disease stage, resulting in a decrease in evaluation accuracy during disease progression; the risk quantification lacks a dynamic threshold adjustment mechanism, and a unified standard is used for patients with different basic health conditions, resulting in insufficient pertinence of the evaluation results.

[0034] Based on this, please refer to Figure 4 The multi-modal risk assessment module includes: a modal alignment unit configured to map feature vectors of different modalities to a unified semantic space through a cross-modal contrast learning method, eliminating the semantic gap between modalities; an attention mechanism unit configured to calculate the attention weight of each modal feature, wherein the weight of the disease-related feature is higher than that of the non-related feature; a risk quantification unit configured to construct a comprehensive risk score system including disease severity, complication risk, and treatment response prediction dimensions based on the attention weight and the semantically aligned feature vector; and a dynamic threshold adjustment unit configured to adaptively adjust the threshold range of the risk score according to the patient's basic health condition and disease progression stage.

[0035] This scheme optimizes the risk assessment process through multi-unit cooperation, solving the problems of insufficient cross-modal fusion, feature correlation ignored, weight static, and threshold fixed in traditional evaluation. The modal alignment unit realizes semantic unification of multi-modal features; the attention mechanism unit dynamically highlights key features; the risk quantification unit constructs a multi-dimensional scoring system; and the dynamic threshold adjustment unit adapts to individual differences, making the risk assessment more accurate and comprehensive. It is worth mentioning that: the modal alignment unit uses a contrast loss function to map the visual features of medical images, the numerical features of laboratory data, and the text features of electronic medical records to a 1024-dimensional unified space, and optimizes the alignment accuracy by calculating the feature cosine similarity; the attention mechanism unit presets the basic weight for different disease types, such as "ECG feature" weight 0.4 in cardiovascular disease and "image feature" weight 0.5 in respiratory disease, and dynamically adjusts according to real-time data; the risk quantification unit constructs a scoring matrix from three dimensions of disease severity, complication probability, and treatment response rate, and generates a comprehensive risk value by weighted summation; the dynamic threshold adjustment unit sets differentiated thresholds for elderly patients and patients with chronic diseases, such as a 20% reduction in the infection risk threshold for diabetic patients compared to healthy people.

[0036] The technical effects achieved by the above embodiments include: modal alignment eliminates semantic gap, making multi-modal feature fusion more efficient, and improving the correlation analysis capability of cross-modal data; the attention mechanism dynamically adjusts the weight to ensure that key features dominate in risk assessment and reduce irrelevant information interference; the multi-dimensional risk quantification system more comprehensively reflects the patient's condition and provides multi-perspective reference for treatment plan development; dynamic threshold adjustment adapts to different patient groups to avoid misjudgment caused by "one-size-fits-all" evaluation, especially for patients with poor basic health conditions, which can provide early warning of risks. The overall accuracy and clinical adaptability of risk assessment are improved to provide a reliable basis for subsequent treatment decisions.

[0037] The traditional technical solution has the following technical problems: the time weight distribution of the existing LoRA fine-tuning mostly adopts linear decay or simple exponential decay, which cannot accurately depict the time effectiveness stratification characteristics of medical data, resulting in insignificant differences in the weights of recent key data and long-term non-key data, affecting the sensitivity of the model to changes in the condition; the data importance evaluation only relies on static clinical coefficients without dynamic correction combined with data freshness, so that outdated high-importance data still dominates the model fine-tuning, reducing the individualized adaptation capability; the urgency difference of diagnosis and treatment scenarios is not considered, and the same weight strategy is used in emergency and routine diagnosis and treatment, resulting in a lag in the response of real-time data in emergency scenarios, affecting the timeliness of decision-making.

[0038] Based on this, in the dynamic weight adjustment LoRA fine-tuning module, the calculation formulas of the time decay factor β, the data importance weight and the modal urgency coefficient are as follows: ; Where t is the time difference between data collection time and current time, is the time decay inflection point threshold, is the decay rate parameter, is the clinical benchmark importance coefficient of the i-th modal data, is the freshness index of the i-th modal data, n is the total number of modal data, is the urgency coefficient of the i-th modal, the vital signs =0.8-1.0 in emergency scenarios, the image data =0.5-0.7, and σ is the emergency degree coefficient of the current diagnosis and treatment stage, is the final modal dynamic weight.

[0039] The formula includes three related expressions of the time decay factor β, the data importance weight and the modal dynamic weight , which are used to dynamically balance the weight distribution of multi-modal medical data in LoRA fine-tuning.

[0040] Time decay factor β: expression is where t represents the interval between data collection time and current time, reflecting the timeliness of data; is the time decay inflection threshold, such as 48 hours, which is the critical point to distinguish recent and long-term data; is the decay rate parameter, such as 24 hours, which controls the steepness of the decay curve. This formula uses the Sigmoid function form, making the β value of recent data close to 1, and the β value of long-term data quickly decays, accurately depicting the timeliness characteristics of key and long-term weight reduction within 48 hours of medical data, solving the problem of traditional linear decay that cannot distinguish the timeliness hierarchy of data.

[0041] Data importance weight : expression is where is the clinical benchmark importance coefficient of the i-th modality data, such as medical imaging λ = 0.4, vital signs λ = 0.2, based on clinical guidelines; is the data freshness index between 0 and 1, the closer to 1 the newer, correcting the static benchmark coefficient; is the total number of modalities. This formula realizes dynamic correction through the product of "benchmark coefficient × freshness", avoiding outdated high importance data still dominating model fine-tuning, and improving the adaptability of the model to the current state of the patient.

[0042] Modality dynamic weight : expression is where is the urgency coefficient of the i-th modality, such as vital signs = 0.8-1.0 in emergency scenarios, image data = 0.5-0.7, reflecting the priority of different modalities in emergency scenarios; is the emergency degree coefficient of the current diagnosis and treatment stage, taking 0.8-1.0 in emergency scenarios, used to amplify the weight influence in emergency scenarios. This formula integrates time decay, dynamic importance, and scene urgency, making the weight distribution of LoRA fine-tuning more consistent with clinical reality, such as prioritizing real-time vital sign data in emergency situations, and balancing image and text data weights in regular diagnosis and treatment.

[0043] The formula optimizes the time weight distribution through the Sigmoid type attenuation function, introduces the freshness index to correct the clinical importance coefficient, and realizes dynamic adaptation combined with the emergency parameter, solves the problem of ignoring the data aging layering and scene emergency difference of traditional LoRA, and improves the response accuracy of the model to emergency scenes. The scheme constructs a dynamic weight mechanism through a three-section formula, the Sigmoid type time decay factor accurately distinguishes the weight difference between recent and long-term data, the freshness index correction makes the clinical importance coefficient dynamically adjust with the data aging, and the emergency parameter adapts to different diagnosis and treatment scenes. The three cooperate to optimize the weight distribution of LoRA fine-tuning, so that the model is more suitable for the characteristics of medical data and the needs of clinical scenes.

[0044] It is worth mentioning that the time decay inflection point threshold is set to 48 hours, and the decay rate parameter is set to 24, so that the β value of the data within 48 hours remains above 0.8, and rapidly decays to below 0.3 after 72 hours, adapting to the characteristics of medical data "key within 48 hours"; the clinical benchmark importance coefficient , among them, medical imaging λ=0.4, laboratory data λ=0.25, vital signs λ=0.2, electronic medical record λ=0.15, freshness index is dynamically refreshed according to the data update frequency, such as real-time vital signs updated every 5 minutes =1.0, daily laboratory data μi=0.8; emergency coefficient In emergency scenes, vital signs =0.9, image data =0.6, σ=0.9, in regular scenes =0.3-0.5, σ=0.5, and the influence of real-time key data is amplified through weight amplification.

[0045] The technical effects achieved by the above embodiments include: the Sigmoid type time decay makes the weight of recent data significantly higher than that of long-term data, improving the model's ability to capture real-time changes in the patient's condition, especially the symptoms of acute patients; the freshness index corrects the clinical importance coefficient, avoiding outdated data from dominating fine-tuning, and enhancing the model's adaptability to the patient's current state; the emergency parameter dynamically amplifies the weight of key data in emergency scenes, so that the model prioritizes responding to real-time data such as vital signs in emergency care, improving decision-making speed; the final modal dynamic weight ωi integrates multiple factors, making LoRA fine-tuning more in line with clinical diagnosis and treatment logic, balancing data aging, importance, and scene emergency, and overall improving the model's individual decision-making accuracy and scene adaptability.

[0046] The traditional technical solutions have the following technical problems: the existing multi-modal risk assessment method adopts single feature weighted summation, only considers the independent influence of each feature, ignores the synergistic correlation between features in medical data, such as the joint indication of high fever + white blood cell increase on the risk of infection, leading to one-sided risk assessment, which does not conform to the logic of symptom-test-image multi-dimensional linkage in clinical diagnosis; the feature transformation function adopts simple linear or single nonlinear function, which cannot accurately depict the complex nonlinear relationship of medical features, affecting the risk quantization accuracy; the weight adjustment mechanism is not designed for the specificity of different modal data, leading to unbalanced contribution between modalities, such as unreasonable weight allocation between image features and text features.

[0047] Based on this, in the multi-modal risk assessment module, the expression of the comprehensive evaluation function is: ; Among them, is the attention weight of the jth modality, is the independent influence coefficient of the kth feature, is the nonlinear transformation function of the kth feature in the jth modality, which adopts a composite function of ReLU and Gaussian function, m is the number of modalities, and p is the number of single modality features, is the interaction influence coefficient of the kth and lth features, which is calculated by mutual information, and the value range is -1 to 1, and the positive correlation feature takes a positive value.

[0048] Outer modality weight : represents the overall contribution of the jth modality, such as medical images and vital signs, which is dynamically updated through cross-modal contrast learning, such as the image modality , the symptom modality , to ensure that the key modality dominates in risk assessment.

[0049] Inner independent feature term : Among them is the independent influence coefficient of the kth feature, such as the independent weight of body temperature on the risk of infection; is the nonlinear transformation function of the kth feature in the jth modality, which adopts a composite function of ReLU and Gaussian function, which is used to capture the nonlinear distribution of the feature, such as the exponential increase of risk after the body temperature exceeds 38℃; p is the number of single modality features. This part describes the independent influence of a single feature on the risk.

[0050] Feature interaction term : Among them is the interaction influence coefficient of the kth and lth features, which is calculated by mutual information, and the value range is -1 to 1, which is used to capture the synergistic effect of clinically strongly correlated features, such as the risk of cough + lung shadow combination being much higher than that of single feature superposition, Take high positive value; l is a feature index different from k. This part solves the problem of ignoring feature correlation in traditional models, making risk assessment more consistent with clinical diagnosis logic, such as high fever + white blood cell elevation combined to suggest infection.

[0051] The formula captures the synergistic effect of symptoms-test results by introducing feature interaction terms, solves the problem that traditional single feature weighting cannot reflect the correlation of medical data, and makes risk assessment more consistent with clinical diagnosis logic.

[0052] This scheme breaks through the limitation of traditional single feature superposition by double-layer weighting mechanism and feature interaction term design, realizes the deep fusion of multi-modal features. The outer layer ωj dynamically adjusts the contribution of each modality, and the inner layer and respectively depict the independent influence and synergistic effect of features, and the composite nonlinear function enhances the feature expression ability, making risk assessment more consistent with the complex characteristics of medical data.

[0053] It is worth mentioning that: the feature interaction term γkl is calculated by mutual information, which gives high positive values to clinically strongly correlated features such as cough + lung shadow and blood glucose elevation + urine sugar positive, and negative values to negatively correlated features such as normal body temperature + bacterial infection, accurately capturing pathological correlations; the nonlinear transformation function uses the composite of ReLU(x)=max(0,x) and Gaussian function exp(-(x-μ)² / 2σ²), which not only retains the nonlinear trend of features, but also suppresses the interference of extreme values; the modality attention weight ωj is dynamically updated through cross-modal contrast learning, such as ωj=0.4 for image modality, ωj=0.3 for symptom modality, and ωj=0.3 for test modality in pneumonia diagnosis.

[0054] The technical effects achieved by the above embodiments include: feature interaction terms effectively capture the synergistic relationship between clinical features, making risk assessment upgrade from isolated feature judgment to associated feature reasoning, more consistent with doctor's diagnosis thinking; double-layer weighting mechanism balances the contribution of multi-modal data, avoiding the one-sidedness caused by single modality dominating evaluation; composite nonlinear function improves feature expression accuracy, making risk characterization of abnormal values and critical values more delicate; the overall risk assessment result is closer to clinical practice, reducing missed diagnosis caused by ignoring feature correlation, and providing more accurate risk basis for treatment plan formulation.

[0055] The traditional technical scheme has the following technical problems: existing treatment scheme adjustment mostly adopts fixed proportion adjustment or linear adjustment based on single efficacy index, without considering the smoothness of efficacy fluctuation, resulting in excessive adjustment amplitude causing patient intolerance; without combining the dynamic relationship between treatment time and patient maximum tolerance duration, making the long-term treatment scheme rigid and unable to adapt to patient tolerance decay; lack of sensitivity design for adjustment interval, frequent adjustment in a short time easily destroys treatment continuity and affects efficacy stability.

[0056] Based on this, the adaptive decision-making module, the calculation formula of the treatment scheme adjustment strategy is: ; Where Δd is the treatment scheme adjustment amount, such as drug dose adjustment percentage, treatment frequency change value, γ is the treatment effect feedback factor, 0.3-0.7, set according to disease type, E is the current efficacy index, such as tumor volume reduction, blood glucose reduction value, is the expected efficacy index, ξ is the effect fluctuation tolerance coefficient, the smaller the value, the more sensitive to deviation, is the patient tolerance parameter, the value range is 0-1, calculated based on pain score and adverse reaction grade, T is the implemented treatment time, is the maximum tolerance time, Δt is the interval time between two adjustments, is the adjustment interval reference value, is the interval sensitivity coefficient. The formula smooths the adjustment amplitude caused by the deviation through the exponential decay term, introduces the time interval sensitivity function to avoid frequent adjustment in a short time, solves the problem that the traditional fixed strategy cannot balance the dynamic changes of treatment effect and patient tolerance, and makes the scheme adjustment more consistent with the clinical stepwise treatment principle.

[0057] The scheme realizes the dynamic adaptation of the treatment scheme through the double adjustment mechanism and the smoothing function design. The first item is based on the efficacy feedback, which suppresses excessive adjustment through the exponential decay term; the second item is based on the tolerance and treatment time, which controls the adjustment frequency through the Sigmoid function, and the two cooperate to balance the efficacy and safety.

[0058] It is worth mentioning that: the effect fluctuation tolerance coefficient ξ is set for different diseases, such as hypertension treatment ξ=0.2 is sensitive to blood pressure fluctuation, and chronic disease rehabilitation ξ=0.5 allows greater fluctuation; the patient tolerance parameter is calculated by visual analog scale (VAS) and adverse reaction grading (CTCAE), VAS score ≥4 points is reduced by 0.3; the adjustment interval reference value τt is set according to the treatment type, the chemotherapy scheme =72 hours, the antibiotic scheme =24 hours, to avoid frequent adjustment; the maximum tolerance time Tmax combines patient age and underlying diseases, such as elderly patients with chemotherapy Tmax=6 cycles, and young patients Tmax=8 cycles.

[0059] The technical effects achieved by the above embodiments include: the exponential decay term smooths the adjustment range caused by the deviation of the efficacy, avoids excessive adjustment caused by single efficacy fluctuation, and protects the patient's tolerance; the time interval sensitive function reduces repeated adjustment in a short time, maintains the continuity and stability of the treatment plan; the dynamic combination of tolerance parameter and treatment time makes the plan self-adaptively optimized with the treatment process, avoiding the rigidity of "using one plan to the end"; the overall adjustment strategy is more in line with the clinical step-by-step treatment principle, which reduces adverse reactions while ensuring efficacy, and improves patient treatment compliance and comfort.

[0060] The traditional technical solution has the following technical problems: the human-computer interaction of the existing medical AI system is mainly based on AI decision, and the doctor's participation is low, which makes it difficult to coordinate when the AI and the clinical judgment conflict; the decision process and basis lack visual display, and the doctor's trust in the AI output result is low and unwilling to adopt; the patient's privacy data lacks targeted protection mechanism in the transmission and storage process in the interaction process, and there is a risk of data leakage, especially in the multi-institutional cooperation scene.

[0061] Based on this, please refer to Figure 5 , the human-computer cooperative interaction module includes: a doctor decision fusion unit configured to weight and fuse the treatment plan adjustment strategy generated by AI and the professional judgment of the clinician, wherein the weight of the doctor's suggestion is higher than that of the AI strategy; a visual interactive interface configured to display the risk assessment results, treatment plan adjustment basis and historical decision records in the form of interactive charts and structured reports; and a privacy protection unit configured to use homomorphic encryption and secure multi-party computation technology to ensure the safety and privacy of patient data during cross-institutional transmission and sharing. The scheme solves the problems of decision conflict, lack of trust and privacy risk in human-computer cooperation through the fusion mechanism of doctor-led + AI-assisted, visual trust building and privacy enhancement technology, and realizes the efficient cooperation of AI and doctors.

[0062] It is worth mentioning that: the doctor decision fusion unit adopts a dynamic weighting mechanism, with a doctor weight of 0.6 and an AI weight of 0.4 in regular cases, and a doctor weight of 0.7-0.8 in difficult cases, to ensure the dominant position of professional judgment; the visual interface adopts hierarchical display, with the final plan and key basis displayed on the surface, and the AI reasoning process such as feature importance ranking and risk assessment curve displayed in the deep layer, supporting the doctor to click to view the original data of each conclusion; the privacy protection unit uses full homomorphic encryption (FHE) to process patient identification information, and multi-party computation ensures that the data is "available but invisible" when multiple institutions cooperate, in line with HIPAA and GDPR specifications.

[0063] The technical effects achieved by the above embodiments include: doctor-led weighted fusion balances AI efficiency and clinical experience, reduces decision-making conflicts, and makes the final solution more suitable for the actual situation of the patient; the visual interface improves the understanding and trust of doctors for AI results by visualizing the reasoning process, and promotes the actual landing of AI-assisted decision-making; homomorphic encryption and multi-party computation technology fully protects patient privacy and eliminates data leakage concerns in cross-institutional collaboration; the overall interaction mechanism not only takes advantage of the data processing of AI, but also retains the clinical dominance of doctors, achieving the collaborative goal of "AI assistance rather than replacement", and improving the safety and acceptability of medical decision-making.

[0064] The traditional technical solution has the following technical problems: existing medical AI model deployment relies on cloud servers, data transmission delay is high, especially in unstable network environments such as primary hospitals or emergency scenes, real-time response is not possible; the model is large in size, requiring high terminal device computing power, which is difficult for primary medical institutions to bear; cloud centralized deployment requires data to be uploaded to the central server, increasing the risk of privacy leakage and not meeting the requirement of local storage of medical data.

[0065] Based on this, the system further includes a lightweight model deployment module configured to compress the personalized decision-making model after LoRA fine-tuning to a mobile terminal device through model pruning and knowledge distillation technology, realizing offline operation and edge computing capability of the medical decision-making Agent system, and solving the delay problem and computing resource bottleneck caused by traditional cloud platform dependence. This solution breaks through the limitations of cloud dependence through model compression and edge deployment technology. Pruning technology removes redundant neurons and connections in the model, and knowledge distillation retains key knowledge through a teacher-student model architecture. The combination of the two achieves model lightweight and adapts to mobile computing power.

[0066] It is worth mentioning that: the model pruning adopts a structured pruning strategy, removes the heads of attention weights <0.1 in the Transformer layer, retains 80% of the core feature extraction capability, and compresses the model size to 30% of the original; knowledge distillation uses a cloud-based large model as a teacher and a mobile small model as a student, and through soft label training with a temperature coefficient T=5, the student model improves the inference speed by 3 times under the premise of accuracy loss <5%; the deployment carrier includes a doctor's tablet terminal and a portable diagnosis and treatment device, supporting Android and iOS systems, with a response time <1 second in offline state, meeting the needs of primary and emergency scenes; local storage of patient data and only uploading desensitized model update parameters meet the data sovereignty requirements.

[0067] The technical effects achieved by the above embodiments include: lightweight deployment makes the system free from cloud dependence and can still respond in real time in unstable network scenarios, especially suitable for emergency and primary medical care; model compression reduces the requirement for terminal computing power, so that primary medical institutions can use it without high equipment investment, promoting the popularization of medical AI; offline operation and localized storage reduce data upload and reduce the risk of privacy leakage from the source, in line with medical data security specifications; edge computing capability enables decision-making on the terminal, making the response faster and saving time for emergency patients, thereby improving the scene adaptability and practicality of the medical AI system as a whole.

[0068] The traditional technical solutions have the following technical problems: the existing medical AI implementation method is fragmented, with data collection, model fine-tuning, risk assessment, and scheme generation running independently, lacking a collaborative optimization mechanism, resulting in poor connection between steps and low overall efficiency; model fine-tuning is not customized for the dynamic characteristics of multi-modal data, risk assessment ignores feature interaction, and treatment scheme adjustment lacks dynamic feedback, with limitations of "each for itself" in each link; the deployment method is single and dependent on the cloud, which cannot adapt to the computing power and network conditions of different medical scenarios, limiting the range of technology landing.

[0069] Based on this, please refer to Figure 2 The embodiment provides an implementation method based on the system of any one of the above, comprising the following steps: collecting multi-modal data of a patient in real time, fine-tuning a pre-trained model based on a time decay factor, data importance weight and modality urgency coefficient through a dynamic weight adjustment LoRA fine-tuning module to generate a personalized decision-making model; fusing feature vectors through a multi-modal risk assessment module and quantifying disease risk using a comprehensive evaluation function containing feature interaction terms; based on the risk assessment result, generating a treatment scheme through an adaptive decision-making module combined with treatment effect feedback, patient tolerance and adjustment interval parameters; fusing AI decision and doctor's advice to form the final medical decision; and realizing edge end operation of the system through a lightweight deployment module.

[0070] This method solves the problems of insufficient model generalization, one-sided risk assessment and rigid treatment scheme in the prior art through dynamic weight layering, feature interaction modeling and adaptive adjustment mechanism, improves clinical adaptability while maintaining decision-making accuracy. This scheme forms a closed loop through collaborative design of the whole process, including data collection, model fine-tuning, risk assessment, decision-making, human-computer fusion and lightweight deployment, and each step is dynamically optimized based on the previous result, realizing end-to-end optimization from data to decision.

[0071] It is worth mentioning that: the timeliness and urgency label is performed immediately after multi-modal data collection, providing pre-processing basis for subsequent fine-tuning module; model fine-tuning and risk assessment share feature interaction data, avoiding repeated calculation and improving overall efficiency; after generating the treatment plan, the historical adjustment record is automatically associated to form an "evaluation-adjustment-feedback" closed loop, supporting iterative optimization of the plan; edge deployment and cloud synchronization are parallel, the cloud updates model parameters when the terminal runs offline, and automatically integrates after networking, balancing real-time and updating; each step of the method can be flexibly cut according to the scene, such as skipping part of the non-key test data collection in the emergency scene to prioritize speed.

[0072] The technical effects achieved by the above embodiments include: the whole-process collaborative mechanism eliminates data barriers at each link, improves overall operation efficiency, and significantly shortens the time from data collection to decision output; dynamic weight and feature interaction run through multiple links to ensure consistency of model fine-tuning, risk assessment and plan generation, and improve the coherence of decision logic; the self-adaptive adjustment mechanism makes the treatment plan evolve dynamically with the patient's condition and state, avoiding the mismatch between "static plan" and "dynamic condition"; edge deployment and cloud cooperation expand the application scenarios of the system, which can be adapted to both tertiary hospitals and primary clinics; the overall method not only maintains the high-efficiency data processing capability of AI, but also integrates the professional judgment of clinicians, realizing the deep integration of technology and clinic and promoting the development of medical decision-making towards individualization and dynamics.

[0073] The above is only a preferred embodiment of the present application, and does not limit the form of the present application. Any person skilled in the art can use the disclosed technical content to make changes or modifications to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the present application.

Claims

1. An AI-based LoRA fine-tuning medical decision-making agent system, characterized in that, include: Multimodal data acquisition module: includes real-time acquisition of patients' vital signs data, medical imaging data, laboratory test data, and electronic medical record text data; LoRA fine-tuning module with dynamic weight adjustment: Based on the multimodal data, by introducing time decay factor, data importance weight and modality urgency coefficient, the pre-trained medical decision model is fine-tuned with low-rank adaptation to generate a personalized decision model for the current patient. The multimodal risk assessment module is configured to integrate the feature vectors output by the fine-tuned personalized decision-making model, and combine the attention mechanism and feature interaction coefficients to construct a comprehensive assessment function, thereby achieving dynamic quantitative assessment of the patient's disease risk. The adaptive decision generation module is used to generate an adaptive treatment plan adjustment strategy based on the risk assessment results, combined with treatment effect feedback factors, patient tolerance parameters, and dynamic response coefficients. The human-computer collaborative interaction module is used to integrate the treatment plan adjustment strategy with the professional judgment of clinicians to form the final medical decision recommendation.

2. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, The multimodal data acquisition module includes: a wearable device interface for real-time acquisition of the patient's heart rate, blood pressure, and blood oxygen saturation vital signs data; Medical imaging interface, including access to CT, MRI and ultrasound medical imaging equipment, and parsing image data based on the DICOM standard protocol; Laboratory data interface, integrated with the hospital LIS system, to obtain laboratory test data such as blood routine, biochemical indicators and pathology reports; The Natural Language Processing (NLP) unit is used to perform entity recognition, relation extraction, and semantic representation learning on electronic medical record text data, generating structured medical record feature vectors.

3. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, The LoRA fine-tuning module for dynamic weight adjustment includes: The data timeliness analysis unit is used to calculate the time decay coefficient of each modality of data, where the decay coefficient of recent data is greater than that of long-term data. The data importance assessment unit is configured to assign importance weights to different modalities of data based on clinical guidelines and expert knowledge, with medical imaging data having a higher weight than routine vital signs data. The low-rank matrix decomposition unit is used to perform singular value decomposition on the weight matrix of the fully connected layer of the pre-trained model, extract the core low-rank subspace, and dynamically update the low-rank matrix through the time decay coefficient and data importance weight. The model fusion unit superimposes the updated low-rank matrix with the original model parameters to generate a personalized medical decision-making model.

4. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, The multimodal risk assessment module includes: The modality alignment unit maps feature vectors from different modalities to a unified semantic space through a cross-modal contrastive learning method, thereby eliminating the semantic gap between modalities. The attention mechanism unit is used to calculate the attention weights of each modality feature, where the weight of disease-related features is higher than that of non-relevant features; The risk quantification unit constructs a comprehensive risk scoring system that includes disease severity, complication risk, and treatment response prediction dimensions based on the attention weights and semantically aligned feature vectors. The dynamic threshold adjustment unit adaptively adjusts the threshold range of the risk score based on the patient's underlying health condition and disease progression stage.

5. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, In the LoRA fine-tuning module for dynamic weight adjustment, the time decay factor β and the data importance weight... and modal urgency coefficient The calculation formula is: ； Where t is the time difference between the data acquisition time and the current time. The threshold for the inflection point of time decay. For decay rate parameters, Let be the clinical benchmark importance coefficient of the i-th modality data. Let be the freshness index of the i-th modality, and n be the total number of modality categories. Let be the urgency coefficient of the i-th modality, representing vital signs in an emergency scenario. =0.8-1.0, image data =0.5-0.7, where σ is the urgency coefficient of the current stage of diagnosis and treatment. This represents the final modal dynamic weights.

6. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, In the multimodal risk assessment module, the expression for the comprehensive assessment function is: ； in, Let j be the attention weights for the j-th modality. Let be the independent influence coefficient of the k-th feature. Let be the nonlinear transformation function of the k-th feature in the j-th mode, where m is the number of modes and p is the number of single-mode features. is the interaction coefficient between the k-th and l-th features.

7. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, In the adaptive decision generation module, the calculation formula for the treatment plan adjustment strategy is as follows: ； in, The treatment regimen adjustment amount is given by γ, the treatment effect feedback factor is given by E, and the current efficacy indicator is given by E. ξ represents the expected efficacy indicator, δ represents the tolerance coefficient for efficacy fluctuations, T represents the patient tolerance parameter, and T represents the duration of treatment. For maximum tolerance time, The interval between the two adjustments. To adjust the interval reference value, is the interval sensitivity coefficient.

8. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, The human-computer collaborative interaction module includes: The physician decision fusion unit weights and integrates the treatment plan adjustment strategies generated by AI with the decision suggestions of clinicians, with the physician suggestions having a higher weight than the AI ​​strategies. A visual interactive interface displays risk assessment results, the basis for treatment plan adjustments, and historical decision records in the form of interactive charts and structured reports; The privacy protection unit employs homomorphic encryption and secure multi-party computation technology to ensure the security and privacy of patient data during cross-institutional transmission and sharing.

9. The AI-based LoRA fine-tuning medical decision-making agent system according to claim 1, characterized in that, It also includes a lightweight model deployment module, which uses model pruning and knowledge distillation techniques to compress the LoRA-tuned personalized decision-making model to mobile devices.

10. An implementation method, applied to an AI-based LoRA fine-tuning medical decision-making agent system as described in any one of claims 1-9, characterized in that, include: Real-time acquisition of patients' multimodal data; through the LoRA fine-tuning module with dynamic weight adjustment, the pre-trained model is fine-tuned based on time decay factor, data importance weight and modality urgency coefficient to generate a personalized decision model. By fusing feature vectors through a multimodal risk assessment module, the risk of disease is quantified using a comprehensive assessment function that includes feature interaction terms; Based on the risk assessment results, a treatment plan is generated by the adaptive decision generation module, which combines treatment effect feedback, patient tolerance, and adjustment interval parameters. The system integrates AI decision-making with doctor recommendations to form the final medical decision; it operates at the edge through lightweight deployment modules.

Images