A method and system for enhancing speech recognition for patients with chronic diseases

Abstract

This application relates to the field of voice interaction technology, and in particular to a voice recognition enhancement method and system for patients with chronic diseases. The method includes: acquiring a voice information set of target Alzheimer's patients; analyzing the ambiguous semantics and potential intentions of the voice information under disease-related language impairments based on the voice information set to obtain an intention hypothesis information set; acquiring an interaction flow dataset; analyzing the joint confidence of each intention based on the interaction flow dataset and the intention hypothesis information set, and proceeding to direct intention confirmation or multimodal assisted intention confirmation according to the joint confidence to obtain a deterministic intention information set; analyzing the decay information of semantic mapping relationships based on the deterministic intention information set, updating the personalized dynamic semantic map corresponding to the patient, retrieving the personalized dynamic semantic map based on the patient's real-time voice information, generating and outputting care warning information, and recording a voice recognition enhancement log. This improves the patient's quality of life and safety.

Description

Technical Field

[0001] This application relates to the field of voice interaction technology, and in particular to a voice recognition enhancement method and system for patients with chronic diseases. Background Technology

[0002] Alzheimer's disease is a common neurodegenerative disease, and patients often have progressive language impairments, such as difficulty finding words, vague references, repetitive sentences, and semantic generalization, which makes their speech expression significantly vague and uncertain.

[0003] In existing technologies, general speech recognition is usually designed based on precise keyword matching and standardized grammar. It is difficult to effectively parse unstructured, low-definition speech content caused by diseases. It often produces a lot of errors or fails to output effective text during the recognition stage, resulting in failure to understand subsequent intent. Summary of the Invention

[0004] This application provides a speech recognition enhancement method and system for patients with chronic diseases to solve the above-mentioned problems.

[0005] In a first aspect, this application provides a speech recognition enhancement method for patients with chronic diseases. The method includes: acquiring a speech information set of a target Alzheimer's disease patient; analyzing the ambiguous semantics and potential intentions of the speech information under disease-related language disorders based on the speech information set to obtain an intention hypothesis information set; acquiring an interaction process dataset; analyzing the joint confidence of each intention based on the interaction process dataset and the intention hypothesis information set, and proceeding to direct intention confirmation or multimodal assisted intention confirmation according to the joint confidence to obtain a deterministic intention information set; analyzing the decay information of semantic mapping relationships based on the deterministic intention information set, updating the personalized dynamic semantic map corresponding to the patient, retrieving the personalized dynamic semantic map according to the patient's real-time speech information, generating and outputting care warning information, and recording a speech recognition enhancement log.

[0006] Through the above technical solutions, disease-related language impairments in Alzheimer's patients can be accurately adapted, significantly improving the accuracy of identifying ambiguous semantics and potential intentions. Personalized dynamic semantic maps can dynamically match individual patient differences and disease progression, ensuring long-term recognition stability. Caregiver alerts enable caregivers to respond to patient needs in a timely manner and avoid risks. Logs can also provide data support for disease assessment, effectively reducing the burden of care and improving patients' quality of life and safety.

[0007] Optionally, the voice information set includes object referential information and demand expression information; based on the object referential information, the matching relationship between the fuzzy referential expression corresponding to the object referential information and the historical high-frequency associated objects is analyzed to obtain an object candidate set; based on the demand expression information, the association and derivation relationship between the broad demand expression corresponding to the demand expression information and the specific demands expressed in the patient's historical interactions, as well as the demand state, is analyzed using knowledge graph analysis technology to obtain a demand type set; the object candidate set and the demand type set are combined to construct multiple intent possibility hypotheses representing the object-demand combination relationship, and the initial probability score is dynamically adjusted for each intent possibility hypothesis to form the intent hypothesis information set.

[0008] Optionally, based on the aforementioned multiple intent probability assumptions and adhering to the principle of minimizing user privacy, differential privacy technology is employed to analyze the identifier information associated with object information and patient information. This identifier information is then filtered and replaced in real-time to obtain interaction context information. Based on this interaction context information, knowledge graph analysis technology is used to analyze successful interaction fragments in similar historical contexts, filtering out portions associated with the current object candidate set and the demand type set to obtain a minimum data call set. Based on this minimum data call set, a weighted scoring mechanism is used for dynamic adjustment. Specifically, for intent assumptions with a success rate higher than a preset threshold in similar historical contexts, the probability is increased proportionally to the success rate; for intent assumptions with a historical confusion frequency higher than a preset threshold, the probability is adjusted downwards based on the number of confusion occurrences to obtain the initial probability score.

[0009] Optionally, the interaction process dataset includes a historical dialogue path library and a multimodal auxiliary record library. Based on the historical dialogue path library and the intent hypothesis information set, the successful clarification paths associated with each intent possibility hypothesis and adapted to the patient's cognitive characteristics are analyzed to obtain path association information. Based on the multimodal auxiliary record library and the path association information, the impact of different clarification paths on the patient's emotions in the current interaction environment is analyzed to obtain path evaluation information. Based on the path evaluation information, the execution order and multimodal feedback intensity of each clarification path are dynamically optimized according to the patient's real-time attention focus to obtain a patient state adaptation scheme set. Based on the patient state adaptation scheme set, the joint confidence of each intent and the corresponding recommendation confirmation process are analyzed, and the process proceeds to the direct intent confirmation or multimodal assisted intent confirmation. The confirmation results are then integrated to obtain the deterministic intent information set.

[0010] Optionally, based on the multimodal auxiliary record library, emotion analysis technology is used to analyze auxiliary record fragments related to the patient's emotional fluctuations in historical interactions to obtain an emotion-related record set; based on the emotion-related record set, combined with the path-related information, behavioral sequence pattern analysis technology is used to analyze the frequency and context of the patient's resistance or confusion feedback in the historical record fragments associated with each clarification path to obtain a path historical risk profile; based on the path historical risk profile, according to the patient's initial emotional state captured in the current interaction environment, the emotional fluctuations caused by each clarification path in the current environment are analyzed to obtain a real-time emotion impact prediction; based on the real-time emotion impact prediction, each clarification path is graded to obtain the path evaluation information; the grade of each clarification path is positively correlated with the patient's emotional state.

[0011] Optionally, based on the path evaluation information and combined with the real-time attention focus, the correlation between the presented content of each clarification path and the patient's current attention focus is analyzed using focus matching analysis technology based on attention mechanisms to obtain focus matching information. Based on the focus matching information and combined with the path historical risk profile, the clarification paths are distinguished into paths with stable historical feedback and paths that have caused emotional fluctuations in the patient through differentiated control technology. Based on this distinction, a priority execution order is assigned to paths with stable historical feedback, while differentiated feedback forms and presentation rhythms are designed for paths that have caused emotional fluctuations in the patient, resulting in a path-level execution strategy. Based on the path-level execution strategy, the impact of executing different strategies on the patient's cognitive state is analyzed, and a personalized plan matching the patient's current state is generated, resulting in the patient state adaptation plan set.

[0012] Optionally, based on the personalized scheme, the co-occurrence frequency changes and offset directions of semantic relationships successfully established in the current interaction and those successfully mapped in the past are analyzed in terms of content and structure to obtain semantic relationship change information; based on the semantic relationship change information, according to the patient's disease progression information, the rate difference characteristics of semantic relationship stability changes under different disease stages are analyzed to obtain disease stage-related decline trajectory information; based on the decline trajectory information, the decline information of semantic mapping relationships revealed by the trajectory information and its impact on the overall map structure are analyzed to obtain a semantic map update strategy; based on the semantic map update strategy, the mapping weights and associated structures of the personalized dynamic semantic map are adjusted, the updated map is searched and matched according to the patient's real-time voice information, the care warning information corresponding to the decline trajectory information is generated and output, and the speech recognition enhancement log is recorded.

[0013] Optionally, based on the aforementioned degradation trajectory information, the trend of decreased referential accuracy and increased generalization of demand expression in the semantic relationships expressed by patients is analyzed using correlation time-series analysis technology to obtain semantic co-degradation information. Based on the semantic co-degradation information, the key semantic mapping set and the secondary semantic associations that can tolerate ambiguity are deduced as necessary for patients to maintain basic communication effectiveness through communication utility evaluation rules, resulting in a map maintenance information set. Based on the map maintenance information set, differentiated operation instructions are integrated to weight and enhance the key semantic mapping set and to selectively simplify or suspend the maintenance of the secondary semantic associations, generating a personalized map update scheme that matches the patient's current cognitive function level, as the semantic map update strategy.

[0014] Optionally, based on the weighted enhancement instruction, the semantic relationships successfully and accurately expressed and confirmed by the patient in recent interactions are analyzed. In the personalized dynamic semantic map, the connection weights of the object referential information and the demand expression information corresponding to the semantic relationships are enhanced to obtain a reinforced core mapping network. Based on the selective simplification or maintenance pause instruction, combined with the semantic collaborative degradation information, the semantic connections in the secondary semantic associations that have not been successfully invoked for a long time due to the ambiguity of patient referentials and the aggravation of demand generalization are located. The dynamic maintenance and weight update of the semantic connections are paused, and the corresponding semantic connections are removed from the high-frequency retrieval paths to obtain a simplified and optimized auxiliary mapping network. The reinforced core mapping network and the simplified and optimized auxiliary mapping network are merged to complete the directional adjustment of the personalized dynamic semantic map.

[0015] Secondly, this application provides a speech recognition enhancement system for patients with chronic diseases. The system includes: an intent hypothesis module, used to acquire a speech information set of a target Alzheimer's disease patient, and based on the speech information set, analyze the fuzzy semantics and potential intents of the speech information under disease-related language impairments to obtain an intent hypothesis information set; an intent confirmation module, used to acquire an interaction process dataset, and based on the interaction process dataset and the intent hypothesis information set, analyze the joint confidence of each intent, and proceed to direct intent confirmation or multimodal assisted intent confirmation according to the joint confidence, to obtain a deterministic intent information set; and a mapping retrieval module, used to analyze the decay information of semantic mapping relationships based on the deterministic intent information set, update the personalized dynamic semantic map corresponding to the patient, and retrieve the personalized dynamic semantic map according to the patient's real-time speech information, generate and output care warning information, and record a speech recognition enhancement log. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of an application scenario provided in an embodiment of this application;

[0018] Figure 2 A flowchart illustrating a speech recognition enhancement method for patients with chronic diseases, provided as an embodiment of this application;

[0019] Figure 3 This is a schematic diagram of a speech recognition enhancement system for patients with chronic diseases, provided as an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0021] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0022] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0023] In the process of caring for patients with Alzheimer's disease, current speech recognition technology is mainly geared towards clear and standardized expressions. It lacks a specific design for the ambiguous and fragmented speech of Alzheimer's patients caused by cognitive decline. It is difficult to extract effective intentions from these speech forms, and the lack of a confidence mechanism leads to frequent interaction failures and a high rate of intention misjudgment. Furthermore, it cannot adapt to the dynamic changes in the patient's condition and evolve in a personalized manner, thus making it difficult to provide stable and reliable long-term support.

[0024] Based on this, this application provides a speech recognition enhancement method and system for patients with chronic diseases, designed specifically for the language impairments of Alzheimer's patients. It can accurately understand ambiguous semantics and potential intentions, and through personalized dynamic semantic maps, it adapts in real time to individual patient differences and changes in their condition, ensuring long-term stable recognition. The system provides care alerts to facilitate timely response to needs and risk prevention, and automatically generates care logs to provide a basis for condition assessment, thereby reducing caregiving burden and improving patient safety and quality of life.

[0025] Figure 1 This application provides an illustration of an application scenario. In the process of caring for patients with Alzheimer's disease, the method provided in this application addresses the communication difficulties of Alzheimer's patients, accurately interpreting their true intentions by penetrating the fog of speech. It can construct a personalized, evolving semantic network for users, closely following changes in individual status and disease progression to ensure continuous and effective comprehension. Its proactive early warning mechanism helps caregivers anticipate needs and mitigate risks, while continuous data recording provides strong support for evaluating treatment effectiveness and adjusting treatment plans, ultimately achieving a dual improvement in nursing burden and patient quality of life.

[0026] Specifically, the method provided in this application can be applied to any server. The server interacts with intelligent audio devices and intelligent interactive platforms to obtain a set of voice information provided by the intelligent audio devices and a dataset of interaction processes provided by the intelligent interactive platforms. This accurately adapts to the disease-related language impairments of Alzheimer's patients, significantly improving the accuracy of recognizing ambiguous semantics and potential intentions. It generates and outputs caregiving warning information to family members and records enhanced voice recognition logs for audio device administrators, effectively reducing the burden of caregiving and improving the patient's quality of life and safety. Specific implementation methods can be found in the following embodiments.

[0027] Figure 2 This is a flowchart illustrating a speech recognition enhancement method for patients with chronic diseases, provided as an embodiment of this application. The method of this embodiment can be applied to servers in the above-described scenarios. Figure 2 As shown, the method includes:

[0028] S201. Obtain the speech information set of target patients with Alzheimer's disease. Based on the speech information set, analyze the ambiguous semantics and potential intentions of the speech information under disease-related language disorders to obtain the intention hypothesis information set.

[0029] The speech information set can be a collection of speech recordings generated by Alzheimer's patients in their daily lives or specific care scenarios, with the data source being a smart audio device worn by the patient. Ambiguous semantics can be the semantic content of speech segments that cannot be accurately interpreted by conventional speech recognition systems due to the decline in language function caused by the disease. Latent intent can be the true needs or purposes that the patient wants to express, hidden behind ambiguous speech expressions. The intent hypothesis information set can be a set of possible intents generated based on the analysis of ambiguity semantics.

[0030] Specifically, patients with chronic diseases like Alzheimer's often experience language impairments, with significant ambiguity and uncertainty in their speech. This makes it difficult to directly apply general speech recognition technology and effectively serve the daily communication and care needs of patients. To address this issue, we specifically collect and analyze speech information sets from target patients, and use natural language processing and machine learning to proactively analyze the ambiguous semantics caused by the disease and uncover the underlying intentions, thereby forming an intention hypothesis information set.

[0031] S202. Obtain the interaction process dataset. Based on the interaction process dataset and the intent hypothesis information set, analyze the joint confidence of each intent. Then, based on the joint confidence, proceed to direct intent confirmation or multimodal assisted intent confirmation to obtain a deterministic intent information set.

[0032] The interaction process dataset can be a collection of data covering common scenarios in chronic disease care, including processes, questions and answers, and historical successful interaction cases, with the intelligent interaction platform as the data source. Joint confidence can be a combination of the probability of the intent hypothesis itself and the degree of matching between that intent and the current interaction process context and historical patient behavior. Direct intent confirmation can be the direct adoption of an intent hypothesis as a deterministic intent when the joint confidence exceeds a preset high threshold. Multimodal assisted intent confirmation can trigger an additional confirmation mechanism when the joint confidence of the intent hypothesis is insufficient. The deterministic intent information set can be a set of highly reliable patient intents obtained after the confirmation process.

[0033] Specifically, in the process of caring for patients with Alzheimer's disease, the intent assumptions derived from a single voice analysis may be ambiguous or lack confidence, and direct use may lead to misjudgment. Therefore, an interaction process dataset is introduced as prior knowledge and contextual constraints. By integrating domain knowledge (interaction process) with real-time analysis results (intent assumptions), a feedback and verification closed loop is formed, which improves the robustness and scene adaptability of intent recognition. It solves the problem of how to make reliable decisions under fuzzy input and is the core link from "guessing" to "confirmation".

[0034] S203. Based on the deterministic intent information set, analyze the decay information of semantic mapping relationship, update the personalized dynamic semantic map corresponding to the patient, and retrieve the personalized dynamic semantic map according to the patient's real-time voice information, generate and output care warning information, and record speech recognition enhancement log.

[0035] Semantic mapping decay information can be the weakening or disordered trajectory of a patient's memory of the association between specific words, concepts, and their corresponding entities or needs. A personalized dynamic semantic map can be a semantic knowledge network stored in a graph structure, specific to a particular patient. Caregiver alert information can be prompts generated by interpreting real-time speech to identify potential risks or unmet needs. A speech recognition augmentation log can record the complete processing flow, intermediate results, and decision-making basis from the original speech input to the final alert generation.

[0036] Specifically, the cognitive state and semantic abilities of Alzheimer's patients are constantly changing, and a static semantic model cannot be applied in the long term. By analyzing the confirmed intentions and their expression, it is possible to continuously track the degeneration or variation patterns of the patient's semantic network and update their personalized model accordingly. This allows for a deeper understanding of the patient at the current stage, improving the accuracy of long-term services. At the same time, by using this dynamically updated semantic map to retrieve and match real-time speech, it is possible to more sensitively detect subtle expression abnormalities or urgent needs, thereby generating proactive care warnings.

[0037] The method provided in this embodiment accurately adapts to the disease-related language impairments of Alzheimer's patients, significantly improving the accuracy of recognizing ambiguous semantics and potential intentions. The personalized dynamic semantic map can dynamically match individual patient differences and disease progression, ensuring long-term recognition stability. Caregiver alerts enable caregivers to respond to patient needs in a timely manner and avoid risks. Logs can also provide data support for disease assessment, effectively reducing the caregiver burden and improving the patient's quality of life and safety.

[0038] In some embodiments, the voice information set includes object referential information and demand expression information. Based on the object referential information, the matching relationship between the fuzzy referential expression corresponding to the object referential information and the historical high-frequency associated objects is analyzed to obtain an object candidate set. Based on the demand expression information, the association and derivation relationship between the broad demand expression corresponding to the demand expression information and the specific demands expressed in the patient's historical interactions, as well as the demand states, are analyzed using knowledge graph analysis technology to obtain a demand type set. The object candidate set and the demand type set are combined to construct multiple intent possibility hypotheses representing the object-demand combination relationship, and the initial probability score is dynamically adjusted for each intent possibility hypothesis to form an intent hypothesis information set.

[0039] Object referential information can be vague descriptions of people, objects, scenes, etc., mentioned in the patient's speech. Need expression information can be the content related to the needs expressed in the patient's speech. The object candidate set can be a set of several objects that may match the patient's referential intent, selected through contextual matching and association analysis of the object referential information. The need type set can be a set of several specific need categories corresponding to the need expression information, representing a precise classification of broad needs. The intent probability hypothesis can be various intent conjectures representing the "object-need" correspondence. The initial probability score can be used to quantify the credibility of different intent hypotheses.

[0040] Specifically, patients with chronic diseases such as Alzheimer's often experience language impairments such as aphasia and difficulty finding words, leading to numerous problems in their speech expressions, including vague references (e.g., inability to name specific objects) and generalized needs (e.g., only being able to express discomfort without specifying a particular need). This poses a significant challenge to existing speech recognition systems based on precise keyword matching, often resulting in interaction failures or misunderstandings due to the inability to understand these vague expressions. For example, the system might misinterpret the patient's reference to "that" as an irrelevant object, or categorize the generalized need of "discomfort" into a single type, failing to meet the patient's true intentions. To address these issues, the system first uses natural language processing technology to analyze the speech information, separating object reference information (e.g., the pronoun "he") from need expression information (e.g., the phrase "feeling panicked"). For object referential information, a combination of contextual similarity calculation and personal historical high-frequency object retrieval is used for analysis. For example, when a patient says "bring that," the semantic similarity between "that" and entities already mentioned in the current dialogue window (such as "water cup," "medicine bottle," and "remote control") is calculated. Simultaneously, the patient's interaction logs for the past month are queried to count frequently associated objects (e.g., "water cup" appears 50 times, "remote control" appears 10 times). The results from both are weighted and fused to generate a probabilistic object candidate set such as {"water cup": 0.7, "medicine bottle": 0.2, "remote control": 0.1}. For demand expression information, a personalized knowledge graph constructed for the patient is invoked for analysis. This graph uses specific demands recorded in historical interactions (e.g., "want to drink water," "feeling dizzy and need to rest," "palpitations and need to check blood pressure") as nodes, and uses co-occurrence relationships and semantic similarity... Similarity is represented by edges. When the input "feeling anxious" is entered, the system retrieves the associated nodes in the graph and derives a set of demand types based on the weights of the associated edges, such as {"measuring blood pressure": 0.6, "taking heart-calming medication": 0.3, "deep breathing to relax": 0.1}. Subsequently, the two sets are combined using a Cartesian product to generate diverse intent probability hypotheses such as "measuring blood pressure with a water cup" and "taking heart-calming medication with a medicine bottle". Finally, dynamic probability scoring is performed based on the patient's historical interaction data: for example, if the proportion of successful confirmation of "measuring blood pressure with a water cup" in historically similar contexts is high (e.g., over 80%), its score is weighted upwards; if "taking heart-calming medication with a medicine bottle" has frequently been confused with "applying ointment to the skin" (e.g., the number of confusions reaches 5 times), its score is corrected downwards, and finally, a structured intent hypothesis information set with probability weights is output.

[0041] The method provided in this embodiment accurately analyzes ambiguous semantics and comprehensively mines potential intentions, significantly improving the accuracy of speech recognition and the completeness of intent understanding. At the same time, it incorporates a privacy protection mechanism to ensure the security of patients' sensitive information and provides a reliable basis for subsequent care warnings. Its dynamic quantitative scoring and personalized adaptation logic make care services more timely and accurate, effectively improving the patient care experience and providing strong technical support for voice interaction and intelligent care for patients with chronic diseases.

[0042] In some embodiments, based on multiple intent probability assumptions and following the principle of minimizing user privacy, differential privacy technology is used to analyze the identification information associated with object information and patient information, and to filter and replace the identification information in real time to obtain interaction context information. Based on the interaction context information, knowledge graph analysis technology is used to analyze successful interaction fragments in similar historical contexts, and to filter out the parts associated with the current object candidate set and demand type set to obtain the minimum data call set. Based on the minimum data call set, a weighted scoring mechanism is used for dynamic adjustment, specifically: for intent assumptions with a success rate higher than a preset threshold in similar historical contexts, the probability is increased by weighting according to the success rate ratio; for intent assumptions with a historical confusion frequency higher than a preset threshold, the probability is adjusted downward according to the number of confusion frequencies to obtain an initial probability score.

[0043] The principle of minimizing user privacy can be applied during the voice intent probability adjustment process by collecting, using, and exposing as little user privacy information as possible, acquiring only the core data necessary for probability adjustment, and avoiding the leakage of irrelevant privacy information. Differential privacy technology can be a technical means to protect data privacy. Identification information can be information associated with object information and patient information, capable of uniquely identifying the patient's identity, health status, or specific interaction scenario. Interaction context information can be a set of non-privacy data, such as interaction scenarios and historical related content, that have been processed by differential privacy technology, removing sensitive identification information and retaining only those related to the current intent hypothesis. Knowledge graph analysis technology can be a technical method for mining data relationships. Successful interaction fragments in historically similar contexts can be past interaction records that are similar to the current patient's object candidate set and demand type set, and ultimately accurately identify the patient's true intent. The minimum data call set can be a minimal subset of data highly related to the current object candidate set and demand type set, selected from historical interaction data. The weighted scoring mechanism can be a calculation rule that sets corresponding weight coefficients based on the performance of different intent hypotheses in historical interaction data, and adjusts the initial probability score upwards or downwards. The success rate preset threshold can be a pre-set critical value used to judge the credibility of the intent hypothesis. The confusion frequency preset threshold can be a pre-set critical value used to judge the degree of confusion of the intent hypothesis.

[0044] Specifically, the speech of Alzheimer's patients has problems such as ambiguous object reference and generalized expression of needs. Fixed initial probability scores are prone to misjudgment of intent, affecting the accuracy of subsequent intent confirmation. When accessing historical data, if privacy protection measures are not taken, sensitive patient information may be leaked, violating medical data security regulations. Furthermore, unfiltered historical data may introduce irrelevant interference, making it impossible to achieve personalized probability adjustment. Ultimately, this leads to deviations in semantic map updates, failure of care warnings, and inability to respond to patients' real needs in a timely manner. To address the aforementioned issues: First, to uphold privacy commitments, differential privacy technology is employed to add or generalize noise in real-time to identifying information (such as the patient's specific term of endearment for their child, "Abao") and patterns that can be reversibly used to infer specific embarrassing events in the current voice recording and subsequent retrieved personal history logs, generating desensitized interaction context information. Then, based on this information, a personalized knowledge graph analysis technique is used to map the current ambiguous reference (such as "that round one") to a patient-specific entity node in the graph (associated with their "favorite ceramic teacup"). Using this as a guide, a limited number of records involving the same node and marked as successful interaction segments (such as conversations where the teacup was successfully accessed three times in the past week) are quickly retrieved and filtered from the patient's entire personal history, forming a highly refined... The system first obtains the minimum data call set. Then, based on this dedicated dataset, the system dynamically adjusts the score using a weighted scoring mechanism. This mechanism uses a preset threshold learned from the patient's long-term data (e.g., the average success rate threshold for the "drinking water" related intent is 0.9). For hypotheses with a success rate higher than this threshold in the call set (e.g., the success rate for "wanting to drink tea" is 0.95), the probability is increased by a weighted percentage (e.g., 5%). At the same time, for hypotheses where the patient's historical confusion frequency is higher than another preset threshold (e.g., 0.3) (e.g., the confusion frequency between "back itching" and "back pain" is 0.4), the probability is adjusted downwards by a corresponding amount based on the specific confusion frequency (e.g., confusion 4 times in the last 10 times). This results in a final initial probability score that deeply integrates the patient's personal lifestyle habits, cognitive characteristics, and privacy protection considerations.

[0045] The approach provided in this embodiment not only protects patient privacy through differential privacy technology, complying with medical data security standards, but also enhances the credibility of intent hypothesis scoring by leveraging precisely selected reference data and a dynamic weighting mechanism, laying a solid foundation for subsequent steps. At the same time, it optimizes the accuracy and efficiency of speech recognition, helping to respond quickly to patient needs, providing support for personalized care alerts, and effectively improving the quality of patient voice interaction and care services.

[0046] In some embodiments, the interaction process dataset includes a historical dialogue path library and a multimodal auxiliary record library. Based on the historical dialogue path library and combined with the intent hypothesis information set, the successful clarification paths associated with each intent possibility hypothesis and adapted to the patient's cognitive characteristics are analyzed to obtain path association information. Based on the multimodal auxiliary record library and combined with the path association information, the impact of different clarification paths on the patient's emotions in the current interaction environment is analyzed to obtain path evaluation information. Based on the path evaluation information, the execution order and multimodal feedback intensity of each clarification path are dynamically optimized according to the patient's real-time attention focus to obtain a patient state adaptation scheme set. Based on the patient state adaptation scheme set, the joint confidence of each intent and the corresponding recommended confirmation process are analyzed, and direct intent confirmation or multimodal assisted intent confirmation is initiated. The confirmation results are then integrated to obtain a deterministic intent information set. The historical dialogue path library can be a database storing the dialogue logic, steps, and methods of the patient's past successful clarification of ambiguous intents. The multimodal auxiliary record library can be a database recording information about the patient's use of multiple modalities to assist in expressing intent in past interactions. The path association information can be information related to the successful clarification paths associated with each intent possibility hypothesis and adapted to the patient's cognitive characteristics. Path assessment information can be the result of a graded assessment of the potential impact of each clarification path on the patient's emotions in the current interactive environment. Real-time attention focus can be the object, topic, or information that the patient is focusing their attention on at the current moment of interaction. Dynamic optimization can be the process of adjusting the execution order of clarification paths and the intensity of multimodal feedback in real time based on the patient's real-time state. The patient state adaptation scheme set can be a collection of optimized clarification path execution strategies and multimodal feedback schemes based on the patient's current cognitive state, emotional state, and attention focus.

[0047] Specifically, in the process of voice interaction with Alzheimer's patients, if there is a lack of integrated analysis of the interaction process dataset and the intent hypothesis information set, if the clarification path is not adapted to the patient's cognitive characteristics, and if the impact of the path on the patient's emotions is not assessed, it is very easy to cause intent confirmation bias, which will cause the patient to become irritable and resistant. At the same time, the confirmation process without dynamic optimization will exacerbate communication barriers, make it impossible to accurately obtain definite intent, and thus affect subsequent semantic map updates and care warnings, seriously reducing the effectiveness of interaction and care quality. To address the aforementioned issues: First, pattern matching technology is used to compare the intent hypothesis information set (e.g., hypotheses generated from the voice segment "that...thing...": {"Want to drink water": probability 0.6, "Want to watch TV": probability 0.3}) with a historical dialogue path database. For each hypothesis, specific dialogue sequences used in similar historical contexts to successfully clarify the intent are associated, forming path association information (e.g., a historical successful path associated with "want to drink water": first ask the general category "Do you need something?", and after the patient nods, show pictures of "water cup and pills" for selection). Next, a multimodal auxiliary recording database is invoked, employing sentiment analysis technology (e.g., analyzing recordings associated with the historical path execution to detect whether the patient's tone is calm and the speaking speed is appropriate) and behavioral sequence pattern analysis technology (e.g., counting the historical frequency of prolonged silence or head shaking after the "show picture" step in the path) to comprehensively assess the historical risk profile of each associated path. Then, combined with the patient's initial emotional state captured by the camera (e.g., whether the face is relaxed), the path's performance in the current environment is predicted. The system analyzes real-time emotional influences to generate path assessment information. Based on this, it accesses the patient's attention focus data in real time (e.g., monitoring the patient's gaze at a table via an eye tracker). Using focus matching analysis technology, it calculates the correlation between the content of each candidate path (e.g., a picture of a "water glass") and the current focus, and performs differentiated adjustments based on the path assessment information: For paths with high focus matching and low historical risk (e.g., the patient is looking at a table where there is often a water glass), it assigns high priority and prepares for direct intent confirmation (e.g., directly asking, "Do you want the water glass on the table?"); For paths with mismatched focus or assessment of emotional risk, it designs buffer schemes, such as first gently calling the patient's name to attract attention, and then using a gentle speech and simplified vocabulary with pictures for multimodal intent confirmation. Finally, through a weighted model, it integrates the initial probability of intent, path matching degree, and emotional risk prediction to calculate the joint confidence of each intent. Based on this confidence and risk threshold, it automatically branches to the corresponding confirmation process and integrates the patient's multimodal feedback in multiple rounds of interaction, such as voice and gestures (e.g., pointing), and finally outputs an accurate and patient-participated set of deterministic intent information.

[0048] The method provided in this embodiment achieves precise matching between intent confirmation and the patient's cognitive and emotional state, significantly reducing misunderstandings of intent and negative emotions in patients; the generated deterministic intent information set provides reliable support for personalized dynamic semantic map updates, ensuring the accuracy of care warnings; it not only improves the efficiency and humanistic care of voice interaction, but also provides thoughtful technical support for daily patient care, helping to optimize the care experience and the patient's quality of life.

[0049] In some embodiments, based on a multimodal auxiliary record library, emotion analysis technology is used to analyze auxiliary record fragments related to the patient's emotional fluctuations in historical interactions to obtain an emotion-associated record set. Based on the emotion-associated record set, combined with path association information, behavioral sequence pattern analysis technology is used to analyze the frequency and context of the patient's resistance or confusion feedback in the historical record fragments associated with each clarification path to obtain a path historical risk profile. Based on the path historical risk profile, according to the patient's initial emotional state captured in the current interaction environment, the emotional fluctuations triggered by each clarification path in the current environment are analyzed to obtain a real-time emotion impact prediction. Based on the real-time emotion impact prediction, each clarification path is graded to obtain path assessment information. The grade of each clarification path is positively correlated with the patient's emotional state.

[0050] Sentiment analysis techniques can extract, identify, and analyze emotional information from interaction data, uncovering patterns in emotional associations to support path risk assessment. Emotional association record sets can provide auxiliary record fragments directly related to patient emotional fluctuations. Behavioral sequence pattern analysis techniques can uncover the correlation patterns and frequency regularities between patient feedback behavior and clarification paths. Path historical risk profiles can characterize the emotional risk features of each clarification path. The current interaction environment can provide the objective scenario and patient state background during the interaction. The patient's initial emotional state can provide the patient's emotional state at the initial stage of the interaction. Real-time emotional impact prediction can predict the degree of emotional fluctuation triggered by each clarification path in the current environment.

[0051] Specifically, in the process of confirming the intent of voice interaction for Alzheimer's patients, if there is a lack of scientific patient state adaptation schemes, a uniform clarification path will be directly used. This may lead to inefficient intent confirmation due to the mismatch between the path and the patient's real-time attention focus. It may also exacerbate patient resistance or confusion by reusing paths that have previously caused emotional fluctuations, thus disrupting the continuity of interaction. Furthermore, it may ignore individual cognitive differences among patients, resulting in a lack of personalized adaptation, which seriously affects the practicality and humanistic adaptability of speech recognition enhancement methods. To address the aforementioned issues: First, sentiment analysis technology is used to deeply mine the multimodal auxiliary record library. This library stores multi-dimensional data from historical interactions, including videos (capturing facial expressions and gestures), audio (analyzing tone of voice changes), and interaction logs (recording response delays). This technology scans and identifies all segments associated with significant emotional fluctuations. For example, it automatically marks historical moments when a patient's volume suddenly increases, accompanied by frowning and waving gestures, forming an emotional association record set. Next, combined with the path association information generated in the current stage (i.e., each alternative clarification path, such as "reiterate confirmation," "provide image options," and "simplify questions"), behavioral sequence pattern analysis technology is used to perform pattern mining on this record set. This analyzes whether a specific negative feedback pattern frequently appears in the patient's behavioral sequence after each path is executed. For example, analysis reveals that "providing multiple (e.g., more than four)" The "Complex Image Options" path, in its historical execution records, has a high probability (for example) of triggering a "confusion-resistance" sequence, where the patient stares at the screen silently for a long time, followed by shaking their head, especially when executed in the evening. The "Repetition Confirmation" path, however, rarely triggers this sequence. Therefore, a quantitative historical risk profile is generated for each path, clarifying its risk context and frequency. Then, the patient's initial emotional state is captured in real-time using sensors such as cameras. For example, analyzing current facial micro-expressions and voice features determines the patient to be "mildly anxious." Based on this, the current emotional state is simulated and matched with the historical risk profiles of each path to generate real-time emotional impact predictions. For example, it is predicted that in the current "mildly anxious" state, executing the aforementioned "Complex Image Options" path has a high probability of escalating anxiety to "moderate confusion," while executing the "Repetition Confirmation" path has minimal impact. Finally, based on these predictions, all alternative clarification paths are ranked and graded according to the estimated risk level of triggering negative emotions (e.g., low, medium, high), producing final path evaluation information that directly guides the selection of the optimal interaction path with the least emotional load.

[0052] The method provided in this embodiment accurately matches the patient's attention focus with the optimized path execution strategy, effectively improving the efficiency and accuracy of intent confirmation, avoiding patient emotional fluctuations caused by mismatched paths, ensuring smooth and harmonious interaction, and at the same time, the personalized solution fully fits the patient's cognitive and emotional state, reflecting humanistic care, making the speech recognition enhancement method more suitable for the special needs of Alzheimer's patients, and laying a solid foundation for subsequent efficient interaction and care alerts.

[0053] In some embodiments, based on path assessment information and combined with real-time attention focus, focus matching analysis technology based on attention mechanisms is used to analyze the correlation between the content presented in each clarification path and the patient's current attention focus, thus obtaining focus matching information. Based on focus matching information and combined with path historical risk profiles, differentiated control technology is used to distinguish between paths with stable historical feedback and paths that have caused emotional fluctuations in patients. Based on this distinction, a priority execution order is assigned to paths with stable historical feedback, while differentiated feedback forms and presentation rhythms are designed for paths that have caused emotional fluctuations in patients, thus obtaining a path-level execution strategy. Based on the path-level execution strategy, the impact of executing different strategies on the patient's cognitive state is analyzed, and a personalized plan matching the patient's current state is generated, thus obtaining a patient state-adapted plan set.

[0054] Attention mechanism focus matching analysis can be used to analyze the close correlation between the content presented in the clarification path and the patient's current focus. Differential modulation technology can be used to implement targeted modulation for different types of clarification paths. A clarification path with stable historical feedback refers to one where, in historical interactions, the patient has shown positive feedback, no resistance or confusion, and a success rate higher than a preset threshold. A clarification path that has triggered emotional fluctuations in the patient refers to one where, in historical interactions, the patient has exhibited resistance or confusion such as frowning, silence, or refusal to respond, and the frequency of such occurrences is higher than a preset threshold. A path-level execution strategy can be a comprehensive strategy developed based on path type differences, including execution order, feedback format, and presentation rhythm. The patient's cognitive state can be the duration of their attention, information comprehension ability, and logical reasoning ability during real-time interaction. A personalized solution can be a clarification path execution plan highly adapted to the patient's current focus of attention, emotional state, and cognitive level.

[0055] Specifically, in the process of confirming intent during interactions with Alzheimer's patients, the lack of a patient-state adaptation solution set construction stage can directly lead to chaotic execution of clarification paths. Paths that have previously caused emotional fluctuations in patients may be presented in a disordered manner, which can easily aggravate patients' resistance. Paths that are out of sync with the patient's real-time focus of attention can distract them, reduce comprehension efficiency, and even interrupt interaction. Furthermore, paths that are not adapted to the patient's cognitive state may lead to errors in intent confirmation, delay caregiver responses, and exacerbate communication difficulties caused by the patient's language impairment. To address the aforementioned issues: First, relying on the device's built-in multi-microphone array and inertial sensors, real-time fusion analysis of the patient's head orientation, speech energy intensity focus, and environmental sound sources is used to non-invasively infer their real-time attention focus. For example, if the device detects that the patient's head is continuously tilted to the right and their speech energy is concentrated in that direction, combined with environmental sound analysis identifying a continuous sound of flowing water on the right, it can be inferred that their attention focus may be related to "drinking water." Subsequently, focus matching analysis technology based on attention mechanisms is used to vectorize and match the speech script content of the clarification path with the inferred focus semantics. For example, the script "Do you want to drink water?" in path A has a high similarity to the focus vector of "drinking water," while the similarity in path B "Do you want to go for a walk?" is low, thus generating quantified focus matching information. Simultaneously, the device locally retrieves the patient's path history risk profile, which records, for example, that "path A (asking at a relatively fast pace) caused the patient to become agitated three times in the afternoon." Based on the risk pattern of "repetitive phrases," differentiated control technology is implemented under the constraint of pure voice interaction: for paths with high matching degree and low historical risk, the device will prioritize their execution and directly broadcast them with synthesized speech at a moderate speed and steady tone; for paths with high matching degree but also high historical risk, differentiated voice feedback design is initiated. For example, the same question (path A) is broken down into first broadcasting a gentle environmental guidance sound (such as a brief and soothing sound of running water), waiting for a pause of about two seconds, and then broadcasting the core question "Do you... want... water...?" at a significantly reduced speed (e.g., half the normal speed) and in a gentler tone. Finally, by combining the matching degree, risk level, and voice control parameters of all paths, a complete set of pure voice interaction schemes that best matches the patient's current auditory attention and emotional tolerance is generated, thereby achieving compliant and fluent intent confirmation on a wearable device without visual interference.

[0056] The method provided in this embodiment accurately adapts to the patient's state, effectively avoiding emotional resistance and interaction interruptions caused by inappropriate clarification paths, ensuring the smoothness and accuracy of intent confirmation. The personalized strategy aligns with the patient's cognitive and attentional characteristics, reducing communication difficulties and minimizing the adverse psychological impact of the interaction on the patient. At the same time, it provides a scientific basis for subsequent intent confirmation, strengthens the adaptability of the speech recognition enhancement method, provides reliable support for semantic map updates and care warnings, and improves the quality and humanistic care level of voice interaction services for patients with chronic diseases.

[0057] In some embodiments, based on a personalized scheme, the co-occurrence frequency changes and offset directions of semantic relationships successfully established in the current interaction and those successfully mapped in the past are analyzed in terms of content and structure to obtain semantic relationship change information. Based on the semantic relationship change information, according to the patient's disease progression information, the rate difference characteristics of semantic relationship stability changes under different disease stages are analyzed to obtain disease stage-related decline trajectory information. Based on the decline trajectory information, the decline information of semantic mapping relationships revealed by the trajectory information and its impact on the overall map structure are analyzed to obtain a semantic map update strategy. Based on the semantic map update strategy, the mapping weights and associated structures of the personalized dynamic semantic map are adjusted, the updated map is retrieved and matched according to the patient's real-time voice information, care warning information corresponding to the decline trajectory information is generated and output, and a speech recognition enhancement log is recorded.

[0058] Semantic relationship change information can be information related to changes in the co-occurrence frequency and offset direction of semantic relationships successfully established in the current interaction and those successfully mapped in the past, in terms of content and structure. Disease progression information can be information reflecting the disease stage and progression rate of Alzheimer's patients. Semantic relationship stability can be the degree of stability of the corresponding associations between patients' semantics. Decline trajectory information can be information that combines the patient's disease stage to present the differences in the rate of change of semantic relationship stability, reflecting the continuity of the decline trend of semantic mapping relationships. Semantic map update strategy can be a specific scheme for adjusting mapping weights and optimizing association structure based on decline trajectory information for personalized dynamic semantic maps. Mapping weight can be a quantitative representation of the strength of association between different semantics in personalized dynamic semantic maps, with higher weights corresponding to tighter associations. Association structure can be the organizational structure formed by various semantics according to association logic in personalized dynamic semantic maps, which determines the matching path of semantic retrieval.

[0059] Specifically, in the daily interactions of Alzheimer's patients, if semantic mapping decay information is not analyzed and personalized dynamic semantic maps are not updated, the maps will fail to adapt to changes in the patient's cognition, leading to frequent misjudgments in real-time voice retrieval and missed critical needs such as medication and water intake. Furthermore, the inability to trace semantic decay trajectories makes it difficult to optimize recognition models and care plans, resulting in care oversights, exacerbating patient discomfort, and severely impacting care quality and patient safety. To address these issues: First, co-occurrence frequency analysis and graph structure comparison techniques are used to compare newly established semantic relationships with historical success records stored in the patient's "personalized dynamic semantic map." For example, analysis reveals that the association weight between "table lamp" and its entity objects has recently shown a downward trend, while the frequency of access to the association edge between "apple" and the "food" category has significantly decreased. Simultaneously, the word "water" has begun to frequently form new weak associations with the vague demand term "that." The quantitative changes in the co-occurrence frequency of these contents, together with the qualitative shifts in the map structure, constitute "semantic relationship change information." Next, "disease progression information" (such as MMSE staging from clinical assessment) is introduced from external input. Using time series analysis and trajectory fitting modeling techniques, this change information is placed on the disease timeline. For example, the analysis reveals that after the disease enters the intermediate stage, the average rate of weakening of core noun referential relationships accelerates several times compared to the earlier stage, while the density of abstract demand generalization significantly increases. This constructs a quantitative and staged "decline trajectory information." Then, based on this trajectory, the impact of these decline patterns on the overall network structure of the existing semantic map is assessed through influence propagation analysis and a strategy inference engine. For example, it is determined that the loosening of the association between core nodes related to "drinking water" may indicate a weakening of the underlying semantic relationship. This physiological need expression will enter a high-risk confusion period, thus automatically generating a "semantic map update strategy" containing specific instructions such as "prioritizing the strengthening of the weights between related nodes" and "adding multiple fuzzy expression mappings for needs such as 'thirst'". Finally, this strategy is executed to complete the targeted update of the semantic map. When the patient speaks real-time speech again, the updated map is used for retrieval and pattern matching. If multiple consecutive failures to understand core words marked as "rapidly declining" are detected, the warning generation module is immediately triggered, outputting a "caregiver warning message" such as "the patient's ability to refer to everyday object nouns is showing an accelerated decline trend, and it is recommended to use more visual aids for communication". All analyses, decisions, actions and results in this cycle are completely recorded in the "speech recognition enhancement log" in a structured format, forming a closed loop.

[0060] The method provided in this embodiment accurately captures the semantic decline trend of patients, allowing personalized dynamic semantic maps to continuously adapt to cognitive changes, improving the accuracy of real-time voice retrieval and avoiding misjudgment of intent; the generated caregiver warnings provide caregivers with clear needs and status references, helping to meet needs in a timely manner and predict the condition; log records provide data support for tracing the decline process and optimizing the technology, improving the patient interaction experience, reducing caregiver pressure, and promoting more humanized speech recognition technology for chronic diseases.

[0061] In some embodiments, based on the degradation trajectory information, the trend of decreased referential accuracy and increased generalization of demand expression in the semantic relationships expressed by patients is analyzed using correlation time-series analysis technology to obtain semantic co-degradation information; based on the semantic co-degradation information, the key semantic mapping set and the secondary semantic associations that can tolerate ambiguity are deduced as necessary for patients to maintain basic communication effectiveness through communication utility evaluation rules to obtain a map maintenance information set; based on the map maintenance information set, differentiated operation instructions are integrated to weight and enhance the key semantic mapping set and to selectively simplify or suspend the maintenance of the secondary semantic associations to generate a personalized map update scheme that matches the patient's current cognitive function level, as the semantic map update strategy.

[0062] Semantic co-degeneration information can characterize the simultaneous development trend of decreased referential accuracy and increased generalization of needs expression in patients' semantic relationships. Communication utility assessment rules can be pre-defined standard rules used to judge the impact of semantic mapping relationships on the effectiveness of patient communication. The key semantic mapping set can be the combination of semantic mappings that must be prioritized to maintain the effectiveness of basic patient communication (such as semantic associations related to daily diet, living arrangements, and emergency needs). Secondary semantic associations can be semantic mappings that have a smaller impact on basic patient communication and can tolerate a certain degree of ambiguity (such as semantic associations related to the names of infrequently used items and descriptions of complex scenarios). The map maintenance information set can be derived based on semantic co-degeneration information and communication utility assessment rules, containing classification information of the key semantic mapping set and secondary semantic associations. Differentiated operation instructions can be different maintenance instructions formulated for the key semantic mapping set and secondary semantic associations. Personalized map update schemes can be schemes adapted to the patient's current cognitive function level after integrating differentiated operation instructions; they are the specific presentation form of the semantic map update strategy.

[0063] Specifically, in the voice interaction process of Alzheimer's patients, their semantic mapping relationship deteriorates with the progression of the disease. Without a targeted update strategy, core semantic associations (such as diet and emergency assistance) weaken, leading to misjudgments of key needs. Secondary semantics consume resources, reducing recognition efficiency, and a uniform update logic cannot adapt to individual differences in deterioration, exacerbating communication barriers, affecting the accuracy of caregiving warnings, and severely delaying responses to patient needs. To address these issues: First, a natural language processing module is invoked to mine the patient's recently stored "deterministic intent information set" and corresponding original voice text. Then, a correlation time-series analysis technique—such as using a sliding time window and a probabilistic statistical model—is employed to calculate... The algorithm calculates the slope of the co-change between the accuracy of referential indicators (such as the percentage of successful identification of specific nouns) and the generalization of needs indicators (such as the proportion of abstract words in expressions) of patients within a continuous time unit (e.g., in "weeks"), thereby quantifying and extracting "semantic co-degradation information." For example, it finds that the accuracy of patients' referential expression of "water cup" has decreased by a certain percentage over the past four weeks, while the frequency of generalized needs expressions such as "help me" has increased by a certain percentage, and the decreasing and increasing trends of the two are statistically significant. Subsequently, it accesses a knowledge base of communication utility assessment rules trained by clinical language therapy rules and historical successful interaction data. This rule base is usually based on decision trees or weighted scoring systems. For example, if a semantic mapping relationship involves core survival safety terms such as "water," "medicine," "pain," and "falling," and its number of successful calls and confirmations in the recent past (e.g., within a month) exceeds a certain threshold (e.g., 6 times), it is classified as a "key semantic mapping set." Conversely, if the mapping involves specific details such as "newspaper brand" or "TV series character name," and the frequency of triggering clarification dialogues is high (e.g., exceeding a certain percentage) or the number of consecutive failures is high, it is classified as a "minor semantic association with tolerable ambiguity." Then, based on this classification result and the current map state, the strategy generation engine automatically generates "differentiated operation instructions" using instruction template filling and parameterization techniques. The instruction is as follows: For mappings in the key set (such as "I-Pain-Help"), the instruction is to increase its connection weight value in the semantic graph database by a specific amount (such as increasing a certain coefficient) and mark its priority as "highest" in the retrieval index; for secondary associations (such as "remote control-TV" and "remote control-air conditioner" are merged because patients are now collectively referred to as "remote control"), the instruction is to perform node merging and link simplification, or to set its "dynamic maintenance flag" to "pause" to stop the update of its weight decay algorithm. Finally, the "personalized map update scheme" that integrates all instructions is serialized and passed to the map update service for execution to complete this strategic maintenance.

[0064] The method provided in this embodiment accurately adapts to the semantic decay pattern of patients, prioritizes the stability of core semantic mapping, and ensures accurate identification of key needs; it simplifies secondary semantic associations, improves speech recognition efficiency, synchronizes semantic maps with patient cognition, provides precise support for intent recognition and care warnings, optimizes doctor-patient / nurse-patient communication experience, and can also indirectly monitor disease progression, providing a basis for personalized care.

[0065] In some embodiments, based on the weighted enhancement instruction, semantic relationships successfully and accurately expressed and confirmed by the patient in recent interactions are analyzed. In the personalized dynamic semantic map, the connection weights of the object referential information and the demand expression information corresponding to the semantic relationships are enhanced to obtain a reinforced core mapping network. Based on the selective simplification or maintenance suspension instruction, combined with the semantic co-degradation information, semantic connections that have not been successfully invoked for a long time due to the ambiguity of patient referentials and the aggravation of demand generalization are located in the secondary semantic associations. The dynamic maintenance and weight update of the semantic connections are suspended, and the corresponding semantic connections are removed from the high-frequency retrieval paths to obtain a simplified and optimized auxiliary mapping network. The reinforced core mapping network and the simplified and optimized auxiliary mapping network are merged to complete the directional adjustment of the personalized dynamic semantic map.

[0066] Connection weights can be numerical values ​​representing the strength of semantic relationships in a personalized dynamic semantic map. The core mapping network can be a core semantic association network composed of semantic relationships that the patient has recently successfully and accurately expressed and confirmed. Selective simplification instructions can be operational instructions that simplify secondary semantic associations to avoid redundant semantic interference. Pause maintenance instructions can be operational instructions that stop dynamically updating secondary semantic connections that have not been successfully invoked for a long time. Semantic connections can be the association links between object references and demand expressions in a personalized dynamic semantic map. Dynamic maintenance can be the process of continuously updating the effectiveness and weights of semantic connections based on the patient's real-time voice interaction. Weight updates can be operations that adjust the semantic connection strength values ​​based on patient interaction feedback. High-frequency retrieval paths can be semantic connection links with high retrieval frequency in a personalized dynamic semantic map, prioritized for matching the patient's real-time voice. Auxiliary mapping networks can be auxiliary semantic networks composed of secondary semantic associations, supplementing core communication needs. Targeted adjustments can be operations that, based on the semantic map update strategy, specifically adjust the mapping weights and association structures in the personalized dynamic semantic map.

[0067] Specifically, during voice interaction with Alzheimer's patients, if the personalized dynamic semantic map is not adjusted, the core semantic connections will be weakened due to lack of reinforcement, and secondary semantic connections that have not been used for a long time will form redundant interference. This will lead to disordered voice retrieval priorities, making it impossible to quickly locate the patient's core intent, reducing recognition accuracy and efficiency, and also causing care warning information to become disconnected from actual needs, exacerbating patient communication barriers and affecting the effectiveness of personalized care. To address the aforementioned issues: First, the instruction set generated by the semantic map update strategy is analyzed. For example, a weighted boosting instruction might require strengthening the mapping between "thirst" and "water cup." Then, the interaction log analysis module is invoked to retrieve all recent (e.g., within the past week) successful interaction fragments. Pattern matching technology is used to identify semantic relationship instances clearly expressed and confirmed by the patient (e.g., the patient repeatedly points to the water cup and says "water," and the caregiver responds successfully each time). For each identified successful instance, a weight boosting algorithm is applied. Based on the instance's recent frequency and the degree of confirmation, a weight increment is calculated and added to the connection weight between the corresponding concept nodes ("thirst" node and "water cup" node) in the personalized dynamic semantic map. After batch processing of a batch of such relationships, a strengthened and consolidated core mapping network is constructed. Simultaneously, for secondary associations that the strategy requires to be simplified or suspended, such as a historical... For historically established but recently untriggered "remote control - watching TV" mappings, the system initiates call frequency analysis and a time decay model to comprehensively evaluate the number of times the connection has been called over a longer period (e.g., within a month) and the most recent call timestamp. If the evaluation results show that its activity is below a preset threshold (e.g., dozens of consecutive interactions have not been triggered), the system performs a pause maintenance operation: it stops the periodic update of the connection weight and removes it from the high-frequency retrieval index of the real-time intent recognition engine, but still retains it in the background archive, thus forming a streamlined and optimized auxiliary mapping network. Finally, through graph fusion and rewiring algorithms, the two networks are merged to ensure that high-weight connections in the core network occupy the central path in the merged map topology, while dormant connections are moved to the edge. This ultimately completes a directional and non-uniform adjustment of the semantic map structure and weights, making it more accurately reflect the patient's current communication patterns and capability boundaries.

[0068] The method provided in this embodiment can strengthen the high-frequency core semantic associations of patients, improve the accuracy of retrieval and matching, simplify invalid and redundant information, improve map retrieval efficiency, enable the semantic map to dynamically adapt to the patient's cognitive decline trajectory, provide reliable support for real-time voice retrieval and care warning, optimize the voice recognition enhancement effect, ensure the patient's basic communication needs, help with precise and personalized care, and improve the patient's interactive experience.

[0069] Figure 3A schematic diagram of the structure of a speech recognition enhancement system for patients with chronic diseases provided in one embodiment of this application is shown below. Figure 3 As shown, a speech recognition enhancement system 300 for patients with chronic diseases in this embodiment includes: an intent hypothesis module 301, an intent confirmation module 302, and a mapping retrieval module 303.

[0070] The intent hypothesis module 301 is used to acquire a set of speech information of the target Alzheimer's patients, and based on the speech information set, analyze the fuzzy semantics and potential intent of the speech information under disease-related language impairment to obtain an intent hypothesis information set; the intent confirmation module 302 is used to acquire an interaction process dataset, and based on the interaction process dataset, combined with the intent hypothesis information set, analyze the joint confidence of each intent, and enter direct intent confirmation or multimodal assisted intent confirmation according to the joint confidence to obtain a deterministic intent information set; the mapping retrieval module 303 is used to analyze the decay information of semantic mapping relationship based on the deterministic intent information set, update the personalized dynamic semantic map corresponding to the patient, and retrieve the personalized dynamic semantic map according to the patient's real-time speech information, generate and output care warning information, and record speech recognition enhancement logs.

[0071] Optionally, when the intent hypothesis module 301 analyzes the fuzzy semantics and potential intents of the speech information under disease-related language disorders based on the speech information set to obtain the intent hypothesis information set, it is specifically used for: the speech information set including object referential information and demand expression information; based on the object referential information, analyzing the matching relationship between the fuzzy referential expression corresponding to the object referential information in the current interaction context and historical high-frequency associated objects to obtain an object candidate set; based on the demand expression information, using knowledge graph analysis technology, analyzing the association and derivation relationship between the broad demand expression corresponding to the demand expression information and the specific demands expressed in the patient's historical interactions, as well as the demand state, to obtain a demand type set; combining the object candidate set and the demand type set to construct multiple intent possibility hypotheses representing the object-demand combination relationship, and dynamically adjusting the initial probability score for each intent possibility hypothesis to form the intent hypothesis information set.

[0072] Optionally, when the intent hypothesis module 301 dynamically adjusts the initial probability score for each intent possibility hypothesis, it is specifically used to: based on the multiple intent possibility hypotheses, following the principle of minimizing user privacy, using differential privacy technology, analyze the identification information associated with object information and patient information, and filter and replace the identification information in real time to obtain interaction context information; based on the interaction context information, using knowledge graph analysis technology, analyze successful interaction fragments in similar historical contexts, and filter out the parts associated with the current object candidate set and the demand type set to obtain a minimum data call set; based on the minimum data call set, use a weighted scoring mechanism for dynamic adjustment, specifically: for intent hypotheses with a success rate higher than a preset threshold in similar historical contexts, increase the probability upward according to the success rate ratio; for intent hypotheses with a historical confusion frequency higher than a preset threshold, adjust the probability downward according to the number of confusion frequencies to obtain the initial probability score.

[0073] Optionally, when the intent confirmation module 302 analyzes the joint confidence of each intent based on the interaction process dataset and the intent hypothesis information set, and enters direct intent confirmation or multimodal assisted intent confirmation according to the joint confidence to obtain a deterministic intent information set, it is specifically used for the following: the interaction process dataset includes a historical dialogue path library and a multimodal assisted record library; based on the historical dialogue path library and the intent hypothesis information set, it analyzes the past successful clarification paths associated with each intent possibility hypothesis and adapted to the patient's cognitive characteristics to obtain path association information; based on the multimodal assisted record library and the path association information, it analyzes the impact of different clarification paths on the patient's emotions in the current interaction environment to obtain path evaluation information; based on the path evaluation information, it dynamically optimizes the execution order and multimodal feedback intensity of each clarification path according to the patient's real-time attention focus to obtain a patient state adaptation scheme set; based on the patient state adaptation scheme set, it analyzes the joint confidence of each intent and the corresponding recommended confirmation process, enters the direct intent confirmation or multimodal assisted intent confirmation, and integrates the confirmation results to obtain the deterministic intent information set.

[0074] Optionally, the intent confirmation module 302, during the construction of the path assessment information, is specifically used for: based on the multimodal auxiliary record library, using sentiment analysis technology, analyzing auxiliary record fragments related to the patient's emotional fluctuations in historical interactions to obtain an emotion-related record set; based on the emotion-related record set, combined with the path-related information, using behavioral sequence pattern analysis technology, analyzing the frequency and context of the patient's resistance or confusion feedback in the historical record fragments associated with each clarification path to obtain a path historical risk profile; based on the path historical risk profile, analyzing the emotional fluctuations caused by each clarification path in the current environment according to the patient's initial emotional state captured in the current interaction environment to obtain a real-time emotional impact prediction; based on the real-time emotional impact prediction, classifying each clarification path to obtain the path assessment information; the level of each clarification path is positively correlated with the patient's emotional state.

[0075] Optionally, the intent confirmation module 302, during the construction of the patient state adaptation scheme set, is specifically used for: based on the path evaluation information, combined with the real-time attention focus, and according to the focus matching analysis technology of the attention mechanism, analyzing the degree of correlation between the presentation content of each clarification path and the patient's current attention focus to obtain focus matching information; based on the focus matching information, combined with the path historical risk profile, and through differentiated control technology, distinguishing the clarification paths into paths with stable historical feedback and paths that have caused emotional fluctuations in the patient, and according to this distinction, assigning a priority execution order to paths with stable historical feedback, while designing differentiated feedback forms and presentation rhythms for paths that have caused emotional fluctuations in the patient, to obtain a path hierarchical execution strategy; based on the path hierarchical execution strategy, analyzing the impact of executing different strategies on the patient's cognitive state, generating a personalized scheme that matches the patient's current state, to obtain the patient state adaptation scheme set.

[0076] Optionally, when the mapping retrieval module 303 analyzes the decay information of semantic mapping relationships based on the deterministic intent information set, updates the personalized dynamic semantic map corresponding to the patient, and retrieves the personalized dynamic semantic map according to the patient's real-time voice information, generates and outputs care warning information, and records the speech recognition enhancement log, it is specifically used for: analyzing the co-occurrence frequency changes and offset directions of semantic relationships successfully established in the current interaction and historically successfully mapped semantic relationships in terms of content and structure based on the personalized scheme, to obtain semantic relationship change information; based on the semantic relationship change information, analyzing the rate difference characteristics of semantic relationship stability changes under different disease stages according to the patient's disease progression information, to obtain decay trajectory information associated with disease stages; based on the decay trajectory information, analyzing the decay information of semantic mapping relationships revealed by the trajectory information and its impact on the overall structure of the map, to obtain a semantic map update strategy; based on the semantic map update strategy, performing directional adjustments to the mapping weights and associated structures of the personalized dynamic semantic map, retrieving and matching the updated map according to the patient's real-time voice information, generating and outputting the care warning information corresponding to the decay trajectory information, and recording the speech recognition enhancement log.

[0077] Optionally, the mapping retrieval module 303, during the construction process of the semantic map update strategy, is specifically used to: based on the decay trajectory information, analyze the trend of decreased referential accuracy and increased generalization of demand expression in the semantic relationships expressed by the patient using correlation time series analysis technology, to obtain semantic co-degradation information; based on the semantic co-degradation information, deduce the key semantic mapping set and the secondary semantic associations that can tolerate ambiguity, which must be prioritized to ensure the patient's basic communication effectiveness, using communication utility evaluation rules, to obtain a map maintenance information set; based on the map maintenance information set, integrate differentiated operation instructions that weight and enhance the key semantic mapping set and selectively simplify or suspend the maintenance of the secondary semantic associations, to generate a personalized map update scheme that matches the patient's current cognitive function level, as the semantic map update strategy.

[0078] Optionally, when performing the directional adjustment of the mapping weights and association structures of the personalized dynamic semantic map, the mapping retrieval module 303 is specifically used for: based on the weighted enhancement instruction, analyzing the semantic relationships that have been successfully and accurately expressed and confirmed by the patient in recent interactions; enhancing the connection weights of the object reference information and the demand expression information corresponding to the semantic relationships in the personalized dynamic semantic map to obtain a reinforced core mapping network; based on the selective simplification or the pause maintenance instruction, combined with the semantic collaborative degradation information, locating semantic connections in the secondary semantic associations that have not been successfully invoked for a long time due to the ambiguity of patient reference and the aggravation of demand generalization; pausing the dynamic maintenance and weight update of the semantic connections; and removing the corresponding semantic connections from the high-frequency retrieval paths to obtain a simplified and optimized auxiliary mapping network; and merging the reinforced core mapping network and the simplified and optimized auxiliary mapping network to complete the directional adjustment of the personalized dynamic semantic map.

[0079] The system in this embodiment can be used to execute the methods of any of the above embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

Claims

1. A speech recognition enhancement method for patients with chronic diseases, characterized in that, include: Acquire a set of speech information from target patients with Alzheimer's disease, and based on the speech information set, analyze the ambiguous semantics and potential intentions of the speech information under disease-related language disorders to obtain an intention hypothesis information set; Obtain the interaction process dataset, and based on the interaction process dataset and the intent hypothesis information set, analyze the joint confidence of each intent, and proceed to direct intent confirmation or multimodal assisted intent confirmation according to the joint confidence, to obtain a deterministic intent information set; The joint confidence score combines the probability of the intent hypothesis itself with the degree of matching between the corresponding intent and the current interaction context and historical patient behavior. Based on the deterministic intent information set, the decay information of semantic mapping relationship is analyzed, the personalized dynamic semantic map corresponding to the patient is updated, and the personalized dynamic semantic map is retrieved according to the patient's real-time voice information to generate and output care warning information, and record speech recognition enhancement logs. The decay information of the semantic mapping relationship is the weakening or disordered trajectory of the patient's memory of the association between specific words, concepts and their corresponding entities or needs. The personalized dynamic semantic map is a semantic knowledge network stored in a graph structure and belonging to a specific patient. Based on the interaction process dataset and the intent hypothesis information set, the joint confidence of each intent is analyzed, and based on the joint confidence, direct intent confirmation or multimodal assisted intent confirmation is performed to obtain a deterministic intent information set, including: The interaction process dataset includes a historical dialogue path library and a multimodal auxiliary record library; Based on the historical dialogue path database and the intent hypothesis information set, the successful clarification paths associated with each intent possibility hypothesis and adapted to the patient's cognitive characteristics are analyzed to obtain path association information. Based on the multimodal auxiliary recording library and combined with the path association information, the impact of different clarification paths on the patient's emotions in the current interactive environment is analyzed to obtain path evaluation information; Based on the path evaluation information, the execution order and multimodal feedback intensity of each clarification path are dynamically optimized according to the patient's real-time attention focus to obtain a set of patient state adaptation schemes. Based on the patient state adaptation scheme set, the joint confidence of each intention and the corresponding recommendation confirmation process are analyzed, and the direct intention confirmation or multimodal assisted intention confirmation is entered. The confirmation results are then integrated to obtain the deterministic intention information set.

2. The method according to claim 1, characterized in that, Based on the speech information set, the ambiguous semantics and potential intentions of the speech information under disease-related language disorders are analyzed to obtain an intention hypothesis information set, including: The voice information set includes object referencing information and demand expression information; Based on the object referential information, the matching relationship between the fuzzy referential descriptions corresponding to the object referential information and the historical high-frequency associated objects is analyzed to obtain a candidate set of objects. Based on the aforementioned demand expression information, knowledge graph analysis technology is used to analyze the association and derivation relationship between the broad demand expression corresponding to the demand expression information and the specific demands expressed in the patient's historical interactions, as well as the demand status, to obtain a set of demand types. The candidate set of objects is combined with the set of demand types to construct multiple possible intention hypotheses representing the object-demand combination relationship, and the initial probability score is dynamically adjusted for each of the possible intention hypotheses to form the intention hypothesis information set.

3. The method according to claim 2, characterized in that, The initial probability score is dynamically adjusted for each said intention probability assumption, including: Based on the aforementioned multiple intent possibilities and following the principle of minimizing user privacy, differential privacy technology is employed to analyze the identifier information associated with object information and patient information, and to filter and replace the identifier information in real time to obtain interaction context information. Based on the interaction context information, knowledge graph analysis technology is used to analyze successful interaction fragments in similar contexts in history, and the parts associated with the current object candidate set and the demand type set are selected to obtain the minimum data call set; Based on the minimum data call set, a weighted scoring mechanism is used for dynamic adjustment. Specifically, for intent hypotheses with a success rate higher than a preset threshold in historical similar contexts, the probability is increased by weighting according to the success rate ratio. For intent hypotheses with a historical confusion frequency higher than a preset threshold, the probability is adjusted downward according to the number of confusion frequencies to obtain the initial probability score.

4. The method according to claim 3, characterized in that, The process of constructing the path evaluation information includes: Based on the aforementioned multimodal auxiliary record library, emotion analysis technology is used to analyze auxiliary record fragments related to the patient's emotional fluctuations in historical interactions, thereby obtaining an emotion-related record set; Based on the emotional association record set and the path association information, the frequency and context of patient resistance or confusion feedback in the historical record fragments associated with each clarification path are analyzed using behavioral sequence pattern analysis technology to obtain a path history risk profile. Based on the historical risk profile of the path, and according to the patient's initial emotional state captured in the current interactive environment, the analysis of the emotional fluctuations caused by each clarification path in the current environment is performed to obtain a real-time emotional impact prediction. Based on the real-time emotion impact prediction, each clarification path is graded to obtain the path evaluation information. The levels of each clarification pathway are positively correlated with the patient's emotional state.

5. The method according to claim 4, characterized in that, The process of constructing the patient state adaptation scheme set includes: Based on the path evaluation information and the real-time attention focus, the correlation between the content presented in each clarification path and the patient's current attention focus is analyzed using focus matching analysis technology of attention mechanism to obtain focus matching information; Based on the focus matching information and the path historical risk profile, the clarification path is divided into paths with stable historical feedback and paths that have caused emotional fluctuations in patients through differentiated control technology. Based on this distinction, a priority execution order is assigned to paths with stable historical feedback, while differentiated feedback forms and presentation rhythms are designed for paths that have caused emotional fluctuations in patients, resulting in a path hierarchical execution strategy. Based on the path-level execution strategy, the impact of executing different strategies on the patient's cognitive state is analyzed, and a personalized solution matching the patient's current state is generated, resulting in the patient state adaptation solution set.

6. The method according to claim 5, characterized in that, Based on the deterministic intent information set, the decay information of semantic mapping relationships is analyzed, the personalized dynamic semantic map corresponding to the patient is updated, and the personalized dynamic semantic map is retrieved according to the patient's real-time voice information to generate and output care warning information, and record speech recognition enhancement logs, including: Based on the aforementioned personalization scheme, the co-occurrence frequency changes and offset directions of semantic relationships successfully established in the current interaction and those successfully mapped in the past are analyzed in terms of content and structure to obtain semantic relationship change information. Based on the semantic relationship change information, and according to the patient's disease progression information, the rate difference characteristics of semantic relationship stability change under different disease stages are analyzed to obtain the decline trajectory information associated with the disease stages. Based on the decay trajectory information, the decay information of the semantic mapping relationship revealed by the trajectory information and its impact on the overall map structure are analyzed to obtain a semantic map update strategy. Based on the semantic map update strategy, the mapping weights and association structures of the personalized dynamic semantic map are adjusted in a targeted manner. The updated map is searched and matched according to the patient's real-time voice information, and the care warning information corresponding to the decline trajectory information is generated and output. The voice recognition enhancement log is also recorded.

7. The method according to claim 6, characterized in that, The process of constructing the semantic map update strategy includes: Based on the aforementioned degradation trajectory information, the trend of decreased referential accuracy and increased generalization of demand expression in the semantic relationships of patient statements is analyzed using correlation time-series analysis technology to obtain semantic collaborative degradation information. Based on the aforementioned semantic co-degradation information, and through communication utility evaluation rules, the key semantic mapping set that must be prioritized to ensure patients' basic communication effectiveness and the secondary semantic associations that can tolerate ambiguity are deduced, thus obtaining the map maintenance information set. Based on the map maintenance information set, differentiated operation instructions are integrated to weight and enhance the key semantic mapping set and to selectively simplify or suspend the maintenance of the secondary semantic associations, thereby generating a personalized map update scheme that matches the patient's current cognitive function level, which serves as the semantic map update strategy.

8. The method according to claim 7, characterized in that, The process of performing targeted adjustments to the mapping weights and association structures of the personalized dynamic semantic map includes: Based on the weighted enhancement instructions, the semantic relationships that were successfully and accurately expressed and confirmed by the patient in recent interactions are analyzed. In the personalized dynamic semantic map, the connection weights of the object referential information and the demand expression information corresponding to the semantic relationships are enhanced to obtain a reinforced and consolidated core mapping network. Based on the selective simplification or maintenance suspension instructions, and combined with the semantic collaborative degradation information, the semantic connections that have not been successfully invoked for a long time due to the ambiguity of patient reference and the aggravation of demand generalization are located in the secondary semantic associations. The dynamic maintenance and weight update of the semantic connections are suspended, and the corresponding semantic connections are removed from the high-frequency retrieval paths to obtain a simplified and optimized auxiliary mapping network. The reinforced core mapping network is merged with the simplified and optimized auxiliary mapping network to complete the orientation adjustment of the personalized dynamic semantic map.

9. A speech recognition enhancement system for patients with chronic diseases, characterized in that, The method applied to any one of claims 1-8 includes: The intent hypothesis module is used to acquire a set of speech information of target patients with Alzheimer's disease, and based on the speech information set, analyze the ambiguous semantics and potential intent of the speech information under disease-related language disorders to obtain an intent hypothesis information set. The intent confirmation module is used to acquire the interaction process dataset, analyze the joint confidence of each intent based on the interaction process dataset and the intent hypothesis information set, and proceed to direct intent confirmation or multimodal assisted intent confirmation according to the joint confidence, so as to obtain a deterministic intent information set. The mapping retrieval module is used to analyze the decay information of semantic mapping relationships based on the deterministic intent information set, update the personalized dynamic semantic map corresponding to the patient, retrieve the personalized dynamic semantic map according to the patient's real-time voice information, generate and output care warning information, and record speech recognition enhancement logs.

Images