Health question and answer method of AI voice assistant and smart watch

Abstract

The application provides a health question and answer method of an AI voice assistant and a smart watch, and the method comprises the following steps: S1, accepting health description input of a user; S2, mapping the health description of the user to a medical concept library to generate a standardized symptom description; S3, inputting the standardized symptom description into a probability association layer, performing dynamic probability flow propagation calculation based on a pre-constructed medical knowledge graph to generate an associated symptom result and a probability of a corresponding disease hypothesis; and S4, combining user portrait data to correct the probability in a portrait adaptation layer. The application changes fuzzy description into a standardized symptom, combines medical knowledge graph reasoning disease probability and optimizes the result according to the user portrait, solves the problem of low question and answer accuracy caused by semantic ambiguity, improves the accuracy of personalized health suggestions, and avoids the disadvantages of traditional systems, such as the influence of forced accurate description on experience and the lack of help of generalized output answers.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a health question-and-answer method for an AI voice assistant and an intelligent hand. Background Technology

[0002] In health consultation scenarios, users often describe their health problems using vague, colloquial natural language, such as unstructured expressions like "my stomach hurts" or "I feel weak all over." These descriptions often lack the precision of medical terminology and contain metaphorical expressions, vague descriptions of feelings, and fragmented information. When processing this type of input, most existing health question-and-answer systems directly output generalized answers based on vague descriptions, failing to provide effective advice tailored to the user's specific situation. The core technical problem is that the semantic ambiguity of unstructured health descriptions makes it difficult to accurately map them to standardized medical concepts during natural language understanding. This leads to a significant decrease in the accuracy of symptom-disease correlation analysis, making it impossible to accurately generate personalized health advice. Furthermore, the lack of effective mining and utilization of the implicit information in vague descriptions makes it difficult to meet users' practical needs for health consultation. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a health-related question-and-answer method for an AI voice assistant and a smartwatch, and the technical solution adopted is as follows:

[0004] A health question-and-answer method for an AI voice assistant, the method comprising:

[0005] S1: Accepts user's health description input;

[0006] S2: Map users' health descriptions to a medical concept library to generate standardized symptom descriptions;

[0007] S3: Input standardized symptom descriptions into the probability association layer, perform dynamic probability flow propagation calculations based on a pre-built medical knowledge graph, and generate associated symptom results and the probabilities of corresponding disease hypotheses;

[0008] S4: Based on user profile data, the probability is corrected in the profile adaptation layer.

[0009] Preferably, S1 specifically includes:

[0010] The user's health description is deduplicated, while the original semantic strength identifier of the ambiguous description is retained.

[0011] Preferably, S2 specifically includes:

[0012] The core feature words of health descriptions are extracted using a natural language processing model, and similarity is calculated with the symptom feature word list in the medical concept library. Medical concepts are then selected as preliminary association results.

[0013] Substitute the preliminary association results into the logical framework of clinical symptom description. If the preliminary association results match the clinical symptoms, retain the preliminary association results. If the preliminary association results do not match the clinical symptoms, expand the similarity calculation range with the symptom feature vocabulary in the medical concept library and perform a second calculation until the preliminary association results match the clinical symptoms.

[0014] Preferably, the core feature words include: metaphorical feature words, location feature words, and degree feature words, and the similarity calculation is based on the feature words to perform weighted vector calculation.

[0015] Preferably, expanding the similarity calculation range with the symptom feature vocabulary in the medical concept library specifically includes:

[0016] The core feature words are extended to related features in the same part within the cross-semantic cluster.

[0017] Preferably, the dynamic probability flow propagation calculation of S3 specifically includes:

[0018] The medical knowledge graph is modeled as a three-layer network: an initial layer, a direct association layer, and an indirect association layer. Probability propagation is performed according to the layers, and propagation terminates when the probability of a node in the indirect association layer decays to a threshold.

[0019] Preferably, the initial layer is used to receive standardized symptom description nodes; the direct association layer is used to connect nodes that have a medical causal relationship with the nodes in the initial layer; and the indirect association layer is used to associate the remaining nodes through intermediate nodes.

[0020] Preferably, S4 specifically includes:

[0021] Establish a user profile matrix, which includes age segments, gender, and underlying disease tags;

[0022] Construct a three-dimensional parameter table based on epidemiological statistics to store the mapping relationship between age segments, gender, underlying disease labels and correction coefficients;

[0023] Based on the user profile matching parameter table, the correction coefficients are accumulated for multiple dimensions that are satisfied simultaneously;

[0024] The final probability is dynamically adjusted based on the sign of the accumulated correction coefficients.

[0025] Preferably, the assignment rule for the correction coefficient is as follows:

[0026] A positive coefficient is assigned when a user profile matches the characteristics of a high-risk group for a disease.

[0027] When a user profile matches the characteristics of a low-risk group for a disease, a negative coefficient is assigned.

[0028] A smartwatch, wherein the smartwatch is used to implement a health question-and-answer method for an AI voice assistant as described in any one of claims 1-9.

[0029] The beneficial effects of this invention are as follows: This invention changes vague descriptions into standardized symptoms, combines medical knowledge graphs to infer disease probabilities and optimizes results based on user profiles, thus solving the problem of low accuracy in question-and-answer due to semantic ambiguity, improving the accuracy of personalized health advice, and avoiding the drawbacks of traditional systems that force precise descriptions, affecting user experience and outputting generalized answers that lack help. Attached Figure Description

[0030] Figure 1 This invention relates to a health question-and-answer method for an AI voice assistant. Detailed Implementation

[0031] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0032] An embodiment of the present invention provides a health question-and-answer method for an AI voice assistant, the method comprising:

[0033] S1: Accepts user's health description input;

[0034] S2: Map users' health descriptions to a medical concept library to generate standardized symptom descriptions;

[0035] S3: Input standardized symptom descriptions into the probability association layer, perform dynamic probability flow propagation calculations based on a pre-built medical knowledge graph, and generate associated symptom results and the probabilities of corresponding disease hypotheses;

[0036] S4: Based on user profile data, the probability is corrected in the profile adaptation layer.

[0037] The working principle and effect of the above technical solution are as follows: First, step S1 receives health information input by the user in natural language (fuzzy, colloquial description), generating an interaction entry point with the user; then, step S2 uses natural language processing technology to parse the unstructured health description input by the user, extracts core features, and matches and maps them with standardized terms in the medical concept library, changing the fuzzy description (such as "stomach discomfort") into a medically recognized standardized symptom description ("abdominal discomfort"), thus resolving the semantic discrepancy between natural language and professional terminology; then, step S3 inputs the standardized symptom description into the probability association layer, based on a pre-built medical knowledge graph, and calculates the probability through dynamic probability flow propagation, transferring and accumulating probabilities along the path of symptom-disease association in the knowledge graph, generating other symptom results associated with this symptom and the initial probability of the corresponding disease hypothesis, realizing the initial reasoning from symptoms to diseases; finally, step S4 calls user profile data in the profile adaptation layer, and combines correction rules based on epidemiological statistics to personalize the initial disease probability generated in step S3;

[0038] This invention transforms vague descriptions into standardized symptoms, combines medical knowledge graphs to infer disease probabilities, and optimizes results based on user profiles. This solves the problem of low accuracy in question-and-answer due to semantic ambiguity, improves the accuracy of personalized health advice, and avoids the shortcomings of traditional systems that force precise descriptions, affecting user experience, and output generalized answers that lack help.

[0039] In one embodiment of the present invention, S1 specifically includes:

[0040] The user's health description is deduplicated, while the original semantic strength identifier of the ambiguous description is retained.

[0041] The working principle and effect of the above technical solution are as follows: After receiving the user's health description, the system identifies and removes duplicate expressions through text comparison, and identifies the original semantic strength of the fuzzy description through semantic analysis tools, which simplifies the input information while retaining key semantic features. By retaining the original semantic strength identifier of the fuzzy description, the system solves the problems of redundant information and loss of fuzzy description strength in the original input, improves the accuracy of subsequent semantic parsing, and avoids key semantic distortion caused by information simplification.

[0042] In one embodiment of the present invention, S2 specifically includes:

[0043] The core feature words of health descriptions are extracted using a natural language processing model, and similarity is calculated with the symptom feature word list in the medical concept library. Medical concepts are then selected as preliminary association results.

[0044] Substitute the preliminary association results into the logical framework of clinical symptom description. If the preliminary association results match the clinical symptoms, retain the preliminary association results. If the preliminary association results do not match the clinical symptoms, expand the similarity calculation range with the symptom feature vocabulary in the medical concept library and perform a second calculation until the preliminary association results match the clinical symptoms.

[0045] The working principle and effect of the above technical solution are as follows: First, the core feature words are extracted from the user's health description using a natural language processing model. Then, the similarity (cosine similarity) of these feature words with the symptom feature word list in the medical concept library is calculated, and medical concepts with high similarity (0.7) are selected as preliminary association results. Next, the preliminary association results are substituted into the symptom expression framework constructed based on clinical medical logic for verification. If they conform to the clinical symptom expression logic, they are retained. If they do not conform, the threshold is reduced to 0.5 for secondary calculation to expand the screening range until an association result that conforms to the clinical symptom logic is found, so that the final standardized symptom description conforms to both medical norms and the user's original semantics.

[0046] This method addresses the challenge of balancing accuracy and coverage in fuzzy description matching under a fixed threshold, improves the consistency between standardized symptom descriptions and the user's original semantics, and ensures that the results conform to clinical logic, thus avoiding subsequent inference errors caused by semantic bias.

[0047] In one embodiment of the present invention, expanding the similarity calculation range with the symptom feature vocabulary in the medical concept library specifically includes:

[0048] The core feature words are extended to related features in the same part within the cross-semantic cluster.

[0049] The working principle and effect of the above technical solution are as follows: When the preliminary association results do not conform to the logic of clinical symptom description, the semantic cluster to which the core feature words extracted from the user's health description belong is taken as the starting point. The original semantic cluster limitation is broken through, and other semantic cluster feature words related to the anatomical part pointed to by the core feature words are retrieved in the medical concept library. These feature words related to the same part are included in the similarity calculation range and are re-matched with the core feature words input by the user, thereby expanding the screening range.

[0050] This approach expands core feature words across semantic clusters to features associated with the same location. When expanding the scope of similarity calculation, the anatomical location is always used as the core anchor point. This avoids the decrease in matching accuracy caused by indiscriminate expansion of the scope. Furthermore, by incorporating cross-semantic cluster feature words of the same location, it solves the problem of insufficient matching of feature words within a single semantic cluster and improves the matching coverage of fuzzy descriptions and medical concepts.

[0051] In one embodiment of the present invention, the dynamic probability flow propagation calculation of S3 specifically includes:

[0052] The medical knowledge graph is modeled as a three-layer network: an initial layer, a direct association layer, and an indirect association layer. Probability propagation is performed according to the layers, and propagation terminates when the probability of a node in the indirect association layer decays to a threshold.

[0053] The working principle and effect of the above technical solution are as follows: the medical knowledge graph is modeled as a three-layer network of initial layer, direct association layer and indirect association layer. During propagation, it starts from the nodes of the initial layer and uses their probability values ​​as the initial probability flow. The probability is then assigned to the nodes of the direct association layer according to the association strength between nodes in the knowledge graph. After receiving the probability, the nodes of the direct association layer propagate to the nodes of the indirect association layer according to the same rules. At the same time, a probability decay mechanism is set. When the probability value received by the nodes of the indirect association layer is lower than the preset threshold (.05), the propagation is terminated. Finally, the cumulative probability of the disease hypothesis and the associated symptom results in each layer node are generated.

[0054] This approach models the medical knowledge graph as a three-layer network and propagates probabilities hierarchically. It then terminates the propagation by using a probability decay threshold. This approach captures the complete chain of associations from initial symptoms to directly and indirectly related diseases through hierarchical propagation, solving the problem of incomplete information in single-level reasoning. Furthermore, the probability decay threshold avoids computational redundancy caused by excessive propagation of irrelevant nodes, thus improving reasoning efficiency.

[0055] In one embodiment of the present invention, S4 specifically includes:

[0056] Establish a user profile matrix, which includes age segments, gender, and underlying disease tags;

[0057] Construct a three-dimensional parameter table based on epidemiological statistics to store the mapping relationship between age segments, gender, underlying disease labels and correction coefficients;

[0058] Based on the user profile matching parameter table, the correction coefficients are accumulated for multiple dimensions that are satisfied simultaneously;

[0059] The final probability is dynamically adjusted based on the sign of the accumulated correction coefficients, and the dynamic adjustment of the final probability is based on the following:

[0060]

[0061] Where P represents the final probability, Let S represent the initial probability generated in step S3, D represent the correction coefficient, and H represent the symptom site binding entropy.

[0062] Furthermore, the binding entropy of the symptom site

[0063] in, This represents the probability of a symptom phrase extracted from a user description matching the i-th anatomical location in a medical knowledge graph, with a value range of [0,1]. This represents the granularity correction factor for the anatomical location, with a value of [value missing]. N represents the total number of anatomical sites in the medical knowledge graph that are related to the symptom phrases extracted from the user description.

[0064] The working principle and effect of the above technical solution are as follows: The mapping relationship between age segmentation, gender, underlying disease labels, and correction coefficients specifically includes: Each correction coefficient is derived from the relative risk ratio (RR) of a certain dimension feature and the risk of disease incidence in authoritative epidemiological data. Specifically, age segmentation is based on WHO standards and divided into children (0-14 years), youth (15-44 years), middle-aged (45-64 years), and elderly (≥65 years). Each segment corresponds to a correction coefficient for a specific disease, such as in "hypertensive headache," the correction coefficient for children is -0.3, for youth -0.1, for middle-aged +0.2, and for elderly +0.4; gender is divided into "male" and "female," and the correction coefficient reflects the deviation of the risk of disease incidence, such as in "hypothyroidism," the correction coefficient for males is -0.2, and for females +0.3; the underlying disease label includes common chronic diseases, and the correction coefficient is set based on the risk of comorbidity, such as in "diabetic nephropathy," the correction coefficient for those without diabetes is -0.5, and for those with diabetes (duration <5 years) is +0.5. The total correction factor D is calculated by adding 0.3 for patients with diabetes (duration ≥ 5 years) and 0.6 for patients with diabetes (duration ≥ 5 years).

[0065] Binding entropy at symptom sites The standardized co-occurrence probability of "symptom phrase - location" from clinical big data is obtained by extracting co-occurrence records of unstructured symptom phrases and anatomical locations from authoritative clinical terminology databases and standardized case databases; the number of co-occurrences of this type of symptom phrase S with each anatomical location is counted, and then the frequency is calculated (frequency = number of co-occurrences of a certain location ÷ total number of co-occurrences of all locations), which is the probability density function. ;

[0066] Existing health question-and-answer systems suffer from a mismatch between ambiguous information and the strength of correction when adjusting disease probabilities based on user profiles:

[0067] The input was clearly linked to the symptom location, but the targeted nature of the profile correction was not enhanced.

[0068] Even with vaguely defined symptom locations, the image correction effect is still amplified according to fixed logic;

[0069] The failure to incorporate the core clinical diagnostic clue of symptom location into the dynamic adjustment mechanism of probability correction leads to a disconnect between the correction logic and the clinical diagnosis and treatment logic.

[0070] When H approaches 0, it indicates that the binding relationship between symptoms and location is highly certain, the corresponding symptom-location association is unambiguous, and it belongs to highly certain information;

[0071] When H approaches the maximum value of 1.58, it indicates that the binding relationship between symptoms and location is highly ambiguous, and there are multiple possibilities for the association between symptoms and location, which belongs to low-deterministic information;

[0072] The median value H indicates that there is some uncertainty regarding the correlation between symptom locations, but there is still a tendency, which is considered moderately certain information.

[0073] In fact, H changes the vague semantics of "symptom-location" in natural language descriptions into quantifiable "uncertainty values", solving the problem of "inappropriate correction of vague information" in existing technologies and improving the accuracy of personalized health recommendations.

[0074] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A health-related question-and-answer method for an AI voice assistant, characterized in that, The method includes: S1: Accepts user's health description input; S2: Map users' health descriptions to a medical concept library to generate standardized symptom descriptions; S3: Input standardized symptom descriptions into the probability association layer, perform dynamic probability flow propagation calculations based on a pre-built medical knowledge graph, and generate associated symptom results and the probabilities of corresponding disease hypotheses; S4: Based on user profile data, the probability is corrected in the profile adaptation layer; S4 specifically includes: Establish a user profile matrix, which includes age segments, gender, and underlying disease tags; Construct a three-dimensional parameter table based on epidemiological statistics to store the mapping relationship between age segments, gender, underlying disease labels and correction coefficients; Based on the user profile matching parameter table, the correction coefficients are accumulated for multiple dimensions that are satisfied simultaneously; The final probability is dynamically adjusted based on the sign of the accumulated correction coefficients; and the dynamic adjustment of the final probability is based on the following: ； Where P represents the final probability, Let S represent the initial probability generated in step S3, D represent the correction coefficient, and H represent the symptom site binding entropy. Furthermore, the binding entropy of the symptom site ; in, This represents the probability of a symptom phrase extracted from a user description matching the i-th anatomical location in a medical knowledge graph, with a value range of [0,1]. This represents the granularity correction factor for the anatomical location, with a value of [value missing]. N represents the total number of anatomical sites in the medical knowledge graph that are related to the symptom phrases extracted from the user description.

2. The health question-and-answer method for an AI voice assistant according to claim 1, characterized in that, S1 specifically includes: The user's health description is deduplicated, while the original semantic strength identifier of the ambiguous description is retained.

3. The health question-and-answer method for an AI voice assistant according to claim 1, characterized in that, S2 specifically includes: The core feature words of health descriptions are extracted using a natural language processing model, and similarity is calculated with the symptom feature word list in the medical concept library. Medical concepts are then selected as preliminary association results. Substitute the preliminary association results into the logical framework of clinical symptom description. If the preliminary association results match the clinical symptoms, retain the preliminary association results. If the preliminary association results do not match the clinical symptoms, expand the similarity calculation range with the symptom feature vocabulary in the medical concept library and perform a second calculation until the preliminary association results match the clinical symptoms.

4. The health question-and-answer method for an AI voice assistant according to claim 3, characterized in that, The core feature words include: metaphorical feature words, location feature words, and degree feature words. The similarity calculation is based on the feature words and a weighted vector calculation.

5. The health question-and-answer method for an AI voice assistant according to claim 3, characterized in that, The expanded similarity calculation range with the symptom feature vocabulary in the medical concept library specifically includes: The core feature words are extended to related features in the same part within the cross-semantic cluster.

6. The health question-and-answer method for an AI voice assistant according to claim 1, characterized in that, The dynamic probability flow propagation calculation of S3 specifically includes: The medical knowledge graph is modeled as a three-layer network: an initial layer, a direct association layer, and an indirect association layer. Probability propagation is performed according to the layers, and propagation terminates when the probability of a node in the indirect association layer decays to a threshold.

7. A health question-and-answer method for an AI voice assistant according to claim 6, characterized in that, The initial layer is used to receive standardized symptom description nodes; the direct association layer is used to connect nodes that have a medical causal relationship with the nodes in the initial layer; and the indirect association layer is used to associate the remaining nodes through intermediate nodes.

8. The health question-and-answer method for an AI voice assistant according to claim 1, characterized in that, The assignment rules for the correction coefficient are as follows: A positive coefficient is assigned when a user profile matches the characteristics of a high-risk group for a disease. When a user profile matches the characteristics of a low-risk group for a disease, a negative coefficient is assigned.

9. A smartwatch, characterized in that, The smartwatch is used to implement the health question-and-answer method of an AI voice assistant as described in any one of claims 1-8.

Images