Man-machine dialogue type health education effect evaluation method for intelligent treatment vehicle

Abstract

The present application relates to the field of artificial intelligence and voice interaction technology, and discloses a man-machine dialogue type health propaganda and education effect evaluation method for a smart treatment vehicle. The method collects patient voice, expression, eye movement, posture and operation feedback data through a multi-modal perception unit, and constructs initial cognitive adaptation parameters in combination with the basic portrait. Feature extraction is performed on each modal data, and a multi-dimensional cognitive state representation vector is generated by fusion, which is input into a pre-trained neural network model to output understanding, acceptance and short-term memory prediction values. According to this, a comprehensive evaluation index is calculated and the propaganda and education strategy is dynamically adjusted. The system includes multi-modal interaction behavior collection, feature extraction, cognitive state fusion, effect evaluation reasoning, strategy dynamic adjustment and historical database management modules. The present application realizes real-time quantitative evaluation and closed-loop optimization of propaganda and education effect, and significantly improves the effectiveness of health information transmission and patient compliance.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and voice interaction technology, specifically relating to a method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle. Background Technology

[0002] With the rapid development of smart healthcare systems, health education, as a crucial component of clinical nursing and chronic disease management, is gradually evolving towards intelligence and personalization. Traditional health education often relies on face-to-face explanations by healthcare professionals or standardized audio-visual presentations. Its content is fixed and the interaction is one-way, making it difficult to adapt to individual differences among patients in terms of cognitive level, learning preferences, and communication habits. Especially in hospitalization, rehabilitation, or community care settings, the wide age range, diverse disease types, and varying educational backgrounds of patients lead to widespread problems with uniform education models, such as low information reception rates, significant misunderstandings, and weak behavioral compliance, severely limiting the actual effectiveness of health interventions.

[0003] Human-computer dialogue-based health education systems, as an emerging technological approach, attempt to enhance patient engagement through natural language interaction. However, existing systems generally employ generic dialogue models, lacking accurate patient identification and dynamic modeling of long-term cognitive states. Even when some systems introduce voice input, it is only used for command recognition, failing to link acoustic features with individual identities or connect them to electronic health records or historical interaction data. This forces each education session to start from "zero cognition," making it impossible to provide targeted reinforcement based on patients' existing knowledge gaps or dynamically adjust the complexity and depth of expression based on their understanding and feedback. Consequently, the education process is repetitive, inefficient, and lacks continuity.

[0004] Existing technologies exhibit a clear disconnect between identity recognition and personalized learning: voiceprint recognition modules operate independently, serving only for identity verification; while cognitive models rely on static questionnaires or one-time assessments to construct initial profiles, failing to achieve incremental updates through continuous dialogue. This fragmented architecture not only creates data silos but also hinders the system from developing a deep understanding of patients, making it impossible to establish trust and cognitive evolution trajectories through multiple interactions. In practical applications, existing solutions often exhibit slow response times, content mismatches, and even the risk of misleading patients when faced with fluctuating cognitive abilities (such as postoperative delirium or early-stage Alzheimer's disease) or changing interaction intentions. Therefore, a health education effectiveness evaluation and optimization mechanism that deeply integrates unique identity identification with dynamic cognitive modeling is urgently needed. Summary of the Invention

[0005] This invention provides a method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle, aiming to solve the technical problem that the effectiveness of general education models is poor due to the significant differences in cognitive abilities and interaction patterns among different patients in existing technologies. The method constructs a dynamic dialogue modeling mechanism oriented towards individual cognitive characteristics, combining multimodal interactive behavior perception and semantic understanding to quantify patients' understanding, acceptance, and memory retention of health education content in real time. This generates personalized evaluation indicators for the effectiveness of education, and based on these indicators, dynamically adjusts subsequent education strategies, achieving closed-loop optimization of the human-computer dialogue-based health education process.

[0006] According to one aspect of the present invention, a method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle is provided, comprising: The multimodal perception unit mounted on the intelligent treatment vehicle collects interactive behavior data of patients during health education dialogues. The interactive behavior data includes voice response content, facial expression change sequence, eye movement trajectory data, body posture information, and operation feedback actions. Based on the preset patient profile information, including age, educational background, medical history, language habits and preliminary cognitive ability assessment level, an initial cognitive adaptation parameter set is constructed. The interactive behavior data is subjected to modal feature extraction. Specifically, after speech recognition of the speech response content, semantic integrity analysis and keyword coverage calculation are performed. Micro-expression recognition is performed on the facial expression change sequence to obtain the emotion tendency index. The eye movement trajectory data is subjected to fixation point density and temporal stability analysis to determine the attention concentration. The body posture information is subjected to open and avoidance posture discrimination to assess the interaction intention intensity. The operation feedback action is subjected to response delay and operation accuracy statistics to measure the instruction comprehension ability. The feature vectors of each modality are weighted and fused with the initial cognitive adaptation parameter set to generate a multidimensional cognitive state representation vector for the current dialogue round. Based on the multidimensional cognitive state representation vector, a pre-trained neural network model for evaluating the effectiveness of education is invoked to output the comprehension score, acceptance score, and short-term memory prediction value of the current education content. Based on the comprehension score, acceptance score, and short-term memory prediction value, a comprehensive education effectiveness evaluation index is calculated, and it is determined whether the comprehensive education effectiveness evaluation index is lower than a preset threshold. If it is lower than the preset threshold, an education strategy reconstruction mechanism is triggered. The education strategy reconstruction mechanism includes adjusting the complexity of the expression of education content, switching the information presentation medium type, introducing an analogical explanation model, or activating a repetition reinforcement module. The interactive behavior data, cognitive state representation vector, and comprehensive education effect evaluation index of this dialogue are stored in the patient's exclusive education history database to update the cognitive adaptation parameter set for subsequent dialogues.

[0007] In one embodiment of the present invention, the multimodal perception unit includes a high-sensitivity microphone array, an infrared depth camera, an eye-tracking sensor, a six-axis inertial measurement unit, and a touchscreen operation recording module. The high-sensitivity microphone array is used to collect the patient's voice signal and suppress environmental noise. The infrared depth camera is used to capture changes in the three-dimensional structure of the face in low-light environments. The eye-tracking sensor uses the pupil-corneal reflection principle to record the coordinates of the gaze point in real time. The six-axis inertial measurement unit is integrated into the handrail of the treatment cart and is used to detect minor posture changes of the patient's upper limbs. The touchscreen operation recording module records the patient's click position, press duration, and sliding trajectory of the interactive buttons in the educational content.

[0008] As one embodiment of the present invention, the semantic integrity analysis is achieved by calculating the edit distance and semantic similarity between the patient's response statement and the standard response template, wherein the semantic similarity is determined based on the cosine angle in the Chinese word vector space; the keyword coverage calculation is based on a preset core knowledge point dictionary for education, and the proportion of the number of core words covered in the patient's response to the total number of core words is statistically analyzed.

[0009] As one embodiment of the present invention, the micro-expression recognition adopts a classification model based on spatiotemporal convolutional neural network. The input is a sequence of facial key point coordinates of 30 consecutive frames, and the output is a probability distribution of seven basic emotions, including happiness, sadness, anger, fear, surprise, disgust and neutrality. The emotion tendency index is obtained by subtracting the sum of the probabilities of the remaining negative emotions from the sum of the probabilities of happiness and neutral emotions.

[0010] As one embodiment of the present invention, the fixation point density is calculated by dividing the propaganda content display area into a nine-square grid area and counting the number of fixations in each sub-region per unit time; the temporal stability is determined by calculating the standard deviation of the Euclidean distance between adjacent fixation points, and the smaller the standard deviation, the more stable the fixation.

[0011] As one embodiment of the present invention, the openness and avoidance posture discrimination is based on the combination rules of shoulder and elbow joint angles. When the shoulder abduction angle is greater than 60 degrees and the elbow flexion angle is less than 90 degrees, it is determined to be an open posture. When the torso leans forward more than 15 degrees and the arms are crossed, it is determined to be an avoidance posture.

[0012] As one embodiment of the present invention, the neural network model for evaluating the effectiveness of education is a three-layer fully connected neural network. The number of nodes in its input layer is equal to the dimension of the multidimensional cognitive state representation vector. The hidden layer uses a modified linear unit activation function, and the output layer contains three independent branches, corresponding to the comprehension score, the acceptance score, and the short-term memory prediction value, respectively. The model is trained under supervision using a historical patient dataset, and the labels are jointly annotated by clinicians based on the results of the immediate question-and-answer test after education and the accuracy of recall during the follow-up within 24 hours.

[0013] In one embodiment of the present invention, the complexity adjustment of the expression is achieved by controlling the average sentence length, the density of professional terms, and the number of nested clauses in the educational text, wherein the average sentence length does not exceed 15 characters, the density of professional terms does not exceed 10%, and the number of nested clauses does not exceed one layer; the information presentation media types include plain text, a combination of text and graphics, animated demonstrations, and interactive three-dimensional organ models; the analogy explanation model retrieves analogies related to patients' life experiences from a pre-constructed medical analogy knowledge base, which contains more than 500 clinically validated analogy mapping pairs; the repetition reinforcement module is automatically activated when the comprehension score is detected to be below the threshold for two consecutive rounds, and the same knowledge point is explained a second time using a changed expression method.

[0014] In one embodiment of the present invention, the updating of the cognitive adaptation parameter set adopts an exponential smoothing algorithm. ,in For the new parameter value, The old parameter value, For smoothing coefficients, The smoothing coefficient is the representation vector of the cognitive state in the current round. The value is 0.8.

[0015] According to another aspect of the present invention, a human-computer dialogue-based health education effectiveness evaluation system for a smart treatment vehicle is provided, comprising: The multimodal interactive behavior acquisition module is used to collect patient interactive behavior data during health education dialogues through the multimodal perception unit mounted on the intelligent treatment vehicle; The patient basic profile construction module is used to generate an initial cognitive adaptation parameter set based on the patient's electronic medical record and the results of the initial interaction assessment; The modal feature extraction module is used to extract features from five types of interactive behavior data: voice, facial expression, eye movement, posture, and operation. The multidimensional cognitive state fusion module is used to weight and fuse the feature vectors of each modality with the initial cognitive adaptation parameter set to generate a multidimensional cognitive state representation vector. The reasoning module of the education effectiveness evaluation model is used to call the pre-trained education effectiveness evaluation neural network model to output comprehension score, acceptance score and short-term memory prediction value; The dynamic adjustment module for missionary strategies is used to trigger and execute corresponding missionary strategy reconstruction operations based on the comprehensive missionary effectiveness evaluation index. The patient education history database management module is used to store and update patient-specific educational interaction records and cognitive state evolution trajectories.

[0016] As one embodiment of the present invention, the modal feature extraction module includes a speech semantic analysis submodule, a micro-expression recognition submodule, an eye-tracking attention analysis submodule, a posture intention discrimination submodule, and an operation response evaluation submodule; each submodule operates independently and sends the output feature vector to the multi-dimensional cognitive state fusion module.

[0017] As one embodiment of the present invention, the propaganda strategy dynamic adjustment module has four built-in strategy adjustment rules, namely content simplification rule, media switching rule, analogy introduction rule, and repetition reinforcement rule; each rule corresponds to different triggering conditions and execution actions, and the combination logic of the comprehensive propaganda effect evaluation index and the scores of each sub-item determines which rule is activated.

[0018] As one embodiment of the present invention, the education history database management module adopts a time-series database structure. Each record includes a timestamp, education topic identifier, original data summary of interactive behavior, cognitive state representation vector, effect evaluation index, and education strategy identifier. The education history database management module supports rapid retrieval and trend analysis based on the patient's unique identifier.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention abandons the traditional "one-size-fits-all" static content delivery model in health education, and for the first time constructs a dynamic effect evaluation and feedback adjustment mechanism centered on the individual cognitive characteristics of patients. By comprehensively capturing patients' non-verbal and verbal behaviors during conversations using multimodal perception technology, it achieves quantitative assessment of comprehension, emotional receptiveness, and memory potential. This method can accurately identify education failure points caused by differences in cognitive abilities and immediately initiate targeted strategy adjustments, significantly improving the effectiveness of health information delivery and patient compliance.

[0020] This invention, through the continuous accumulation of individualized health education historical data, enables the system to possess long-term cognitive adaptability, autonomously optimizing the health education model as the patient's condition progresses or cognitive state changes. Compared to existing evaluation methods that rely solely on post-hoc questionnaires or simple question-and-answer accuracy rates, the real-time, multi-dimensional, and objective evaluation system provided by this invention solves the fundamental technical challenge of general health education models failing to adapt to individual differences, offering an engineering-ready technical path for human-machine collaborative health education in smart healthcare scenarios. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the human-computer dialogue-based health education effect evaluation method for intelligent treatment vehicles proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of dynamic dialogue modeling and multimodal interaction behavior fusion based on individual cognitive characteristics in this invention. Figure 3 This is a logical flowchart of the multimodal interaction behavior data acquisition and submodal feature extraction in this invention; Figure 4 This is a logical flowchart of the reasoning process for the multidimensional cognitive state representation generation and education effect evaluation model in this invention. Figure 5 This is a logical flowchart of the mechanism for dynamic adjustment and reconstruction of missionary strategies in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the intelligent treatment vehicle and the patient in this invention. Detailed Implementation

[0022] This invention provides a method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle, aiming to solve the technical problem that the effectiveness of general education models is poor due to the significant differences in cognitive abilities and interaction patterns among different patients in existing technologies. The method constructs a dynamic dialogue modeling mechanism oriented towards individual cognitive characteristics, combining multimodal interactive behavior perception and semantic understanding to quantify the patient's understanding, acceptance willingness, and memory retention level of health education content in real time. This generates personalized evaluation indicators for the effectiveness of education, and based on these indicators, dynamically adjusts subsequent education strategies, achieving closed-loop optimization of the human-computer dialogue-based health education process. The following will be combined with the appendix... Figures 1 to 6 The specific implementation methods of each component of the system are described in detail.

[0023] As one embodiment of the present invention, the method includes the following steps: S1, the multimodal perception unit mounted on the intelligent treatment vehicle collects interactive behavior data of patients during health education dialogues. The interactive behavior data includes voice response content, facial expression change sequence, eye movement trajectory data, body posture information and operation feedback actions. S2, Based on the preset patient profile information, which includes age, educational background, past medical history, language habits and preliminary cognitive ability assessment level, an initial cognitive adaptation parameter set is constructed; S3, perform modal feature extraction on the interactive behavior data, including performing speech recognition on the speech response content and performing semantic integrity analysis and keyword coverage calculation, performing micro-expression recognition on the facial expression change sequence to obtain the emotion tendency index, performing fixation point density and temporal stability analysis on the eye movement trajectory data to determine the attention concentration, performing open and avoidance posture discrimination on the body posture information to assess the interaction intention intensity, and performing response delay and operation accuracy statistics on the operation feedback action to measure the instruction comprehension ability. S4, weighted fusion of each modality feature vector with the initial cognitive adaptation parameter set to generate a multidimensional cognitive state representation vector for the current dialogue round; S5. Based on the multidimensional cognitive state representation vector, call the pre-trained propaganda effect evaluation neural network model to output the comprehension score, acceptance score and short-term memory prediction value of the current propaganda content. S6. Based on the comprehension score, acceptance score, and short-term memory prediction value, calculate the comprehensive education effect evaluation index and determine whether the comprehensive education effect evaluation index is lower than a preset threshold. If it is lower than the preset threshold, trigger the education strategy reconstruction mechanism. The education strategy reconstruction mechanism includes adjusting the expression complexity of the education content, switching the information presentation medium type, introducing an analogy explanation model, or activating the repetition reinforcement module. S7 stores the interactive behavior data, cognitive state representation vector, and comprehensive education effect evaluation index of this dialogue into the patient's exclusive education history database, which is used to update the cognitive adaptation parameter set for subsequent dialogues.

[0024] In the above method, step S1 involves collecting patient interaction data during health education dialogues using a multimodal sensing unit mounted on a smart treatment vehicle. The multimodal sensing unit includes a high-sensitivity microphone array, an infrared depth camera, an eye-tracking sensor, a six-axis inertial measurement unit, and a touchscreen operation recording module. The high-sensitivity microphone array is located in a ring-shaped area on the top of the treatment vehicle, with a sampling frequency set to 48,000 Hz. It features beamforming and adaptive noise suppression capabilities, enabling accurate acquisition of patient voice signals in a reverberant ward environment while suppressing background noise interference from air conditioning, monitors, etc. The infrared depth camera is mounted in the center above the front display screen of the treatment vehicle, operating at a wavelength of 850 nanometers and a frame rate of 30 Hz. It can continuously capture changes in the three-dimensional structure of the patient's face in a nighttime ward environment with illumination below 5 lux, outputting a temporal data stream containing coordinates of 68 facial key points.

[0025] The eye-tracking sensor employs a combination of dual infrared light sources and a high-speed image sensor. Based on the principle of pupil-corneal reflection, it records the two-dimensional coordinates of the patient's gaze points within the educational content display area in real time, with a sampling frequency of 120 Hz and a spatial resolution of 0.5 degrees of visual angle. A six-axis inertial measurement unit, integrated into the armrests on both sides of the treatment cart, includes a three-axis accelerometer and a three-axis gyroscope. It detects minute posture changes of the patient's upper limbs during interaction, with a data output frequency of 50 Hz and an angle measurement error of less than 1 degree. A touchscreen operation recording module is embedded below the interactive display screen on the front of the treatment cart, supporting multi-touch and recording the patient's click positions, press durations, swipe trajectories, and gesture types for interactive buttons within the educational content, with timestamp accuracy down to the millisecond level. Data streams from all sensing units are synchronously transmitted to the central processing unit via the cart's in-vehicle controller LAN bus, ensuring strict temporal alignment of data from each modality.

[0026] In the above method, step S2 constructs an initial cognitive adaptation parameter set based on preset patient basic profile information. The patient basic profile information comes from the hospital's electronic medical record system interface and is automatically imported when the patient first establishes a connection with the intelligent treatment vehicle. The basic profile information includes the patient's age, maximum years of education, primary diagnostic disease code, number of comorbidities, native language, frequency of dialect use, and the preliminary cognitive ability assessment level completed by nursing staff during the admission assessment stage.

[0027] The initial cognitive ability assessment level was generated using a score mapping of the Mini-Mental State Examination (MMSE), divided into four levels: Level 1 corresponds to a score of ≥24, indicating normal cognitive function; Level 2 corresponds to a score of 21-23, indicating mild cognitive impairment; Level 3 corresponds to a score of 17-20, indicating moderate cognitive impairment; and Level 4 corresponds to a score of ≤16, indicating severe cognitive impairment. The initial cognitive adaptation parameter set was stored in vector form with 7 dimensions. Each dimension corresponds to the normalized value of age, the encoded value of educational background, the disease complexity index, the language habit matching coefficient, the cognitive level weighting factor, the interaction patience threshold, and the information receiving preference identifier.

[0028] The age normalization value is obtained by dividing (actual age - 60) by 40, with a range of -1.5 to +1.5; the education background coding value is divided into five levels based on the highest level of education: primary school, junior high school, high school, junior college, and undergraduate and above, with values ​​of 1 to 5 respectively; the disease complexity index is generated by weighted summation of the number of major diagnoses and comorbidities followed by logarithmic compression; the language habit matching coefficient is set according to the consistency between the patient's native language and the system's default educational language, with a value of 1 when they are consistent, 0.7 when they are inconsistent but dialect translation is supported, and 0.4 otherwise; the cognitive level weight factor is directly taken as the reciprocal of the preliminary cognitive ability assessment level; the interaction patience threshold is set to 300 seconds × cognitive level weight factor; the information receiving preference identifier is set to a combined text and image mode by default when historical interaction data is missing.

[0029] In the above method, step S3 extracts modal features from the interactive behavior data. The voice response content is first converted into a text sequence by a speech recognition engine. The engine uses an end-to-end Chinese speech recognition model, with a vocabulary covering 120,000 entries from both medical terminology and everyday language databases, and a word error rate controlled within 5%. Subsequently, semantic integrity analysis is performed, calculated by determining the edit distance and semantic similarity between the patient's response and the standard response template. The standard response template is pre-written by clinical experts for each educational knowledge point, containing declarative sentences with complete semantic structures. The edit distance is calculated using a dynamic programming algorithm, with a maximum allowed length of 100 characters. Semantic similarity is determined based on the cosine angle in the Chinese word vector space. The word vectors are generated using a context-sensitive embedding model pre-trained on 1 billion medical texts, and the sentence vectors are reduced to 100 dimensions through principal component analysis after weighted averaging of the word vectors.

[0030] Keyword coverage is calculated based on a pre-defined dictionary of core medical knowledge points for public education. This dictionary contains 3 to 5 essential core medical concepts for each educational topic. The proportion of covered core terms in patient responses relative to the total number of core terms is calculated, and this proportion, after smoothing, is used as the keyword coverage feature. Facial expression change sequences are input into a micro-expression recognition model, which is a classifier based on a spatiotemporal convolutional neural network. The input is a sequence of 30 consecutive frames of facial key point coordinates, with each frame containing 68 two-dimensional coordinate points. After normalization, a 60×68 spatiotemporal tensor is formed. The model output is the probability distribution of seven basic emotions: happiness, sadness, anger, fear, surprise, disgust, and neutrality. The emotion tendency index is obtained by subtracting the sum of the probabilities of other negative emotions from the sum of the probabilities of happiness and neutral emotions, with a value ranging from -1 to +1. Eye-tracking data is first used to divide the educational content display area into a nine-square grid, with each sub-region being half the width and height of the screen.

[0031] Fixation density is calculated by counting the number of fixations in each sub-region per unit time and normalizing it to the total number of fixations, generating a 9-dimensional density vector. Temporal stability is determined by calculating the standard deviation of the Euclidean distance between adjacent fixations; a smaller standard deviation indicates more stable fixation. This value, after inverse transformation, is used as a stability feature. Limb posture information is obtained by parsing the shoulder and elbow joint angle sequences output by the six-axis inertial measurement unit. Openness and avoidance posture discrimination are based on the combination rules of shoulder and elbow joint angles. When the shoulder abduction angle is greater than 60 degrees and the elbow flexion angle is less than 90 degrees, it is judged as an open posture with a confidence level of 0.9; when the trunk forward tilt angle is greater than 15 degrees and the arms are crossed, it is judged as an avoidance posture with a confidence level of 0.85. The interaction intention intensity feature is generated by subtracting the avoidance posture duration from the open posture duration. Operation feedback action features include response delay and operation accuracy. Response delay is defined as the time interval from the system issuing an operation prompt to the patient's first touch of the screen, normalized after logarithmic transformation. The accuracy rate is calculated as the percentage of patients who correctly complete interactive tasks within a specified time, such as clicking the correct button or sliding in the indicated order. The number of incorrect operations is included in the denominator.

[0032] In the above method, step S4 weightedly fuses the feature vectors of each modality with the initial cognitive adaptation parameter set to generate a multidimensional cognitive state representation vector for the current dialogue turn. Each modality feature vector includes a speech semantic feature vector (3 dimensions), an facial expression emotion feature vector (1 dimension), an eye-tracking attention feature vector (10 dimensions), a posture intention feature vector (1 dimension), and an operation response feature vector (2 dimensions), for a total dimension of 17. The initial cognitive adaptation parameter set has a dimension of 7. The fusion process employs a gated weighting mechanism, with weights determined by the current dialogue context and the patient's historical performance. Specifically, firstly, the 17-dimensional behavioral feature vector and the 7-dimensional profile parameter vector are concatenated into a 24-dimensional input vector, which is then input into a two-layer fully connected network, outputting 7 gate weights, each corresponding to a profile parameter dimension. After Softmax normalization, the gate weights are multiplied by the original profile parameters to obtain an adaptively adjusted cognitive adaptation vector. Subsequently, this cognitive adaptation vector is concatenated again with the behavioral feature vector and passed through a residual connection module to finally output a 32-dimensional multidimensional cognitive state representation vector. This multidimensional cognitive state representation vector retains the joint information of the original behavioral details and the individual's cognitive priors, serving as input for subsequent evaluation models.

[0033] In the above method, step S5 invokes a pre-trained neural network model for evaluating the effectiveness of educational outreach based on a multidimensional cognitive state representation vector, outputting a comprehension score, an acceptance score, and a short-term memory prediction value. The neural network model for evaluating the effectiveness of educational outreach is a three-layer fully connected neural network with 32 nodes in the input layer and 64 nodes in the hidden layer. It uses a modified linear unit activation function, and the output layer contains three independent branches, corresponding to the comprehension score, acceptance score, and short-term memory prediction value, respectively. Each branch has an output dimension of 1. The comprehension score reflects the patient's grasp of the logical structure and core concepts of the educational content, ranging from 0 to 100; the acceptance score reflects the patient's emotional identification with and behavioral intention regarding the educational content, ranging from 0 to 100; and the short-term memory prediction value predicts the probability that the patient can accurately recall the key points of the educational outreach within 24 hours, ranging from 0 to 1.

[0034] The model was trained under supervision using a historical patient dataset. The training set contained interaction records from 5000 hospitalized patients. Each record was labeled independently by two attending physicians, with labels including the immediate question-and-answer test score after patient education (for comprehension), the self-assessment result on a five-point Likert scale (for acceptability), and the percentage of correctly recalled knowledge points during a 24-hour telephone follow-up (for short-term memory). The model was trained using the mean squared error loss function, with Adam as the optimizer and a learning rate of 0.001. The training run consisted of 200 epochs. On the validation set, the mean absolute errors of the three metrics were controlled to be within 5, 6, and 0.08, respectively.

[0035] In the above method, step S6 calculates a comprehensive education effectiveness evaluation index based on comprehension score, acceptance score, and short-term memory prediction value, and determines whether the comprehensive education effectiveness evaluation index is lower than a preset threshold. The formula for calculating the comprehensive education effectiveness evaluation index is as follows:

[0036] in, Rate comprehension. Rate it for acceptability. This is a short-term memory prediction value. The preset threshold is set to 70. If... If the score is less than 70, the education strategy reconstruction mechanism is triggered. This mechanism includes four operations: adjusting the complexity of the education content's expression, switching the information presentation medium type, introducing an analogical explanation model, or activating a repetition reinforcement module. The complexity of the expression is adjusted by controlling the average sentence length, the density of technical terms, and the number of nested clauses in the education text. The average sentence length should not exceed 15 characters, the technical terminology density should not exceed 10%, and the number of nested clauses should not exceed one layer. Information presentation medium types include plain text, a combination of text and images, animated demonstrations, and interactive 3D organ models. The system automatically selects the optimal medium based on the patient's historical preferences and current attentional state.

[0037] The analogy explanation model retrieves analogies related to patients' life experiences from a pre-built medical analogy knowledge base containing over 500 clinically validated analogy pairs, such as "a blocked blood vessel is like a rusty water pipe" and "insulin is like a key that opens a cell door." The repetition reinforcement module is automatically activated when the comprehension score is detected to be below 60 for two consecutive rounds. It then provides a secondary explanation of the same knowledge point using a modified representation, including synonym substitution, perspective shift, or context reconstruction.

[0038] In the above method, step S7 stores the interactive behavior data, cognitive state representation vector, and comprehensive education effectiveness evaluation index of this dialogue into a patient-specific education history database. The database adopts a time-series structure, with each record containing a timestamp, education topic identifier, original interactive behavior data summary, cognitive state representation vector, effectiveness evaluation index, and the identifier of the education strategy used. The original interactive behavior data summary is stored in hash value form to ensure privacy and security. The cognitive adaptation parameter set is updated using an exponential smoothing algorithm. ,in For the new parameter value, The old parameter value, For smoothing coefficients, The smoothing coefficient is the representation vector of the cognitive state in the current round. The value is set to 0.8. The updated parameter set is used for the initial adaptation of the next round of dialogue, enabling the continuous evolution of the cognitive model.

[0039] As one embodiment of the present invention, the human-computer dialogue-based health education effectiveness evaluation system for intelligent treatment vehicles includes a multimodal interactive behavior acquisition module, a patient basic profile construction module, a submodal feature extraction module, a multidimensional cognitive state fusion module, an education effectiveness evaluation model reasoning module, an education strategy dynamic adjustment module, and an education history database management module. The multimodal interactive behavior acquisition module is integrated into the intelligent treatment vehicle hardware platform and is responsible for simultaneously acquiring five types of raw signals: voice, facial expression, eye movement, posture, and operation. The patient basic profile construction module obtains static patient information through the hospital information system interface and generates an initial cognitive adaptation parameter set based on the initial interaction assessment.

[0040] The modal feature extraction module includes sub-modules for speech and semantic analysis, micro-expression recognition, eye-tracking attention analysis, posture and intention discrimination, and operational response evaluation. Each sub-module operates independently and sends its output feature vector to the multi-dimensional cognitive state fusion module. The multi-dimensional cognitive state fusion module performs weighted fusion and residual enhancement to generate a unified representation vector. The education effectiveness evaluation model inference module loads a pre-trained neural network model, performs forward inference, and outputs three scores. The education strategy dynamic adjustment module has four built-in strategy adjustment rules: content simplification, media switching, analogy introduction, and repetition reinforcement. Each rule corresponds to different triggering conditions and execution actions, and the combination logic of the comprehensive education effectiveness evaluation index and each sub-score determines which rule is activated. The education history database management module adopts a time-series database structure, supports rapid retrieval and trend analysis by patient's unique identifier, and provides data support for long-term cognitive adaptation.

[0041] This embodiment, through the aforementioned method and system, achieves real-time, multi-dimensional, and quantitative assessment of an individual patient's cognitive state during health education, and establishes a closed-loop feedback mechanism, significantly improving the effective delivery of educational content and patient compliance. The entire technical solution, without relying on subjective questionnaires, utilizes objective behavioral data to drive personalized strategy generation, solving the fundamental technical challenge that general health education models cannot adapt to individual differences.

Claims

1. A method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle, characterized in that, include: The multimodal perception unit mounted on the intelligent treatment vehicle collects interactive behavior data of patients during health education dialogues. The interactive behavior data includes voice response content, facial expression change sequence, eye movement trajectory data, body posture information, and operation feedback actions. Based on the preset patient profile information, including age, educational background, medical history, language habits and preliminary cognitive ability assessment level, an initial cognitive adaptation parameter set is constructed. Modal feature extraction is performed on the interactive behavior data; The feature vectors of each modality are weighted and fused with the initial cognitive adaptation parameter set to generate a multidimensional cognitive state representation vector for the current dialogue round. Based on the multidimensional cognitive state representation vector, a pre-trained neural network model for evaluating the effectiveness of education is invoked to output the comprehension score, acceptance score, and short-term memory prediction value of the current education content. Based on the comprehension score, acceptance score, and short-term memory prediction value, calculate the comprehensive education effectiveness evaluation index, and determine whether the comprehensive education effectiveness evaluation index is lower than a preset threshold. If the content falls below a preset threshold, a mission strategy restructuring mechanism is triggered. This mechanism includes adjusting the complexity of the mission content, switching the information presentation medium type, introducing an analogical explanation model, or activating a repetition reinforcement module. The interactive behavior data, cognitive state representation vector, and comprehensive education effect evaluation index of this dialogue are stored in the patient's exclusive education history database to update the cognitive adaptation parameter set for subsequent dialogues.

2. The method for evaluating the effectiveness of human-computer dialogue-based health education for intelligent treatment vehicles according to claim 1, characterized in that, The multimodal sensing unit includes a high-sensitivity microphone array, an infrared depth camera, an eye-tracking sensor, a six-axis inertial measurement unit, and a touchscreen operation recording module. The high-sensitivity microphone array is used to acquire the patient's voice signal and suppress environmental noise; The infrared depth camera is used to capture changes in the three-dimensional structure of the face in low-light environments; The eye-tracking sensor uses the principle of pupil-corneal reflection to record the coordinates of the gaze point in real time. The six-axis inertial measurement unit is integrated into the handrail of the treatment cart and is used to detect minute posture changes of the patient's upper limbs; The touchscreen operation recording module records the patient's click location, press duration, and sliding trajectory for interactive buttons in the educational content.

3. The method for evaluating the effectiveness of human-computer dialogue-based health education for intelligent treatment vehicles according to claim 1, characterized in that, After performing speech recognition on the voice response content, semantic integrity analysis and keyword coverage calculation are performed, including: The semantic integrity score is obtained by jointly calculating the edit distance and semantic similarity between the patient's response statement and the standard response template. Based on the pre-set dictionary of core knowledge points for health education, the proportion of core words covered in patient responses to the total number of core words is calculated to obtain the keyword coverage. The semantic similarity is determined based on the cosine angle in the Chinese word vector space, and the word vectors are generated by a context-sensitive embedding model pre-trained on medical texts.

4. The method for evaluating the effectiveness of human-computer dialogue-based health education for intelligent treatment vehicles according to claim 1, characterized in that, Micro-expression recognition is performed on facial expression change sequences to obtain an emotion tendency index, including: Input the sequence of facial key point coordinates from 30 consecutive frames into a micro-expression recognition model based on a spatiotemporal convolutional neural network, and output the probability distribution of seven basic emotions. The emotional tendency index is obtained by subtracting the sum of the probabilities of other negative emotions from the sum of the probabilities of happy and neutral emotions.

5. The method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle according to claim 1, characterized in that, Eye-tracking trajectory data is analyzed for fixation density and temporal stability to determine attentional concentration, including: The area for displaying the missionary content is divided into a nine-square grid, and the number of fixations in each sub-area is counted per unit time to calculate the fixation point density. Calculate the standard deviation of the Euclidean distance between adjacent fixation points, and take the reciprocal of the standard deviation to obtain the temporal stability index; The fixation point density and temporal stability index are combined to form an attention concentration feature vector.

6. The method for evaluating the effectiveness of human-computer dialogue-based health education for intelligent treatment vehicles according to claim 1, characterized in that, To assess the intensity of interaction intention, openness and avoidance posture discrimination is performed on body posture information, including: An open posture is defined as a shoulder abduction angle greater than 60 degrees and an elbow flexion angle less than 90 degrees. When the torso leans forward at an angle greater than 15 degrees and the arms are crossed, it is considered an avoidance posture. The interaction intention intensity feature value is generated by subtracting the avoidance posture duration from the open posture duration.

7. The method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle according to claim 1, characterized in that, Statistical analysis of response latency and accuracy of operational feedback actions is performed to measure instruction comprehension ability, including: The response delay is defined as the time interval from when the system issues an operation prompt to when the patient first touches the screen, and this time interval is normalized after logarithmic transformation; The percentage of patients who correctly complete the interactive task within a specified time is used as the operation accuracy rate.

8. The method for evaluating the effectiveness of human-computer dialogue-based health education for intelligent treatment vehicles according to claim 1, characterized in that, The feature vectors of each modality are weighted and fused with the initial cognitive adaptation parameter set to generate a multidimensional cognitive state representation vector for the current dialogue round, including: The speech semantic feature vector, facial expression emotion feature vector, eye movement attention feature vector, posture intention feature vector, and operation response feature vector are concatenated into a behavior feature vector; The behavioral feature vector is concatenated with the initial cognitive adaptation parameter set and then input into a gating weight network to generate an adaptively adjusted cognitive adaptation vector. The adaptively adjusted cognitive adaptation vector is concatenated with the behavioral feature vector again and passed through the residual connection module to output a 32-dimensional multidimensional cognitive state representation vector.

9. The method for evaluating the effectiveness of human-computer dialogue-based health education for intelligent treatment vehicles according to claim 1, characterized in that, The pre-trained neural network model for evaluating the effectiveness of evangelism is invoked, outputting comprehension scores, acceptance scores, and short-term memory prediction values ​​for the current evangelistic content, including: The multidimensional cognitive state representation vector is input into a three-layer fully connected neural network, and the hidden layer of the neural network uses a modified linear unit activation function. The output layer outputs comprehension scores, acceptance scores, and short-term memory prediction values ​​through three independent branches. The comprehension score and acceptance score range from 0 to 100, and the short-term memory prediction value ranges from 0 to 1.

10. The method for evaluating the effectiveness of human-computer dialogue-based health education in a smart treatment vehicle according to claim 1, characterized in that, Calculate the comprehensive education and outreach effectiveness evaluation index and determine whether the comprehensive education and outreach effectiveness evaluation index is lower than a preset threshold, including: According to the formula: Calculate the comprehensive publicity and education effectiveness evaluation index; When the comprehensive education effectiveness evaluation index is below 70, the education strategy reconstruction mechanism is triggered.

Images