A health monitoring method based on elderly care robot and elderly care robot

Abstract

This invention relates to the field of artificial intelligence technology and provides a health monitoring method based on a senior care robot and the senior care robot. The method comprises: obtaining historical posts posted by a user on a social platform using the senior care robot; extracting depressive symptom features, emotional history features, and text history features based on the historical posts using a preset large language model; integrating and concatenating the depressive symptom features, emotional history features, and text history features to construct a final feature vector; inputting the final feature vector into a pre-trained prediction model, outputting a predicted value after multiple iterations of the prediction model, and generating an explanatory output based on the predicted value. This invention improves the accuracy of depression diagnosis and reduces prediction costs.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a health monitoring method based on an elderly care robot and the elderly care robot. Background Art

[0002] As society ages, the number of elderly people increases, while young people spend less and less time with them. Consequently, over time, elderly people may develop psychological problems, such as depression. Current methods of monitoring the mental health of the elderly rely primarily on their behavior or through specialized medical examinations. However, subjective observation of elderly behavior is inherently subjective, resulting in low accuracy in diagnosing depression. Consulting with specialized medical examiners to determine if there are any psychological issues increases costs. Summary of the Invention

[0003] In response to the technical problems existing in the prior art, the present invention provides a health monitoring method based on an elderly care robot and the elderly care robot thereof, so as to improve the accuracy of depression judgment and reduce the prediction cost.

[0004] In a first aspect, the present invention provides a health monitoring method based on an elderly care robot, comprising:

[0005] Obtain historical posts published by users on social platforms based on the elderly care robot;

[0006] Based on historical posts, a preset large language model is used to extract depressive symptom features, emotional process features, and text history features;

[0007] Integrate and concatenate the depressive symptom features, emotional process features, and text history features to construct a final feature vector;

[0008] The final feature vector is input into a pre-trained prediction model, and the prediction model outputs a prediction value after multiple iterations, and an explanatory output is generated based on the prediction value.

[0009] According to a health monitoring method based on an elderly care robot provided by the present invention, the extraction of depressive symptom features based on historical posts specifically includes:

[0010] Based on historical posts, the post content and symptom templates are embedded into text using a preset text embedding model;

[0011] Calculating the average similarity between each historical post and the symptom template, and annotating the historical posts corresponding to the average similarity exceeding a first preset value using a preset large language model for annotation;

[0012] The depression symptom vector is calculated based on the symptom template corresponding to the historical posts annotated by the large language model used for annotation, and the depression symptom features are obtained.

[0013] According to a health monitoring method based on an elderly care robot provided by the present invention, the extraction of emotional process features based on historical posts specifically includes:

[0014] Obtain historical posts and embed the post content with the sentiment template using the preset sentiment embedding model;

[0015] Calculating the similarity between all historical posts and each emotion template, obtaining historical posts whose similarity exceeds a second preset value, and forming an emotion set;

[0016] Based on the emotion set, a user's emotion process description is synthesized through a preset large language model for emotion description to obtain emotion process features.

[0017] According to a health monitoring method based on an elderly care robot provided by the present invention, the extraction of text history features based on historical posts specifically includes:

[0018] Extract text embeddings from historical posts through a pre-set large language model for text extraction;

[0019] A historical post collection is constructed based on the text embedding amount to generate text history features.

[0020] According to a health monitoring method based on an elderly care robot provided by the present invention, the depressive symptom features, emotional process features, and text history features are integrated and spliced to construct a final feature vector, specifically including:

[0021] Directly integrating and adding the emotional process features and the text history features to generate intermediate features;

[0022] The intermediate features are concatenated with the depressive symptom features to generate a final feature vector.

[0023] According to a health monitoring method based on an elderly care robot provided by the present invention, the final feature vector is input into a pre-trained prediction model, and the prediction model outputs a predicted value after multiple iterations, and an explanatory output is generated based on the predicted value, specifically including:

[0024] Inputting the final feature vector into a pre-trained prediction model, and constructing a decision tree set based on the pre-trained prediction model through a gradient boosting tree;

[0025] Select the optimal splitting criterion within the decision tree set to automatically perform feature interaction, complete feature fusion, and output the predicted value;

[0026] Based on the prediction value, an explanatory output of the text type is generated by a preset large language model for text interpretation.

[0027] In a second aspect, the present invention further provides a senior care robot, comprising a robot control center, a feature acquisition module, a feature extraction module, a feature aggregation module, and a prediction module, wherein the robot control center is connected to the feature acquisition module, the feature extraction module, the feature aggregation module, and the prediction module, respectively, to control and manage each module;

[0028] The feature acquisition module is used to obtain historical posts published by users on social platforms based on the elderly care robot;

[0029] A feature extraction module is used to extract depressive symptom features, emotional process features, and text history features based on historical posts using a preset large language model;

[0030] A feature aggregation module is used to integrate and splice the depression symptom features, emotional process features, and text history features, and generate a final feature vector by constructing a decision tree set;

[0031] A prediction module is used to input the final feature vector into a pre-trained prediction model, output a prediction value after multiple iterations of the prediction model, and generate an explanatory output based on the prediction value.

[0032] In a third aspect, the present invention also provides an electronic device comprising: a memory for storing a computer software program; and a processor for reading and executing the computer software program, thereby implementing any of the above-mentioned health monitoring methods based on the elderly care robot.

[0033] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium, characterized in that a computer software program is stored in the storage medium, and when the computer software program is executed by a processor, it implements any of the above-mentioned health monitoring methods based on the elderly care robot.

[0034] In a fifth aspect, the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned health monitoring methods based on the elderly care robot.

[0035] The health monitoring method based on the elderly care robot provided by the present invention utilizes the high intelligence of the elderly care robot and the powerful text comprehension ability of the large language model to evaluate whether each user's high-risk text reflects depressive symptoms, thereby constructing expert features for each user. In addition, in order to model the user's emotional journey, a text embedding model is used to filter posts with high emotional intensity, and a large language model is used to summarize these texts into descriptions of the user's emotional journey. These high-emotional-intensity texts and emotional journey descriptions are then vectorized to generate a representation of the emotional change process. In order to address the challenge of achieving both high accuracy and high interpretability, features from different spaces are combined and a final judgment is generated using a gradient boosting tree, achieving high accuracy, realizing accurate prediction of depression and reducing prediction costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the health monitoring method based on the elderly care robot provided by the present invention;

[0037] Figure 2 This is a structural diagram of the elderly care robot provided by the present invention;

[0038] Figure 3 A schematic diagram of an electronic device according to an embodiment of the present invention;

[0039] Figure 4 A schematic diagram of an embodiment of a computer-readable storage medium provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0041] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0042] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0043] Reference Figure 1 , Figure 1 This is a flow chart of the health monitoring method based on the elderly care robot provided by the embodiment of the present invention. Figure 1 The health monitoring method based on the elderly care robot provided by the embodiment of the present invention may include steps 10 to 40:

[0044] Step 10: Obtain historical posts published by users on social platforms based on the elderly care robot.

[0045] Step 20: Extract depressive symptom features, emotional process features, and text history features based on historical posts using a preset large language model.

[0046] In the present invention, the possibility of medically diagnosed depression is predicted through expressions in social media posts. Depression symptom features, emotional process features, and text history features are constructed based on the user's historical social media posts, and the user's depressive symptoms are predicted based on the above three features.

[0047] The present invention aims to assess the risk of depression based on an individual's historical posts. Users often use posts to record events around them and express their feelings, making these frank expressions a valuable resource for measuring the likelihood of depression. For a user u, the historical posts that can be collected are denoted as P = {p1, p2, ..., p n}. Each post has a corresponding timestamp, recorded as t1, t2, ..., t n The task is formalized as determining the depression label y of user u, given a post history P, where y∈{0,1}.

[0048] Specifically, depressive symptom features are extracted based on historical posts, including:

[0049] Based on historical posts, the post content and the symptom template are text-embedded through a preset text embedding model; the average similarity between each historical post and the symptom template is calculated, and the historical posts corresponding to the average similarity exceeding the first preset value are annotated through a preset large language model for annotation; the depressive symptom vector is calculated based on the symptom template corresponding to the historical posts annotated by the large language model for annotation, and the depressive symptom characteristics are obtained.

[0050] While established methods for diagnosing depression are already available in the medical field, the potential of this established medical knowledge for detecting depression on social media has yet to be fully explored. By combining these diagnostic criteria and supporting decision-making with medical knowledge, the accuracy and credibility of these decisions can be improved.

[0051] Specifically, user features are constructed based on medical diagnostic criteria. The present invention utilizes DSM-5, a widely recognized tool for the assessment and diagnosis of psychological disorders. It provides a series of clinical criteria for depression, including depressed mood, loss of interest, abnormal changes in weight, abnormal changes in sleep duration, nervous agitation or retardation, irritability, guilt, decreased attention, suicidal thoughts, etc. Large language models have demonstrated excellent semantic understanding capabilities, and studies have shown that they have the potential to replace human annotators in certain tasks. Here, a large language model is used to annotate text, in particular to identify whether and which self-expressed depressive symptoms exist in the post. The prompt words are as follows:

[0052] Imagine you are a psychiatrist specializing in depression. Given a piece of text, determine whether the message contains the following author status:

[0053] A. Depressed mood B. Loss of interest / pleasure... I. Suicidal thoughts.

[0054] If it exists, please answer in the format of enclosed letters, for example, (A,B,C). If it does not exist, please answer None.

[0055] By post-processing the LLM output, a 9-dimensional vector can be generated for each post, where each element is 0 or 1, indicating the absence or presence of a specific depressive symptom. For example, if the output of the large language model annotating a post p is (G, I), then the corresponding vector E p is (0,0,0,0,0,0,1,0,1).

[0056] Since using large language models requires a lot of money and energy costs, and most posts on social media platforms are not related to depression, it would be extremely wasteful to annotate all posts. To solve this problem, the present invention designs an efficient annotation method that first filters high-risk texts and then annotates only these texts. In my experiments, I found that annotating only selected high-risk texts can achieve performance comparable to annotating all texts while greatly reducing the number of requests to LLM. Specifically, for each symptom, a symptom template containing potential first-person text expressions is designed. For example, for the symptom "B. Loss of interest / pleasure", the template yes:

[0057] I lost interest, felt apathetic, bored, unconcerned, lacked enthusiasm, had no motivation, no interest in activities, no motivation, no interest in almost everything, my motivation dropped drastically, I couldn't feel happy, I felt the world was dark, and I couldn't concentrate all day.

[0058] Then, use the text embedding model to get the text embeddings for all symptom templates:

[0059] For i=A to I.

[0060] For each post p, its embedding is calculated as follows:

[0061] H p = encoder (p).

[0062] Next, calculate the average similarity between post p and each symptom template:

[0063] Sim p =mean(Sim(H p ,H i )), for i=A to I.

[0064] Among them, Sim p represents the depression risk level of post p. Only LLM annotations with the highest k%Sim are used p The score of posts, while all other posts have a depressive symptom vector E p Directly set to zero vector. Here k is a hyperparameter. Finally, for each user u, average all depression symptom vectors to obtain the diagnostic standard features

[0065]

[0066] Where N is the total number of posts of user u.

[0067] Extract emotional process features based on historical posts, including:

[0068] Obtain historical posts and embed the post content with the sentiment template using the preset sentiment embedding model;

[0069] Calculating the similarity between all historical posts and each emotion template, obtaining historical posts whose similarity exceeds a second preset value, and forming an emotion set;

[0070] Based on the emotion set, the user's emotional process description is synthesized through a preset large language model for emotion description to obtain the emotional process characteristics.

[0071] The mood course, defined as the temporal pattern and progression of mood states, is crucial for diagnosing clinical depression. It describes the onset, duration, and recurring patterns of mood episodes, providing crucial information about the nature of the illness. Accurately modeling the mood course is crucial for distinguishing depressive disorders from transient mood swings, aiding in early detection and appropriate intervention for depression.

[0072] Previous work on depression detection has largely ignored the emotional journey, focusing instead on static sentiment analysis within a single text. This research addresses this shortcoming by explicitly modeling the emotional journey and integrating it into elderly care robots. The following details the relevant methods:

[0073] Not all social media posts are emotional. First, we screened out posts with high emotional content. Following the classification of previous research, we categorized emotions into five main types: 1) anger, 2) disgust, 3) anxiety, 4) happiness, and 5) sadness. For each of these emotion categories, we created a template for emotional expression. For example, the template for sadness, Defined as:

[0074] "I feel sad, grief, melancholy, pain, loss, depression, pessimism, tears, mourning, melancholy, depression, suicidal, heartbroken, devastated, uneasy, crying, deep sorrow, disappointment, frustration, lamentation, desolate, gloomy, mourning, crying, despair, heartbroken, indignant."

[0075] For each sentiment template, a pre-trained text embedding model is used to generate a representation:

[0076] For j=1,2,…,5.

[0077] For each post p, get its representation H p . Calculate the similarity between post p and each sentiment template:

[0078] For j=1,2,…,5.

[0079] For each sentiment j, retain the posts ranked in the top m% by similarity to form a set Sj Here m is a hyperparameter. The final set of posts S with high sentiment concentration is all S j For each user u, the historical post set P u Take the intersection with the high sentiment concentration post set S and get Based on the subset of emotion-expressing posts, LLM is used to synthesize the emotional journey description T of user u MC , use the following tips:

[0080] “If you were a psychiatrist specializing in treating depression, please summarize the blogger’s emotional journey during this period based on their self-expression:

[0081] Time: t1, post: p1, time: t2, post: p2…”.

[0082] Then calculate T MC Embedding:

[0083] H MC = encoder (T MC ).

[0084] User u's emotional journey representation The calculation is as follows:

[0085]

[0086] in yes The total number of posts in .

[0087] Extract text history features based on historical posts, including:

[0088] Extract text embeddings from historical posts through a pre-set large language model for text extraction;

[0089] Build a historical post collection based on text embeddings and generate text history features.

[0090] In this paper, diagnostic criteria features and emotional journey features are constructed from a user's post history through filtering and tagging. While these filters highlight key criteria for diagnosing depression in medicine, they can also lead to information loss. To address this issue, a representation of a user's post history is constructed as follows:

[0091]

[0092] Among them H p is the embedding of the p-th post, and N is the total number of posts of the user.

[0093] Step 30: Integrate and concatenate the depressive symptom features, emotional process features, and text history features to construct the final feature vector.

[0094] Specifically, it includes: directly integrating and adding the emotional process features and text history features to generate intermediate features;

[0095] The final feature vector is generated by concatenating the intermediate features with the depressive symptom features.

[0096] In the present invention, for F MC (emotional process characteristics) and F PH (Text history features) These two features share the same space and are directly added to avoid increasing the dimension and increasing the risk of overfitting. DC (Depression symptom feature) exists in a different space, so this feature is concatenated with the sum of the first two features. The final feature vector F is obtained as follows:

[0097] F=Concat(F MC +F PH ,F DC ).

[0098] Step 40: Input the final feature vector into the pre-trained prediction model, output the predicted value after multiple iterations of the prediction model, and generate an explanatory output based on the predicted value.

[0099] Specifically, the final feature vector is input into the pre-trained prediction model, and a decision tree set is constructed based on the pre-trained prediction model through gradient boosting tree;

[0100] Each subsequent tree in the decision tree corrects the error of the previous tree, selects the optimal splitting standard in the decision tree, automatically performs feature interaction, completes feature fusion, and outputs the predicted value;

[0101] Generates text-type explanatory output based on the predicted value through a preset large language model for text interpretation.

[0102] Gradient Boosting Decision Tree (GBDT) is a powerful and commonly used algorithm in the field of machine learning. The specific working process includes:

[0103] Initialization: First, GBDT uses a constant (usually the mean of the target variable) as the initial prediction value. This initial prediction value represents the initial guess of the target variable.

[0104] Iterative training: GBDT is an iterative algorithm that gradually improves the model through multiple iterations. In each iteration, GBDT trains a new decision tree with the goal of reducing the residual (or error) of the previous model. The residual is the difference between the actual observation and the current model's prediction, and the new tree will learn how to correct for these residuals.

[0105] Integration: Ultimately, GBDT adds the predictions of all decision trees together to obtain the final integrated prediction. This process enables the model to capture complex relationships in the data, thereby improving prediction accuracy.

[0106] During the initialization process, at the beginning of training, GBDT uses an initial prediction value to represent the average of the entire data. This initial prediction value can be the mean of the target variable or other appropriate initial values. The initial prediction value represents the model's initial estimate of the entire data.

[0107] GBDT is an iterative algorithm that usually includes multiple rounds of training. In each round of training, the model trains a new decision tree whose goal is to reduce the residual error of the previous round of model. The specific steps are as follows:

[0108] Calculate the residual. At the beginning of each iteration, calculate the residual between the current model's predicted value of the training data and the actual observed value. This residual represents the part that the previous round model failed to predict correctly.

[0109] Train a new decision tree using the calculated residuals as the new target variable. This tree will try to correct the errors of the previous model to reduce the residuals.

[0110] Update the model by combining the newly trained decision tree with the previous model. Specifically, add the predictions of the new tree to the predictions of the previous model to get the updated model.

[0111] After multiple rounds of iterative training, GBDT accumulates the predictions of all decision trees to produce a final ensemble prediction. This ensemble process fully leverages the contributions of each tree, enabling the model to better fit the data and improve prediction accuracy. GBDT's working mechanism ensures that the model's predictive power gradually improves, with each iteration attempting to correct errors from the previous round. This ensemble learning approach enables GBDT to excel in various data mining tasks.

[0112] High-precision prediction capability GBDT can handle complex nonlinear relationships and high-dimensional data. It can usually achieve higher accuracy than a single decision tree or linear model in classification and regression tasks.

[0113] GBDT demonstrates the power of ensemble learning. It combines multiple weak learners (typically decision trees) into a powerful model. This ensemble learning approach effectively reduces the model's bias and variance, thereby improving prediction accuracy. Each decision tree focuses on correcting the prediction errors of the previous tree. Through multiple rounds of iteration, the model gradually approaches the distribution of the real data.

[0114] Many practical problems involve data with complex nonlinear relationships, which traditional linear models often struggle to capture. GBDT's nonlinear modeling capabilities enable it to better fit data and improve the model's predictive accuracy. For example, in image recognition tasks, the visual features of objects often exhibit nonlinear variations, and GBDT can effectively capture these features, improving recognition accuracy.

[0115] GBDT is adaptable to various types of data (numerical, categorical, text, etc.) and does not require special preprocessing of the data. This makes it easier to use in practical applications.

[0116] In real-world data mining tasks, mixed data types are often encountered. For example, in the housing price prediction problem, features include numerical data (such as house area and number of bedrooms), categorical data (such as house location and building type), and text data (such as house description). GBDT can directly process this mixed data without converting it into a unified format. This simplifies the data preprocessing step and saves modeling time.

[0117] Unlike some machine learning algorithms, such as support vector machines and neural networks, GBDT does not require feature scaling or normalization. This means that differences in feature scale do not affect model performance. In some algorithms, inconsistent feature scales can prevent the model from learning correctly and require tedious feature scaling. However, GBDT processes raw features directly, reducing the burden of data preprocessing.

[0118] Advantages in the case of data imbalance: In many practical problems, the dataset is often imbalanced, that is, the number of samples in some categories is far less than that in other categories. GBDT performs well in handling imbalanced data.

[0119] The weighted loss function used by GBDT allows different weights to be assigned to samples of different categories. This allows the model to focus more on the minority class, improving its ability to handle imbalanced data. For example, in medical disease diagnosis tasks, diseased samples are typically far fewer than healthy samples. By assigning higher weights to diseased samples, GBDT can better identify disease conditions and improve diagnostic accuracy.

[0120] GBDT's iterative training method gradually corrects errors made by the model in the previous round. When dealing with imbalanced data, the model often focuses on difficult-to-classify minority class examples over multiple rounds. By gradually correcting these errors, the model gradually improves its ability to classify the minority class, thereby improving prediction results.

[0121] In this paper, we use the gradient boosted tree (GBT) method for classification. This method constructs a collection of decision trees in a sequential manner, where each subsequent tree is designed to correct the errors of its predecessor. It automatically performs feature interaction by selecting the optimal splitting criterion within its decision tree, thereby effectively fusing the components of F.

[0122] The final prediction y is the binary classification result derived from the ensemble model. The process is as follows:

[0123] First, the ensemble model G is initialized, and then the prediction ability is gradually enhanced by adding decision trees:

[0124] G m (x) = G m-1 (x)+v·h m (x).

[0125] Each tree h m (x) Adapted to G m-1 The negative gradient of the loss function evaluated, which aims to minimize:

[0126]

[0127] The prediction y is the output G after M iterations M The sign of (x) is given by:

[0128] y=sign(G M (x)).

[0129] Here, L represents the loss function, N is the number of samples, M is the total number of trees, and ν is the learning rate.

[0130] Depression detection is a critical and safety-sensitive task that requires accuracy and interpretability. Several research papers have demonstrated the potential of large language models for interpretable mental health analysis. However, directly using large language models to analyze raw text often results in less accurate classification than domain-specific models, and the predictions are highly sensitive to cue words and lack stability. To ensure accuracy and stability, our approach avoids relying on the large language model for direct judgment and instead uses a traditional classifier. However, we also leverage the power of the large language model to provide reasonable explanations for the model's decisions.

[0131] In this invention, a large language model is used to annotate texts posted by users on social media platforms that show symptoms of depression. In addition, a description of the user's emotional journey is generated. These outputs not only serve as intermediate variables in the operation of the elderly care robot, but also as part of the final output, enhancing the user's understanding of the elderly care robot's results. In addition, a new output T is generated using the large language model. Exp , providing explanations for the model’s classification results. The prompt words used are as follows:

[0132] Imagine you are a psychiatrist specializing in treating depression. Here is a user's emotional journey: MC ; The following are posts showing this user exhibiting depressive symptoms and the types of symptoms they exhibit: ...; This user has been judged as depressed / normal by the elderly care robot. Please consider the user's emotional journey and posts to generate an explanation for this judgment. This explanatory output T Exp , together with the elderly care robot's annotations on depression symptoms in tweets, the user's emotional journey description, and the elderly care robot's classification results, constitute the final output.

[0133] In a specific embodiment, cosine similarity is used to calculate the similarity between embeddings. In order to improve the usability of the method in a low computing resource environment, a low-resource pre-trained embedding model gte-small-zh is used, which only requires 1GB of memory to run smoothly. The embedding model in the elderly care robot can be easily switched to other high-performance models to further improve performance. For the large language model (LLM), GPT-3.5-Turbo-1103 is used, which only requires an Internet connection to interact through the API service provided by OpenAI. The LLM in the elderly care robot can also use open source models, such as Mentallama, to further reduce financial costs. In fact, if a user wants to use the present invention to detect depression based on social media text, he only needs to obtain the calling permission or capability of any large language model that can be used for conversation, such as the API account and private key of OpenAI's ChatGPT, and then he can use the Python simulation code written by the present invention to complete the desired function. In general, the elderly care robot can run at a low computing cost, enhancing its usability. XGBoost is used to efficiently implement the gradient boosting tree.

[0134] The embodiment of the present invention utilizes the high intelligence of the intelligent elderly care robot and the powerful text comprehension capabilities of the large language model to evaluate whether each user's high-risk text reflects depressive symptoms, thereby constructing expert features for each user. In addition, in order to model the user's emotional journey, a text embedding model is used to filter posts with high emotional intensity, and a large language model is used to summarize these texts into a description of the user's emotional journey. These texts with high emotional intensity and emotional journey descriptions are then vectorized to generate a representation of the emotional change process. In order to address the challenge of achieving both high accuracy and high interpretability, features from different spaces are combined and a final judgment is generated using a gradient boosting tree, achieving high accuracy, realizing accurate prediction of depression and reducing prediction costs.

[0135] The elderly care robot provided by the present invention is described below. The elderly care robot described below and the health monitoring method based on the elderly care robot described above can be referenced to each other.

[0136] like Figure 2 As shown, Figure 2 This is a structural diagram of the elderly care robot provided by the present invention. The elderly care robot includes a robot control center, a feature acquisition module, a feature extraction module, a feature aggregation module and a prediction module. The robot control center is connected to the feature acquisition module, the feature extraction module, the feature aggregation module and the prediction module respectively to control and manage each module.

[0137] Feature acquisition module 201, used to acquire historical posts published by users on social platforms based on the elderly care robot;

[0138] A feature extraction module 202 is configured to extract depression symptom features, emotional process features, and text history features based on historical posts using a preset large language model;

[0139] A feature aggregation module 203 is used to integrate and combine the depression symptom features, emotion process features, and text history features, and generate a final feature vector by constructing a decision tree set;

[0140] The prediction module 204 is configured to input the final feature vector into a pre-trained prediction model, output a prediction value through the prediction model after multiple iterations, and generate an explanatory output based on the prediction value.

[0141] This embodiment uses the high intelligence of the intelligent elderly care robot and the powerful text comprehension capabilities of the large language model to evaluate whether each user's high-risk text reflects depressive symptoms and build expert features for each user. In addition, in order to model the user's emotional journey, a text embedding model is used to filter posts with high emotional intensity, and a large language model is used to summarize these texts into descriptions of the user's emotional journey. These high-emotional intensity texts and emotional journey descriptions are then vectorized to generate a representation of the emotional change process. In order to address the challenge of achieving both high accuracy and high interpretability, features from different spaces are combined and a final judgment is generated using a gradient boosting tree, achieving high accuracy, realizing accurate prediction of depression and reducing prediction costs.

[0142] In one embodiment, the feature extraction module 202 is further configured to:

[0143] Based on historical posts, the post content and symptom templates are embedded into text using a preset text embedding model;

[0144] Calculating the average similarity between each historical post and the symptom template, and annotating the historical posts corresponding to the average similarity exceeding a first preset value using a preset large language model for annotation;

[0145] The depression symptom vector is calculated based on the symptom template corresponding to the historical posts annotated by the large language model used for annotation, and the depression symptom features are obtained.

[0146] In one embodiment, the feature extraction module 202 is further configured to:

[0147] Obtain historical posts and embed the post content with the sentiment template using the preset sentiment embedding model;

[0148] Calculating the similarity between all historical posts and each emotion template, obtaining historical posts whose similarity exceeds a second preset value, and forming an emotion set;

[0149] Based on the emotion set, a user's emotion process description is synthesized through a preset large language model for emotion description to obtain emotion process features.

[0150] In one embodiment, the feature extraction module 202 is further configured to:

[0151] Extract text embeddings from historical posts through a pre-set large language model for text extraction;

[0152] A historical post collection is constructed based on the text embedding amount to generate text history features.

[0153] In one embodiment, the feature aggregation module 203 is further configured to:

[0154] Directly integrating and adding the emotional process features and the text history features to generate intermediate features;

[0155] The intermediate features are concatenated with the depressive symptom features to generate a final feature vector.

[0156] In one embodiment, the prediction module 204 is further configured to:

[0157] Inputting the final feature vector into a pre-trained prediction model, and constructing a decision tree set based on the pre-trained prediction model through a gradient boosting tree;

[0158] Select the optimal splitting criterion within the decision tree set to automatically perform feature interaction, complete feature fusion, and output the predicted value;

[0159] Based on the prediction value, an explanatory output of the text type is generated by a preset large language model for text interpretation.

[0160] See also Figure 3 , Figure 3 Schematic diagram of an embodiment of an electronic device provided by an embodiment of the present invention. Figure 3 As shown, an embodiment of the present invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, the following steps are implemented:

[0161] Obtain historical posts published by users on social platforms based on the elderly care robot;

[0162] Based on historical posts, a preset large language model is used to extract depressive symptom features, emotional process features, and text history features;

[0163] Integrate and concatenate the depressive symptom features, emotional process features, and text history features to construct a final feature vector;

[0164] The final feature vector is input into a pre-trained prediction model, and the prediction model outputs a prediction value after multiple iterations, and an explanatory output is generated based on the prediction value.

[0165] See also Figure 4 , Figure 4 Schematic diagram of an embodiment of a computer-readable storage medium provided in an embodiment of the present invention. Figure 4 As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 411 is stored. When the computer program 411 is executed by a processor, the following steps are implemented:

[0166] Obtain historical posts published by users on social platforms based on the elderly care robot;

[0167] Based on historical posts, a preset large language model is used to extract depressive symptom features, emotional process features, and text history features;

[0168] Integrate and concatenate the depressive symptom features, emotional process features, and text history features to construct a final feature vector;

[0169] The final feature vector is input into a pre-trained prediction model, and the prediction model outputs a prediction value after multiple iterations, and an explanatory output is generated based on the prediction value.

[0170] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0171] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0172] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0173] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0174] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0175] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0176] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A health monitoring method based on an elderly care robot, characterized in that: include: Obtain historical posts published by users on social platforms based on the elderly care robot; Based on historical posts, a preset large language model is used to extract depressive symptom features, emotional process features, and text history features; Integrate and concatenate the depressive symptom features, emotional process features, and text history features to construct a final feature vector; Inputting the final feature vector into a pre-trained prediction model, outputting a predicted value through the prediction model after multiple iterations, and generating an explanatory output based on the predicted value; The extraction of depressive symptom features based on historical posts specifically includes: Based on historical posts, the post content and symptom templates are embedded into text using a preset text embedding model; Calculating the average similarity between each historical post and the symptom template, and annotating the historical posts corresponding to the average similarity exceeding a first preset value using a preset large language model for annotation; Calculate the depression symptom vector based on the symptom template corresponding to the historical posts annotated by the large language model used for annotation, and obtain the depression symptom features; Use the text embedding model to get text embeddings for all symptom templates: For i = A to I; For each post p, its embedding is calculated as follows: H p = encoder(p); Calculate the average similarity between post p and each symptom template: Sim p =mean(Sim(H p ,H i )), for i=A to I; Among them, Sim p represents the depression risk level of post p; using LLM annotations with the highest k%Sim p Score posts, all other posts depressive symptoms vector E p Directly set to zero vector; k is a hyperparameter; for each user u, average all depression symptom vectors to obtain the diagnostic standard features Where N is the total number of posts of user u; The extraction of emotional process features based on historical posts specifically includes: Obtain historical posts and embed the post content with the sentiment template using the preset sentiment embedding model; Calculating the similarity between all historical posts and each emotion template, obtaining historical posts whose similarity exceeds a second preset value, and forming an emotion set; Based on the emotion set, a user's emotion process description is synthesized using a preset large language model for emotion description to obtain emotion process features; For each sentiment template, a pre-trained text embedding model is used to generate a representation: For j = 1, 2, ..., 5; For each post p, get its representation H p ; Calculate the similarity between post p and each sentiment template: For j = 1, 2, ..., 5; For each sentiment j, retain the posts ranked in the top m% by similarity to form a set S j ; m is a hyperparameter; the final set of posts S with high sentiment concentration is all S j The union of; For each user u, the historical post set P u Take the intersection with the high sentiment concentration post set S and get Based on the subset of emotion-expressing posts, LLM is used to synthesize the emotional journey description T of user u MC ; Calculate T MC Embedded: H MC = encoder (T MC ); User u's emotional journey representation The calculation is as follows: in yes Total number of posts in; The integration and splicing of the depression symptom features, the emotional process features, and the text history features to construct a final feature vector specifically includes: Directly integrating and adding the emotional process features and the text history features to generate intermediate features; Generate a final feature vector based on the concatenation of the intermediate features and the depressive symptom features; Constructed a representation of the user's post history: Among them H p is the embedding of the p-th post, and N is the total number of posts of the user; Emotional process characteristics F MC and text history features F PH are features that share the same space, the depressive symptom feature F DC Existing in different spaces, the final eigenvector F is obtained as follows: F=Concat(F MC +F PH ,F DC )。 2. The health monitoring method based on the elderly care robot according to claim 1 is characterized in that: The extraction of text history features based on historical posts specifically includes: Extract text embeddings from historical posts through a pre-set large language model for text extraction; A historical post collection is constructed based on the text embedding amount to generate text history features.

3. The health monitoring method based on the elderly care robot according to claim 1 is characterized in that: The final feature vector is input into a pre-trained prediction model, and the prediction model outputs a prediction value after multiple iterations, and an explanatory output is generated based on the prediction value, specifically including: Inputting the final feature vector into a pre-trained prediction model, and constructing a decision tree set based on the pre-trained prediction model through a gradient boosting tree; Select the optimal splitting criterion within the decision tree set to automatically perform feature interaction, complete feature fusion, and output the predicted value; Based on the prediction value, an explanatory output of the text type is generated by a preset large language model for text interpretation.

4. A robot for elderly care, characterized in that: The elderly care robot includes a robot control center, a feature acquisition module, a feature extraction module, a feature aggregation module and a prediction module. The robot control center is connected to the feature acquisition module, the feature extraction module, the feature aggregation module and the prediction module respectively to control and manage each module; The feature acquisition module is used to obtain historical posts published by users on social platforms based on the elderly care robot; A feature extraction module is used to extract depressive symptom features, emotional process features, and text history features based on historical posts using a preset large language model; A feature aggregation module is used to integrate and splice the depression symptom features, emotional process features, and text history features, and generate a final feature vector by constructing a decision tree set; A prediction module, configured to input the final feature vector into a pre-trained prediction model, output a predicted value after multiple iterations of the prediction model, and generate an explanatory output based on the predicted value; The extraction of depressive symptom features based on historical posts specifically includes: Based on historical posts, the post content and symptom templates are embedded into text using a preset text embedding model; Calculating the average similarity between each historical post and the symptom template, and annotating the historical posts corresponding to the average similarity exceeding a first preset value using a preset large language model for annotation; Calculate the depression symptom vector based on the symptom template corresponding to the historical posts annotated by the large language model used for annotation, and obtain the depression symptom features; Use the text embedding model to get text embeddings for all symptom templates: For i = A to I; For each post p, its embedding is calculated as follows: H p = encoder(p); Calculate the average similarity between post p and each symptom template: Sim p =mean(Sim(H p ,H i )), for i=A to I; Among them, Sim p represents the depression risk level of post p; using LLM annotations with the highest k%Sim p Score posts, all other posts depressive symptoms vector E p Directly set to zero vector; k is a hyperparameter; for each user u, average all depression symptom vectors to obtain the diagnostic standard features Where N is the total number of posts of user u; The extraction of emotional process features based on historical posts specifically includes: Obtain historical posts and embed the post content with the sentiment template using the preset sentiment embedding model; Calculating the similarity between all historical posts and each emotion template, obtaining historical posts whose similarity exceeds a second preset value, and forming an emotion set; Based on the emotion set, a user's emotion process description is synthesized using a preset large language model for emotion description to obtain emotion process features; For each sentiment template, a pre-trained text embedding model is used to generate a representation: For j = 1, 2, ..., 5; For each post p, get its representation H p ; Calculate the similarity between post p and each sentiment template: For j = 1, 2, ..., 5; For each sentiment j, retain the posts ranked in the top m% by similarity to form a set S j ; m is a hyperparameter; the final set of posts S with high sentiment concentration is all S j The union of; For each user u, the historical post set P u Take the intersection with the high sentiment concentration post set S and get Based on the subset of emotion-expressing posts, LLM is used to synthesize the emotional journey description T of user u MC ; Calculate T MC Embedded: H MC = encoder (T MC ); User u's emotional journey representation The calculation is as follows: in yes Total number of posts in; The integration and splicing of the depression symptom features, the emotional process features, and the text history features to construct a final feature vector specifically includes: Directly integrating and adding the emotional process features and the text history features to generate intermediate features; Generate a final feature vector based on the concatenation of the intermediate features and the depressive symptom features; Constructed a representation of the user's post history: Among them H p is the embedding of the p-th post, and N is the total number of posts of the user; Emotional process characteristics F MC and text history features F PH are features that share the same space, the depressive symptom feature F DC Existing in different spaces, the final eigenvector F is obtained as follows: F=Concat(F MC +F PH ,F DC )。 5. An electronic device, characterized in that: include: Memory for storing computer software programs; A processor is used to read and execute the computer software program, thereby implementing the health monitoring method based on the elderly care robot as described in any one of claims 1 to 3.

6. A non-transitory computer-readable storage medium, characterized in that The storage medium stores a computer software program, which, when executed by a processor, implements the health monitoring method based on the elderly care robot according to any one of claims 1 to 3.

7. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the health monitoring method based on the elderly care robot as described in any one of claims 1 to 3 is implemented.

Images