Psychological health assessment and intervention method and system for cancer patients

Abstract

The application discloses a cancer patient psychological health evaluation and intervention method and system, and the method comprises the following steps: acquiring community interaction behavior data and clinical diagnosis and treatment characteristic data of a target user on a platform; performing natural language processing and voice emotion recognition on the community interaction behavior data, and extracting original emotion characteristics; generating a current physiological state baseline and a treatment side effect expected window of the target user based on the clinical diagnosis and treatment characteristic data; inputting the original emotion characteristics into a joint analysis model, weighting and correcting original emotion characteristic values by using the physiological state baseline and the treatment side effect expected window, and calculating a psychological health index; comparing the dynamic psychological health index with a preset risk threshold, determining a psychological health risk level of the user, and triggering a corresponding grading intervention mechanism. The application can accurately distinguish between physiological distress and pathological depression, and realizes dynamic monitoring and accurate management of the psychological state of a cancer patient in an off-hospital rehabilitation period.

Description

Technical Field

[0001] This invention belongs to the field of Internet medical technology, and specifically relates to methods and systems for assessing and intervening in the mental health of cancer patients. Background Technology

[0002] Cancer, a serious disease that severely threatens human health, often causes immense psychological stress for patients during diagnosis and the long recovery process, making them highly susceptible to anxiety, depression, and other mental health issues. Studies have shown that a positive mental state can help improve immune function, alleviate treatment side effects, and prolong survival. Therefore, timely and effective psychological health assessment and intervention for cancer patients is an indispensable part of comprehensive cancer treatment.

[0003] Traditional methods of patient psychological assessment typically rely on manual observation and evaluation by physicians and mental health professionals. While effective in some cases, these methods have limitations. First, they often require extensive manual intervention, depending on the experience of healthcare workers, potentially leading to subjectivity and inconsistencies in the results. Second, traditional methods are often limited to a single assessment dimension (e.g., relying solely on interviews or questionnaires), failing to comprehensively evaluate the patient's psychological state. Third, because a patient's psychological state changes continuously with disease progression and treatment, traditional methods often struggle to monitor these changes in real time. This approach exhibits significant lag and subjectivity, making it difficult to cover the long recovery period outside the hospital.

[0004] With the development of information technology, the use of auxiliary systems for psychological monitoring has become a research hotspot. For example, Chinese invention patent application CN120823969A discloses a psychological health assessment and intervention system for lung cancer patients based on a virtual platform. This system includes a virtual environment creation module, a psychological state monitoring module, a psychological health assessment module, and a psychological health intervention module. It creates an immersive environment using virtual reality (VR) technology, simulating hospital and rehabilitation scenarios to induce emotions; integrates biosensors to monitor patients' physiological responses in real time; combines speech recognition and motion capture technologies to obtain comprehensive data; and utilizes machine learning to build models to identify emotion types. Finally, it provides psychological guidance through interaction with virtual characters. This solution, to a certain extent, addresses the problem of traditional assessment methods lacking objective data support, and enhances the dimensions of assessment through the fusion of multimodal data.

[0005] The above scheme still has the following significant limitations in practical application: negative emotions of cancer patients often have two sources, one is pathological emotional disorders, and the other is physiological pain caused by treatment side effects. The above scheme also lacks the ability to decouple the complex physiological-psychological relationship, resulting in a high false alarm rate.

[0006] Therefore, there is an urgent need for a method for assessing and intervening in the mental health of cancer patients that can deeply integrate clinical medical data and daily behavioral data, effectively eliminate the interference of physiological factors, and provide a closed-loop service throughout the entire process. Summary of the Invention

[0007] This invention provides a method and system for assessing and intervening in the mental health of cancer patients, aiming to solve the problems of existing technologies being unable to effectively distinguish between physiological pain caused by treatment side effects and pathological psychological and emotional disorders when assessing the mental state of cancer patients, resulting in a high false alarm rate for psychological crisis warnings and a lack of targeted intervention measures.

[0008] To address the aforementioned technical problems, this invention proposes a method for assessing and intervening in the mental health of cancer patients, applied to an online rehabilitation management platform for cancer patients, comprising the following steps:

[0009] Acquire target users' community interaction behavior data and clinical diagnosis and treatment characteristic data on the platform;

[0010] Natural language processing and speech emotion recognition are performed on the community interaction behavior data to extract the original emotional features; based on the clinical diagnosis and treatment feature data, a baseline of the target user's current physiological state and a treatment side effect expectation window are generated;

[0011] The original emotional characteristics are input into the patho-psychological joint analysis model. The original emotional characteristic values ​​are weighted and corrected using the physiological state baseline and the expected window of treatment side effects, and the mental health index is calculated.

[0012] The dynamic mental health index is compared with a preset risk threshold to determine the user's mental health risk level and trigger a corresponding graded intervention mechanism.

[0013] Preferably, the community interaction behavior data includes the text content posted by users, the acoustic characteristics of voice messages, and the distribution of community activity time; the clinical diagnosis and treatment characteristic data includes cancer type, TNM stage, current treatment plan, medication cycle, and side effect records.

[0014] Preferably, the generation of the expected treatment side effect window specifically includes:

[0015] A pre-built knowledge base of anti-tumor drug side effects is provided, which includes the probability distribution of adverse reactions of various chemotherapy drugs, targeted drugs and immunotherapies under different dosing cycles.

[0016] Based on the target user's current medication regimen and the start time of medication, the target user is calculated to be in a high-incidence period for specific side effects based on the knowledge base, and the high-incidence period is marked as the expected window for treatment side effects.

[0017] Preferably, the pathology-psychology joint analysis model adopts a multimodal deep neural network architecture, including three parallel processing branches and a feature fusion layer:

[0018] The text semantic encoding branch uses a pre-trained BERT model or Transformer model to vectorize and embed the target user's community posts, and extract the text sentiment feature vector.

[0019] The speech acoustic feature branch uses a convolutional neural network combined with a long short-term memory network to extract time-frequency graph features from the target user's speech data and generate a speech emotion feature vector.

[0020] The clinical state mapping branch inputs the target user's structured medical data into a fully connected neural network, mapping it into a high-dimensional physiological stress feature vector;

[0021] The feature fusion layer uses the physiological stress feature vector as the query vector and the text emotion feature vector and the voice emotion feature vector as key-value pairs. It calculates the attention weight of clinical physiological features on emotional expression through a multi-head attention mechanism.

[0022] Preferably, the weighted correction is performed using the following method:

[0023] Obtain the user's pain score or physical condition score on the day they post negative emotional content;

[0024] If the daily pain score exceeds the preset threshold or the physical condition score shows a severe decline in physical fitness, then the negative emotion is determined to be dominated by physiological pain, and an attenuation coefficient is introduced to reduce the weight of the corresponding negative emotion on the mental health index.

[0025] If the daily pain score and physical condition score are both within the normal range, but the emotional characteristic value remains negative, then the negative emotion is determined to be dominated by a pathological emotional disorder, and its influence weight on the mental health index is maintained or increased.

[0026] Preferably, the weighted correction is performed using the following method:

[0027] Based on the user's cancer type, TNM stage, current medication regimen, and time since diagnosis, the platform's historical database is searched and matched with K nearest neighbor users with similar clinical characteristics to construct a homogeneous reference cohort.

[0028] Historical emotional data and side effect records of the homogeneous reference cohort within the same treatment cycle were extracted. Time series regression analysis was used to fit and generate the expected physiological emotional fluctuation curve of the same type of patients under a specific treatment cycle, which is used to reflect the average trend of the change of emotions with physical discomfort under the influence of pharmacokinetics in the same patient group.

[0029] Project the target user's current original emotional feature value onto the expected physiological emotional fluctuation curve, and calculate the deviation between the user's actual emotional value and the expected baseline value; iteratively adjust the weight of the corresponding emotional feature in the mental health index based on a preset step size until the deviation is within a preset tolerance range.

[0030] Preferably, the method for calculating the mental health index is as follows:

[0031]

[0032] In the formula, To assess the mental health index at time t on a given day; These are the normalized scores for negative emotions based on text and voice, respectively; C represents the user's medical compliance score. The somatization symptom score is derived by weighting the numerical pain score and the severity of side effects. This is a correction factor that varies with the treatment cycle; These are the weights after weighted adjustment.

[0033] Preferably, the correction coefficient is calculated as follows:

[0034]

[0035]

[0036] In the formula, is the cumulative fatigue regulator; n is the cumulative number of chemotherapy or immunotherapy courses currently being administered to the patient; The time interval between the current time t and the end of the last administration; The time interval between the current time t and the next scheduled drug administration; is the preset time to peak side effects for this type of drug; m is a morphological factor used to control the length of the fading tail of side effects; k is the anxiety sensitivity coefficient. is the fatigue accumulation constant; A is the baseline coefficient for the drug-induced depressive intensity.

[0037] Preferably, the tiered intervention mechanism specifically includes:

[0038] In response to low-risk areas, the system automatically pushes relevant rehabilitation science information or nutritional support plans based on the user's cancer type.

[0039] In response to medium-risk areas, the system automatically matches patients with rehabilitation mentors or peer support groups for community guidance.

[0040] In response to high-risk zones, an early warning report is generated that includes the user's recent emotional curve and clinical data, and the intervention process of artificial medical experts or psychological experts is triggered.

[0041] In another aspect, the present invention provides a mental health assessment and intervention system for cancer patients, the system being used to implement the method as described in the first aspect of the present invention, comprising:

[0042] The data acquisition module is used to acquire community interaction data and clinical diagnosis and treatment characteristic data of target users on the communication platform;

[0043] The feature processing and baseline generation module is used to perform natural language processing and voice emotion recognition on the community interaction behavior data, extract the original emotion features, and generate the target user's current physiological state baseline and treatment side effect expectation window based on the clinical diagnosis and treatment feature data.

[0044] The mental health assessment module is used to input the original emotional characteristics into the patho-psychological joint analysis model, and to use the physiological state baseline and the expected window of treatment side effects to weight and correct the original emotional characteristic values ​​to calculate the mental health index.

[0045] The graded intervention control module is used to compare the dynamic mental health index with a preset risk threshold to determine the user's mental health risk level and trigger the corresponding graded intervention mechanism.

[0046] The database and knowledge base storage module is used to store the community interaction behavior data, clinical diagnosis and treatment feature data, a pre-built knowledge base of anti-tumor drug side effects, and homogeneous reference queue data of historical users of the platform.

[0047] Compared with the prior art, the present invention has the following technical effects:

[0048] 1. The mental health assessment and intervention method proposed in this invention introduces a treatment side effect expectation window and a physiological baseline, and uses a patho-psychological joint analysis model to weight and correct emotional characteristics. The system can identify and eliminate physiological emotional fluctuations caused by pharmacokinetics, thereby accurately identifying true pathological emotional disorders and effectively reducing the false alarm rate of psychological crisis warnings.

[0049] 2. The dynamic correction coefficient algorithm proposed in this invention's mental health assessment and intervention method combines a biphasic pharmacokinetic-psychological coupling model. It not only considers the metabolic lag effect after a single dose but also introduces a cumulative fatigue modulatory factor, ensuring that the system's evaluation criteria for patient mental health are not static but adaptively adjust as the patient's treatment duration increases and physical function declines. This dynamic modeling, consistent with clinical rehabilitation principles, addresses the limitation of traditional static scales in adapting to long-term cancer rehabilitation monitoring.

[0050] 3. The mental health assessment and intervention method proposed in this invention classifies risks into three levels: low, medium, and high. The system can intelligently allocate medical resources: low-risk users are maintained using low-cost popular science information, while high-risk users are precisely targeted with scarce expert resources. This tiered mechanism greatly improves the operational efficiency of internet healthcare platforms.

[0051] 4. The mental health assessment and intervention method proposed in this invention constructs a homogeneous reference cohort. By comparing an individual's emotions with the expected emotional baseline of a group with the same disease type, it can eliminate background noise caused by specific social events or group empathy. This calibration mechanism based on large-scale group data ensures that the system maintains high robustness and reference value even in out-of-hospital rehabilitation scenarios lacking real-time physician involvement. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the method described in this invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0054] Example 1

[0055] Methods for assessing and intervening in the mental health of cancer patients, applied to online rehabilitation management platforms for cancer patients, such as... Figure 1 As shown, it includes the following steps one through four:

[0056] Step 1: Acquire community interaction behavior data and clinical diagnosis and treatment characteristic data of target users on the platform. In this embodiment, a front-end data acquisition module deployed on the user terminal (such as a smartphone APP or WeChat mini-program) communicates with a back-end data aggregation center deployed on a cloud server. The back-end server adopts a distributed database architecture, including a relational database for storing structured clinical data and a NoSQL database for storing unstructured community interaction data.

[0057] The community interaction data includes the text content posted by users, the acoustic characteristics of voice messages, and the distribution of community activity times. The specific implementation is as follows:

[0058] The system monitors the API interface of the platform's community section to capture text data submitted by users in real time. It then cleans and desensitizes this text data, such as removing HTML tags, special emoticons, and irrelevant advertising links. Furthermore, it uses named entity recognition technology to automatically mask real names, phone numbers, and specific home addresses in the text, retaining only text information related to disease descriptions and emotional expressions.

[0059] When a user posts a voice update in the community or sends a voice message in the consultation section, the audio processing engine is invoked to analyze the audio and extract the following key acoustic features: fundamental frequency (F0) and its fluctuation rate, which reflect the intonation; a flat tone is usually associated with depression; jitter and flicker, which reflect the periodicity and amplitude of vocal cord vibration, respectively, and are used to detect hoarseness and weakness in the voice; speech rate, which is marked as psychomotor retardation if the speech rate is significantly lower than the user's historical baseline.

[0060] The system also records the server timestamps for each time a user opens the application, refreshes a page, or publishes content.

[0061] The clinical diagnostic and treatment characteristic data includes cancer type, TNM stage, current treatment plan, medication cycle, and side effect records. This data determines the medical context of the subsequent evaluation model, and is obtained through a dual-mode approach of proactive data entry and intelligent analysis.

[0062] When users register or complete their profiles, the system provides a structured disease selection tree (ICD-10 encoding mapping). For users unfamiliar with medical terminology, the system integrates an Optical Character Recognition (OCR) module, allowing users to upload photos of discharge summaries or pathology reports. After recognizing the text in the image, the OCR module automatically parses out key fields using a medical entity extraction model, such as:

[0063] {"Cancer_Type":"Lung_Adenocarcinoma","Stage_T":"2","Stage_N":"1","Stage_M":"0"}.

[0064] For tracking dynamic treatment data, the system has a built-in calendar function, allowing users to set their current treatment plan and first-dose date. The backend calculates the following parameters based on the current system time and first-dose date: the current cumulative treatment cycle number n, i.e., how many chemotherapy sessions the user has already undergone; and the current day of the treatment cycle. This is used to calculate subsequent correction coefficients to determine whether the user is in a high-risk window for side effects.

[0065] In addition, data can be collected through daily pop-up mini-questions. For example, pain scores: using the Numerical Rating Scale (NRS) corresponding to the Visual Analogue Scale (VAS), users can slide a slider to enter their daily pain value; another example is adverse reaction grading: based on the symptoms selected by the user, the data is automatically mapped to grading values ​​under the Common Terminology Criteria for Adverse Events (CTCAE) and directly stored in the clinical feature vector database.

[0066] After the above data collection and processing, the system constructs a multimodal feature tensor for each target user that is updated according to the time series. An example of its JSON data structure is as follows:

[0067] {

[0068] "User_ID":"MJ_102488",

[0069] "Timestamp":"2025-05-20T23:45:00Z",

[0070] "Clinical_Context":{

[0071] "Diagnosis":"Breast_Cancer_Her2+",

[0072] "Treatment_Phase":"Chemotherapy_Cycle_3",

[0073] "Days_Post_Chemo":4, / / Corresponding parameter τ

[0074] "Physical_Status":{

[0075] "NRS_Pain_Score":6, / / Pain score

[0076] "CTCAE_Vomiting":2 / / Side effect classification

[0077] }

[0078] },

[0079] "Behavior_Metrics":{

[0080] "Text_Content":"This level is too hard to get through, my whole body aches, and I want to give up."

[0081] "Audio_Features":{

[0082] "Pitch_Variance":0.02, / / flat tone

[0083] "Jitter": 0.15, / / Sound jitter

[0084] "Speech_Rate": 2.1 / / Slow speech rate

[0085] },

[0086] "Activity_Flags":["Late_Night_Active"]

[0087] }

[0088] }

[0089] This step yields all the input variables needed for subsequent mental health assessments.

[0090] After acquiring multi-source data, the system proceeds to step two: performing natural language processing and voice emotion recognition on the community interaction behavior data to extract original emotional features; and generating the target user's current physiological baseline and expected treatment side effects window based on the clinical diagnosis and treatment feature data.

[0091] For feature extraction from community interaction data, the system establishes and runs a parallel deep learning inference pipeline to process text and speech data separately, outputting standardized raw sentiment features. The system loads a BERT (Bidirectional Encoder Representations from Transformers) model pre-trained on a medical vertical domain corpus. The text content posted by the user in step one is input into the BERT model, transforming each sentence into a 768-dimensional sentence vector. A fully connected layer (such as a Softmax classifier) ​​is then connected after the BERT output layer to output the sentiment polarity probability distribution. Simultaneously, keyword matching technology is used to identify high-risk semantic features in the text, generating Boolean feature vectors. The output text sentiment feature value is normalized to [-1, 1], where -1 represents extremely negative and 1 represents positive.

[0092] For acoustic feature extraction of voice messages, the voice signal acquired in step one is subjected to a short-time Fourier transform to generate a Mel-frequency cepstral coefficient spectrogram. Spatial features and temporal pitch variations are extracted from the spectrogram to generate voice emotion feature values.

[0093] The generation of the expected treatment side effects window specifically includes:

[0094] The system backend stores and maintains a pre-built knowledge base of anti-tumor drug side effects. This knowledge base contains the probability distribution of adverse reactions to various chemotherapy drugs, targeted drugs, and immunotherapies under different dosing cycles. Below is an example of data from the knowledge base:

[0095] {

[0096] "Drug_ID":"D_Cisplatin",

[0097] "Side_Effects":[

[0098] {

[0099] "Type":"Nausea_Vomiting",

[0100] "Probability_Distribution":"Gamma_Function", / / Gamma distribution

[0101] "Peak_Day":3, / / Peak peak is reached on the 3rd day after administration.

[0102] "Window_Start":1, / / Starting from day 1

[0103] "Window_End":5, / / Day 5 ends

[0104] "Severity_Weight": 0.9 / / Severity weight

[0105] },

[0106] {

[0107] "Type":"Fatigue" (fatigue)

[0108] "Window_Start":3,

[0109] "Window_End":10

[0110] } ]

[0112] }

[0113] The knowledge base's data comes from drug instructions, NCI (National Cancer Institute) clinical data, and real-world patient data accumulated on the platform.

[0114] Based on the target user's current medication regimen and the start time of medication, the target user is calculated to be in a high-incidence period for specific side effects based on the knowledge base, and the high-incidence period is marked as the expected window for treatment side effects.

[0115] Specifically, the example calculation process is as follows:

[0116] The system reads the user's profile and obtains the current regimen as "cisplatin (D1) + etoposide (D1-D3)" and the most recent first-dose administration time. The date is May 18, 2025. A search of the knowledge base revealed that the emetic window for cisplatin is... That is, from May 19th to May 23rd. Further searching for the bone marrow suppression window of etoposide revealed... That is, from May 25 to June 1.

[0117] The system adds timestamp labels to the user's timeline:

[0118] [2025-05-19, 2025-05-23]: Marked as the expected window for strong ejaculatory side effects;

[0119] [2025-05-25, 2025-06-01]: Marked as the expected window for myelosuppressive fatigue.

[0120] Step 3: Input the original emotional characteristics into the patho-psychological joint analysis model, and use the physiological state baseline and the expected window of treatment side effects to weight and correct the original emotional characteristic values ​​to calculate the mental health index.

[0121] In this embodiment, the pathology-psychology joint analysis model adopts a multimodal deep neural network architecture, including three parallel processing branches and a feature fusion layer:

[0122] The text semantic encoding branch employs a pre-trained BERT or Transformer model to vectorize and embed the target user's community posts, extracting text sentiment feature vectors. The system input is the user post text preprocessed in step two. After processing by BERT's 12-layer Transformer Encoder, the output vector of the [CLS] marker is extracted as the global semantic representation of the sentence, denoted as... This vector not only contains the semantics of sadness, but also implicitly contains the semantics of pain description.

[0123] The speech acoustic feature branch employs a convolutional neural network combined with a long short-term memory network to extract time-frequency graph features from the target user's speech data, generating a speech emotion feature vector; the input is the Mel spectrogram of the speech. A 2D-CNN layer is responsible for extracting texture features in the frequency domain, and a Bi-LSTM layer is responsible for capturing dynamic features in the temporal domain. Finally, a low-dimensional dense vector is output through a fully connected layer, denoted as... .

[0124] The clinical state mapping branch inputs the target user's structured medical data into a fully connected neural network, mapping it into a high-dimensional physiological stress feature vector. The input features include cancer type unique heat encoding, TNM stage value, current treatment side effect CTCAE grade, and the number of days since the last dose. These discrete and continuous variables are concatenated and input into a 3-layer MLP network, which maps them to a latent space vector of the same dimension as the text / speech vector, denoted as . This is used to characterize the user's current level of physical pain.

[0125] The feature fusion layer uses the physiological stress feature vector as the query vector and the text emotion feature vector and voice emotion feature vector as key-value pairs. It calculates the attention weights of clinical physiological features on emotional expression through a multi-head attention mechanism; that is... , Using formulas Calculate attention weights. When When exhibiting extremely high physiological stress, the model automatically lowers its correlation with... The association weights of negative emotional characteristics. That is, the model learned to ignore some negative expressions in the patient's speech when the patient is in extreme physical pain, and to regard it as physiological rather than psychological.

[0126] To further improve the accuracy of the evaluation, the system introduces post-processing corrections based on statistics or rules, in addition to the model output. This invention provides two specific implementation methods for these corrections:

[0127] Method 1: The weighted correction adopts the following method:

[0128] Based on a user's cancer type, TNM stage, current medication regimen, and time since diagnosis, the platform's historical database is searched and matched with K nearest neighbor users who have similar clinical characteristics to construct a homogeneous reference cohort. Specifically, using the KNN (K-Nearest Neighbors) algorithm, with K=500, de-identified data from 500 historical users are retrieved using the label "lung adenocarcinoma + cisplatin chemotherapy + 3rd cycle of treatment".

[0129] Historical emotional data and side effect records of the homogeneous reference cohort within the same treatment cycle were extracted. Time series regression analysis was used to fit and generate the expected physiological emotional fluctuation curve of the same type of patients under a specific treatment cycle, which is used to reflect the average trend of the change of emotions with physical discomfort under the influence of pharmacokinetics in the same patient group.

[0130] Project the target user's current original emotional feature value onto the expected physiological emotional fluctuation curve, and calculate the deviation between the user's actual emotional value and the expected baseline value; iteratively adjust the weight of the corresponding emotional feature in the mental health index based on a preset step size until the deviation is within a preset tolerance range.

[0131] Method 2, the weighted correction adopts the following method:

[0132] Obtain the user's pain score or physical condition score on the day they post negative emotional content;

[0133] If the daily pain score exceeds the preset threshold or the physical condition score shows a severe decline in physical fitness, then the negative emotion is determined to be dominated by physiological pain, and an attenuation coefficient is introduced to reduce the weight of the corresponding negative emotion on the mental health index.

[0134] If the daily pain score and physical condition score are both within the normal range, but the emotional characteristic value remains negative, then the negative emotion is determined to be dominated by a pathological emotional disorder, and its influence weight on the mental health index is maintained or increased.

[0135] After the above corrections, the system outputs the final quantitative index using the following mathematical model. The calculation method for the mental health index is as follows:

[0136]

[0137] In the formula, To assess the mental health index at time t on a given day, a lower score indicates a worse mental state; C represents the normalized score of negative emotion based on text and voice, respectively; C represents the user's medical compliance score calculated based on the on-time medication attendance rate. The somatization symptom score is derived by weighting the numerical pain score and the severity of side effects. This is a correction factor that varies with the treatment cycle; These are the weights after weighted adjustment.

[0138] The correction coefficient is calculated as follows:

[0139]

[0140]

[0141] In the formula, The cumulative fatigue modulator is used to simulate the decline in psychological tolerance caused by long-term treatment; n is the cumulative number of chemotherapy or immunotherapy courses currently being administered to the patient. The time interval between the current time t and the end of the last administration; The time interval between the current time t and the next scheduled drug administration; is the preset time to peak side effects for this type of drug; m is a morphological factor used to control the length of the fading tail of side effects; k is the anxiety sensitivity coefficient. is the fatigue accumulation constant; A is the baseline coefficient for the drug-induced depressive intensity.

[0142] Step 4: Compare the dynamic mental health index with the preset risk threshold to determine the user's mental health risk level and trigger the corresponding graded intervention mechanism.

[0143] This embodiment's risk stratification logic is based on a pre-set risk threshold table derived from clinical psychology scale norms. Let... The value range is normalized to [0, 100], where a higher score represents a higher degree of psychological distress.

[0144] like It was determined to be Level 1: Low risk / psychologically stable period;

[0145] like It is classified as Level 2: Medium Risk / Psychological Fluctuation Period;

[0146] like It was determined to be Level 3: High Risk / Precursor to Crisis Intervention.

[0147] Additionally, keyword detection can be performed; for example, if extreme keywords are detected in the text, regardless of... The value will trigger a direct jump to the Level 3 processing flow.

[0148] In this embodiment, the tiered intervention mechanism specifically corresponds to:

[0149] In the low-risk zone (Level 1), the system automatically pushes relevant rehabilitation information or nutritional support plans based on the user's cancer type. Specifically, it can utilize the recommendation engine, with user tags input as shown in the example below:

[0150] User_Tags={Type:"Lung_Cancer",Stage:"Chemo_Recovery",Symptom:"Loss_of_Appetite"}.

[0151] Retrieve articles or videos with the highest tag matching in the content management system. Example action: automatically updating related articles in the app's homepage recommendation cards. The intervention goal in this area is to address specific rehabilitation pain points, maintain the patient's sense of control, and prevent anxiety.

[0152] In response to the medium-risk zone (Level 2), the system automatically matches users with rehabilitation mentors or patient support groups specializing in the same disease for community guidance. Specifically, the platform can activate a social graph matching algorithm to search the database for typical users who meet the following criteria: suffering from the same type of cancer, with a survival period exceeding a certain number of years (e.g., 5 years), and exhibiting high recent community activity and positive emotions. Additionally, it can search for the most popular disease-based support groups with the same treatment methods. An example action is the system sending a push notification to the target user: "We found a patient with a similar experience to yours, diagnosed 5 years ago, currently in excellent condition. Would you like to invite her to share her recovery experience?" In the community feed, the algorithm enhances the exposure of positive recovery diaries, alleviating users' loneliness through social recognition.

[0153] Responding to the high-risk zone (Level 3) primarily involves two actions: first, generating an early warning report containing the user's recent emotional curve and clinical data, and triggering intervention by a human medical expert or psychologist. In this embodiment, the early warning report is generated by calling a visualization rendering engine to produce an assessment report. The core of the report includes an overlay graph of emotion and disease course, with the X-axis representing the time axis (e.g., the last 30 days) and the Y1 axis representing the mental health index. The Y2 axis is marked by clinical events, such as the day of chemotherapy administration, the day of CT examination, and the peak of the NRS pain score.

[0154] Secondly, the intervention process is triggered. The system sends a work order to the experts signed up on the platform via the WeChat Work API or other work order systems. The work order includes a link to the aforementioned PDF report and a one-click call button for the user.

[0155] In some other embodiments, if the user has signed a patient-doctor data sharing agreement, the system will send an alert to their attending physician's mobile app. If the assessment involves suicide risk, an interactive voice response system can be activated to make a voice call to a pre-registered emergency contact.

[0156] Example 2

[0157] This embodiment is a psychological health assessment and intervention system for cancer patients. The system is used to implement the method described in Embodiment 1, including:

[0158] The data acquisition module is used to acquire community interaction data and clinical diagnosis and treatment characteristic data of target users on the communication platform;

[0159] The feature processing and baseline generation module is used to perform natural language processing and voice emotion recognition on the community interaction behavior data, extract the original emotion features, and generate the target user's current physiological state baseline and treatment side effect expectation window based on the clinical diagnosis and treatment feature data.

[0160] The mental health assessment module is used to input the original emotional characteristics into the patho-psychological joint analysis model, and to use the physiological state baseline and the expected window of treatment side effects to weight and correct the original emotional characteristic values ​​to calculate the mental health index.

[0161] The graded intervention control module is used to compare the dynamic mental health index with a preset risk threshold to determine the user's mental health risk level and trigger the corresponding graded intervention mechanism.

[0162] The database and knowledge base storage module is used to store the community interaction behavior data, clinical diagnosis and treatment feature data, a pre-built knowledge base of anti-tumor drug side effects, and homogeneous reference queue data of historical users of the platform.

[0163] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for assessing and intervening in the mental health of cancer patients, applied to an online rehabilitation management platform for cancer patients, characterized by: Includes the following steps: Acquire target users' community interaction behavior data and clinical diagnosis and treatment characteristic data on the platform; Natural language processing and speech emotion recognition are performed on the community interaction data to extract the original emotional features; Based on the aforementioned clinical diagnostic and treatment feature data, a baseline of the target user's current physiological state and a window for expected treatment side effects are generated. The original emotional characteristics are input into the patho-psychological joint analysis model. The original emotional characteristic values ​​are weighted and corrected using the physiological state baseline and the expected treatment side effects window to calculate the mental health index. The weighting correction adopts the following method: based on the user's cancer type, TNM stage, current medication regimen, and time to diagnosis, K nearest neighbor users with similar clinical characteristics are retrieved and matched in the platform's historical database to construct a homogeneous reference cohort. Historical emotional data and side effect records of the homogeneous reference cohort within the same treatment cycle are extracted. Time series regression analysis is used to fit and generate the expected physiological emotional fluctuation curve of similar patients under a specific treatment cycle, which is used to reflect the average trend of emotional changes with physical discomfort under the influence of pharmacokinetics in similar patient groups. The target user's current original emotional characteristic value is projected onto the expected physiological emotional fluctuation curve to calculate the deviation between the user's actual emotional value and the expected baseline value. The weights of corresponding emotional features in the mental health index are iteratively adjusted based on a preset step size until the deviation is within a preset tolerance range. The mental health index is compared with a preset risk threshold to determine the user's mental health risk level and trigger a corresponding graded intervention mechanism.

2. The method according to claim 1, characterized in that, The community interaction behavior data includes the text content posted by users, the acoustic characteristics of voice messages, and the distribution of community activity time; the clinical diagnosis and treatment characteristic data includes cancer type, TNM stage, current treatment plan, medication cycle and side effect records.

3. The method according to claim 1, characterized in that, The generation of the expected treatment side effects window specifically includes: A pre-built knowledge base of anti-tumor drug side effects is provided, which includes the probability distribution of adverse reactions of various chemotherapy drugs, targeted drugs and immunotherapies under different dosing cycles. Based on the target user's current medication regimen and the start time of medication, the target user is calculated to be in a high-incidence period for specific side effects based on the knowledge base, and the high-incidence period is marked as the expected window for treatment side effects.

4. The method according to claim 1, characterized in that, The pathology-psychology joint analysis model adopts a multimodal deep neural network architecture, including three parallel processing branches and one feature fusion layer: The text semantic encoding branch uses a pre-trained BERT model or Transformer model to vectorize and embed the target user's community posts, and extract the text sentiment feature vector. The speech acoustic feature branch uses a convolutional neural network combined with a long short-term memory network to extract time-frequency graph features from the target user's speech data and generate a speech emotion feature vector. The clinical state mapping branch inputs the target user's structured medical data into a fully connected neural network, mapping it into a high-dimensional physiological stress feature vector; The feature fusion layer uses the physiological stress feature vector as the query vector and the text emotion feature vector and the voice emotion feature vector as key-value pairs. It calculates the attention weight of clinical physiological features on emotional expression through a multi-head attention mechanism.

5. The method according to claim 1, characterized in that, The weighted correction is performed using the following method: Obtain the user's pain score or physical condition score on the day they post negative emotional content; If the daily pain score exceeds the preset threshold or the physical condition score shows a severe decline in physical fitness, then the negative emotion is determined to be dominated by physiological pain, and an attenuation coefficient is introduced to reduce the weight of the corresponding negative emotion on the mental health index. If the daily pain score and physical condition score are both within the normal range, but the emotional characteristic value remains negative, then the negative emotion is determined to be dominated by a pathological emotional disorder, and its influence weight on the mental health index is maintained or increased.

6. The method according to claim 1, characterized in that, The method for calculating the mental health index is as follows: In the formula, To assess the mental health index at time point t of the day; These are the normalized scores for negative emotions based on text and voice, respectively; C represents the user's medical compliance score. The somatization symptom score is derived by weighting the numerical pain score and the severity of side effects. This is a correction factor that varies with the treatment cycle; These are the weights after weighted adjustment.

7. The method according to claim 6, characterized in that, The correction factor is calculated as follows: In the formula, is the cumulative fatigue regulator; n is the cumulative number of chemotherapy or immunotherapy courses currently being administered to the patient; The time interval between the current time t and the end of the last administration; The time interval between the current time t and the next scheduled drug administration; is the preset time to peak drug side effects; m is a morphological factor used to control the length of the side effect's fading tail; k is the anxiety sensitivity coefficient. is the fatigue accumulation constant; A is the baseline coefficient for the drug-induced depressive intensity.

8. The method according to claim 1, characterized in that, The tiered intervention mechanism is specifically as follows: In response to low-risk areas, the system automatically pushes relevant rehabilitation science information or nutritional support plans based on the user's cancer type. In response to medium-risk areas, the system automatically matches patients with rehabilitation mentors or peer support groups for community guidance. In response to high-risk zones, an early warning report is generated that includes the user's recent emotional curve and clinical data, and the intervention process of artificial medical experts or psychological experts is triggered.

9. A psychological health assessment and intervention system for cancer patients, characterized in that, The system is used to implement the method as described in any one of claims 1-8, comprising: The data acquisition module is used to acquire community interaction data and clinical diagnosis and treatment characteristic data of target users on the communication platform; The feature processing and baseline generation module is used to perform natural language processing and voice emotion recognition on the community interaction behavior data, extract the original emotion features, and generate the target user's current physiological state baseline and treatment side effect expectation window based on the clinical diagnosis and treatment feature data. The mental health assessment module is used to input the original emotional characteristics into the patho-psychological joint analysis model, and to use the physiological state baseline and the expected window of treatment side effects to weight and correct the original emotional characteristic values ​​to calculate the mental health index. The graded intervention control module is used to compare the mental health index with a preset risk threshold to determine the user's mental health risk level and trigger the corresponding graded intervention mechanism. The database and knowledge base storage module is used to store the community interaction behavior data, clinical diagnosis and treatment feature data, a pre-built knowledge base of anti-tumor drug side effects, and homogeneous reference queue data of historical users of the platform.

Images