A depression patient interview data scoring method based on large model semantic understanding

Abstract

This invention proposes a scoring method for interview data of depressed patients based on large-scale model semantic understanding, comprising the following steps: collecting interview audio data and scoring it as clinical annotation scores; transcribing and preprocessing the collected audio data; deploying a large language model, inputting the preprocessed text sequence along with three layers of structured cue words into the large language model to generate item-level summary text; dividing the item-level summary text dataset into training, validation, and test sets, and constructing an item-level scoring network model based on a lightweight Transformer; while the model outputs the scores for each item and the total score, the corresponding item-level summary is also output as evidence text along with the results. This invention, through a cascaded inference architecture that coordinates large and small models, confines the large language model to the item-level evidence extraction and structured summary generation stages, while the lightweight small model is responsible for accurate regression scoring, thus decoupling semantic understanding and scoring decisions.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence and clinical scale assessment technology, specifically a scoring method for interview data of depressed patients based on large model semantic understanding. Background Technology

[0002] Depressive disorders are a common mental disorder with a high global burden. In clinical practice, semi-structured interviews combined with scale assessments remain an important basis for depression evaluation. Among these, scales such as the Hamilton Depression Rating Scale (HAMD), which are led by clinicians, are the most widely used. However, their implementation requires professional training, is time-consuming, and is easily influenced by the subjectivity of the rater.

[0003] With the rapid development of artificial intelligence, especially large language models, mental health assessment based on natural language understanding has ushered in a new technological opportunity. Research and evaluation show that large language models have potential in screening, conversational assessment and clinical support. However, existing depression assessment schemes based on large models still have several key technical pain points before clinical implementation. (1) Long conversations lack semantic utilization across rounds. Clinical interviews often span multiple rounds of questions and answers, and clues are scattered. Existing methods tend to cover only some elements and it is difficult to stably align and aggregate scattered answers with scale items. (2) End-to-end scoring by LLM is not interpretable and accurate enough. Directly letting the large model give item scores presents a "black box" decision-making form, which is difficult to provide item-level evidence. This limits clinicians' trust in the results and makes it difficult to achieve the accuracy required for scale-level fine assessment. (3) Lack of modeling of correlation and dependency between items. Some items in the HAMD scale are clinically correlated (such as the three sleep-related items "difficulty falling asleep", "light sleep" and "early awakening", "mental anxiety" and "physical anxiety"). Existing methods mostly treat each item as an independent task and predict them separately, ignoring the semantic dependency between items, resulting in insufficient overall consistency of scores. (4) Lack of score consistency verification and quality control loop. Even in manual assessment, it is necessary to use consistency rules and statistical outlier alerts to find potential misscoring and drift. However, most AI solutions lack a systematic quality control and feedback mechanism for comparison with clinical scores. (5) Insufficient adaptation to Chinese medical scenarios. Professional terminology, doctor-patient dialogue style and interview process in the Chinese context pose additional challenges to model understanding and alignment, requiring targeted evaluation and optimization on Chinese tasks. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a scoring method for patient interview data based on large-scale model semantic understanding. Targeting scenarios involving physician-led rating scales such as the Hamilton Depression Rating Scale, this method comprehensively utilizes automatic Chinese medical speech recognition, large-scale language model semantic understanding and structured summarization based on three-layer structured cue words, and a Transformer item-level scoring network based on multi-task learning to achieve automatic scoring and quality control of doctor-patient interview data. The technical solution provided by this invention is as follows:

[0005] A scoring method for depression patients based on large-scale model semantic understanding includes the following steps:

[0006] Step 1: Collect interview audio data and score it as a clinical annotation score;

[0007] Step 2: The collected audio data is transcribed using an automatic speech recognition module optimized for Chinese medical scenarios. The transcribed text is then de-identified and segmented into doctor's turns and patient's turns to form a structured text sequence.

[0008] Step 3: Deploy the large language model. Input the text sequence that has been de-identified and processed by turn-taking along with the three-layer structured prompts into the large language model to generate item-level summary text. The three-layer structured prompts include a symptom definition layer, a consultation pattern layer, and an expert rule layer.

[0009] Step 4: Divide the item-level summary text dataset into training, validation and test sets, construct an item-level scoring network model based on lightweight Transformer multi-task learning, train the scoring network model through multi-task joint loss function, and select the model with the smallest loss on the validation set as the deployment version.

[0010] Step 5: Input the text to be scored into the trained scoring network model. The model outputs the score for each item and the total score, and outputs the corresponding item-level summary as evidence text along with the results to generate a report.

[0011] Preferably, the symptom definition layer provides a clear clinical definition and scoring boundary description for each symptom item; the consultation pattern layer summarizes the typical consultation patterns of psychiatrists for each symptom item and embeds these patterns into prompt words as identification anchors for diagnostic signals; the expert rule layer corrects common symptom classification errors in the form of explicit rules.

[0012] Preferably, the large language model is deployed as an in-hospital service or a private network API, with access permissions and auditing mechanisms set. During decoding, low temperature and appropriate top-p parameters are used to reduce output fluctuations, and a JSON Schema validation and automatic retry module is integrated on the server side.

[0013] Preferably, the server performs structured validation on the output item-level summary, fills in missing fields, merges and deduplicates results from different blocks, and finally summarizes them by item and stores them in JSON format.

[0014] Preferably, the scoring network includes the following modules:

[0015] Shared Semantic Encoder: An L-layer Transformer encoder is used as a shared backbone network to learn deep semantic representations of the input item-level summary sequence. The output of the shared semantic encoder is the context-aware semantic vector of each item, denoted as... Where K is the number of entries;

[0016] Inter-item interaction attention layer: An inter-item interaction attention layer is introduced after the shared encoder to explicitly model the semantic dependencies between items;

[0017] Item-level rating headers: An independent rating header is set for each item, and each rating header consists of two fully connected layers (…). and It consists of a semantic vector and a ReLU activation function, which maps the semantic vector of the corresponding item to the predicted score of that item. ;

[0018] Total Score Aggregation Module: The total score is obtained by directly summing the predicted scores of each item. .

[0019] Preferably, the scoring network model employs an embedding layer with an input embedding dimension of 768, followed by a 4-layer shared Transformer encoder. Each layer contains 8 attention heads and a feedforward network with a hidden dimension of 2048. Based on the output of the shared encoder layer, a 2-layer inter-item interaction attention layer is introduced to process the semantic vector sequence of the items. Multi-head self-attention modeling and residual fusion are performed to obtain semantic vectors of inter-entity dependencies between fused entries. At the output end, an independent two-layer fully connected regression head is set for each item, the hidden dimension of the item scoring head is 256, and a Dropout rate of 0.1 is uniformly used in the network.

[0020] Preferably, the joint loss function for multiple tasks is:

[0021]

[0022]

[0023] Where K is the number of items and B is the mini-batch size. Let be the predicted score of the nth sample in the jth entry. This corresponds to the clinical label score.

[0024] Preferably, the scoring network model is trained using a multi-task joint loss function, and the Adam optimizer is used for parameter updates. Learning rate decay and early stopping strategies are introduced to ensure training stability and generalization performance. The specific formula is as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] in, Denotes the model parameters in the t-th iteration. The learning rate; Indicates in loss function The gradient; and These are the exponential moving averages of the first and second momentum of the gradient, respectively. This represents element-wise multiplication; and This is the correction amount for the momentum deviation; Control the decay of first-order and second-order momentum separately. To avoid the denominator being zero, the numerical stability constant is used. Achieve adaptive step size control.

[0032] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0033] This invention designs a three-layer structured cue word framework, organically combining a symptom definition layer, a consultation mode layer, and an expert rule layer. It explicitly injects clinical knowledge into the item-level summary generation process of a large language model, solving problems such as cross-item symptom attribution ambiguity and complex symptom splitting. The item-level scoring network based on multi-task learning treats each item's score as an independent but related sub-task. It models inter-item dependencies through an inter-item interaction attention layer and uses the sum of MAE values ​​for each item as the joint loss function, improving the overall consistency and clinical rationality of the scoring. The large language model is limited to item-level evidence extraction and structured summary generation, with a lightweight Transformer model responsible for accurate regression scoring, achieving decoupling of semantic understanding and scoring decision-making. By outputting item-level summaries as evidence text along with the scoring results, the traditional "black box" scoring is transformed into interpretable and traceable text evidence-driven decision-making. Attached Figure Description

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0035] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] The proposed scoring method for depression patient interview data based on large-model semantic understanding is applicable to various psychiatric peer-rating scale scenarios that primarily rely on interviews. Specifically, in hospital outpatient and inpatient settings, this method can be embedded into a doctor's workbench system to automatically transcribe, generate item-level summaries, and score doctor-patient interview audio. It also generates a review report containing item scores, total scores, and supporting evidence text, allowing clinicians to quickly review and archive the data, thereby reducing doctors' workload and improving assessment efficiency and scoring consistency. In internet hospitals and follow-up centers, this method can automatically generate scale scoring results and review reports during remote audio consultations or follow-ups, assisting in remote diagnosis and treatment and disease tracking.

[0038] In quality control and research scenarios, the method of this invention can perform consistency assessment and drift monitoring of scoring results from large-scale clinical data, providing data support for the standardization of scoring within institutions and the iteration of AI models. Through batch analysis of the differences between model scores and clinical scores, the impact of changes in scoring standards and data distribution on model performance can be identified, and corresponding calibration strategies can be developed accordingly. In educational and training scenarios, the item-level evidence text can also serve as teaching examples, helping resident physicians and visiting physicians learn standardized questioning methods and scoring criteria, thereby improving their scale assessment capabilities.

[0039] Because this invention uses item semantic templates and item-level question-and-answer as key intermediate representations, it can be easily transferred to various other-rating scale assessment scenarios, such as anxiety, bipolar disorder, and negative symptoms of schizophrenia, simply by replacing specific scale items and their labeled data. It possesses good versatility and scalability, providing a unified AI-assisted scoring and quality control technical framework for the scaled assessment of various mental illnesses, and has broad prospects for clinical application and industrial transformation. To make the above-mentioned objectives, features, and effects of this invention more apparent and understandable, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Example 1: A scoring method for interview data of depressed patients based on large model semantic understanding, such as... Figure 1 As shown, it includes the following steps:

[0041] Step 1: Collect interview audio data and score it as a clinical annotation score. In the specific implementation process, this invention first collects doctor-patient interview audio data in real clinical scenarios, under the premise of ethical approval and informed consent from the subjects. The audio is mono, with a sampling rate of 16kHz, and the duration of each case is approximately 20-40 minutes. In this embodiment, a total of 383 interview audios were collected, and two qualified psychiatrists completed item-level scoring according to HAMD-17. When the difference between the two scores is greater than 1 point, a third senior physician arbitrates the score to form a clinical annotation score consisting of an item score and a total score.

[0042] Step 2: Audio Data Transcription and Preprocessing. First, an automatic speech recognition module optimized for Chinese medical scenarios is used to transcribe the collected audio data. This module incorporates a medical lexicon, abbreviation list, and error correction dictionary during the language modeling stage, prioritizing the coverage of key information such as disease names, drug names, and scale terms to improve the transcription accuracy of medical-related terminology. Then, the transcribed text undergoes de-identification processing, replacing or masking direct or indirect identity indicators such as names, addresses, contact information, ID numbers, and medical institutions, while preserving timestamps and sentence boundary information. Finally, based on speaker recognition results and pause features, the text is segmented into doctor turns and patient turns, forming a structured sequence. If necessary, acoustic adaptation can be performed based on on-site noise, accents, and other conditions, without affecting the implementation of the main process.

[0043] Step 3 involves using a three-layer structured cue word framework to constrain the large language model, and then deploying the large language model to generate item-level summary text. The three-layer structured cue word framework includes a symptom definition layer, a consultation pattern layer, and an expert rule layer, to achieve accurate capture and item-level semantic alignment of scattered symptom information in long-term doctor-patient dialogues.

[0044] Symptom Definition Layer: For each assessment item in the HAMD scale, this invention provides a clear clinical definition and scoring boundary description for each symptom item. The symptom definition layer sets the following for each item:

[0045] Core meaning of symptoms: Clearly define the clinical manifestations of the symptom, such as defining "difficulty falling asleep" as "the time from preparing to fall asleep to actually falling asleep exceeds 30 minutes, or the subjective feeling of difficulty falling asleep"; Severity grading criteria: According to the scale rating scale (e.g., 0-4 points or 0-2 points), give a description of the typical manifestations corresponding to each level; Exclusion criteria: Clearly define the boundaries between this symptom and similar symptoms to avoid confusion.

[0046] Consultation Pattern Layer: This invention analyzes a large amount of real clinical interview data to summarize and categorize typical consultation patterns used by psychiatrists for various symptom items, and embeds these patterns into prompt words as identification anchors for diagnostic signals. The consultation pattern layer includes:

[0047] Direct inquiry mode: such as for sleep disorders, "How has your sleep been lately? What time do you go to bed and wake up at night? Do you wake up during the night?"; follow-up questioning mode: such as for symptom frequency and duration, "How long has this been going on? How many days a week does this happen?"; situational trigger mode: such as for anxiety symptoms, "What things make you particularly nervous or worried? Do you feel any physical discomfort?"; negative confirmation mode: such as, "Besides XX, are there any other problems you are experiencing?"

[0048] Expert Rule Layer: This invention incorporates the experiential knowledge of psychiatric clinical experts to correct and constrain common symptom classification errors through explicit rules. The expert rule layer primarily addresses the following issues:

[0049] Ambiguity in symptom attribution across categories: For example, it is clearly stipulated that "urinary system problems (such as urinary frequency and urgency)" should be classified under the "somatic anxiety" category rather than the "reproductive system symptoms" category; "palpitations and chest tightness" should be classified under the "somatic anxiety" category rather than the "hypochondria" category; rules for splitting compound symptoms: For example, when a patient describes "loss of appetite and weight loss" at the same time, it should be mapped to the "gastrointestinal symptoms" and "weight loss" categories respectively; time window constraints: The assessment time window for each symptom should be clearly defined, and symptom descriptions outside the time range should be marked rather than directly included in the score; severity calibration rules: Standardized mapping rules for patients using vague degree terms (such as "somewhat", "occasionally", "frequently").

[0050] The three-layer structured cue word framework, through top-down hierarchical constraints, explicitly injects clinical knowledge into the reasoning process of the large language model, enabling it to possess symptom identification and classification capabilities similar to professional assessors when generating item-level summaries, thereby improving the accuracy and stability of cross-round information aggregation. The specific design of the three-layer structured cue words is as follows, using the item "somatic anxiety" as an example:

[0051] Example of symptom definition layer:

[0052] Somatic anxiety is defined as the physical manifestations of anxiety, including but not limited to:

[0053] Cardiovascular system symptoms: palpitations, rapid heartbeat, chest tightness;

[0054] Respiratory system symptoms: difficulty breathing, shortness of breath, chest tightness;

[0055] Gastrointestinal system symptoms: dry mouth, bloating, diarrhea, hiccups;

[0056] Urinary system: frequent urination, urgency;

[0057] Autonomic nervous system: sweating, dizziness, headache, goosebumps, dilated pupils.

[0058] Scoring criteria:

[0059] 0 points: No signs of somatic anxiety;

[0060] 1 point: Mild, with occasional symptoms but no impact on daily functioning;

[0061] 2 points: Moderate, the above symptoms occur frequently and have some impact on daily functioning;

[0062] 3 points: Severe, with frequent occurrence of multiple symptoms, significantly affecting daily function;

[0063] 4 points: Extremely severe, with persistent and severe physical symptoms that significantly affect daily functioning.

[0064] Example of a consultation mode layer:

[0065] Typical consultation process: "Do you have any discomfort? For example, palpitations, chest tightness, dizziness?" "What physical reactions do you experience when you are nervous?" "Do you experience excessive sweating, hand tremors, or frequent urination?" "How often do these physical discomforts occur? How long do they last?"

[0066] Example of expert rule layer:

[0067] Rule 1: Urinary symptoms (frequent urination, urgency, etc.) should be classified as [somatic anxiety], not [reproductive symptoms];

[0068] Rule 2: If the patient describes both psychological stress and physical symptoms, the physical symptoms are classified under this entry, and the psychological symptoms are classified under [Psychological Anxiety].

[0069] Rule 3: Gastrointestinal symptoms (such as diarrhea and constipation) that are clearly related to anxiety should be classified under this item; if they are simply appetite problems, they should be classified under [Gastrointestinal Symptoms].

[0070] Rule 4: Headaches and dizziness accompanied by anxiety or stressful situations are classified under this entry; if they are simply physical discomfort, they are classified under [Systemic Symptoms].

[0071] The Chinese medical language model is deployed as an in-hospital service or a private network API, with access permissions and auditing mechanisms set. Low temperature and appropriate top-p parameters are used during decoding to reduce output fluctuations, and a JSON Schema validation and automatic retry module is integrated on the server side.

[0072] In the item-level summary generation stage, the transcribed text, after de-identification and turn segmentation, along with the aforementioned three-layer structured cue words, is input into the large language model. For extremely long conversations, processing can be done by time segments or semantic blocks. The large language model aggregates and summarizes related turns across turns based on the semantic key points of the items, generating item-level summaries consisting of the doctor's summary of the question and the patient's summary of the answer. The server performs structured validation on the output results, fills in missing fields, merges and deduplicates results from different blocks, and finally summarizes them by item and stores them in JSON format. An example key-value pair can be represented as {item, item-level summary: {doctor's summary of the question, patient's summary of the answer}, source turn index / time information (optional)}, so that the subsequent scoring network can directly use it and present it as evidence in the review report.

[0073] Step 4: Divide the item-level summary text dataset into training, validation, and test sets. Construct an item-level scoring network based on lightweight Transformer multi-task learning. Treat each item score as an independent but related sub-task. Through shared representation learning and dependency modeling between items, jointly optimize the scores of each item. The total score is obtained by summing and aggregating the scores of each item. This scoring network adopts a multi-task learning architecture with shared encoding and multi-head output, and mainly includes the following modules:

[0074] Shared Semantic Encoder: An L-layer Transformer encoder is used as the shared backbone network to learn deep semantic representations of the input item-level summary sequence. The network uses the item-level summary text as the sole semantic input, encodes the doctor's summary questions and the patient's summary answers, and then concatenates them into a unified sequence after adding positional and source type encodings, for example, according to... The input is constructed sequentially. The output of the shared encoder is the context-aware semantic vector of each entry, denoted as... , where K is the number of entries.

[0075] Inter-item interaction attention layer: Considering the clinical relevance between some items in the HAMD scale (such as "depressive mood" and "work and interests", "difficulty falling asleep" and "light sleep" and "early awakening"), this invention introduces an inter-item interaction attention layer after the shared encoder to explicitly model the semantic dependencies between items. By using inter-item interaction attention, the model can refer to the semantic information of other related items when predicting the score of a certain item, thereby improving the overall consistency and clinical rationality of the scoring.

[0076] Item-level scoring headers: An independent scoring header is set for each HAMD entry, and each scoring header consists of a two-layer fully connected network. and It consists of a semantic vector and a ReLU activation function, which maps the semantic vector of the corresponding item to the predicted score of that item. .

[0077] Total Score Aggregation Module: The total score is obtained by directly summing the predicted scores of each item, without the need for an additional total score prediction header, ensuring logical consistency between the total score and the item scores. This design ensures the interpretability of the total score—clinicians can directly trace the contribution of each item in the total score.

[0078] In its implementation, this invention employs an embedding layer with an input embedding dimension of 768, followed by a four-layer shared Transformer encoder. Each layer contains eight attention heads and a feedforward network with a hidden dimension of 2048 to extract item-level semantic representations. To further characterize the dependencies between different items, two inter-item interaction attention layers are introduced based on the output of the shared encoder layer to process the item semantic vector sequence. Multi-head self-attention modeling and residual fusion are performed to obtain semantic vectors of inter-entity dependencies between fused entries. At the output end, a separate two-layer fully connected regression head is set for each item. The hidden dimension of the item scoring head is 256, and a dropout rate of 0.1 is uniformly used in the network to improve the model's generalization ability. Each regression head outputs the predicted score for its corresponding item, and the total score is obtained by summing and aggregating the scores of all items. .

[0079] This invention designs a multi-task joint loss function based on mean absolute error (MAE), treating each item rating task as an independent regression subtask, and achieving joint optimization by summing the MAE of each task: The loss of the j-th entry Defined as the mean absolute error between the predicted and actual values ​​of this entry:

[0080]

[0081] Where B is the small batch size. Let be the predicted score of the nth sample in the jth entry. This corresponds to the clinical label score.

[0082] During the training phase, item-level question-and-answer sessions and clinically labeled scores obtained through independent ratings by two psychiatrists and arbitration by a third physician were used as supervision signals. The Adam optimizer was used for parameter updates, and learning rate decay and early stopping strategies were introduced to ensure training stability and generalization performance. The specific formula is as follows:

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] in, Denotes the model parameters in the t-th iteration. The learning rate; Indicates in loss function The gradient; and These are the exponential moving averages of the first and second momentum of the gradient, respectively. This represents element-wise multiplication; and This is the correction amount for the momentum deviation; Control the decay of first-order and second-order momentum separately. To prevent the denominator from being zero, the numerical stability constant is used. Achieve adaptive step size control.

[0090] The default hyperparameters in this embodiment can be set to... , , Initial learning rate If there is no improvement in the validation set for 3 epochs, then the decay is multiplied by 0.5 (minimum). The early stopping strategy is to stop training if there is no improvement in the validation set for 5 consecutive epochs; finally, the model archive point with the smallest loss in the validation set is selected as the deployment version.

[0091] Step 5: During online inference, the model outputs the scores for each item and the total score, along with the corresponding item-level summaries as evidence text, generating a report to provide clinicians with item-level semantic evidence and a traceable link. The physician workbench can display the model's predicted scores and evidence text at the item level in the review report, and can optionally display the source conversation and time period for quick review and consistency assessment. Below is a snippet of the final review report:

[0092] Total score: 19 (Doctor 20, Δ=−1, not exceeding the threshold);

[0093] Item: Early awakening: Model=2, Doctor=2, Δ=0;

[0094] Evidence text:

[0095] The doctor summarized the questions as follows: Have you woken up in the middle of the night in the past two weeks? How often / for how long?

[0096] The patient summarized their experience: They woke up almost every day around 2 or 3 a.m., and then fell back asleep after about 30 minutes, for a period of half a month.

[0097] Item: Gastrointestinal symptoms: Model=1, Doctor=0, Δ=+1 (alarm);

[0098] Evidence text:

[0099] The doctor summarized the questions as follows: Has your appetite decreased in the past two weeks?

[0100] The patient summarized their response: They have been eating less than before during this period, and often have no appetite at noon.

[0101] To meet latency and compliance requirements, speech recognition and summary generation can be executed in parallel or in a pipelined manner, with item-level scoring completed within seconds. The large language model and scoring network are primarily deployed within the institute's dedicated network environment, coupled with access control and audit logs to meet data security and compliance requirements.

[0102] To comprehensively evaluate the effectiveness of the method, the following indicators were used: mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE). These indicators are widely used in medical scale scoring to measure accuracy.

[0103]

[0104]

[0105]

[0106] Where N is the number of samples; The model prediction for the i-th sample; This corresponds to the actual annotation.

[0107] To verify the effectiveness of the method of this invention, two schemes were compared under the same data and training strategy. The test set n=43. Comparison method A only uses a small scoring network model for scoring; comparison method B uses an end-to-end large model for direct scoring. The results are shown in the table below.

[0108]

[0109] As can be seen from the data in Table 1, the method proposed in this invention, which combines a large model with a small model, achieves error reduction at both the item and total score levels: compared to comparative method A, which only scores the small model, the item MAE decreases by 24.4%, the item MSE decreases by 44.1%, the item RMSE decreases by 25.2%, and the total score MAE decreases by 49.1%; compared to comparative method B, which uses end-to-end large model scoring, the item MAE decreases by 35.9%, the item MSE decreases by 69.8%, the item RMSE decreases by 45.0%, and the total score MAE decreases by 73.9%. Overall, the results show that the method of this invention is superior to both comparative methods in terms of accuracy and stability.

[0110] Example 2: The computer-readable storage medium of this example stores a computer program that, when executed by a processor, implements the steps in the method for scoring interview data of depressed patients based on large model semantic understanding in Example 1.

[0111] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0112] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0113] Example 3: The computer device of this example includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for scoring interview data of depressed patients based on large model semantic understanding in Example 1.

[0114] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0115] Those skilled in the art will clearly understand that each implementation can be achieved using software plus the necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0116] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A scoring method for interview data of depressed patients based on large-scale model semantic understanding, characterized in that, Includes the following steps: Step 1: Collect interview audio data and score it as a clinical annotation score; Step 2: The collected audio data is transcribed using an automatic speech recognition module optimized for Chinese medical scenarios. The transcribed text is then de-identified and segmented into doctor's turns and patient's turns to form a structured text sequence. Step 3: Deploy the large language model. Input the text sequence that has been de-identified and processed by turn-taking along with the three-layer structured prompts into the large language model to generate item-level summary text. The three-layer structured prompts include a symptom definition layer, a consultation pattern layer, and an expert rule layer. Step 4: Divide the item-level summary text dataset into training, validation and test sets, construct an item-level scoring network model based on lightweight Transformer multi-task learning, train the scoring network model through multi-task joint loss function, and select the model with the smallest loss on the validation set as the deployment version. Step 5: Input the text to be scored into the trained scoring network model. The model outputs the score for each item and the total score, and outputs the corresponding item-level summary as evidence text along with the results to generate a report.

2. The method for scoring interview data of depressed patients based on large-model semantic understanding according to claim 1, characterized in that, The symptom definition layer provides a clear clinical definition and scoring boundary description for each symptom item; the consultation pattern layer summarizes the typical consultation patterns of psychiatrists for each symptom item and embeds these patterns into prompt words as identification anchors for diagnostic signals; the expert rule layer corrects and constrains common symptom classification errors in the form of explicit rules.

3. The method for scoring interview data of depressed patients based on large-model semantic understanding according to claim 2, characterized in that, The large language model is deployed as an in-hospital service or a private network API, with access permissions and auditing mechanisms set. During decoding, low temperature and appropriate top-p parameters are used to reduce output fluctuations, and a JSONSchema validation and automatic retry module is integrated on the server side.

4. The method for scoring interview data of depressed patients based on large-model semantic understanding according to claim 3, characterized in that, The server performs structured validation on the output item-level summaries, fills in missing fields, merges and deduplicates results from different blocks, and finally summarizes them by item and stores them in JSON format.

5. The method for scoring interview data of depressed patients based on large-model semantic understanding according to claim 1, characterized in that, The scoring network includes the following modules: Shared Semantic Encoder: An L-layer Transformer encoder is used as a shared backbone network to learn deep semantic representations of the input item-level summary sequence. The output of the shared semantic encoder is the context-aware semantic vector of each item, denoted as... Where K is the number of entries; Inter-item interaction attention layer: An inter-item interaction attention layer is introduced after the shared encoder to explicitly model the semantic dependencies between items; Item-level rating headers: An independent rating header is set for each item, and each rating header consists of two fully connected layers (…). and It consists of a semantic vector and a ReLU activation function, which maps the semantic vector of the corresponding item to the predicted score of that item. ; Total Score Aggregation Module: The total score is obtained by directly summing the predicted scores of each item. .

6. The method for scoring interview data of depressed patients based on large-model semantic understanding according to claim 5, characterized in that, The rating network model employs an embedding layer with an input embedding dimension of 768, followed by a 4-layer shared Transformer encoder. Each layer contains 8 attention heads and a feedforward network with a hidden dimension of 2048. Two inter-item interaction attention layers are introduced on top of the shared encoder output to process the semantic vector sequence of the items. Multi-head self-attention modeling and residual fusion are performed to obtain semantic vectors of inter-entity dependencies between fused entries. At the output end, an independent two-layer fully connected regression head is set for each item, the hidden dimension of the item scoring head is 256, and a Dropout rate of 0.1 is uniformly used in the network.

7. The method for scoring interview data of depressed patients based on large-model semantic understanding according to claim 1, characterized in that, The joint loss function for multiple tasks is: ； ； Where K is the number of items and B is the mini-batch size. Let be the predicted score of the nth sample in the jth entry. This corresponds to the clinical label score.

8. The method for scoring interview data of depressed patients based on large-model semantic understanding according to claim 7, characterized in that, The scoring network model is trained using a multi-task joint loss function, and the parameters are updated using the Adam optimizer. Learning rate decay and early stopping strategies are introduced to ensure training stability and generalization performance. The specific formulas are as follows: ； ； ； ； ； ； in, Denotes the model parameters in the t-th iteration. The learning rate; Indicates in loss function The gradient; and These are the exponential moving averages of the first and second momentum of the gradient, respectively. This represents element-wise multiplication; and This is the correction amount for the momentum deviation; Control the decay of first-order and second-order momentum separately. To avoid the denominator being zero, the numerical stability constant is used. Achieve adaptive step size control.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for scoring interview data of depressed patients based on large model semantic understanding as described in any one of claims 1-8.

10. A computer device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for scoring interview data of depressed patients based on large model semantic understanding as described in any one of claims 1-8.

Images