Evidence condition optimization method and system for multi-disorder mental health screening

Abstract

The application discloses an evidence condition optimization method and system for multi-disorder mental health screening, and belongs to the technical field of multi-modal mental health screening and reinforcement learning. The application constructs a reinforcement learning framework of evidence conditions, including a protocolized psychological evidence space construction mechanism and a multi-component reward training strategy of mask perception; the evidence space construction mechanism extracts four-dimensional interpretable psychological clues of facial dynamics, physiological signals, body movements and voices from videos through a domain-specific perception model, standardizes the clues into readable text evidence, decouples multi-modal perception and decision learning, and realizes structured reasoning with clinical interpretability; the mask perception reward strategy converts the reinforcement learning paradigm from full-label supervision to heterogeneous partial-label adaptation, avoids false supervision caused by missing labels through a mask mechanism, and realizes output control in compliance with clinical regulations in combination with a safety constraint reward.

Description

Technical Field

[0001] This invention relates to the field of multimodal mental health screening and reinforcement learning, and in particular to a method and system for optimizing evidence conditions for multi-disorder mental health screening. Background Technology

[0002] In recent years, video-based mental health screening has become a research hotspot at the intersection of mental health and artificial intelligence. With the increasing accessibility of audio and video data, the demand for large-scale, non-invasive mental health assessments continues to grow. Video data can simultaneously capture multi-dimensional behavioral signals highly correlated with mental disorders, including facial dynamics, physiological cues, body movements, and voice patterns. These signals have all been proven to be significantly correlated with mental disorder states, providing a data foundation for multimodal data-based mental disorder screening.

[0003] Currently, the mainstream technical solutions for video-based mental health screening can be mainly divided into three categories:

[0004] The first category is multimodal fusion schemes based on traditional supervised learning. Most existing research transforms mental health analysis into supervised prediction tasks, estimating potential emotional dimensions from multimodal features by regressing continuous risk scores. Some schemes extend this paradigm to multi-disorder prediction through supervised multi-task learning. While these schemes can achieve basic risk prediction, the output often lacks direct clinical interpretability, requiring post-hoc mapping to match diagnostic categories. This can easily lead to discrepancies between model predictions and the clinical definitions and expressions of mental disorders. Furthermore, multi-task frameworks typically assume that all samples have complete and homogeneous labels across all disorder dimensions and datasets. This assumption is completely inconsistent with the characteristics of real-world psychiatric data—labels in real clinical data are heterogeneous and often missing. Simply defaulting to negative samples with missing labels introduces serious spurious supervision, limiting the model's scalability.

[0005] The second category is inference schemes based on Multimodal Large Language Models (MLLMs). Leveraging their powerful reasoning capabilities, MLLMs have been increasingly applied to mental health-related tasks. However, most existing schemes only operate during the inference phase, achieving functionality through cue word engineering with fixed parameters, without any learning optimization during the training phase. This cue word-only approach lacks a fundamental mechanism to optimize the model's obstacle-level reasoning ability, cannot handle some supervised scenarios, and cannot force the model to adhere to clinically applicable output constraints, making it difficult to adapt to the specific requirements of clinical scenarios.

[0006] The third category is optimization schemes for large language models based on reinforcement learning. Advances in reinforcement learning, especially the Group Relative Policy Optimization (GRPO) algorithm, have significantly improved the performance of multimodal large language models in complex general reasoning tasks. Existing research has verified the technical effectiveness of reinforcement learning in improving the interpretability and reasoning reliability of large language models. Some studies have also applied reinforcement learning to general psychology-related scenarios such as psychological support dialogue and emotional empathy. However, currently, no existing research has applied the GRPO algorithm to multi-disorder mental health screening tasks, nor has any relevant scheme used reinforcement learning to construct an end-to-end optimization framework adapted to clinical scenarios, addressing the three core pain points of multimodal mental health screening: heterogeneous annotation, interpretability, and clinical compliance.

[0007] In summary, existing technologies still have three major insurmountable technical shortcomings in the practical clinical application of multi-disorder mental health screening:

[0008] First, the interpretability of decisions is insufficient, making it unable to meet clinical compliance requirements. Existing end-to-end solutions mostly optimize directly based on raw audio and video inputs without setting protocolized intermediate evidence variables. They cannot decouple multimodal perception from clinical decision learning, and the model decision-making process is a black box, failing to provide traceable and verifiable clinical decision-making evidence, thus making it difficult to meet the compliance requirements of clinical applications.

[0009] Second, it cannot adapt to heterogeneous labeled data, easily introducing false supervision and resulting in insufficient generalization ability. Existing solutions generally assume that training samples have complete annotations across all obstacle dimensions. However, real clinical data has high label heterogeneity and is often missing. Defaulting missing labels as negative samples will introduce serious false supervision, leading to inefficient model training, poor generalization performance, and an inability to achieve large-scale stable training across heterogeneous datasets.

[0010] Third, the lack of closed-loop optimization and clinical safety constraint mechanisms during the training phase poses a high risk to clinical application. Existing solutions based on multimodal large language models mostly achieve this only through prompt word engineering during the inference phase, without introducing targeted closed-loop optimization mechanisms during the training phase. This makes it impossible to optimize the model's obstacle-level decision-making capabilities through feedback. At the same time, the absence of effective safety constraint mechanisms makes it impossible to force the model to output non-diagnostic results that conform to clinical norms, which can easily lead to non-compliant outputs such as definitive diagnostic statements, posing significant risks to clinical application safety. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention provides an evidence-conditional reinforcement learning optimization method (ECRL-MH) for multi-disorder mental health screening. This invention can enhance the decision reliability and clinical interpretability of multimodal models in mental health screening tasks through interpretable evidence-mediated reasoning and a reward mechanism that adapts to heterogeneous annotations.

[0012] The technical solution of this invention is: a method for optimizing evidence conditions for multi-disorder mental health screening, comprising the following steps:

[0013] S1. Construction of Protocol-Based Psychological Evidence Space: Through a domain-specific perception model, extract four-dimensional interpretable psychological cues from the input video x, including facial dynamics, physiological signals, body movements, and speech, and standardize them into readable text evidence.

[0014] S2, Evidence Condition Policy Generation: Based on standardized text evidence, the policy network generates structured multi-label decision outputs, which are constrained by a predefined JSON format;

[0015] S3, Mask Awareness Reward Calculation: Introducing a binary supervised mask function, the reward is calculated only for obstacle dimensions with valid annotations, and multi-component reward signals are generated by combining structural compliance rewards and security awareness rewards;

[0016] S4. Group Relative Policy Optimization: Based on the multi-component reward signal, the relative quality score is calculated through the group relative policy optimization algorithm to complete the policy network parameter update.

[0017] Furthermore, in step S1, multi-dimensional psychological cues are extracted using the evidence extraction function E(x):

[0018]

[0019] In the formula, E_{face}(x) represents dynamic evidence of facial micro-expressions, capturing the intensity, frequency, and temporal variability of micro-movements; E_{phys}(x) represents physiological activity evidence estimated from remote photoplethysmography (PPG), reflecting heart rate-related arousal patterns and abnormalities. As evidence of body posture and behavior, describe the openness, tension, and range of motion of the posture; The evidence is presented as acoustic features of speech, representing prosodic attributes such as energy, rhythm, and vocalization state. The extracted evidence variables are converted into standardized natural language descriptions and aggregated into single-text contexts that conform to a unified paradigm, serving as conditional inputs for policy optimization.

[0020] Furthermore, the strategy generation formula in step S2 is:

[0021]

[0022]

[0023] In the formula, For standardized evidence representation, For policy networks Let y be a K-dimensional binary vector, where each dimension corresponds to a screening indicator for a mental disorder. The large language model is based on... Natural language screening descriptions are generated in conjunction with evidence; policy output is constrained to a predefined JSON format, with fixed barrier key names and binary values, and penalties are imposed for deviations from the format.

[0024] Furthermore, the binary check mask function in step S3 is:

[0025]

[0026] In the formula For a set of obstacle indexes with effective annotations, The mask value corresponds to the dimension, and the mask encodes the label availability.

[0027] Furthermore, the relative policy optimization in step S4 includes: for the input evidence condition query p, sampling G candidate decision outputs from the policy distribution, and calculating the relative quality score through standardization:

[0028]

[0029] In the formula, This represents the standardized advantage value of the g-th candidate output within the group.

[0030] An evidence condition optimization system that implements the above evidence condition optimization method includes:

[0031] The protocol-based psychological evidence space construction module is used to execute step S1, which extracts and standardizes four-dimensional psychological evidence through a domain-specific perception model.

[0032] The evidence condition strategy generation module is used to execute step S2, which generates structured multi-label decision output based on standardized text evidence;

[0033] The mask-aware reward training module is used to execute step S3. It avoids false supervision caused by missing labels through the mask mechanism and achieves clinical compliance output control by combining safety constraint rewards.

[0034] The group-relative strategy optimization module is used to execute step S4, which completes the strategy optimization update based on the multi-component reward signal.

[0035] Furthermore, the protocol-based psychological evidence space construction module includes: a facial evidence extraction unit, a physiological evidence extraction unit, a posture evidence extraction unit, a voice evidence extraction unit, and an evidence standardization unit.

[0036] Furthermore, the mask-aware reward training module includes a mask processing unit and a multi-component reward calculation unit.

[0037] Furthermore, the group relative strategy optimization module includes: a candidate output sampling unit, a reward evaluation unit, a relative advantage calculation unit, and a strategy update unit.

[0038] The beneficial effects of this invention are as follows:

[0039] 1. This invention converts multimodal perception signals into standardized readable evidence by protocolizing the intermediate evidence space, which can be directly used as conditional variables for language model strategies. This decouples domain-specific perception from downstream decision learning, realizes structured reasoning with clinical interpretability, and overcomes the limitations of existing methods that rely on latent feature associations and have insufficient interpretability.

[0040] 2. The mask-aware reinforcement learning objective designed in this invention can handle heterogeneous and partially labeled clinical data, avoid false supervision caused by missing labels, achieve large-scale stable training across heterogeneous datasets, and adapt to the labeling characteristics of real-world psychiatric data.

[0041] 3. This invention integrates multi-component reward functions and combines group relative strategy optimization (GRPO) to ensure screening accuracy while constraining the model to output non-diagnostic results that conform to clinical norms, thereby improving the model's decision reliability and deployment security.

[0042] 4. This invention achieves performance improvements on multiple classic mental health benchmark datasets such as DAIC-WOZ, E-DAIC, and D-Vlog, and has the ability to generalize across disorder types, effectively adapting to screening scenarios for various mental disorders such as depression and PTSD. Attached Figure Description

[0043] Figure 1 This is an example diagram comparing the inference output of the ECRL-MH method of this invention with that of the baseline model;

[0044] Figure 2 This is a diagram illustrating the overall architecture and workflow of the ECRL-MH model of this invention;

[0045] Figure 3 This is a bar chart comparing the F1 performance of the present invention with existing baseline methods on the DAIC-WOZ and E-DAIC datasets;

[0046] Figure 4 This is a visual example of the application of the method of the present invention in multimodal video mental health screening. Detailed Implementation

[0047] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0048] Example 1

[0049] like Figure 1 , Figure 2 As shown, this embodiment proposes an evidence-conditional reinforcement learning optimization method (ECRL-MH) for multi-disorder mental health screening. Figure 1 To compare the inference process of this method with the baseline model, Figure 2 This outlines the overall architecture and workflow of the method. The core of this method comprises five main modules: a protocol-based psychological evidence space construction mechanism, an evidence condition strategy generation module, a mask-aware supervision module, a multi-component reward design module, and a GRPO-based strategy optimization module. These modules are interconnected to jointly complete interpretable reasoning and strategy optimization for multi-disorder mental health screening. The specific implementation methods are as follows.

[0050] In this embodiment, for the input video x, multi-dimensional psychological evidence is first extracted using an evidence extraction function:

[0051]

[0052] Among them, facial evidence extracts the intensity, frequency, and temporal variability of micro-movements through facial landmark detection and micro-expression recognition models; physiological evidence estimates heart rate-related arousal patterns and abnormalities from facial videos through remote photoplethysmography (PPG) technology; posture evidence extracts posture openness, tension, and range of motion through human posture estimation models; and speech evidence extracts acoustic features under eGeMAPS configuration using the OpenSmile tool to obtain prosodic attributes such as speech energy, rhythm, and vocalization state.

[0053] The continuous features from the aforementioned dimensions are converted into standardized natural language descriptions and aggregated into single-text contexts conforming to a unified paradigm, serving as conditional inputs for policy optimization. The evidence extraction process is executed online during both the training and inference phases, ensuring consistency in evidence distribution between the policy optimization and deployment phases.

[0054] In this embodiment, based on standardized textual evidence, strategy \pi_{\theta} generates structured multi-label decisions:

[0055]

[0056]

[0057] in, Let y be a K-dimensional binary vector, where each dimension corresponds to a screening indicator for a mental disorder. The large language model is based on... Natural language screening descriptions are generated in conjunction with evidence. The policy output is strictly constrained to a predefined JSON format, with fixed barrier key names and binary values, and penalties are imposed for outputs with missing keys or invalid value types.

[0058] In this embodiment, a multi-component complementary reward function is designed, the core of which includes:

[0059] 1. Masking barrier-level accuracy bonus:

[0060]

[0061] 2. Structure Compliance Rewards: Positive rewards are given for outputs that conform to JSON format requirements, while penalties are imposed for outputs that violate formatting requirements;

[0062] 3. Safety perception reward: Positive rewards are given for non-diagnostic screening statements that conform to clinical standards, while penalties are given for definitive diagnostic statements.

[0063] The total reward is a weighted combination of the above components.

[0064] In this embodiment, policy optimization is performed based on the Group Relative Policy Optimization (GRPO) algorithm, and the overall process is similar to... Figure 2 The model architecture shown is completely consistent; please follow these steps:

[0065] 1. Candidate Decision Sampling: For each standardized evidence condition input, the policy network... Sample and generate G groups of candidate decision outputs, parse each group of candidate outputs into a structured multi-obstacle binary vector, and complete the compliance verification in the preset JSON format;

[0066] 2. Multi-component reward calculation: Based on the aforementioned multi-component reward function, the corresponding comprehensive reward value is calculated for each group of candidate decision outputs. A masking mechanism is used to calculate obstacle-level accuracy rewards only for obstacle dimensions with valid annotations, simultaneously evaluating structural compliance rewards and safety awareness rewards, ultimately obtaining the comprehensive reward value for each group of candidate outputs. ;

[0067] 3. Within-group relative advantage calculation: Based on the comprehensive reward value of each group's candidate output, the within-group relative advantage value is calculated to eliminate within-group benchmark bias and achieve more stable policy updates. The calculation formula is as follows:

[0068]

[0069] 4. Strategy Parameter Update: Based on the calculated relative advantage value, the parameters of the strategy network are updated through gradient ascent, optimizing the strategy towards higher reward and higher clinical compliance decisions. The parameter update formula is as follows:

[0070]

[0071] Example 2

[0072] This embodiment validates the performance of the method in Example 1 on three classic mental health screening benchmark datasets, including DAIC-WOZ, E-DAIC, and D-Vlog, covering clinical interview scenarios and real-world vlog scenarios. It also validates the model's generalization ability in screening for multiple disorders, including depression and PTSD. Detailed information for each dataset is as follows:

[0073] 1. DAIC-WOZ and E-DAIC datasets: Both are classic audio-visual interview datasets for depression screening, collected from semi-structured clinical interviews between subjects and virtual interviewers. They provide time-synchronized audio and video signals, covering facial and upper body behavioral data. The publicly available version of DAIC-WOZ contains 189 interview sessions, each lasting 7-33 minutes with an average duration of 16 minutes. It uses a fixed method of independent subject partitioning, with training / validation / test set sample sizes of 107 / 35 / 47. E-DAIC is an extended dataset of DAIC-WOZ, containing 275 interview sessions, using the PHQ-8 score as a marker for depression severity. The training / validation / test set sample sizes are 163 / 56 / 56. This embodiment uses both pure visual and audio-visual fusion settings, excluding transcribed text and any text-derived features, and follows the above standard dataset partitioning method to ensure comparability with existing studies.

[0074] 2. D-Vlog Dataset: A non-verbal depression recognition dataset for real-world scenarios, collected from publicly available YouTube vlog videos. It contains 961 videos, totaling 160 hours, covering 816 different subjects. All videos were manually labeled with binary classification for depression / non-depression according to standardized specifications. To protect privacy, D-Vlog releases de-identified multimodal features instead of the raw audio and video data: acoustic features were extracted using OpenSmile with eGeMAPS configuration, generating 25-dimensional descriptors segmented by second; visual features are 136-dimensional vectors composed of the coordinates of 68 facial key points extracted per second using dlib.

[0075] In this embodiment, all sequences are truncated or padded with zeros according to the training requirements to ensure uniform input length.

[0076] This embodiment is based on the Qwen2.5-Omni model for training. In the pre-training stage, cold start training, preliminary reinforcement learning adaptation and multimodal reasoning ability enhancement are completed in sequence. In the specialized reinforcement learning stage, policy optimization is completed for the target dataset.

[0077] The experimental results of this embodiment on three benchmark tests validate the effectiveness of ECRL-MH in the multi-disorder mental health screening task, and the results are shown in Table 1.

[0078] Table 1. Performance comparison of Example 1 model on DAIC-WOZ and E-DAIC datasets.

[0079]

[0080] As shown in Table 1, Example 1 outperforms existing baseline methods across multiple datasets and modal settings. In the pure video setting of the DAIC-WOZ dataset, the F1 score reaches 0.666, exceeding the optimal baseline by 4.6 percentage points, with an accuracy of 0.800, effectively reducing false positives. In the audio-video fusion setting, the F1 score further improves to 0.714, exceeding the optimal baseline by 4.4 percentage points. On the E-DAIC dataset, where the distribution shift is more significant, this example also achieves stable performance improvements, with an F1 score of 0.595 in the audio-video fusion setting, exceeding the optimal baseline by 3.5 percentage points.

[0081] Table 2 shows the performance comparison of the model in Example 1 on the D-Vlog dataset.

[0082]

[0083] As can be seen from Table 2, on the real-world D-Vlog dataset, the F1 score of this embodiment reached 0.9075, while achieving a high recall rate of 0.9515, which meets the core requirement of reducing missed diagnoses in screening scenarios.

[0084] To further verify the effectiveness of the core design of this invention, an ablation experiment was conducted in this embodiment to compare the performance differences between textual evidence and original feature input. The results are shown in Table 3.

[0085] Table 3 Ablation experiment results for different evidence input methods

[0086]

[0087] As shown in Table 3, textual evidence input consistently outperforms the original feature input across all datasets, validating the effectiveness of protocolized intermediate evidence variables. By standardizing textual evidence to align the reasoning and instruction-following capabilities of the language model, more stable credit allocation and more calibrated decision-making are achieved.

[0088] This embodiment also conducted a zero-sample ablation experiment for PTSD identification to verify the model's cross-barrier generalization ability. The results are shown in Table 4.

[0089] Table 4 Results of zero-sample ablation experiments for PTSD identification

[0090]

[0091] As can be seen from Table 4, this embodiment still outperforms the existing baseline in recognizing the new mental disorder type PTSD under the zero-sample setting, verifying that the method of the present invention has good cross-barrier generalization ability.

[0092] To visually and intuitively evaluate the performance of the method of this invention in mental health screening tasks, a qualitative comparison was made between Example 1 and the baseline model on clinical interview samples in the DAIC-WOZ dataset, such as... Figure 4 As shown. Qualitative results indicate that the model in Example 1 can generate evidence-based, clinically compliant, and interpretable screening conclusions based on multimodal behavioral cues, significantly improving the model's clinical suitability.

[0093] Evidence condition optimization systems for multi-disability mental health screening include:

[0094] The protocol-based psychological evidence space construction module extracts and standardizes four-dimensional psychological evidence through a domain-specific perception model.

[0095] The protocol-based psychological evidence space construction module includes: a facial evidence extraction unit, a physiological evidence extraction unit, a posture evidence extraction unit, a voice evidence extraction unit, and an evidence standardization unit.

[0096] The evidence condition strategy generation module generates structured multi-label decision outputs based on standardized text evidence;

[0097] The mask-aware reward training module avoids false supervision caused by missing labels through a mask mechanism, and combines safety constraint rewards to achieve clinical compliance output control.

[0098] The mask-aware reward training module includes a mask processing unit and a multi-component reward calculation unit.

[0099] The group-based strategy optimization module optimizes and updates strategies based on multi-component reward signals.

[0100] The relative strategy optimization module includes: a candidate output sampling unit, a reward evaluation unit, a relative advantage calculation unit, and a strategy update unit.

[0101] Figure 3This is a bar chart comparing the F1 performance of the present invention with existing baseline methods on the DAIC-WOZ and E-DAIC datasets. The chart shows the F1 performance comparison results of the present invention's method with existing mainstream baseline methods such as HumanOmniV2 and David et al. on the two classic depression screening benchmark datasets, DAIC-WOZ and E-DAIC. The blue bars represent the test results on the DAIC-WOZ dataset, and the orange bars represent the test results on the E-DAIC dataset. The results show that the present invention's method significantly outperforms existing baseline methods in F1 scores on both datasets, validating the performance advantages and generalization ability of the present invention in multi-disorder mental health screening tasks.

[0102] Figure 4 This is a visual example of the application of the method of the present invention in multimodal video mental health screening. The figure shows the actual screening effect of the method of the present invention under real multimodal video input. For the input video data containing audio and visual information, the method of the present invention can extract multi-dimensional psychological evidence such as facial expressions, body movements, and voice features. Combined with contextual rewards and inference logic rewards, it generates evidence-based, clinically compliant non-diagnostic screening conclusions, intuitively demonstrating the decision interpretability and clinical scenario adaptability of the method of the present invention.

[0103] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A method for optimizing evidence conditions for multi-disorder mental health screening, characterized in that, Includes the following steps: S1. Construction of Protocol-Based Psychological Evidence Space: Through a domain-specific perception model, extract four-dimensional interpretable psychological cues from the input video x, including facial dynamics, physiological signals, body movements, and speech, and standardize them into readable text evidence. S2, Evidence Condition Policy Generation: Based on standardized text evidence, the policy network generates structured multi-label decision outputs, which are constrained by a predefined JSON format; S3, Mask Awareness Reward Calculation: Introducing a binary supervised mask function, the reward is calculated only for obstacle dimensions with valid annotations, and multi-component reward signals are generated by combining structural compliance rewards and security awareness rewards; S4. Group Relative Policy Optimization: Based on the multi-component reward signal, the relative quality score is calculated through the group relative policy optimization algorithm to complete the policy network parameter update.

2. The method for optimizing evidence conditions for multi-disorder mental health screening according to claim 1, characterized in that, In step S1, multi-dimensional psychological cues are extracted using the evidence extraction function E(x): In the formula, E_{face}(x) represents dynamic evidence of facial micro-expressions, capturing the intensity, frequency, and temporal variability of micro-movements; E_{phys}(x) represents physiological activity evidence estimated from remote photoplethysmography (PPG), reflecting heart rate-related arousal patterns and abnormalities. As evidence of body posture and behavior, describe the openness, tension, and range of motion of the posture; The evidence is presented as acoustic features of speech, representing prosodic attributes such as energy, rhythm, and vocalization state. The extracted evidence variables are converted into standardized natural language descriptions and aggregated into single-text contexts that conform to a unified paradigm, serving as conditional inputs for policy optimization.

3. The method for optimizing evidence conditions for multi-disorder mental health screening according to claim 1, characterized in that, The strategy generation formula in step S2 is: In the formula, For standardized evidence representation, For policy networks Let y be a K-dimensional binary vector, where each dimension corresponds to a screening indicator for a mental disorder. The large language model is based on... Natural language screening descriptions are generated in conjunction with evidence; policy output is constrained to a predefined JSON format, with fixed barrier key names and binary values, and penalties are imposed for deviations from the format.

4. The method for optimizing evidence conditions for multi-disorder mental health screening according to claim 1, characterized in that, The binary check mask function in step S3 is: In the formula For a set of obstacle indexes with effective annotations, The mask value corresponds to the dimension, and the mask encodes the label availability.

5. The method for optimizing evidence conditions for multi-disorder mental health screening according to claim 1, characterized in that, Step S4, relative policy optimization, includes: for the input evidence condition query p, sampling G candidate decision outputs from the policy distribution, and calculating the relative quality score through standardization. In the formula, This represents the standardized advantage value of the g-th candidate output within the group.

6. An evidence condition optimization system for implementing the evidence condition optimization method according to any one of claims 1 to 5, characterized in that, include: The protocol-based psychological evidence space construction module is used to execute step S1, which extracts and standardizes four-dimensional psychological evidence through a domain-specific perception model. The evidence condition strategy generation module is used to execute step S2, which generates structured multi-label decision output based on standardized text evidence; The mask-aware reward training module is used to execute step S3. It avoids false supervision caused by missing labels through the mask mechanism and achieves clinical compliance output control by combining safety constraint rewards. The group-relative strategy optimization module is used to execute step S4, which completes the strategy optimization update based on the multi-component reward signal.

7. The evidence condition optimization system according to claim 6, characterized in that, The protocol-based psychological evidence space construction module includes: a facial evidence extraction unit, a physiological evidence extraction unit, a posture evidence extraction unit, a voice evidence extraction unit, and an evidence standardization unit.

8. The evidence condition optimization system according to claim 6, characterized in that, The mask-aware reward training module includes a mask processing unit and a multi-component reward calculation unit.

9. The evidence condition optimization system according to claim 6, characterized in that, The relative strategy optimization module includes: a candidate output sampling unit, a reward evaluation unit, a relative advantage calculation unit, and a strategy update unit.

Images