An Emotion Training System and Method Based on Multi-Sensory Intervention and AI Dialogue
By combining multi-sensory intervention with AI dialogue, and integrating simultaneous visual, auditory, and tactile stimulation, a proactive emotional training closed-loop system is constructed. This addresses the shortcomings of existing panic attack intervention tools and achieves personalized and safe training results.
Patent Information
- Application Number
- CN202610196411.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-10-15
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-30
AI Technical Summary
Existing panic attack intervention tools lack active training mechanisms, have limited sensory stimulation, weak situational relevance, insufficient intelligent interaction, and lack data recording and feedback, resulting in limited training effectiveness and insufficient privacy and security.
Employing multi-sensory intervention and AI dialogue, this system provides personalized training and feedback through simultaneous visual, auditory, and tactile stimulation, combined with real-time data collection and analysis. It constructs a closed-loop system for proactive emotion training, including modules for multimodal perception, AI dialogue generation, and data security.
It enables multi-sensory proactive training, enhances counter-terrorism capabilities, improves training effectiveness, provides personalized feedback and data security, and expands the coverage and depth of digital psychological intervention.
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer applications and artificial intelligence technology, and in particular to an emotion training system and method based on multi-sensory intervention and AI dialogue for the intervention of panic attacks and acute anxiety attacks. Background Technology
[0002] With the development of mobile internet and artificial intelligence technologies, more and more users are using mobile applications, self-help programs, or smart wearable devices for daily emotion regulation and mental health training. Among these, digital intervention tools for panic attacks (acute anxiety attacks) are gradually appearing in the public eye.
[0003] Current intervention methods primarily include traditional relaxation training, breathing exercises, mindfulness meditation, and cognitive behavioral therapy (CBT). These methods can help users calm their emotions and alleviate physical discomfort during panic to some extent. For example, some mobile applications guide users through diaphragmatic breathing, muscle relaxation, or focus exercises using animations or voice prompts to help them regain calm during anxiety. However, most of these methods primarily aim to "relieve symptoms," focusing on calming users down during an attack rather than proactively utilizing the peak of panic for systematic neuromodulation training. Meanwhile, some digital psychological counseling products (such as chatbots or AI psychological assistants) can provide general advice through text or voice dialogue, but they often use fixed scripts or simple question-and-answer logic, making it difficult to provide personalized responses and dynamic guidance based on the user's real-time state, specific context, and time constraints. Therefore, in practice, users often report that these systems "don't understand them well enough" and lack genuine empathy and targeted approach.
[0004] Overall, current panic attack support tools on the market suffer from the following major technical shortcomings and defects: 1. Lack of active training mechanisms: Most existing methods only help users passively wait for panic to subside, failing to guide them to actively train during peak moments and promote the relearning of the brain's emotional system. Users often "endure" the attack rather than learn to regulate it. 2. Limited sensory stimulation and insufficient immersion: Most applications only provide single-sensory guidance (voice or visual), which is difficult to attract attention or establish a sense of security during periods of high-intensity fear. The lack of multimodal sensory stimulation combining visual, auditory, and tactile senses limits the training effect. 3. Weak situational relevance: Current content is mostly general relaxation exercises, without differentiated guidance for different scenarios (such as airplanes, enclosed elevators, being alone at night, etc.), making it difficult for users to obtain support tailored to their current environment. 4. Emphasis on avoidance rather than facing fear: Traditional methods often emphasize shifting attention or temporarily escaping the situation, which is effective in the short term but may reinforce avoidance psychology in the long term, lacking training strategies to guide users to face fear in a safe environment. 5. Insufficient intelligent interaction and low personalization: Existing AI tutoring is mostly based on pre-set scripts, unable to dynamically identify changes in user emotions or intentions, and lacks keyword extraction, emotion analysis, and adaptive questioning mechanisms. User feedback is often vague and lacks substantial help. 6. Lack of data recording and feedback mechanisms: Existing tools rarely record the entire process of an attack or generate training reports. Users cannot quantify their progress, such as whether fear intensity has decreased or recovery speed has increased, leading to insufficient motivation and hindering systematic data accumulation. Furthermore, existing products also have data compliance risks. Some applications lack the "minimum necessity principle" and "edge processing mechanism" when collecting voice or location information, resulting in insufficient data security and privacy protection.
[0005] Therefore, proposing an active, quantifiable, and learnable closed-loop system and method for emotion training that can overcome the limitations of traditional psychological relaxation applications such as "single sensory experience, passive relief, and lack of feedback" is of great significance for the application of computer and artificial intelligence technologies in the field of emotion training. Summary of the Invention
[0006] To address the aforementioned shortcomings, this invention proposes a method for emotional training in panic attack scenarios based on multi-sensory intervention and AI dialogue, utilizing computer and artificial intelligence technologies. The proposed method integrates multi-modal perception (visual, auditory, tactile, and physiological proxy signals) with AI dialogue generation technology through multi-source data acquisition, threshold / rule and simplified regression judgment, and a closed loop of reporting and incentive feedback. This enables personalized intervention and training for users in non-medical settings, providing real-time guidance, state recognition, and quantitative assessment from the peak of an attack to recovery. Furthermore, the method includes a data security module to ensure user privacy. This method overcomes the limitations of traditional psychological relaxation applications—"single-sensory, passive relief, and lack of feedback"—and constructs an active, quantifiable, and learnable closed-loop emotional training system. It is suitable for individual users' daily self-management of emotions and can also serve as a tool for professional training, such as psychological counselors. Simultaneously, this invention also provides a situational emotional training system based on multi-sensory intervention and AI dialogue, which enables emotional training in panic attack scenarios using computer and artificial intelligence technologies.
[0007] This invention is achieved through the following technical solution: An emotion training method based on multi-sensory intervention and AI dialogue, characterized by the following steps: Step 1: The user initiates an AI dialogue for training and collects data; Step 2: Dynamic visual effects are presented to the user, along with audio and vibration; Step 3: Real-time collection of interactive data and physiological proxy data generated by the user through AI dialogue and obtained by real-time calculation of the emotion intensity curve I(t), calculation of multi-source evidence scores, and obtaining multi-source evidence data. The dynamic visual effects, audio, and vibration frequencies are adjusted according to the multi-source evidence data: when the multi-source evidence score meets the PEAK peak entry condition, the dynamic visual effects, audio, and vibration frequencies automatically... The system automatically switches to peak dynamic visual effects, peak audio, and peak vibration frequency. Users interact with the AI through dialogue to enter the security center of the dynamic visual effects screen. When the multi-source evidence score meets the RECOV non-peak entry condition, the dynamic visual effects, audio, and vibration automatically switch to non-peak dynamic visual effects, non-peak audio, and non-peak vibration frequency, and the dynamic visual effects, audio, and vibration frequency adjust according to the multi-source evidence data. Step 4: After the AI dialogue ends, the system calculates the amygdala sensitivity index (ASI), prefrontal control index (PCI), growth score (G), and courage index (CI) based on the multi-source evidence data, and generates a training report.
[0008] Furthermore, the user interaction data and physiological agent data collected in real time through AI dialogue in step 3 include: A4 engagement data elements: user interaction duration (touch_duration), click rate per unit time (tap_rate), touch area ratio (touch_area_ratio), and the pulse signal triggered by user clicks (tap_impulse); cumulative time for users to follow voice guidance (tts_follow_secs), peak phase duration (peak_phase_secs), the proportion of time spent following voice guidance during the peak phase (tts_ratio), the proportion of time spent maintaining interaction during the PEAK peak phase (peak_follow_ratio), and the number of times users actively interacted during the peak phase (tap). A6 physiological data: Heart rate (HR), respiratory rate, and skin conductance response (GSR); A7 speech feature data: tremor index; The calculation method for the emotion intensity curve I(t) in step 3 is as follows: I(t) = λ1z(HR(t)) + λ2z(GSR(t)) + λ3Q3(t); where: z is the z-score standardized function; HR(t) is the real-time heart rate data, GSR(t) is the real-time skin conductance response data, and Q3(t) is the subjective intensity score given by the user in real time; λ1, λ2, and λ3 are preset values, all ≥ 0, and λ1 + λ2 + λ3 = 1; The interaction data and physiological proxy data obtained by calculating the emotional intensity curve I(t) include: A2 recovery refinement data: time_to_50% for symptom intensity to decrease by 50%, time_to_safe for symptom recovery to a safe level, and recovery time reduction rate drop(time_to_safe); time_to_50% is the time required for I(t) to decrease from its peak to 50%, time_to_safe is the time required for I(t) to decrease to a preset safe value, drop(time_to_safe) = (baseline recovery time - time_to_safe) / baseline recovery time, where the baseline recovery time is the arithmetic mean of time_to_safe from multiple training sessions for the user; and the episode duration Q2.
[0009] Furthermore, the user interaction data and physiological agent data collected in real time through AI dialogue in step 3 also include: A1 timeline data: training start time start_ts, recovery confirmation time recovery_ts, and total session duration session_duration; A4 Participation Factor Data: The guidance mode used in this training (guidance_mode_used); A5 scene context data: user location category coarse_location_tag, without precise geographic coordinates; A7 speech feature data: maximum volume (max_db), average volume, speech rate, and pause rate.
[0010] Furthermore, the method for calculating the multi-source evidence score in step 3 is as follows: Etap is obtained by standardizing the tap rate per unit time (tap_rate), user interaction duration (touch_duration), and touch area ratio (touch_area_ratio) using the z-score function, i.e.: Etap = z(tap_rate + λ4·touch_duration + μ·touch_area_ratio), where λ4 and μ are preset values, both ≥ 0 and summed to 1; Evoice is obtained by standardizing the tremor index (tremor_index) using the z-score function, i.e., Evoice = z(tremor_index); Ehr and Egsr are obtained by z-score normalization of heart rate (HR) and skin conductance response (GSR), respectively, i.e., Ehr = z(HR), Egsr = z(GSR); In step 4, when the multi-source evidence score meets the PEAK peak entry condition, the method for automatically switching to the dynamic visual effect, audio, and vibration frequency of the PEAK peak is as follows: it is triggered when Score_in ≥ T_in and the duration ≥ τ_in, where Score_in = w1·Etap + w2·Ehr + w3·Egsr + w4·Evoice + w5·peak_mark; In step 5, when the multi-source evidence score meets the RECOV non-peak entry condition, the method for switching to the dynamic visual effects, audio, and vibration frequency during the non-peak period is as follows: it is triggered when Score_out ≥ T_out and the duration ≥ τ_out, where Score_out = u1·(-ΔEhr) + u2·(-ΔEgsr) + u3·(-ΔEtap) + u4·(-ΔEvoice) + u5·better_mark; Where peak_mark is the peak value actively marked by the user, peak_mark∈{0,1}; better_mark is the improvement mark value actively marked by the user; T_in is the preset time threshold range representing entering the peak state; τ_in is the preset duration range of the time threshold for entering the peak state; T_out is the preset time threshold range for leaving the peak state; τ_out is the preset duration range of the time threshold for leaving the peak state; Δ represents the use of sliding window difference; the weights w1-w5 and u1-u5 are preset values, both ≥ 0, and the sum of w1-w5 and u1-u5 is equal to 1.
[0011] Furthermore, the dynamic visual effects, audio, and vibration frequencies in steps 3 and 5 are adjusted according to the multi-source evidence data through the following method: When the dynamic visual effect diagram is a dynamic storm diagram, the speed of the storm vortex and the size of the eye are adjusted according to the multi-source evidence data. When the multi-source evidence score meets the PEAK peak entry condition, it automatically switches to the peak stage, increases the vortex angular velocity, enlarges the eye, and enhances the audio and vibration intensity, guiding the user to actively interact through AI dialogue to enter the safe center of the dynamic storm diagram, which is displayed as the eye. When the multi-source evidence score meets the RECOV non-peak entry condition, it switches to the non-peak stage, and the vortex speed, eye size, audio, and vibration intensity are adjusted according to the multi-source evidence data. The dynamic storm map changes according to the following response model: r(t+1)=r(t)·(1-β), ω(t+1)=ω(t)+γ·tap_impulse, where r is the eye radius, r(t) represents the current eye radius, ω is the vortex angular velocity, ω(t) represents the current vortex angular velocity, β and γ are response coefficients, both taking values ∈(0,1]; the tempo of the audio is proportional to the score of the multi-source evidence data; the vibration frequency is proportional to the score of the multi-source evidence data.
[0012] The dynamic visual effect diagram described above is used to transform abstract emotional states into visual spatial metaphors. The spatial metaphor model can be a dynamic storm diagram, or it can use other dynamic scenes with a central stable region (such as waves, flames, tunnels, etc.), but its core training mechanism is the same.
[0013] Furthermore, the calculation methods for the ASI, PCI, G, and CI indicators in step 6 are as follows: The Amygdala Sensitivity Index (ASI) is calculated as: ASI = z(Q3_peak) × norm(Q2), where Q3_peak is the peak fear intensity score input by the user based on AI dialogue prompts, Q2 is the duration of the attack, z(Q3_peak) is the result of standardizing Q3_peak using the z-score function, and norm(Q2) is the result of normalizing Q2 using the norm function. Normalization can be performed based on the individual baseline: the arithmetic mean of the Q2 values collected each time the user uses the system is summed and divided by the number of times the system is used is the individual baseline value. The L2 norm of the individual baseline value is calculated, the current data vector Q2 is subtracted from the individual baseline value, and the difference is divided by the L2 norm of the individual baseline value to obtain the normalized result. The prefrontal control index (PCI) is calculated as Q4 × Engagement × RS, where Q4 is the user's self-input rating of their perceived control during the conversation; Engagement = σ(α1·Etap + α2·tts_ratio + α3·peak_follow_ratio), where σ is the Sigmoid function; RS = 1 / (1+time_to_safe); and α1, α2, and α3 are preset values, all ≥ 0 and satisfying α1 + α2 + α3 = 1. The growth score G = 0.4·Q7 + 0.4·drop(time_to_safe) + 0.2·Q4, where Q7 is the user's self-efficacy rating input based on AI dialogue prompts, and drop(time_to_safe) is the recovery time reduction rate calculated from the emotion intensity curve I(t); The Courage Index (CI) is calculated as: CI = norm(tap) + norm(High-Intensity Completion) + norm(Q7). Here, tap represents the number of proactive interactions during the peak phase, High-Intensity Completion is the user's self-assessment of their performance in completing tasks during the conversation, and Q7 is the user's self-assessment of their self-efficacy. The norm function normalizes the values, which can be based on a personal baseline: summing the values collected each time the user uses the system and dividing by the number of times the system is used, obtaining the arithmetic mean as the personal baseline value, calculating the L2 norm of the personal baseline value, subtracting the personal baseline value from the current value, and dividing the difference by the L2 norm to obtain the normalized result.
[0014] Furthermore, the following privacy protection and data security mechanisms are adopted: Minimal identifiability principle: Only coarse_location_tag category labels are stored, and GPS precise coordinates are not saved; only speech acoustic features are stored, and the original audio is not saved; Edge-side feature extraction: Speech processing and emotion classification are completed locally on the mobile device, and the cloud only receives anonymized tag text; Separate table encrypted storage: Personally Identifiable Information (PII) and training logs are stored in separate tables, sensitive fields are encrypted with AES-256, and access permissions are minimized; Compliance interface: Provides a channel for user data withdrawal and export that complies with PIPL / GDPR standards.
[0015] An emotion training system based on multi-sensory intervention and AI dialogue includes the following modules: a multi-sensory intervention module, used to provide multimodal synchronous guidance and active interactive training through visual, auditory, and tactile senses during panic or anxiety attacks; an AI dialogue system, used to provide cognitive counseling and structured intervention based on natural language processing during non-peak attack periods; and a user data recording and feedback subsystem, used to collect, store, and analyze user interaction data, physiological proxy data, and subjective report data, generate quantitative assessment reports, and perform calculations based on collected multi-source evidence, judging and displaying the user's panic or anxiety attack state based on the scoring results. The multi-sensory intervention module maps the panic attack process into multiple stages using a dynamic visual effect model, and adjusts the parameters of visual, audio, and tactile stimuli in real time based on the calculation results of multi-source evidence from the user data recording and feedback subsystem, achieving active neuromodulation training throughout the entire attack cycle.
[0016] The present invention has the following beneficial technical effects: 1. This method enables multi-sensory active training to confront fear and enhance coping abilities: By introducing simultaneous visual, auditory, and tactile stimulation during peak periods, combined with active interaction, it guides users to complete practical exposure training and break through negative patterns; and through repeated training, it can establish new connections, reduce amygdala hypervigilance, and enhance prefrontal cortex regulation capabilities.
[0017] 2. Utilizing computer and artificial intelligence technologies, it can be applied instantly on mobile devices and computers without additional hardware, making it easy to deploy and promote: The tone and depth of the conversation can be determined by combining "goals / available time"; it improves the accessibility and effectiveness of panic attack intervention; users can receive intervention and support without appointments, covering common anxiety scenarios and expanding the breadth and depth of digital psychological intervention.
[0018] 3. Data-driven feedback loop promotes improvement and awareness: The entire process is recorded and presented quantitatively, forming a data loop. Users can intuitively see changes in key indicators; trends and achievement incentives enhance training motivation; adaptive optimization: user feedback feeds back into the weights, and content and parameters are gradually personalized.
[0019] 4. Engineering advantages: Explainable throughout the entire process: Each stage provides clear information, with clearly defined fields, formulas, thresholds, durations, and parameter ranges designed for each stage of data collection, judgment, indicators, and reporting, facilitating verification and reproduction; Stable peak identification: Integrating evidence state machines enhances robustness and reduces misjudgments; Privacy-friendly: Balancing privacy principles with feasibility, facilitating deployment in dual scenarios.
[0020] Compared to existing solutions that are "single-sensory, fixed-script, and lack quantification and personalization," this invention achieves synergistic advantages in four dimensions: intervention effect, personalization fit, interpretability, and engineering feasibility through multi-sensory active training, AI-customized dialogue, closed-loop data quantification, and low-cost, easy deployment. This can significantly improve the digital intervention effect for panic attacks and daily anxiety management. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the specific implementation methods and systems of this invention are further described in detail below with reference to specific embodiments. It should be understood that these embodiments are for illustration and explanation only and are not intended to limit the invention. Those skilled in the art, after reading this specification, can make appropriate adjustments to the order of steps, parameter values, etc., described herein without departing from the principles of this invention; such adjustments should all be considered to fall within the protection scope of this invention.
[0022] Example 1 This embodiment provides a complete hardware deployment scheme for an emotion training system based on multi-sensory intervention and AI dialogue. The system adopts a hybrid architecture of client-edge computing node-cloud service layer to ensure real-time data processing and privacy compliance.
[0023] The multi-sensory intervention module maps the panic attack process into multiple stages using a dynamic storm diagram model, and adjusts the parameters of visual, audio, and tactile stimuli in real time based on a peak detection state machine, thereby achieving active neuromodulation training throughout the entire attack cycle.
[0024] The multi-sensory intervention module is used to provide simultaneous multimodal guidance and active interactive training involving vision, hearing, and touch during panic or anxiety attacks. It is deployed on user terminal devices, including but not limited to smartphones, tablets, or wearable devices. Specific hardware configurations include: Visual presentation unit: It adopts a high refresh rate OLED display (refresh rate ≥90Hz) to render dynamic visual effects such as dynamic storm diagrams. The screen size is 5.5-6.7 inches and the resolution is no less than 1920×1080 to ensure the delicacy and immersion of visual stimulation.
[0025] Audio playback unit: Integrated dual speakers or Bluetooth headphones supporting stereo output, with a frequency response range of 20Hz-20kHz and dynamic volume adjustment range of 0-85dB. Audio files are pre-stored locally in lossless compression format (FLAC) to avoid network latency.
[0026] Haptic feedback unit: Built-in linear motor (LRA or Z-axis linear motor), adjustable vibration frequency range of 50Hz-300Hz, vibration intensity supports 256 levels of micro-level control. High frequency and high intensity vibration is output during peak periods (200-300Hz, intensity ≥200 levels), and low frequency and low intensity mode is switched during off-peak periods (50-100Hz, intensity ≤100 levels).
[0027] The AI dialogue system is used to provide cognitive counseling and structured intervention based on natural language processing during non-peak periods. It utilizes a lightweight natural language processing model (such as DistilBERT or TinyLlama) with fewer than 300M parameters, ensuring a response latency of ≤500ms on edge computing nodes (such as mobile NPUs or cloud inference servers). The dialogue system incorporates structured intervention scripts for cognitive behavioral therapy (CBT), supporting intent recognition, sentiment analysis, and empathic response generation.
[0028] The user data recording and feedback subsystem is used to collect, store, and analyze user interaction data, physiological proxy data, and subjective report data; generate quantitative assessment reports; and calculate based on the collected multi-source evidence, determine and display the user's panic or anxiety attack state based on the scoring results; including: Physiological data acquisition unit: Connects to an external heart rate monitoring device (sampling rate ≥1Hz) or integrates a PPG sensor via Bluetooth; acquires skin conductance response data through the device's built-in GSR sensor (sampling rate ≥4Hz). All physiological data undergoes edge computing processing at the terminal, and only the desensitized statistical characteristic values are uploaded.
[0029] Interactive behavior acquisition unit: Records raw data of user touch events, including A4 engagement factor data: user interaction duration (touch_duration), tap rate per unit time (tap_rate), touch area ratio (touch_area_ratio), and tap_impulse triggered by user taps; cumulative time (tts_follow_secs) for following voice guidance, peak phase duration (peak_phase_secs), peak phase time ratio for following voice guidance (tts_ratio), peak phase time ratio for maintaining interaction (peak_follow_ratio), peak phase number of active interactions (tap), and attack duration (Q2); A7 voice feature data: tremor index, etc.
[0030] Computing processor: Uses an ARM Cortex-A78 or higher level CPU, or integrates a dedicated AI acceleration chip, responsible for real-time calculation of the emotion intensity curve I(t), multi-source evidence scores, and peak detection status.
[0031] The multi-sensory intervention module maps the panic attack process into multiple stages using dynamic visual effect diagrams, such as dynamic storm diagram models. Based on the calculation results of multi-source evidence by the user data recording and feedback subsystem, it adjusts the parameters of visual, audio and tactile stimuli in real time to achieve active neuromodulation training throughout the entire attack cycle.
[0032] Example 2 An emotion training method based on multi-sensory intervention and AI dialogue includes the following steps: Step 1: The user initiates an AI dialogue for training to collect data; Step 2: Present the user with dynamic visual effects, play audio, and vibrate. Step 3: Real-time collection of user interaction data generated through AI dialogue and real-time calculation of the emotion intensity curve I(t) and physiological proxy data; calculation of multi-source evidence score; and obtaining multi-source evidence data. The dynamic visual effect image, audio, and vibration frequency are adjusted according to the multi-source evidence data: When the multi-source evidence score meets the PEAK peak entry condition, the dynamic visual effect image, audio, and vibration automatically switch to peak dynamic visual effect image, peak audio, and peak vibration frequency, and the user interacts through AI dialogue to enter the security center of the dynamic visual effect image; when the multi-source evidence score meets the RECOV non-peak entry condition, the dynamic visual effect image, audio, and vibration automatically switch to non-peak dynamic visual effect image, non-peak audio, and non-peak vibration frequency, and the dynamic visual effect image, audio, and vibration frequency are adjusted according to the multi-source evidence data. Step 4: After the AI dialogue ends, calculate the amygdala sensitivity index (ASI), prefrontal control index (PCI), growth score (G), and courage index (CI) based on multi-source evidence data, and generate a training report.
[0033] The process for calculating the interaction data and physiological proxy data obtained from the collected interaction data and physiological proxy data, and the emotion intensity curve I(t) in step 3 is as follows: A1 Timeline Data: The system clock synchronization uses the NTP protocol to ensure that the timestamp error of all events is <50ms. The training start time start_ts records the precise moment when the user clicks the "Start Training" button; the recovery confirmation time recovery_ts is detected when the emotional intensity I(t) is below the preset safety threshold for 3 consecutive minutes and the user actively confirms; the total session duration session_duration = recovery_ts - start_ts.
[0034] The emotional intensity curve I(t) is calculated as follows: I(t) = λ1z(HR(t)) + λ2z(GSR(t)) + λ3Q3(t); where: z is the z-score standardized function; HR(t) is the real-time heart rate data, GSR(t) is the real-time skin conductance response data, and Q3(t) is the subjective intensity score given by the user in real time; λ1, λ2, and λ3 are preset values, all ≥ 0, and λ1 + λ2 + λ3 = 1.
[0035] Based on the real-time calculation results of the emotion intensity curve I(t), the refined A2 recovery data is obtained: When I(t) reaches its peak value I_peak, the system starts a timer. `time_to_50%` records the precise time it takes for I(t) to drop from I_peak to 0.5 × I_peak; `time_to_safe` records the time it takes for I(t) to drop to the preset safety value I_safe (usually set to 0.3 × I_max, where I_max is the historical peak baseline). Both metrics are accurate to the millisecond level. `drop(time_to_safe) = (Baseline recovery time - Current recovery time) / Baseline recovery time`. The current recovery time is the time `time_to_safe` calculated for the symptoms to recover to a safe level in the current session, i.e., the time required for the emotional intensity I(t) to drop to the preset safety value. The baseline recovery time is the arithmetic mean of `time_to_safe` from multiple training sessions (the default baseline for the first session is 600 seconds).
[0036] The duration of an episode, Q2, in the A4 participation data can also be calculated: Q2 is the time from when a user's emotional / anxiety / panic episode is recognized by the system to when they recover. In this method, it is a time-based index calculated based on the emotional intensity curve I(t), defined as the length of time that I(t) is continuously higher than a preset episode threshold T_attack. The episode start point t_start is the time when I(t) first exceeds T_attack and remains above it for at least a first preset time period Δt_start; the episode end point t_end is the time when I(t) drops below the recovery threshold T_recover and remains below it for at least a second preset time period Δt_end. Q2 = t_end - t_start. In this embodiment, based on the available survey data on heart rate and skin conductance during panic attacks, the user's rating range is assessed. The preset episode threshold T_attack is set to 2.2, the first preset time period Δt_start to 100 seconds, the recovery threshold T_recover to 1, and the second preset time period Δt_end to 100 seconds, resulting in a calculated Q2 of 5 minutes.
[0037] Other A4 participation factor data: The guidance mode, guidance_mode_used, uses an enumeration value: it provides five scenarios for users to choose from before entering an AI dialogue: transportation, public places, alone, airplane, and night; the differences lie in the background color, wording, and rhythm template, while the training logic is consistent.
[0038] The total number of user clicks (tap_count) is the total number of valid click events during the session (press duration > 100ms and touch area > 20 pixels); the user interaction duration (touch_duration) is the sum of the press durations of all click events; the click rate per unit time (tap_rate = tap_count / session_duration) is standardized to times / minute.
[0039] Touch area ratio = average touch area / total interactive area of the screen (usually the central 60% area of the screen).
[0040] The pulse signal tap_impulse is calculated as the product of touch pressure (in Newtons) and press duration (in seconds). When there is no pressure sensing device, it is estimated by weighting the touch area and press duration.
[0041] The cumulative time (tts_follow_secs) for users to follow the voice guidance: When the AI dialogue system plays the guidance, a timer is started simultaneously; the system detects whether the user reads along or responds during the guidance playback through voice recognition (voice activity is detected and the volume is >40dB), and accumulates all valid following time periods, accurate to the second.
[0042] Peak phase duration peak_phase_secs: The time difference between the start of the countdown from the moment the state transitions to PEAK and the moment the state leaves PEAK (enters a non-peak state) is peak_phase_secs.
[0043] The percentage of time the user follows the voice guidance during peak phases (tts_ratio) = Follow time within peak phase / peak_phase_secs. Follow time within peak phases is the cumulative time the user actually spends following the voice guidance during the peak_phase_secs period (a subset of tts_follow_secs during peak phases).
[0044] The peak-phase interaction time ratio peak_follow_ratio: peak_follow_ratio = effective interaction time within the peak phase / peak_phase_secs. The effective interaction time is the cumulative time of user clicks, touches, or voice interactions during the peak_phase_secs period (interaction events are detected and the duration is >500ms), which is determined by touch event timestamps and voice activity detection.
[0045] A5 Scene Context Data: Coarse-grained location tags (coarse_location_tag) are obtained through device GPS or base station positioning. Geohash encoding is used to retain the first four characters (accuracy approximately 20km), without collecting precise coordinates. These location tags are used for subsequent analysis of training performance differences in different scenarios (e.g., home, workplace, public places).
[0046] A6 physiological proxy data: Heart rate HR(t) was extracted using the PPG signal peak detection algorithm and denoised by 5-point median filtering; Skin conductance response GSR(t) was recorded as skin conductance value (in micro Siemens) and trend term was eliminated using the baseline drift correction algorithm.
[0047] A7 speech feature data: User speech is collected through the device's microphone (sampling rate 16kHz). The maximum volume max_db and average volume avg_db are calculated using RMS energy per 20ms frame and converted to decibels. Speech rate is estimated using the frame synchronization rate of the speech recognition decoding graph, in words per minute. Pause rate: The percentage of silent segments (energy below the threshold) is detected. Tremor index: The fundamental frequency (F0) jitter parameter and zero-crossing rate change rate are extracted and calculated using the formula tremor_index = 0.6 × jitter + 0.4 × Δzero_crossing_rate.
[0048] Emotional intensity curve I(t) calculation: The system updates the I(t) value at a frequency of 1Hz during the session. Preset weights λ1, λ2, and λ3 are calibrated based on the user's baseline, with default configurations of λ1=0.4 (heart rate weight), λ2=0.4 (GSR weight), and λ3=0.2 (subjective rating weight). z-score standardization uses a sliding window statistic with a window length of 300 seconds, dynamically calculating the mean and standard deviation to ensure the standardization process adapts to individual differences.
[0049] Example 3 Compared with Example 2, this example differs in that it includes the implementation method of the core algorithm described in step 3.
[0050] Multi-source evidence score calculation: Etap, the interactive evidence, employs a weighted linear combination with preset coefficients λ4=0.6 (interaction duration weight) and μ=0.4 (touch area weight). The formula is Etap = z(tap_rate + 0.6·touch_duration + 0.4·touch_area_ratio).
[0051] Physiological evidence Ehr and Egsr: Direct z-score normalization of real-time physiological data using the overall distribution parameters of user historical session data.
[0052] Voice evidence Evoice: Only the tremor index is used in the main score calculation, and the remaining voice features are used for confidence correction.
[0053] Entering Peak State (PEAK): Triggered when Score_in ≥ T_in and duration ≥ τ_in. Score_in is calculated as weighted evidence fusion: Score_in = w1·Etap + w2·Ehr + w3·Egsr + w4·Evoice + w5·peak_mark. Default weight configuration: w1=0.25, w2=0.25, w3=0.25, w4=0.15, w5=0.1. T_in∈[1.0,1.5], τ_in=7.5-8.5 seconds. In this embodiment, the threshold T_in is initially set to 1.0 seconds, and τ_in=8 seconds.
[0054] Non-peak state (RECOV): Triggered when Score_out ≥ T_out and duration ≥ τ_out. Score_out = u1·(-ΔEhr) + u2·(-ΔEgsr) + u3·(-ΔEtap) + u4·(-ΔEvoice) + u5·better_mark, where Δ is the 30-second sliding window difference operator, representing the decreasing trend of evidence. Default weights: u1=0.3, u2=0.3, u3=0.2, u4=0.1, u5=0.1, T_out∈[0.8,1.2], τ_out=8-11 seconds. In this embodiment, the thresholds are T_out=1 second and τ_out=10 seconds.
[0055] The user data recording and feedback subsystem outputs control signals to the multi-sensory intervention module to achieve seamless switching.
[0056] Example 4 Compared with Example 3, the difference in this example is that: T_in∈[1.0,1.5], τ_in=7.5-8.5 seconds, and in this example, the threshold T_in is initially set to 1.5 seconds, and τ_in=8.5 seconds. T_out∈[0.8,1.2], τ_out=8-11 seconds, and in this example, the threshold T_out=1.2 seconds, and τ_out=11 seconds.
[0057] Example 5 Compared with Example 3, the difference in this example is that: T_in∈[1.0,1.5], τ_in=7.5-8.5 seconds, and in this example, the threshold T_in is initially set to 1.2 seconds, and τ_in=7.5 seconds. T_out∈[0.8,1.2], τ_out=8-11 seconds, and in this example, the threshold T_out=0.8 seconds, and τ_out=8 seconds.
[0058] Example 6 Compared with Example 3, this example differs in that it includes the specific implementation of the multi-sensory dynamic adjustment mechanism described in step 3.
[0059] The purpose of the dynamic visual effect diagram is to transform abstract emotional states into visual spatial metaphors. The spatial metaphor model can be a dynamic storm diagram, or it can use other dynamic scenes with a central stable region (such as waves, flames, tunnels, etc.), but its core training mechanism is the same.
[0060] In this embodiment, the dynamic visual effect diagram used is a dynamic storm diagram. The speed of the vortex and the size of the eye change according to the multi-source evidence score. When the multi-source evidence score meets the PEAK peak entry condition, it automatically switches to the peak stage, increases the vortex angular velocity, enlarges the eye, and enhances the audio and vibration intensity, guiding the user to actively interact through AI dialogue to enter the safe center of the dynamic storm diagram, which is displayed as the eye. When the multi-source evidence score meets the RECOV non-peak entry condition, it switches to the non-peak stage, and the vortex speed, eye size, audio, and vibration intensity are adjusted according to the multi-source evidence data.
[0061] The dynamic storm map changes according to the following response model: r(t+1)=r(t)·(1-β), ω(t+1)=ω(t)+γ·tap_impulse, where r is the viewing radius, r(t) represents the current eye viewing radius, ω is the vortex angular velocity, ω(t) represents the current vortex angular velocity, and β, γ∈(0,1] are response coefficients; the tempo of the audio is proportional to the score of the multi-source evidence data; the vibration frequency is proportional to the score of the multi-source evidence data.
[0062] The specific implementation method is as follows: Visual animation response model: The storm map is rendered using a pseudo-3D particle system, with the core parameters being the view radius r(t) and the vortex angular velocity ω(t). Initial state: r(0) = 0.8 (normalized radius), ω(0) = 0.5 rad / s. During state transition: Entering the PEAK peak state: β=0.15, γ=0.4, achieving eye contraction (r decreases) and vortex acceleration (ω increases), simulating the storm intensification effect.
[0063] Entering RECOV non-peak state: β=-0.1 (negative value indicates radius expansion), γ=-0.3, the eye of the storm expands, the vortex slows down, and a calm signal is transmitted.
[0064] The particle colors are dynamically mapped using the HSL color space: during the peak period, H=240° (deep blue-violet) and saturation S=90%, conveying a sense of tension; during the recovery period, H=180° (cyan-green) and S=60%, creating a soothing atmosphere. The number of particles is proportional to Score_in, with the particle density increasing to 200% during the peak period.
[0065] Phased audio guidance: Pre-set three-segment audio library: (1) Non-peak period: When the Score_out value is stable, a soothing piano melody is superimposed on a white noise base (40dB) at a tempo of 60BPM to guide the user to breathe steadily; when the Score_out value is unstable but gradually weakens, the intensity of the binaural beat is linearly reduced according to the Score_out value, and natural sound sampling (sea waves, forest wind) is increased, with the tempo gradually decreasing to 50BPM to guide the heart rate to slow down; (2) Peak period (PEAK): A binaural beat is added, with alpha waves (8-12Hz) and theta waves (4-7Hz) alternating, and the volume linearly increases to 65dB with the Score_in, with the tempo accelerating to 80-100BPM to enhance wakefulness and focus. The audio switching uses crossfading technology, with a fade-in and fade-out duration of 2 seconds to avoid abruptness.
[0066] Vibrational tactile stimulation: A sinusoidal modulated vibration signal is used. The peak frequency is f = 200 + 100 Score_in Hz, and the intensity is A = 200 + 55 Score_in. The off-peak frequency is f = 50 + 30 (1 - Score_out) Hz, and the intensity is A = 50 + 50 (1 - Score_out). The vibration duration adopts an intermittent mode: 2 seconds on and 1 second off during the peak period, and 1 second on and 3 seconds off during the recovery period.
[0067] Example 7 Compared with Example 6, this example differs in that it includes the precise calculation method and clinical interpretation of the four core indicators described in step 4.
[0068] The Amygdala Sensitivity Index (ASI) is calculated as ASI = z(Q3_peak) × norm(Q2). Specifically: Q3_peak is the integer value entered by the user when prompted "Please assess the current peak fear intensity (0-10 points)," with a range of 0-10; z(Q3_peak) is calculated using the cumulative distribution of the user's historical Q3_peak values; norm(Q2) can be normalized to a personal baseline: the arithmetic mean of the Q2 values collected each time the user uses the system is summed and divided by the number of times the system is used is the personal baseline value. The L2 norm of the personal baseline value is calculated, and the current data vector Q2 is subtracted from the personal baseline value. The difference is then divided by the L2 norm of the personal baseline value to obtain the normalized result.
[0069] The Prefrontal Control Index (PCI) is calculated as PCI = Q4 × Engagement × RS. Specifically, Q4 is the user's self-reported "sense of control during conversation," with a score ranging from 0 to 10; Engagement = σ(α1·Etap + α2·tts_ratio + α3·peak_follow_ratio), where σ is the Sigmoid function mapped to the (0,1) interval, with default values of α1=0.5, α2=0.3, and α3=0.2. tts_ratio represents the percentage of time spent guiding and following during the peak phase, peak_follow_ratio represents the percentage of time spent maintaining interaction during the peak phase, and RS = 1 / (1 + time_to_safe), normalizing time_to_safe (seconds) to the 0-1 range to reflect recovery speed. The PCI value reflects the user's prefrontal cortex's ability to inhibit the limbic system; a target value >5 indicates good control establishment.
[0070] It should be noted that although the above indicators are named "Amygdala Sensitivity Index" and "Prefrontal Control Index", these are functional training assessment indicators and do not constitute a medical diagnosis.
[0071] Growth score G: The calculation formula is G = 0.4·Q7 + 0.4·drop(time_to_safe) + 0.2·Q4. Implementation details: Q7 is the user's self-efficacy score (0-10); drop(time_to_safe) = (baseline recovery time - current recovery time) / baseline recovery time. Current recovery time is the time (time_to_safe) required for symptoms to recover to a safe level in the current session, i.e., the time required for emotional intensity I(t) to drop to a preset safe value. Baseline recovery time is the arithmetic mean of time_to_safe from multiple training sessions, with the baseline set to 600 seconds for the first session. Q4 has a lower weight to avoid excessive influence of a sense of control on growth assessment. The G score is used to track long-term progress.
[0072] Courage Index (CI): The calculation formula is norm(tap) + norm(High-Intensity Completion) + norm(Q7), where tap represents the number of active interactions during the peak phase, High-Intensity Completion is the user's self-reported score on the degree to which the user completed the task in the dialogue, and Q7 is the user's self-reported score on their self-efficacy, with a value of (0-10). The norm function normalizes the values, which can be normalized to a personal baseline: the arithmetic mean of the values collected each time the user uses the system is summed and divided by the number of times the system is used is the personal baseline value. The L2 norm of the personal baseline value is calculated, the current value is subtracted from the personal baseline value, and the difference is divided by the L2 norm to obtain the normalized result. This index balances the user's willingness to be exposed and sensitivity, guiding the dynamic adjustment of training difficulty. When CI < 0.5, the system automatically reduces the exposure level; when CI > 1.5, the challenge can be appropriately increased.
[0073] Example 8 Compared to Example 7, this example differs in that it describes the privacy protection measures implemented by the system in data collection and storage to comply with personal information protection regulations.
[0074] Minimal identifiability principle: Only store the coarse_location_tag category label, do not save the precise GPS coordinates; only store the speech acoustic features, do not save the original audio; On-device feature extraction: Speech processing and emotion classification are completed locally on the mobile device, while the cloud only receives anonymized labeled text; Separate table encrypted storage: Personal identity information (PII) and training logs are stored in separate tables, sensitive fields are encrypted with AES-256, and access permissions are minimized. Compliance Interface: Provides a channel for user data withdrawal and export that complies with PIPL / GDPR standards.
[0075] Example 9 User A, female, 28 years old, diagnosed with panic disorder. Upon first use, the system guided her through baseline calibration: wearing a heart rate bracelet and sitting quietly for 5 minutes to collect resting heart rate (HR) and gamma retrieval rate (GSR). Training then commenced. 1. Triggering Phase: The user experiences palpitations and rapid breathing, and actively launches the application. The AI dialogue system asks, "How are you feeling right now?", and the user replies via voice, "Very scared, my heart is beating very fast." The system detects max_db=78, tremor_index=2.3, and initially marks peak_mark=0.
[0076] 2. Peak Detection: After 30 seconds, HR rises from the baseline of 72 to 105, GSR increases by 180%, and tap_rate rises to 45 times / minute. Score_in is calculated to be 1.2, lasting for 8 seconds, and the state transitions to PEAK. The visual storm map's eye shrinks to r=0.3, angular velocity ω=1.2 rad / s, and the color changes to dark purple; audio initiates binaural beats; vibration begins with high-frequency pulses.
[0077] 3. Interactive Guidance: The AI plays the TTS guidance, "Follow my rhythm, inhale for 4 seconds, hold your breath for 4 seconds, exhale for 6 seconds." The user follows the breathing, accumulating 45 seconds of tts_follow_secs, increasing the touch area, and enhancing tap_impulse. The system updates the Score_in in real time, maintaining it between 1.2 and 1.4.
[0078] 4. Resumption Switching: After 120 seconds, HR drops to 85, GSR decreases by 40%, and ΔEhr and ΔEgsr become strongly negative. Score_out = 0.9 for 9 seconds, and the state transitions to RECOV. The storm graph shows the eye expanding and the color turning green; the audio gradually increases to a natural sound; the vibration shifts to a lower frequency.
[0079] 5. Session End: The user confirms that the panic has subsided and clicks "End". The recovery_ts record is at second 380. The system calculates ASI=6.2, PCI=5.8, G=6.5, CI=0.85, generates a report, and suggests increasing the exposure time by 5 minutes for the next training session.
[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Any modifications, equivalent substitutions, or improvements made by those skilled in the art within the scope of the claims of the present invention should be included within the protection scope of the present invention.
Claims
1. An emotion training method based on multi-sensory intervention and AI conversation, characterized in that: Includes the following steps: Step 1: The user initiates an AI dialogue for training to collect data; Step 2: Present the user with dynamic visual effects, play audio, and vibrate. Step 3: Real-time collection of user interaction data generated through AI dialogue and real-time calculation of the emotion intensity curve I(t) and physiological proxy data; calculation of multi-source evidence score; and obtaining multi-source evidence data. The dynamic visual effect image, audio, and vibration frequency are adjusted according to the multi-source evidence data: When the multi-source evidence score meets the PEAK peak entry condition, the dynamic visual effect image, audio, and vibration automatically switch to peak dynamic visual effect image, peak audio, and peak vibration frequency, and the user interacts through AI dialogue to enter the security center of the dynamic visual effect image; when the multi-source evidence score meets the RECOV non-peak entry condition, the dynamic visual effect image, audio, and vibration automatically switch to non-peak dynamic visual effect image, non-peak audio, and non-peak vibration frequency, and the dynamic visual effect image, audio, and vibration frequency are adjusted according to the multi-source evidence data. Step 4: After the AI dialogue ends, calculate the amygdala sensitivity index (ASI), prefrontal control index (PCI), growth score (G), and courage index (CI) based on multi-source evidence data, and generate a training report.
2. The emotion training method based on multi-sensory intervention and AI dialogue as described in claim 1, characterized in that: The interactive data and physiological agent data generated by the user through AI dialogue in real time in step 3 include: A4 engagement data elements: user interaction duration (touch_duration), click rate per unit time (tap_rate), touch area ratio (touch_area_ratio), and the pulse signal triggered by user clicks (tap_impulse); cumulative time for users to follow voice guidance (tts_follow_secs), peak phase duration (peak_phase_secs), the proportion of time spent following voice guidance during the peak phase (tts_ratio), the proportion of time spent maintaining interaction during the PEAK peak phase (peak_follow_ratio), and the number of times users actively interacted during the peak phase (tap). A6 physiological data: Heart rate (HR), respiratory rate, and skin conductance response (GSR); A7 speech feature data: tremor index; The calculation method for the emotion intensity curve I(t) in step 3 is as follows: I(t) = λ1z(HR(t)) + λ2z(GSR(t)) + λ3Q3(t); where: z is the z-score standardized function; HR(t) is the real-time heart rate data, GSR(t) is the real-time skin conductance response data, and Q3(t) is the subjective intensity score given by the user in real time; λ1, λ2, and λ3 are preset values, all ≥ 0, and λ1 + λ2 + λ3 = 1; The interaction data and physiological proxy data obtained by calculating the emotion intensity curve I(t) include: A2 recovery refinement data: time_to_50% for symptom intensity to decrease by 50%, time_to_safe for symptom recovery to a safe level, and recovery time reduction rate drop(time_to_safe); the time_to_50% value is the time required for I(t) to decrease from the peak to 50%, and the time_to_safe value is the time required for I(t) to decrease to the preset safe value. drop(time_to_safe) = (baseline recovery time - time_to_safe) / baseline recovery time, where the baseline recovery time is the arithmetic mean of time_to_safe from multiple training sessions of the user; and the duration of the episode Q2 of the A4 participation element data.
3. The emotion training method based on multi-sensory intervention and AI conversation of claim 2, wherein: The interactive data and physiological agent data generated by the user through AI dialogue in real time in step 3 also include: A1 timeline data: training start time start_ts, recovery confirmation time recovery_ts, and total session duration session_duration; A4 Participation Factor Data: The guidance mode used in this training (guidance_mode_used); A5 scene context data: user location category coarse_location_tag, without precise geographic coordinates; A7 speech feature data: maximum volume (max_db), average volume, speech rate, and pause rate.
4. The emotion training method based on multi-sensory intervention and AI conversation of claim 2, wherein: The method for calculating the multi-source evidence score in step 3 is as follows: Etap is obtained by standardizing the tap rate per unit time (tap_rate), user interaction duration (touch_duration), and touch area ratio (touch_area_ratio) using the z-score function, i.e.: Etap = z(tap_rate + λ4·touch_duration + μ·touch_area_ratio), where λ4 and μ are preset values, both ≥ 0 and summed to 1; Evoice is obtained by standardizing the tremor index (tremor_index) using the z-score function, i.e., Evoice = z(tremor_index); Ehr and Egsr are obtained by z-score normalization of heart rate (HR) and skin conductance response (GSR), respectively, i.e., Ehr = z(HR), Egsr = z(GSR); In step 4, when the multi-source evidence score meets the PEAK peak entry condition, the method for automatically switching to the dynamic visual effect, audio, and vibration frequency of the PEAK peak is as follows: it is triggered when Score_in ≥ T_in and the duration ≥ τ_in, where Score_in = w1·Etap + w2·Ehr + w3·Egsr + w4·Evoice + w5·peak_mark; In step 5, when the multi-source evidence score meets the RECOV non-peak entry condition, the method for switching to the dynamic visual effects, audio, and vibration frequency during the non-peak period is as follows: it is triggered when Score_out ≥ T_out and the duration ≥ τ_out, where Score_out = u1·(-ΔEhr) + u2·(-ΔEgsr) + u3·(-ΔEtap) + u4·(-ΔEvoice) + u5·better_mark; Where peak_mark is the peak mark value actively marked by the user, peak_mark∈{0,1}; better_mark is the improvement mark value actively marked by the user; T_in is the preset time threshold range representing the time threshold for entering the peak state; τ_in is the preset time threshold range for entering the peak state; T_out is the preset time threshold range for leaving the peak state; τ_out is the preset time threshold range for leaving the peak state; Δ represents the sliding window difference method; the weights w1-w5 and u1-u5 are preset values, both ≥0, and the sum of w1-w5 and u1-u5 is equal to 1.
5. The emotion training method based on multi-sensory intervention and AI conversation of claim 4, wherein: The dynamic visual effects, audio, and vibration frequencies in steps 3 and 5 are adjusted according to the multi-source evidence data through the following methods: When the dynamic visual effect diagram is a dynamic storm diagram, the speed of the storm vortex and the size of the eye are adjusted according to the multi-source evidence data. When the multi-source evidence score meets the PEAK peak entry condition, it automatically switches to the peak stage, increases the vortex angular velocity, enlarges the eye, and enhances the audio and vibration intensity, guiding the user to actively interact through AI dialogue to enter the safe center of the dynamic storm diagram, which is displayed as the eye. When the multi-source evidence score meets the RECOV non-peak entry condition, it switches to the non-peak stage, and the vortex speed, eye size, audio, and vibration intensity are adjusted according to the multi-source evidence data. The dynamic storm map changes according to the following response model: r(t+1)=r(t)·(1-β), ω(t+1)=ω(t)+γ·tap_impulse, where r is the eye radius, r(t) represents the current eye radius, ω is the vortex angular velocity, ω(t) represents the current vortex angular velocity, β and γ are response coefficients, both taking values ∈(0,1]; the tempo of the audio is proportional to the score of the multi-source evidence data; the vibration frequency is proportional to the score of the multi-source evidence data.
6. The emotion training method based on multi-sensory intervention and AI conversation of claim 5, wherein, The calculation methods for the ASI, PCI, G, and CI indicators in step 6 are as follows: The Amygdala Sensitivity Index (ASI) is calculated as: ASI = z(Q3_peak) × norm(Q2), where Q3_peak is the peak fear intensity score input by the user based on AI dialogue prompts, Q2 is the duration of the attack, z(Q3_peak) is the result of standardizing Q3_peak using the z-score function, and norm(Q2) is the result of normalizing Q2 using the norm function. Normalization can be performed based on the individual baseline: the arithmetic mean of the Q2 values collected each time the user uses the system is summed and divided by the number of times the system is used is the individual baseline value. The L2 norm of the individual baseline value is calculated, the current data vector Q2 is subtracted from the individual baseline value, and the difference is divided by the L2 norm of the individual baseline value to obtain the normalized result. The prefrontal control index (PCI) is calculated as Q4 × Engagement × RS, where Q4 is the user's self-input rating of their perceived control during the conversation; Engagement = σ(α1·Etap + α2·tts_ratio + α3·peak_follow_ratio), where σ is the Sigmoid function; RS = 1 / (1+time_to_safe); and α1, α2, and α3 are preset values, all ≥ 0 and satisfying α1 + α2 + α3 = 1. The growth score G = 0.4·Q7 + 0.4·drop(time_to_safe) + 0.2·Q4, where Q7 is the user's self-efficacy rating input based on AI dialogue prompts, and drop(time_to_safe) is the recovery time reduction rate calculated from the emotion intensity curve I(t); The Courage Index (CI) is calculated as: CI = norm(tap) + norm(High-Intensity Completion) + norm(Q7). Here, tap represents the number of proactive interactions during the peak phase, High-Intensity Completion is the user's self-assessment of their performance in completing tasks during the conversation, and Q7 is the user's self-assessment of their self-efficacy. The norm function normalizes the values, which can be based on a personal baseline: summing the values collected each time the user uses the system and dividing by the number of times the system is used, obtaining the arithmetic mean as the personal baseline value, calculating the L2 norm of the personal baseline value, subtracting the personal baseline value from the current value, and dividing the difference by the L2 norm to obtain the normalized result.
7. The emotion training method based on multi-sensory intervention and AI conversation of claim 6, wherein: The following privacy protection and data security mechanisms are adopted: Minimal identifiability principle: Only store the coarse_location_tag category label, do not save the precise GPS coordinates; only store the speech acoustic features, do not save the original audio; On-device feature extraction: Speech processing and emotion classification are completed locally on the mobile device, while the cloud only receives anonymized labeled text; Separate table encrypted storage: Personal identification information (PII) and training logs are stored in separate tables, sensitive fields are encrypted with AES-256, and access permissions are minimized. Compliance Interface: Provides a channel for user data withdrawal and export that complies with PIPL / GDPR standards.
8. An emotion training system based on multi-sensory intervention and AI conversation, applying the emotion training method based on multi-sensory intervention and AI conversation according to claims 1-7, characterized in that: Includes the following modules: The multi-sensory intervention module is used to provide simultaneous multimodal guidance and active interactive training through vision, hearing, and touch during panic or anxiety attacks in users; An AI dialogue system is used to provide cognitive tutoring and structured intervention based on natural language processing during non-peak periods. The user data recording and feedback subsystem is used to collect, store and analyze user interaction data, physiological proxy data and subjective report data, generate quantitative assessment reports, and perform calculations based on the collected multi-source evidence. Based on the scoring results, it judges and displays the user's state of panic or anxiety attacks. The multi-sensory intervention module maps the panic attack process into multiple stages through a dynamic visual effect model, and adjusts the parameters of visual, audio and tactile stimuli in real time based on the calculation results of multi-source evidence by the user data recording and feedback subsystem, so as to realize active neuromodulation training throughout the entire attack cycle.
