A sound stimulation adaptive regulation system based on abnormal state identification

Abstract

The application discloses a sound stimulation self-adaptive regulation system based on abnormal state identification, wherein a signal acquisition and calculation module calculates an anxiety index by acquiring various physiological and motion sensing data of a user in real time; a panic identification and alarm module first corrects a motion state according to the motion sensing data, then runs a rapid threshold path and an unsupervised model path in parallel, and outputs a panic attack alarm signal and an abnormal intensity score; a sound stimulation control module selects a sound stimulation mode according to the output of the panic identification and alarm module and the anxiety index, and outputs corresponding sound stimulation according to an individualized file of the user; and a parameter optimization learning module evaluates the effect of sound stimulation in real time during the stimulation intervention process of the sound stimulation control module each time, optimizes stimulation parameters in the next step, and files and learns whole-process data after the intervention ends. According to the application, early and accurate identification and timely intervention of a panic attack state can be realized.

Description

Technical Field

[0001] This invention belongs to the field of mental health monitoring and intervention technology, and in particular relates to an adaptive control system for sound stimulation based on abnormal state recognition. Background Technology

[0002] Panic attacks are a common manifestation of acute autonomic dysfunction, often occurring suddenly and reaching their peak within 10 minutes. Typical symptoms include strong palpitations, shortness of breath, chest tightness, dizziness, and other physical reactions, as well as intense fear (such as a feeling of impending death or loss of control). Currently, cognitive behavioral therapy (CBT) and drug therapy are the main clinical interventions for panic attacks, but these methods have significant limitations: (1) Drugs have a slow onset of action (usually several tens of minutes), making it difficult to control the intense symptoms in the early stages of an attack in a timely manner, and long-term use may lead to side effects such as addiction, drowsiness, and tolerance; (2) CBT requires patients to learn coping strategies when not having an attack, but patients often find it difficult to calmly implement the plan during an actual attack; (3) Existing protocols lack a real-time feedback and regulation mechanism for individual physiological states, and cannot dynamically adjust intervention strategies according to the real-time changes in physiological parameters of different patients. The lack of individualization leads to controversy regarding the effectiveness of the intervention.

[0003] Sound stimulation, as a non-pharmacological intervention, has been shown to have a certain regulatory effect on autonomic nervous system function and anxiety. For example, soothing music and natural environmental sounds (such as ocean waves and rain) can improve parasympathetic nerve activity, but they usually take 5–8 minutes to take effect, which is insufficient to meet the need for rapid relief in the early stages of a panic attack. In contrast, some spontaneous sensory meridian response (ASMR) sounds (such as whispers, brush strokes, and finger tapping) can induce relaxation responses in a short time: studies have shown that ASMR stimulation can enhance brain electrical activity. Wave, reduction ASMR sounds promote the activation of the parasympathetic nervous system and inhibit the sympathetic nervous system, thereby significantly relieving tension and anxiety within one minute. Therefore, using ASMR sounds for early intervention in panic attacks has the potential for rapid effectiveness.

[0004] However, current research also indicates that the long-term sustainability of the relaxation effect following such brief sensory stimulation remains uncertain, and the relaxation triggered by a single ASMR session is often transient. Without follow-up consolidation measures, the individual's autonomic nervous system may rebound to sympathetic dominance within minutes, thus reducing the overall intervention effect. Therefore, it is necessary to supplement ASMR with secondary interventions that can stabilize the rhythm and prolong the parasympathetic dominance state after rapid ASMR relief. Slow-paced sounds with a rhythmic frequency of 4–8 Hz are believed to help stabilize emotions. The "slow-beat" sounds used in this invention (e.g., low-frequency rhythms simulating a heartbeat or gentle drumbeats) are designed to harmonize the user's breathing and cardiovascular rhythms: a slow, fluctuating sound of 4–8 beats per minute gradually guides breathing and heart rate down to a resonant frequency of approximately 0.1 Hz, thereby further enhancing parasympathetic activity, inhibiting excessive sympathetic excitation, and prolonging the relaxation state after initial relief. Furthermore, studies have found that low-frequency pulsed sounds synchronized with the heart rhythm, when suddenly stopped, can induce a significant decrease in heart rate and autonomic balance adjustment within one cardiac cycle (<1 second) through cardiovascular reflexes. This mechanism is faster than most central nervous system regulatory processes, equivalent to a direct action on the acute reflex pathway of the autonomic nervous system. If a heart-synchronized pulse is introduced during the peak of a panic attack and interrupted at the appropriate time, it is expected to momentarily interrupt excessive sympathetic activation, creating favorable physiological conditions for subsequent rhythm stabilization interventions. In summary, the multi-stage sound intervention strategy of rapid ASMR induction + beat consolidation + heart rate synchronization pulse has the potential to achieve both rapid onset and sustained stability, with the two complementing each other's advantages.

[0005] Despite the potential shown by sound stimulation intervention, existing sound-based panic attack intervention methods still have shortcomings. First, most methods use fixed sound types and stimulation parameters, lacking a mechanism for hierarchical dynamic adjustment based on real-time physiological feedback. For example, patent document CN110743077A proposes a sound wave relaxation device that collects user physiological signals and sets individualized target parameters. During relaxation training, it monitors changes in physiological signals and controls the switching of music between multiple tracks in real time to guide the user towards gradual relaxation. This method utilizes physiological feedback to adjust music output, avoiding reliance on a human psychologist. However, such methods are mainly used for routine psychological relaxation training, not for immediate emergency intervention during panic attacks, and the sound stimulation used is singular, lacking hierarchical stimulation strategies for different stages of the attack. Second, while some wearable devices on the market can collect multiple physiological signals, the accuracy of multimodal data fusion and real-time application are still insufficient, limiting their role in automatic intervention triggering decisions. Many products only monitor and alarm for a single indicator, which can easily lead to delayed intervention or frequent false alarms. For example, existing systems can use wearable devices to predict panic attacks by analyzing various physiological data such as heart rate, skin conductance, and breathing patterns, and issue warnings to users before an attack. However, such systems typically only provide warning prompts or guidance for users to relax themselves, without integrating automatic stimulation intervention functions. Furthermore, the lack of effective recognition of the user's movement state makes it difficult for the system to distinguish between the physiological changes caused by intense physical movement and the autonomic nervous system response during a true panic attack, easily leading to false triggers and reducing user experience and trust. Single physiological signals often cannot reliably distinguish this difference; multi-sensor information must be combined. Literature shows that fusing signals from different modalities can significantly improve the accuracy of emotional state detection: for example, the accuracy of single-modal emotion recognition based on voice or ECG is only 70–80%, while the recognition rate can be increased to over 90% after fusing voice + ECG and other multi-modal signals. Therefore, systems lacking multi-parameter fusion and motion pseudo-signal correction often struggle to balance sensitivity and specificity, resulting in either missed or false alarms. These shortcomings mean that existing sound intervention systems remain unsatisfactory in terms of quickly and accurately identifying panic attacks and implementing effective interventions.

[0006] To address these issues, the industry has begun exploring new technologies integrating closed-loop detection and intervention. For example, patent document US11235156B2 proposes a wearable headphone device with built-in physiological sensors that continuously monitor the user's state. Once abnormal events such as anxiety or panic attacks are detected, it automatically provides electrical stimulation to the vagus nerve in the ear. However, a multimodal, high-precision monitoring system combined with personalized sound stimulation closed-loop control designed specifically for panic attacks is currently lacking. Summary of the Invention

[0007] This invention provides an adaptive sound stimulation control system based on abnormal state recognition, which can use multi-source physiological signals to assess the user's autonomic nervous state in real time and dynamically output graded sound stimulation based on the assessment results, so as to achieve early and accurate identification and timely intervention of panic attack state.

[0008] An adaptive sound stimulus control system based on abnormal state recognition includes a signal acquisition and calculation module, a fear recognition and alarm module, a sound stimulus control module, and a parameter optimization and learning module.

[0009] The signal acquisition and calculation module acquires various physiological and motion sensing data of the user in real time through wearable sensors, and synchronously fuses the acquired physiological data to calculate the anxiety index.

[0010] The panic detection and alarm module first corrects motion state based on motion sensor data, then runs a fast thresholding pathway and an unsupervised model pathway in parallel, and outputs a panic attack alarm signal. and abnormal intensity score Among them, panic attack alarm signals It is a Boolean quantity. A value of 1 indicates that the user is currently experiencing a panic attack. =0 indicates no seizure alarm or the alarm has been cleared; Abnormal intensity score The risk levels are further categorized into low risk, medium risk, and high risk.

[0011] The sound stimulation control module selects a sound stimulation mode based on the output of the panic recognition alarm module and the anxiety index, and outputs corresponding sound stimulation according to the user's individual profile under the selected sound stimulation mode.

[0012] The parameter optimization learning module evaluates the effect of sound stimulation in real time during each stimulation intervention by the sound stimulation control module and optimizes the next stimulation parameters; at the same time, it archives and learns the entire data after the intervention.

[0013] This invention uses wearable sensors as an entry point, integrating multi-source physiological and motor information to generate a real-time anxiety index and abnormal intensity score that intuitively reflect autonomic nervous system tension. Sound intervention is triggered under a recognition mechanism of "motor state correction + dual-pathway abnormality detection." The intervention employs a three-level strategy of "rapid—rhythmic—maintaining," and achieves safe, comfortable, and stable closed-loop control through online optimization and self-learning driven by individual acoustic profiles. After the intervention, the system encrypts and archives all data and incrementally updates the model to continuously improve individualized effects.

[0014] Furthermore, the physiological signals acquired in the signal acquisition and calculation module include: electrocardiogram (ECG) signal, captured heart rate (HR), heart rate variability (HRV), respiratory waveform, and skin conductance (EDA).

[0015] The collected motion sensing data includes: triaxial accelerometer data and gyroscope data.

[0016] Furthermore, after acquiring various physiological and motion sensing data, the signal acquisition and calculation module needs to calculate the signal quality index of each channel signal. The formula is as follows:

[0017] ;

[0018] In the formula, For signal-to-noise ratio, For data missing rate, The degree of the wake; , , These are the weighting coefficients; The range of values ​​is ,when Once the signal quality is deemed satisfactory, the system's panic detection and alarm module will activate. This is a preset signal quality threshold.

[0019] Furthermore, the specific process for calculating the anxiety index in the signal acquisition and calculation module is as follows:

[0020] Physiological signals collected by different sensors are preprocessed to obtain a synchronization signal sequence. Each second in this synchronization signal sequence contains a set of synchronized physiological signal values ​​for feature extraction.

[0021] With an update step of 1 second, a sliding time window is selected to calculate various physiological characteristics on the synchronization signal sequence;

[0022] Calculate different weights for each physiological characteristic. Standardized bias of all physiological characteristics According to the corresponding weight Summing yields the overall deviation score. ;

[0023] The composite deviation score is mapped to an anxiety index in the range of 0–100 using a sigmoid function. During mapping, the midpoint θ and slope k of the Sigmoid curve are adjusted according to the individual user situation so that 50 points correspond to the user's resting state.

[0024] The instantaneous anxiety index calculated every second is smoothed using an exponential moving average, resulting in the final output time. Corresponding smoothed anxiety index for:

[0025] ;

[0026] In the formula, This is the smoothing coefficient.

[0027] Furthermore, the panic recognition alarm module performs motion state correction based on motion sensor data, specifically including:

[0028] Calculate the user's average acceleration per unit time based on motion sensor data. and step frequency ;

[0029] if Exceeding the preset threshold and lasting for more than 5 seconds, or step frequency If the user reaches the brisk walking / running level, the user's exercise status flag is set to TRUE; otherwise, the exercise status flag is set to FALSE.

[0030] When the motion state flag is set to TRUE, the anomaly detection conditions for the fast thresholding path and the unsupervised model path are corrected; when the internal motion state flag is set to FALSE, no correction is made.

[0031] Furthermore, the working process of the fast thresholding pathway is as follows:

[0032] Direct threshold judgment rules are set for physiological features extracted from physiological data, for each physiological feature. Set an abnormal threshold ;

[0033] Whenever there is a deviation in physiological characteristics Exceeding the abnormal threshold If the direction of change is consistent with the pattern of a panic attack, then this physiological characteristic is marked as a suspicious abnormality at the current moment. ;

[0034] Multiple characteristic deviations After normalizing each MDC, the weighted summation yields the joint effect size. If the combined effect size Exceeding the set threshold for combined effect size Then the suspicious anomaly will be considered. Perform a count to obtain an anomaly count. ;

[0035] Based on joint effect size With joint effect size threshold The relationship between the numbers and the anomaly count The value, the output boolean exception flag. .

[0036] Furthermore, the working process of the unsupervised model pathway is as follows:

[0037] Anomaly detection of physiological feature patterns is performed using pre-trained single-class support vector machines and isolated forest models, respectively; the normalized anomaly probabilities output are respectively... and .

[0038] Furthermore, the panic detection alarm module outputs a panic attack alarm signal. and abnormal intensity score The specific process is as follows:

[0039] The comprehensive anomaly strength score is calculated based on the outputs of the fast thresholding pathway and the unsupervised model pathway. The formula is:

[0040] ;

[0041] In the formula, , , For the weighting coefficients, satisfying ;

[0042] The trigger condition is set using m-of-k logic, within a decision window of the most recent k seconds. If at least m moments meet the condition... This will trigger an alarm and output a panic attack alarm signal. ; This is the preset alarm threshold.

[0043] Furthermore, the sound stimulation control module includes a rapid intervention mode, a rhythmic calming mode, and a maintenance and consolidation mode for its sound stimulation modes;

[0044] The rapid intervention mode is activated when the user is in a high-risk state, at which point the panic detection alarm module outputs a panic attack alarm signal. = 1 and overall anomaly intensity score Or anxiety index This mode plays fast-paced, relaxing sounds; the duration will not exceed 60 seconds.

[0045] The condition for entering rhythm-easing mode is: the user's status is in the medium-risk range, at which point... or In this mode, theta rhythm beats are played for 3–5 minutes.

[0046] The entry condition for maintaining consolidation mode is: the user's status is at a low-risk level. < 0.30 and In this mode, soothing background sounds or soft music play for 3-5 minutes.

[0047] During the playback of each sound stimulus pattern, the system continuously monitors the real-time changes in the user's physiological indicators and obtains a comprehensive abnormality intensity score. and anxiety index To assess the effectiveness of the intervention, after each sound stimulation pattern ends, the next phase pattern is selected based on the intervention effect, or the sound stimulation is discontinued.

[0048] Furthermore, a user's personalized profile includes individualized sound stimulation thresholds, sound preferences, and safety restrictions.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This invention utilizes a "dual-pathway" identification mechanism combining multimodal physiological signal fusion and motion state correction to achieve early, accurate, and low-false-alarm detection of panic attacks. Based on a graded sound stimulation strategy, it provides rapid-acting acoustic stimulation at the onset of an attack, achieving second-level intervention, followed by a steady-state transition with rhythmic sound and consolidation with ambient sound, significantly shortening relief time and prolonging the relaxation effect. It establishes individualized acoustic profiles and adaptively adjusts parameters based on real-time physiological feedback during intervention, achieving precise control and a comfortable experience tailored to the individual and the time. The invention incorporates self-learning and historical optimal solution hot-start, ensuring continuous optimization and increasing effectiveness with use. Smooth transitions and safety boundaries are set throughout the process, controlling intensity and switching seamlessly, balancing effectiveness and safety. It provides interpretable state quantities such as anxiety indices and intervention logs for easy effect tracking and long-term management in clinical or home settings. The invention employs a wearable and mobile integrated approach, making it non-drug, non-invasive, low-barrier, and easy to deploy, while local processing and encrypted archiving ensure privacy and data security. It possesses robustness and redundancy strategies for sensor deficiencies and complex environments, ensuring stable and reliable closed-loop control capabilities even under everyday use conditions. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the workflow of an adaptive sound stimulus control system based on abnormal state recognition, according to an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0055] like Figure 1 As shown, an adaptive sound stimulus control system based on abnormal state recognition operates as follows: Using a wearable sensor as the entry point, data is first acquired and monitored via a "signal acquisition—signal quality" link. If the quality is unsatisfactory, the data is returned for acquisition; otherwise, it enters the "anxiety index calculation" channel, continuously outputting the status for downstream use. Subsequently, "motion state correction" uses motion information to eliminate motion artifacts. The corrected features are then sent to "fear recognition (dual-pathway)" for parallel rapid rule and unsupervised model judgment. If the result is "no," it returns to the anxiety index channel for continued monitoring; if it is "yes," "sound stimulus" is triggered. Sound stimulus is executed according to a three-level strategy of "rapid intervention—rhythm easing—maintenance and consolidation," and is monitored by a "termination judgment": intervention continues until stability is achieved, ending the process and transitioning to "parameter optimization and self-learning." This unit archives the data from this round, updates the model, and improves individual acoustic profiles. Simultaneously, it sends real-time parameter adjustment commands back to "sound stimulus" for online fine-tuning, forming a closed-loop collaboration of "detection—intervention—evaluation—optimization," making subsequent triggering faster, more accurate, and more individualized.

[0056] The system of this invention mainly includes a signal acquisition and calculation module, a fear recognition and alarm module, a sound stimulus control module, and a parameter optimization and learning module. These four modules are described in detail below.

[0057] Module 1: Signal Acquisition and Calculation Module

[0058] This module acquires various physiological and behavioral data from users in real time through a wearable sensor network, and simultaneously fuses the multi-source data to calculate an anxiety index (AI). It ensures that various signals are collected synchronously with sufficient sampling accuracy to capture subtle changes in autonomic neural activity during a panic attack, and transforms complex multi-channel physiological changes into an intuitive and quantifiable anxiety index for use by subsequent modules.

[0059] 1.1 Signal Acquisition Submodule

[0060] The signal acquisition submodule synchronously controls various sensors using a unified time reference, continuously and in parallel acquiring raw signals output from various physiological sensing devices worn by the user, including but not limited to: electrocardiogram (ECG) signals, capturing key data such as heart rate (HR) and heart rate variability (HRV), with a sampling rate setting ≥250 Hz; respiratory waveform monitoring of respiratory frequency and depth changes, with a sampling rate ≥25 Hz; electrical conductance of skin (EDA) reflecting minute changes in skin sweat gland activity, with a sampling rate ≥4 Hz; skin temperature – indicating changes in peripheral blood flow, an optional parameter; electromyography (EMG) capturing changes in muscle tension, an optional parameter; blood oxygen saturation (SpO2) reflecting blood oxygen content, an optional parameter; electroencephalography (EEG) recording brain activity as an optional reference signal; motion sensing signals including triaxial accelerometer and gyroscope data, used to monitor user activity status and posture changes; and auxiliary information input (optional), such as facial expression / body posture changes captured by the camera, voice features collected by the microphone, and subjective reports obtained by the mobile app.

[0061] The signal acquisition submodule ensures that all sensor data is recorded synchronously using a unified clock: to meet real-time assessment requirements, the main physiological signals use a specified high sampling rate (as described above), and a unified timestamp or sequence number is attached to each data sample to achieve precise alignment across channels. During operation, the module also performs basic data quality monitoring: it performs preliminary filtering on signals from each channel (e.g., removing ECG power frequency interference, identifying and marking motion artifacts), eliminating obviously distorted data segments; and it detects sensor status (e.g., poor electrode contact, device disconnection) and displays signal quality indicators. During acquisition, specific sensors can be dynamically activated / deactivated as needed (e.g., temporarily turning off the microphone in noisy environments to save energy).

[0062] Based on this, the signal acquisition submodule further calculates the signal quality index for each channel. This metric is used to measure the reliability of current physiological signals. It is based on a multi-factor weighted calculation, including signal-to-noise ratio. Data missing rate and the degree of motion trail The calculation formula is as follows:

[0063] ;

[0064] Its value ranges from [0,1], with a larger value indicating a more reliable signal. When (like The system will only trigger the alarm logic when it considers the signal quality to be up to standard.

[0065] The signal acquisition submodule outputs a multimodal raw data stream that has been time-synchronized and verified. Specifically, this includes physiological signal data from each channel arranged in time series, motion state data, and possible auxiliary information inputs (all with a unified timestamp). This data will serve as the foundational input for subsequent fusion and recognition modules. The main physiological signals are used to extract features and calculate the anxiety index; motion sensor data is directly provided to the panic detection and alarm module to identify motion artifacts and dynamically adjust thresholds; additional visual, audio, and subjective feedback information can be used to corroborate the recognition results or for manual verification. The signal acquisition submodule operates continuously from system startup, providing uninterrupted raw data support for continuous monitoring and closed-loop control.

[0066] 1.2 Anxiety Index Calculation Submodule

[0067] The anxiety index calculation submodule synchronously fuses and extracts features from multi-source physiological data from the signal acquisition submodule to calculate the anxiety index (AI), a comprehensive indicator reflecting the user's current level of anxiety / tension, thereby achieving real-time quantitative assessment of the user's autonomic nervous system state. The anxiety index ranges from 0 to 100, with 50 corresponding to a normal, quiet state, 0 indicating extreme relaxation, and 100 indicating extreme anxiety. This index will serve as a crucial basis for the identification and control modules.

[0068] The data source for the anxiety index calculation submodule is the synchronous multi-channel physiological signal sequence provided by the signal acquisition submodule, as well as the baseline statistical values ​​(resting mean and standard deviation of each indicator) and initial values ​​of channel weights of the user in the resting state.

[0069] The calculation process for the anxiety index includes the following steps:

[0070] Multi-channel alignment and preprocessing: Due to the different sampling rates of different sensors and the potential for data loss, asynchronous signals need to be aligned and interpolated. A multimodal Kalman filter algorithm is used, treating each channel signal as an observation and estimating the "optimal" signal value on a unified time axis in real time, thus outputting aligned multi-channel signal samples per second. Simultaneously, significant noise (such as power line interference and motion artifacts in ECG signals) is filtered out to improve signal reliability. After this processing step, a set of synchronized physiological signal values ​​can be obtained every second for feature extraction.

[0071] Sliding window feature extraction: A sliding time window of a certain length (e.g., 30 seconds) is selected with an update step of 1 second to calculate various physiological characteristics on the synchronized signal sequence. Typical features include: current heart rate (HR) (estimated from the ECG RR interval), respiratory rate (RR), time-domain indices of heart rate variability (e.g., RMSSD, root mean square of the difference between adjacent heartbeats), frequency-domain indices of HRV (e.g., low-frequency / high-frequency power ratio LF / HF), mean skin conductance (SCL), and transient peak response count (SCR) of skin conductance. In addition, shorter windows (e.g., 5 or 10 seconds) can be used to calculate transient characteristics, such as the instantaneous rate of heart rate rise and sudden increases in EDA. The calculated series of feature values ​​are then compared with the user's resting baseline and converted into standardized deviation values ​​(Z-scores): the difference between the current value of each feature and the resting mean divided by the resting standard deviation, representing the degree of deviation of that feature relative to the individual's resting state.

[0072] Feature threshold discrimination and weighted fusion: To improve robustness, a change threshold dead zone is set: when the deviation of a feature does not exceed its minimum detectable change (MDC), the change is considered to be likely submerged in noise and can be regarded as "non-contributing" to the overall anxiety state, and the feature deviation is recorded as 0; when the feature deviation exceeds the MDC, it is considered to have a significant change. Given that different physiological indicators have different sensitivities and reliability to anxiety states, the module assigns different weights to each feature. First, the reliability of the instantaneous signal is assessed based on the observation error covariance of this channel obtained by Kalman filtering. (The smaller the error, the higher the reliability), and then combine it with the feature importance coefficients obtained from prior settings or online learning. Calculate the unnormalized weights and normalize them to obtain This ensures a weighted vector that sums to 1. The calculation formula is as follows:

[0073] ;

[0074] Then, standardize the deviations of all features. The overall deviation score is obtained by summing the scores according to their corresponding weights:

[0075] ;

[0076] The score reflects the overall deviation of the current autonomic activation level from the resting baseline. The more significant the deviation and the more likely it is to originate from a high-confidence channel, the greater its contribution.

[0077] Mapping to Anxiety Index: The composite deviation score is mapped to an anxiety index in the range of 0–100 using a sigmoid function. During mapping, the midpoint θ and slope k of the Sigmoid curve are adjusted according to individual user circumstances so that 50 points correspond to the user's resting state, and the sensitivity of the score to physiological changes conforms to empirical judgment. The mapping formula is as follows:

[0078] ;

[0079] Where θ is the midpoint parameter (usually the mean or median of the score at rest), and k is the slope parameter (usually between 0.5 and 2.0, which can be set according to individual resting fluctuations) to control the sensitivity of the index to physiological deviations.

[0080] Smoothing the output anxiety index: The instantaneous anxiety index calculated every second is smoothed using an exponential moving average to reduce misjudgments caused by instantaneous fluctuations. Assuming a smoothing coefficient α≈0.2, the final smoothed anxiety index output is:

[0081] ;

[0082] This smoothing process allows the historical data from the past few seconds to still have some influence on the current index, thus resulting in a stable anxiety index. .

[0083] Output: Anxiety Index AI (0–100). This index is updated in near real-time (with a delay in the order of seconds) as an objective measure of the user's current autonomic nervous system tension. In addition to outputting the AI ​​value, the module also retains information such as the features of each component and their Z-scores, which can be accessed and referenced by other modules.

[0084] The signal acquisition and calculation module provides two main outputs to the downstream: one is the anxiety index time series and key physiological characteristics used by the panic recognition and alarm module for abnormal state determination; the other is real-time status feedback used by the sound stimulus control module and parameter optimization learning module to determine intervention triggers and evaluate intervention effects (for example, the level of AI can determine whether emergency intervention is needed, and changes in AI are used to evaluate the effectiveness of sound stimuli, etc.).

[0085] 2. Panic Recognition Alarm Module

[0086] The panic detection and alarm module is the core decision-making unit of the system, responsible for detecting panic attacks from real-time physiological data and triggering alarm signals to initiate intervention. This module achieves high sensitivity and low false alarms through a combined mechanism of "dual-pathway anomaly detection + motion state correction": first, it detects whether the user is engaged in strenuous activity and dynamically adjusts the judgment threshold to eliminate interference from drastic physiological changes caused by movement; then, it runs two algorithms in parallel—fast feature threshold filtering and an unsupervised anomaly detection model—to identify abnormal patterns in the autonomic nervous system. After fusing the results from the two pathways, the final alarm signal is obtained through m-of-k rolling decision and hysteresis logic. This module is not only used for detection and triggering at the initial stage of an attack but also continuously monitors the intensity of abnormalities during intervention to determine when to end the intervention or whether a secondary attack has occurred.

[0087] The main inputs to the panic recognition alarm module come from multiple physiological feature time series (HR, RMSSD, EDA, etc.), motion sensor data (accelerometer, gyroscope signals), and signal quality indicators from the signal acquisition and calculation module. (Indicates the reliability of the current data). The module also refers to the threshold parameters of various features when the user is at rest (such as feature MDC, resting mean, etc.).

[0088] 2.1 Motion State Correction Submodule

[0089] The motion state correction submodule first uses acceleration and gyroscope data to calculate the user's motion intensity index, such as the average acceleration per unit time. Step frequency These parameters are used to determine whether the user is currently in a state of vigorous activity. For example:

[0090] Mean acceleration The calculation formula is as follows: For time windows Mean internal acceleration.

[0091] ;

[0092] Step frequency The calculation formula is as follows: For time windows The step frequency per minute converted from the number of steps detected internally.

[0093] ;

[0094] if If the preset threshold is exceeded and lasts for more than 5 seconds, or if the cadence reaches the level of brisk walking / running, the internal motion status flag will be set to TRUE, indicating that the current physiological signal may be affected by strenuous exercise.

[0095] When the exercise indicator is TRUE, the system assumes that the increase in heart rate and respiration may be due to exercise rather than a panic attack. Therefore, the criteria for judging a panic attack are dynamically modified: (1) Increase the threshold for judging abnormality: for example, the requirement that the increase in heart rate above the resting baseline be increased from "+2 bpm" to "+5 bpm" or more is considered abnormal; the decrease in HRV needs to be more significant (e.g., RMSSD decreases by more than 25% instead of 10%) to be considered abnormal; (2) Increase the requirement for multiple indicators to be abnormal at the same time: when multi-channel data is available, other physiological indicators, such as a significant increase in EDA and abnormal changes in respiratory variability, must appear at the same time to trigger an alarm; if a certain auxiliary channel is missing, the threshold of the remaining channels is further increased to compensate for insufficient information and ensure that there is no false alarm due to fluctuation of a single indicator during exercise.

[0096] If the exercise status flag is FALSE (user at rest or with slight activity), the system uses the individual's resting state sensitivity threshold for judgment. Exercise status detection itself does not directly output an alarm; instead, it serves as a dynamic gating mechanism to adjust the sensitivity of subsequent anomaly detection in real time. To avoid false alarms caused by residual high heart rate immediately after exercise, the system sets a threshold recovery period: when the exercise status changes from TRUE to FALSE, the judgment threshold linearly and gradually returns to the resting level over the next 1–2 minutes. For example, the heart rate abnormality threshold... As time t progresses, the motion threshold gradually transitions back to the resting threshold:

[0097] ;

[0098] in The end of the exercise. This is the threshold recovery time. This measure prevents alarm judgment jitter caused by a sharp drop in the threshold.

[0099] 2.2 Dual-path anomaly detection submodule

[0100] After eliminating interference from motion factors, the dual-pathway anomaly detection submodule performs true panic attack anomaly detection on multi-channel physiological characteristics. It employs parallel analysis of two pathways simultaneously, with each pathway independently providing anomaly indications.

[0101] (1) Rapid Threshold Filtering Pathway: This pathway sets direct threshold judgment rules for several key physiological characteristics, characterized by simple calculation and extremely low latency (<1 second). Typically, four indicators most sensitive to panic attacks are selected, such as heart rate (HR), HRV time-domain indicators (e.g., RMSSD), skin conductance level (SCL), and instantaneous heart rate change rate, to form a rapid threshold filtering path. For each feature Set an abnormal threshold This threshold is taken as an empirical value. Set the desired direction symbol For example, HR: HRV: Based on the empirical threshold of this indicator. The maximum value is taken from the minimum detectable change value (MDC) of this indicator to ensure that the threshold change is statistically detectable.

[0102] ;

[0103] Whenever a certain feature deviation If the value exceeds the threshold and the direction of change is consistent with the typical pattern of a panic attack (e.g., increased heart rate and decreased RMSSD), then this feature is marked as "suspicious abnormality" at the current moment:

[0104] ;

[0105] in This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise.

[0106] However, the anomaly labeling of a single feature may be affected by noise. For robustness, a weighted effect size similar to that used in the aforementioned effective response determination is introduced: the biases of multiple features are normalized according to their respective MDCs and then weighted and summed.

[0107]

[0108] Among them, weight The parameters are set based on the signal-to-noise ratio, reliability (ICC), and historical stability of each feature. If the combined effect size exceeds the threshold... (If a value of 1.0 is used, the overall deviation reaches the minimum detectable level; usually, a value of 1.0 is used, but it can also be individually adjusted within the range of 0.8–1.5), then a significant abnormality is considered to have occurred, and the abnormality count is recorded. .

[0109]

[0110] In signal quality Provided the criteria are met, the final determination of the fast track will adopt a parallel "voting + fusion" principle:

[0111] ;

[0112] in, The minimum number of abnormal features threshold (e.g.) ), For the threshold of the combined effect size (e.g.) The corresponding overall deviation reaches the MDC level.

[0113] (2) Unsupervised model approach: Anomaly detection of feature patterns is performed using pre-trained machine learning models. The module runs two unsupervised anomaly detection models in parallel: One-Class SVM and IsolationForest. These models have been trained on system initialization or normal user historical data to represent the "normal state" distribution.

[0114] One-Class SVM: Uses the RBF kernel to map multidimensional physiological features to a high-dimensional space, learns decision boundaries around normal data, and its discriminant function outputs the distance from the sample to the boundary. The format is:

[0115]

[0116] in, For support vector weights, The kernel function output represents the similarity between the input sample and the normal samples in the training set. This is the boundary constant. When When this occurs, it is considered abnormal.

[0117] Isolation Forest: This method randomly selects a subset of features to construct a large number of decision trees. The degree of anomaly is measured by how easily a sample is separated within the tree. The average path length for a given sample is also considered. (The number of branches required to find x on each tree in the forest), calculate the normalized outlier score based on the training data size N. :

[0118] ;

[0119] in This is a normalization constant for the data size. A value closer to 1 indicates a higher likelihood of isolation (more anomalous), while a value closer to 0 indicates a more normal pattern. The system threshold is set to the 99th percentile of the anomaly score distribution in the training set (this can be adjusted appropriately within the 95%-99% range based on the actual cost of false positives / false negatives in the application). If the module outputs anomaly scores... If the value is greater than this quantile, it is considered abnormal.

[0120] The unsupervised model pathway outputs two values ​​at each time step: the SVM discriminant function. (Can be converted to probability or distance) and outlier scores in Isolation Forest To facilitate fusion, they are normalized to the 0–1 interval respectively.

[0121] 2.3 Result Fusion and Alarm Decision Submodule

[0122] The results from the fast path and the model path need to be comprehensively evaluated to provide the final alarm decision. During fusion, the reliability of each path is weighted, and the signal quality index Q(t) is used to control alarm activation and deactivation. (like The system will only trigger the alarm logic when it considers the signal quality to be up to standard.

[0123] Comprehensive anomaly intensity calculation: Let the Boolean anomaly label (or effect size probability) output by the fast pathway be... The normalized outlier probabilities of the unsupervised model path output are respectively and The comprehensive anomaly intensity score can then be expressed as:

[0124] ;

[0125] in, , , satisfy The default weighting is average, but it can also be dynamically adjusted based on signal quality; for example, the weight of a channel can be reduced when the quality of a channel it depends on deteriorates. The calculated... A larger value indicates a more significant abnormal pattern.

[0126] Triggering condition (m-of-k logic): To improve robustness, a sliding window counting method is used to determine alarms. The condition is set to be met within the most recent k-second judgment window if at least m conditions are met. (The overall abnormal intensity exceeds the alarm threshold, such as...) = 0.5), then an alarm is triggered. Taking typical values ​​of k=5 and m=3 as an example, if there is an anomaly for 3 seconds within a continuous 5 seconds, it is determined to be a continuous anomaly, and an alarm trigger signal is output. This m-of-k logic can filter out individual isolated spike interference. When An alarm will be triggered at that time.

[0127]

[0128] Alarm hysteresis hold: Once the alarm is triggered as TRUE, the system initiates a hysteresis mechanism: Requires Drop below the threshold and remain at least The alarm will be deactivated after a few seconds (e.g., 5 seconds). This mechanism avoids frequent triggering / deactivation caused by repeated jitter at the threshold.

[0129] Model Adaptive Update: To maintain detection accuracy, the module periodically updates the anomaly detection model in the background using incremental new data. Specifically, this includes appending normal physiological data accumulated during intervention intervals to the training sets of One-Class SVM and Isolation Forest, retraining or fine-tuning the model boundaries to reflect the user's latest normal physiological range (ensuring the model learns the normal envelope over time, reducing false positives). Simultaneously, feature thresholds in the fast thresholding pathway are adjusted. The system updates based on the latest resting data:

[0130] ;

[0131] in The update coefficient is (0.7–0.9). This represents the average change in the new baseline measurement. This update ensures smooth threshold adjustment, unaffected by short-term fluctuations. When the system detects that certain channels are missing for an extended period (e.g., a user stops wearing a sensor), the module switches to a pre-trained dimensionality reduction model (redistributing weights after removing missing channels) to maintain the continuity of the recognition function.

[0132] The output results of the result fusion and alarm decision submodule are as follows:

[0133] Panic attack warning signals : Boolean value indicating whether a high-probability panic attack event was detected. When =1, it indicates that the user is very likely experiencing a panic attack, and the intervention process is triggered with high priority. = 0 indicates that no alarm has occurred or the alarm has been cleared. In implementation, Determined by the above m-of-k decision logic, and taking into account signal quality control, it is triggered only when the data is reliable.

[0134] Abnormal intensity score A continuous numerical value (0-1 range) is used to quantify the severity of the current panic state. This score is essentially a comprehensive assessment of the anomaly intensity. Based on the model confidence scores, risk levels can be further subdivided: for example, 0.0–0.3 low risk (only slight abnormal signs), 0.3–0.6 intermediate risk (moderate abnormality), and 0.6–1.0 high risk (significant abnormality). This score will be used in the sound stimulation control module to determine the intensity of intervention (a higher score results in a stronger stimulus), and can also be used to assess the effectiveness of the intervention (this value should decrease after the episode subsides). In addition, the module can output abnormality type information (which indicators triggered the alarm) in the system backend for recording and analysis, and supports providing auxiliary judgment criteria (such as combining typical symptom subjective reports from DSM-5 as auxiliary verification).

[0135] Alternative solutions: The dual-path mechanism adopted in this module has fully combined the advantages of real-time evaluation and unsupervised learning. However, in specific scenarios or when data is limited, some simplified alternative solutions can also be adopted, such as: (1) Fixed threshold method: Pre-set absolute thresholds (such as heart rate > 120 bpm or RMSSD < 20 ms) and simply judge if it exceeds the threshold to trigger an alarm; (2) Sliding window statistical method: Calculate the short-term mean and standard deviation of features in real time. When the current value deviates from the mean by more than ±k (e.g., 3 times the standard deviation) is considered abnormal; (3) Template matching method: Dynamic time warping (DTW) or cosine similarity is used to compare the current feature curve with the historical typical panic attack pattern. If the similarity is high, it is considered abnormal; (4) Subjective report trigger: Combined with the clinical diagnostic criteria (DSM-5), if the user reports a sudden strong fear with typical symptoms (palpitation, shortness of breath, etc.) through the App, it is directly used as one of the alarm trigger conditions; (5) Group model comparison: Principal component analysis (PCA) model is established using the feature distribution of a large number of normal people to determine whether the current data falls outside the normal range of the group. These schemes can be used as redundancy or verification methods of this module and can be selected according to the application requirements. Overall, this identification module achieves early and accurate identification and reliable alarm of panic attack state through multi-indicator fusion and dynamic adjustment, providing a decision basis for the timely initiation of subsequent sound intervention.

[0136] 3. Sound Stimulation Control Module

[0137] The sound stimulation control module is the execution unit of the system's closed-loop regulation. After a panic attack is identified, it outputs appropriate sound stimulation to regulate the user's autonomic nervous system state based on the decision results of the identification module and the user's individual profile. This module implements a three-level intervention strategy, managing the switching between different intervention modes through a state machine, gradually guiding the user from the peak of panic to a stable state with progressively increasing sound types and intensities. The module also includes intervention termination judgment logic to determine when to safely end the sound stimulation based on the user's physiological recovery.

[0138] Inputs include alarm trigger and deactivation signals from the panic detection alarm module. = 1 or 0), user thresholds and preference parameters from individual acoustic profiles (threshold sound pressure level EMR, preference degree, initial recommendation strength, safety limit, etc. for each sound type), and real-time physiological feedback from the acquisition / fusion module (anxiety index AI, abnormal intensity score). (HR, HRV, EDA, etc.). In addition, the parameter optimization learning module provides real-time parameter adjustment instructions to guide this module in fine-tuning specific parameters of the sound output (such as sound intensity and rhythm) in a given mode.

[0139] The sound stimulation control module is divided into three continuous modes based on the developmental stages of panic attacks: rapid intervention, rhythmic calming, and maintenance consolidation. The mode switching logic is managed through a finite state machine. Each mode has its specific function: the rapid mode is used for emergency symptom suppression, the rhythmic mode is used for consolidation and stabilization, and the maintenance mode is used for recovery and transition. The entry conditions, stimulation strategies, durations, and switching conditions for each mode are explained below:

[0140] 3.1 Rapid Intervention Submodule

[0141] Entry condition: Alarm triggered by the identification module ( = 1), and the overall anomaly intensity score reaches high risk (e.g., (or anxiety index) This mode also automatically enters when there is no alarm but the user manually triggers emergency intervention. At this time, the user is in the early or peak stage of a panic attack and needs immediate and strong intervention.

[0142] Stimulation Action: The module immediately plays selected fast-relaxing sounds from the user's acoustic profile. Prioritizing ASMR sounds that have been verified to rapidly induce a parasympathetic response (such as whispers, rustling, tapping, etc.), the volume is set to the user's effective minimum threshold EMR for that sound type plus approximately 3 dB(A) (slightly higher than the threshold to ensure effectiveness). If this information is lacking in the user profile, the system's default low-frequency, gentle ASMR sound is used as the initial stimulus (frequency 50–500 Hz, sound pressure level approximately 45–55 dB(A)). During sound playback, the module continuously monitors real-time changes in the user's heart rate, skin conductance, and other physiological indicators, and obtains the data calculated by the recognition module. And AI values ​​to assess the effectiveness of the intervention.

[0143] Emergency Intervention Sub-process: If an extremely abnormal and rapidly deteriorating user condition is detected (e.g., HR spikes to ≥120 bpm or AI spikes to ≥90), the module will trigger an emergency intervention sub-process that overlays a heartbeat-synchronized pulse sound. Specifically, it acquires the user's current heart rate (determined by the ECG RR interval) and inserts a short, low-frequency pulse sound (frequency 0.5–4 Hz, pulse width 50–100 ms, duty cycle ≤30%) at the peak of each heartbeat, continuously outputting several heartbeat cycles (usually 3–5 consecutive pulses within a few seconds, with a total duration not exceeding 10 seconds). This rapid pulse, by simulating the heartbeat rhythm and intervening slightly ahead of time, can trigger a vagal reflex within approximately 1 second, causing a sudden drop in heart rate of 1–2 bpm, thereby quickly interrupting the overactivation of the sympathetic nervous system and preventing further symptom deterioration. The emergency sub-process and ASMR sound can be superimposed, but the duration is strictly limited to avoid excessive stimulation.

[0144] Duration: The rapid intervention mode typically lasts no more than 60 seconds. If symptoms do not improve during this period, it may be extended appropriately, but not exceeding 120 seconds (including emergency sub-processes that may be triggered multiple times). The design aims to pull the user back from a high-risk state as quickly as possible; if 1 minute is not effective, it may be extended to 2 minutes.

[0145] Mode switching: When the rapid intervention mode ends, the system determines the next mode path based on the degree of improvement in the user's condition.

[0146] If, after approximately 60 seconds of rapid intervention, the user's condition has significantly recovered to a low-risk level (e.g., ...), If the low-risk state is less than 0.30 and AI(t) < 50, and the low-risk state has been stable for at least 20 seconds by the m-of-k criterion, and the signal quality Q(t) meets the standard, then the system determines that the user's autonomic nervous system has been greatly calmed, skips the second stage and directly switches to the maintenance and consolidation mode (believing that there is no need to go through rhythm relaxation again).

[0147] If the user's condition has improved but is still in the medium-risk range (e.g., 0.30≤ If the condition is <0.60 or 50≤AI(t)<70), and there has been some improvement in indicators relative to the peak of the attack (e.g., the heart rate has decreased by more than its MDC from the peak, but is still more than 5% higher than the baseline), and the above condition is maintained stably with good Q(t), then the system considers the rapid intervention to be initially effective but still needs to be consolidated, and switches to the rhythm relaxation mode to further stabilize the autonomic nervous function.

[0148] If the user remains in the high-risk zone ( If no significant decrease is observed (≥0.60 or AI(t)≥70), the system determines that the attack has not been effectively controlled, continues to maintain the rapid intervention mode (extending the total duration to a maximum of 120 seconds), and triggers the emergency pulse subprocess again if necessary.

[0149] If the rapid mode persists for 120 seconds without significantly reducing the abnormal intensity (indicators remain high-risk for most of the time), the system will consider the intervention insufficient and automatically issue a prompt (e.g., via user terminal interface) suggesting seeking manual intervention or medical assistance. Simultaneously, the system may prepare to enter a rhythm calming mode to attempt further intervention (unless the user terminates the intervention). For safety reasons, a rapid mode lasting longer than 2 minutes within a panic attack cycle is considered abnormal, and other intervention methods should be evaluated promptly.

[0150] 3.2 Rhythm Moderation Submodule

[0151] Entry conditions: When the rapid intervention ends and the user's state enters the "transitional / moderate risk" zone described above, the system automatically switches to rhythmic calming mode. This mode aims to further reduce physiological arousal using specific rhythmic sounds, guiding the user from a moderate level of tension to a completely calm state. If a rollback is detected in the rapid mode (such as a rollback due to a deterioration in the maintenance mode described later), re-entering rhythmic calming follows the same procedure.

[0152] Stimulation Action: In this mode, the module outputs a theta-rhythmic beat as the primary sound stimulus. Implementation: A continuous background sound (e.g., a soft, low-pitched drumbeat, a simulated heartbeat "thump-thump," etc.) is played, with a fundamental frequency in the range of 50–500 Hz, and a low-frequency amplitude modulation of 4–8 Hz is applied, creating a slow, undulating rhythm (equivalent to 240–480 micro-oscillations per minute, corresponding to a respiratory rate of 4–8 breaths / min). This design aims to guide the user's breathing rhythm to gradually slow down to 4–8 breaths / min, matching the cardiovascular resonant frequency (approximately 0.1 Hz, or 6 breaths / min), maximizing parasympathetic nerve activity. Additionally, the theta band (4–8 Hz) rhythmic stimulation helps stabilize the user's low-frequency brain electrical activity (enhancing alpha / theta waves and inhibiting high-frequency beta / gamma waves), further alleviating anxiety. During playback, the system continuously monitors changes in the user's HR, EDA, and other indicators, and calculates... And AI to assess whether the user continues to improve.

[0153] Duration: The calming mode typically lasts approximately 3–5 minutes. This time is used to consolidate the effects of the rapid intervention and ensure that the user's physiological indicators have fully recovered and stabilized. If the user is completely calm within 3 minutes, the session can be ended earlier; if the user has not fully recovered after 5 minutes, it may need to be extended or other measures may be required (see switching conditions below).

[0154] Mode switching: The rhythm easing mode decides whether to continue, end early, or revert based on changes in the user's state.

[0155] If user metrics have entered a low-risk range during the rhythm slowing mode (e.g.) If the value of AI(t) is less than 0.30 and AI(t) is less than 50, and the value of Q(t) is met for at least 20 seconds by the m-of-k criterion, then the system considers that the user's autonomic nervous system state has basically recovered to a stable state and switches to the maintenance and consolidation mode to enter the final stage of the intervention.

[0156] If the user's condition deteriorates again to a high-risk range during the rhythm easing process ( A value ≥0.60 or AI(t)≥70 indicates a possible new round of panic attacks or that a previous attack was not fully controlled. In this case, the system immediately reverts to the rapid intervention mode to suppress the symptoms again with stronger stimulation (considered a reprocessing of a secondary attack). To avoid repeated mode oscillations, the reversion from the rhythmic calming mode to the rapid mode should not exceed 2 times in a single attack intervention cycle; if it exceeds this, manual intervention is prompted.

[0157] If the rhythm-easing mode continues for more than 5 minutes and the user still spends a significant proportion of time (e.g., >80%) in the medium-risk zone (without further improvement), the system will prompt for manual intervention or other auxiliary measures to prevent the user from being "stuck" in this state for an extended period. Simultaneously, the system may attempt to adjust stimulation parameters (such as slightly increasing sound intensity or changing tone) based on suggestions from the optimization module to seek a breakthrough.

[0158] 3.3 Maintenance and Consolidation Submodule

[0159] Entry conditions: When the rhythm relaxation mode ends and the user's physiological state has entered a low-risk, stable zone, the system switches to maintenance and consolidation mode. Additionally, if the recovery criteria are reached directly after the rapid mode (as in the case of skipping rhythm relaxation as mentioned above), the system will also directly enter maintenance mode. This mode aims to consolidate the intervention effect, prevent short-term physiological rebound, and allow the user to fully return to a relaxed state.

[0160] Stimulation Action: Upon entering maintenance mode, the module switches the output sound type. It stops the previously used ASMR or gentle beat stimulation and instead plays soothing background ambient sounds or soft music. These sounds can be selected based on the user's acoustic profile preferences, such as: natural sounds like ocean waves, rain, wind chimes, or a campfire (rich in frequency components but generally gentle, not stimulating tension, typical volume 45–60 dB(A)); or gentle music therapy tracks, calming guided voices, etc. (slow and steady rhythm, providing a sense of safety and relaxation). The focus of sound stimulation in the maintenance phase is to provide a calm sound environment to help the user maintain the established parasympathetic dominance and gradually reduce residual physiological stress responses. Because the user's physiology has largely stabilized after the first two phases, no special sound effects are needed in this phase; it is only necessary to prevent excessive external stimulation or a resurgence of internal anxiety.

[0161] Duration: The consolidation mode typically lasts 3–5 minutes. However, the specific time can be adjusted based on the user's actual recovery status: if the user's indicators have remained stable for an extended period after 3 minutes, it can be considered to end the session early; conversely, if the user is still slightly anxious, it can be extended to a maximum of approximately 10 minutes. If the user subjectively feels completely calm, they can also actively terminate the sound through the terminal interface (the system allows early termination after confirming that their indicators are normal).

[0162] Mode switching / termination: The maintenance mode is the final stage of intervention, mainly facing two scenarios: normal termination or abnormal regression.

[0163] If the maintenance mode continues for a period of time (at least 3 minutes) and the user's physiological indicators remain stable in the low-risk zone, for example... If the heart rate, HRV, and EDA remain consistently below 0.25 and the AI(t) remains consistently below 45, while the heart rate, HRV, and EDA are close to the resting baseline and meet the set stability duration (e.g., meeting the above criteria for 3 consecutive minutes), the system determines that the intervention has been successful and the user has returned to calm. At this point, the intervention termination process is triggered: the module smoothly fades out the sound output (usually the volume gradually decreases to zero within about 2 seconds), then stops all sound stimulation, declaring the intervention complete. If the maintenance mode has lasted for 5 minutes but the indicators, although within the low-risk range, have not fully met the exit criteria (e.g., ...), the intervention will terminate. If the value is less than 0.30 and AI(t) is less than 50 (not yet lower), the system can extend the mode for up to 10 minutes. If the target is not met after 10 minutes, the mode will fade out (to avoid over-reliance on stimulation) and prompt the user to take a short break or consider other assistance.

[0164] If, during the maintenance mode, the user's status fluctuates: the metric rises again to the transition zone (e.g., 0.30 ≤ <0.60 or 50≤AI(t)<70) or high-risk zone ( If the value is ≥0.60 or AI(t)≥70, it indicates that instability may have re-emerged. The system will immediately revert to the corresponding mode based on the degree of elevation: moderate elevation will revert to the rhythm-easing mode for enhanced intervention, and severe elevation will directly revert to the rapid mode for emergency treatment. Similar to the above, to avoid infinite loops, the reversion from the maintenance mode will not exceed 2 times within one intervention cycle; if frequent recurrences occur, manual intervention or termination of the intervention is recommended.

[0165] Intervention Termination Decision Logic: To ensure the intervention ends at the appropriate time, the module has a built-in termination decision mechanism that assesses in real time whether the user's physiological state has stabilized enough to safely stop the auditory stimulation. The termination decision logic starts after the intervention begins and checks key indicators at fixed intervals:

[0166] Judgment criteria: A sliding window statistical method is used to continuously compare the user's current physiological indicators with the individual's resting baseline. Specifically, a time window of 60 seconds is taken, and the average and standard deviation of the core indicators (heart rate HR, HRV RMSSD, and EDA) within the window are calculated and compared with the mean and fluctuation of the resting baseline. The physiological state is considered to have returned to stability when the following conditions are met: (1) The sliding window mean of indicators such as heart rate and HRV returns to within ±10% of the resting baseline, and the EDA level does not increase significantly relative to resting; (2) The fluctuation (standard deviation) of these indicators within the window is not greater than 1.2 times the fluctuation in the resting state; (3) The above state is maintained continuously for at least 3 window cycles (about 3 minutes). Only when all conditions are met will the termination judgment module output a "return to stability" signal.

[0167] Normal termination procedure: Once a stable recovery signal is detected, the module triggers an intervention termination command to the sound control unit, executing a sound fade-out and stopping output (as mentioned above, fade-out for 2 seconds and then stop). Since the termination of intervention usually means the attack has subsided for the user, the system can notify the user of the intervention termination via the interface while fading out and record the termination timestamp.

[0168] User-initiated termination process: When a user initiates termination by clicking "End / Pause" on the interface or by long-pressing the hardware button, the system first performs a confirmation to prevent accidental touches. If the current condition has not yet stabilized and physiological indicators are still within the abnormal range, a secondary confirmation and "Continue Intervention / End Immediately / Seek Help" options will pop up. After confirmation, the voice control unit will fade out linearly for the same 2 seconds. The system will simultaneously record the termination timestamp, termination type "User-Initiated", reason code (e.g., "Discomfort / Time's Up / Ineffective / Other"), immediate subjective score, and remarks, and mark the event as "Initiated Termination" for subsequent effect analysis and parameter weight adjustment. If the duration of this session does not reach the minimum effective duration, the intervention efficacy evaluation will be treated as "partial data" or included with reduced weight. The system will provide brief closing instructions on the termination interface (e.g., maintain slow breathing for 30–60 seconds) and will only be able to restart after the cooldown time is set according to the safety policy. If significant abnormalities are still detected after the voluntary termination (e.g., HR / EDA exceeds the safety threshold), the system will indicate the risk and suggest switching to rapid relief mode or contacting professional support.

[0169] Forced termination upon exceeding time limit: If the intervention continues for the preset maximum duration (e.g., 15 minutes) and still fails to meet the stability standard, the system will forcefully enter the termination process to prevent discomfort or fatigue that may be caused by prolonged sound stimulation: the sound intensity will be gradually reduced to a safe level (e.g., reduced to below 45 dB(A) and maintained for several tens of seconds), while a prompt will appear suggesting that the user consider seeking professional medical help or implementing other relaxation methods, and then the sound will fade out and stop. The forced termination still follows a smooth transition to avoid causing sudden unease to the user.

[0170] The termination decision logic and mode switching mechanism jointly determine the end of the intervention: it usually ends when the maintenance mode meets the conditions, but if the user reaches stability earlier, it may end in the rhythm easing mode (in which case the maintenance mode will be skipped and the intervention will end directly). The decision signal of this module is also provided to the parameter optimization and data archiving modules to mark whether the intervention is effective and to process the data at the end time. The termination module ensures that the intervention is neither interrupted too early nor delayed too much, making the entire closed loop more intelligent and reliable.

[0171] Output: The sound stimulation control module directly outputs digital audio signals to the user's headphones or speakers to achieve physical stimulation. This includes the specific sound content played at each stage, as well as its dynamically changing parameters over time (such as sound pressure level, frequency components, rhythm patterns, etc.). Simultaneously, the module outputs the current intervention mode status (fast / rhythmic / maintained) and related parameters (current sound type, sound pressure level, rhythm frequency, etc.) in real time for other modules to access. On one hand, the parameter optimization module adjusts its algorithm based on this information and the user's physiological feedback; on the other hand, the data archiving module records the sound control module's behavior log (including mode switching times, sound types used, parameter adjustment trajectories, etc.).

[0172] Module Interaction: The sound stimulus control module cooperates with other modules in the following ways: when the alarm module outputs... When the value is 1, this module immediately responds and enters the intervention process; after the alarm is cleared or a termination signal is issued, this module executes the termination action. When initiating intervention, the control module queries the user's acoustic profile to obtain individualized parameters (e.g., preferred sound type, threshold sound pressure level (EMR), safety constraints, etc.) to formulate the initial stimulation plan. During intervention, the module continuously receives physiological feedback (e.g., heart rate, EDA, and anxiety index AI) from the acquisition / fusion module. This feedback is either directly used for simple internal adjustments (e.g., emergency pulse triggering conditions) or processed by the optimization module to form adjustment instructions. The parameter optimization module and the control module have a closed-loop cooperative relationship: the optimization module calculates the optimal parameter combination for the next step based on feedback, but the control module checks and executes it according to the current intervention mode and safety constraints (the optimization module cannot directly change the mode, but only provides parameter fine-tuning suggestions within that mode). When executing external adjustment instructions, the control module ensures a smooth transition in output changes (e.g., sound fade-in / fade-out to avoid abrupt changes). Throughout the intervention process, the control module acts as both an "actuator" and a "monitor": it completes the sound output and simultaneously senses changes in the user's physiological response and makes mode-level decisions. After the intervention ends, the control module stops the sound and notifies other modules to close the current round of intervention. Through the above mechanism, the sound stimulation control module realizes the concrete implementation of the intervention strategy, acting as a bridge between the upper and lower layers of the closed-loop system: connecting the recognition decision trigger at the top and the physiological effects generated by the user at the bottom, forming a virtuous cycle with the optimization and self-learning modules.

[0173] 4. Parameter Optimization Learning Module

[0174] The parameter optimization learning module is the system's adaptive adjustment and continuous improvement unit. On the one hand, it evaluates the effect of sound stimulation in real time during each intervention and optimizes the stimulation parameters for the next step, so that the intervention process is not limited to a fixed plan, but is adjusted according to the individual and the time. On the other hand, after the intervention, it archives and learns from the entire data, updates the individualized model, and improves the intervention effect in the future.

[0175] 4.1 Individual Acoustic Profile Submodule

[0176] Upon initial system use or when the user is in a calm, non-ictal state, the parameter optimization module first executes an individualized sound threshold measurement procedure to create a personalized acoustic parameter profile for the user. This profile can be considered as calibration data for the system, providing individualized sound stimulus thresholds, preferences, and safety limits, thus laying the foundation for subsequent real-time optimization.

[0177] The input to the individual acoustic profile submodule comes from the multimodal physiological signals of the user in a resting state provided by the signal acquisition submodule, as well as various test sound signal libraries built into the system. The sound library contains different types of sound materials, including various ASMR sounds (such as whispers, brush rubbing, paper rustling, etc., at least five or six types), natural ambient sounds, soft music clips, etc. All test audios have undergone uniform loudness and spectral normalization processing for comparability.

[0178] Resting baseline recording: Before the measurement begins, instruct the user to relax in a quiet environment and wear the system-specified audio output device (headphones). The system records physiological data for at least 3 minutes at rest, calculates the mean and standard deviation of key indicators (such as heart rate (HR), RMSSD of HRV, skin conductance, etc.), and calculates the minimum detectable change (MDC) for each indicator through permutation tests or resampling to assess baseline stability. If baseline fluctuations are too large, the relaxation time is extended until the data stabilizes. MDC is typically defined at a 95% confidence level, for example:

[0179] ;

[0180] Where SD is the resting standard deviation of the index, and ICC is the reliability coefficient calculated by repeated measures.

[0181] Sound stimulation threshold determination: Several typical sound source types were selected from the sound library for testing (to avoid excessive testing and fatigue, low-frequency, gentle sounds that easily elicit autonomic nervous system responses are usually preferred, such as selecting three different types of ASMR sounds: whisper, brush sound, and paper sound, one segment each). For each test sound, the effective stimulation threshold was determined using the rising-falling step method: initially played at a low volume (e.g., 35 dB(A)), if no obvious physiological changes were observed, the volume was gradually increased by approximately +3 dB each step until a clear physiological response was observed or the user subjectively reported feeling obvious stimulation; then the volume was reduced to find the critical sound pressure level between a response and no response. Each sound pressure level was usually played continuously for about 30 seconds (including a 1-second fade-in and fade-out, a continuous sound for about 9 seconds, followed by about 20 seconds of silence to observe the residual effect), while simultaneously recording changes in the user's multimodal physiological indicators.

[0182] The determination of "effective physiological response" is based on a comprehensive assessment of changes in multiple indicators, constructed using a combined effect size (Ẽ). This includes the decrease in heart rate relative to baseline (-ΔHR), the increase in RMSSD of HRV (ΔRMSSD), the decrease in skin conductance (-ΔEDA), and the user's subjective relaxation / pleasure score (S). Each change is standardized by dividing by its minimum detectable change value, and then weighted accordingly. Weighted summation:

[0183] + + + ;

[0184] Weight To balance the contributions of different channels, a threshold can be set empirically (e.g., 25% each) or adaptively adjusted based on the signal-to-noise ratio and reliability of each indicator, ensuring that the judgment considers individual differences while also being robust. A threshold is typically set for this purpose. (That is, the combined effect size reaches the sum of the MDC levels of each indicator) to determine that an effective response has occurred. At this point, the sound pressure level is considered to have elicited a significant relaxation response. The sound pressure level is then adjusted repeatedly around the threshold to more precisely find the lowest sound pressure level that just triggers the response (called the EMR, Effective Minimum Response level). For example, the intensity of the initial response is lowered by one level to check for any lack of response; if no response is detected, this level is considered the threshold; otherwise, another level is lowered, and this process is repeated multiple times until a stable value is obtained after more than six intensity reversals. If the difference between two measured thresholds exceeds 3 dB, an additional test is automatically performed. Finally, the posterior median of the multiple measurements is taken as the EMR for this sound type, and its 95% confidence interval is recorded for reference.

[0185] Preference Assessment: After determining the threshold EMR for each sound, to assess the user's subjective preference, the system plays each test sound for approximately 10 seconds at an intensity of "threshold + 3 dB(A)," and allows the user to subjectively rate or rank the comfort or liking of each sound. The module collects these ratings and statistically analyzes the user's preference level for different sound types. This step helps the system select the sounds that the user is more willing to accept from a variety of available sounds for intervention.

[0186] Acoustic Profile Generation: Based on the test results above, the module establishes an individual acoustic profile for each user. The profile includes: the effective minimum sound pressure level (EMR) threshold for each test sound type; the user's preference for that sound (e.g., like, neutral, dislike); the suggested initial intervention intensity (usually EMR + 3 dB as the initial SPR); and the user's safety limits (e.g., maximum single adjustment range, maximum maximum SPR for that sound). For example, safety parameters can be set: sound intensity variation should not exceed ±6 dB every 30 seconds, and the intensity of any sound type should not exceed the type's EMR + 9 dB(A). The profile also records the user's physiological response sensitivity indicators during the test (e.g., maximum HR drop), for the optimization module's reference. After the acoustic profile is established, it is stored in the system database and can be used for individualized configuration of subsequent interventions. During actual intervention, the control module will prioritize the sound type with the highest preference in the profile that is within the safety range as the initial stimulus and adjust the intensity according to the thresholds and safety parameters in the profile. The profile data also provides a customized search space for the parameter optimization module (e.g., the optimization algorithm only adjusts among sounds the user likes, without trying sounds the user explicitly dislikes). The data can be updated gradually as the user uses it (e.g., if the user has not used it for a long time or their physical condition changes, the data can be re-measured).

[0187] The individual acoustic profile measurement submodule typically runs once when a user first uses the system, or is initiated by the user when the system is idle. The data obtained from the measurement is mainly used by the sound stimulus control module and the parameter optimization module to achieve personalized intervention. During real-time monitoring and intervention, the profile measurement module itself does not participate in calculations, but its results are frequently reviewed: for example, the control module uses the profile to determine sound selection and initial volume, and the optimization module uses the profile to limit the optimization range and adjust weights. If the user's physiological condition or preferences change significantly, the system should rerun the measurement to update the profile to ensure the model parameters remain accurate.

[0188] 4.2 Real-time parameter optimization

[0189] During each intervention, the parameter optimization module runs an optimization algorithm at fixed intervals (e.g., every 60 seconds) to adjust the current sound stimulus parameters. Its workflow is a "evaluation-prediction-optimization-execution" cycle.

[0190] Intervention utility assessment: At the end of each optimization cycle, the module calculates an intervention utility score U(t) based on changes in the user's physiological state in the most recent cycle, used to quantify the effectiveness of the current sound stimulation. The utility U is defined by comprehensively considering the improvement of multiple key indicators, such as decreased heart rate, increased HRV, and decreased EDA. A linear weighted model can be used.

[0191] ;

[0192] in This indicates the change in heart rate compared to the start of the previous cycle (a decrease to a negative value indicates improvement). This indicates a change in the RMSSD of the HRV index (an increase to a positive value indicates improvement). This indicates changes in skin conductance levels (a decrease to a negative value indicates improvement). For the weighting coefficients, satisfying The weight w can be dynamically set based on the user's acoustic profile and historical intervention effects (e.g., the user is more sensitive to certain metrics, and therefore receives greater weight). If A value greater than 0 indicates that the overall condition has improved compared to the beginning of the cycle. A value less than 0 may result in poor performance or even worsening of the effect. The module provides this utility score to the optimization algorithm to determine the direction of adjustment, and records it in the database for post-analysis.

[0193] Candidate parameter generation: At the beginning of each new optimization cycle, the module generates a set of candidate parameters for the next cycle based on the current intervention mode and user profile. This step limits the search space for optimization, ensuring that adjustments do not deviate from the current context or exceed safety boundaries. For example:

[0194] Rapid intervention mode: = {ASMR type, impulse sound, beat sound, heartbeat synchronized ultra-low frequency impulse sound}, SPL∈ [EMR, EMR+6 dB], r ∈ [0.8,1.2].

[0195] Rhythm slowing mode: = { Rhythmic beats}, SPL ∈ [45,65] dB(A), r ∈ [0.5,1.0].

[0196] Maintain consolidation mode: = {natural sounds, meditative speech, soft music}, SPL ∈ [45,60] dB(A), r ∈ [0.8,1.2]

[0197] The module generates several candidate parameter combinations based on the above rules, each... This includes the selection of specific sound types, target sound pressure levels, and rhythm-related parameters. These candidate parameters all meet the pattern requirements and safety constraints, laying the foundation for the next step of utility prediction.

[0198] Bayesian Optimization: The core of the parameter optimization module employs a Bayesian optimization strategy to find the optimal solution in the parameter space. This involves using a prior Gaussian Process (GP) model to predict the utility of each candidate parameter and selecting the optimal balance between exploration and utilization through a data acquisition function. Specific steps: The module considers each of the above candidate parameters... Input the GP model to obtain a mean predicted utility. and prediction uncertainty (GP infers from previously observed parameter-utility data, providing estimated utility and confidence intervals for currently untried parameters.) Subsequently, the acquisition function value for each candidate is calculated. To evaluate its optimization value. Commonly used acquisition functions include the upper confidence bound UCB strategy:

[0199] ;

[0200] in To control exploration - utilizing the balance factor ( Larger areas tend to explore uncertain regions, while smaller areas tend to utilize high-mean regions. Alternatively, Thompson sampling can be used: a possible utility function is randomly sampled from the posterior distribution of the generalized system (GP), and then the optimal parameters are selected on that sampled function, essentially weighing exploration / utilization probabilistically. Regardless of the strategy, the module selects the parameter combination that maximizes the sampling function as the optimal parameter θ* for the next cycle. θ* contains specific instructions for the sound output (such as switching the sound type to X, setting the volume to Y dB, adjusting the tempo factor to r times, etc.).

[0201] Parameter Adjustment and Safety Control: After obtaining θ*, the module compares the current parameters with the recommended parameters and determines the execution plan considering safety constraints: if the change is within the safe range, it is adopted directly; if it exceeds the range, it is adjusted step by step. Main control rules:

[0202] Sound pressure level variation limit: The volume adjustment range shall not exceed ±6 dB within every 30 seconds. If the required increment of θ* is too large, only a portion will be adjusted in this cycle (e.g., within 6 dB), and the remaining portion will be adjusted in the next cycle to avoid sudden volume changes that may cause user discomfort.

[0203] Rhythm adjustment limits: The rate of change of rhythm frequency or tempo multiple r shall not exceed ±0.05 (dimensionless) per minute to prevent excessively rapid changes from interfering with the user's adapted breathing / heartbeat rhythm.

[0204] Smooth sound type switching: If θ* suggests changing the sound type (e.g., from a whisper to rain), the control module performs a smooth fade-in / fade-out transition: the current sound fades out within 1–2 seconds, while the new sound fades in to the target volume, ensuring a natural and unobtrusive transition. If the recommended type does not match the current mode (e.g., a pulse sound is recommended in sustain mode), the unreasonable instruction is rejected or delayed, and the type is only considered during mode switching.

[0205] The final execution parameters are obtained after applying the above strategy. The data is then sent to the sound stimulation control module for execution. The control module updates the sound output accordingly and records the changes. The parameter optimization module then enters the next monitoring and evaluation phase, repeating the above cycle. It's worth noting that if the utility U is negative for several consecutive cycles (e.g., U(t) < 0 for three consecutive cycles), it indicates that the current stimulation protocol may be ineffective or even counterproductive. In this case, the optimization module triggers an early warning mechanism: instructing a reduction in stimulation intensity (e.g., a slight decrease in volume) or suggesting termination of the intervention / manual intervention to avoid overstimulation. The module can also incorporate predictive models to proactively adjust for future trends: for example, if a rebound in HR or AI is detected, a slight reduction of 2 dB or a slowing of the pace by 0.05 is initiated to prevent physiological indicators from deteriorating again.

[0206] Output: In each optimization cycle, the parameter optimization module outputs new sound parameter configurations (type, sound pressure level, rhythm, etc.) for the control module to execute. This is equivalent to setting a "recipe" for the intervention in the next period. Simultaneously, it outputs the utility score U calculated for that cycle and other monitoring data (such as changes in various indicators) for recording. For significant optimization decisions (such as changing the sound type or significantly adjusting the volume), the module can record the reasons for the decision in the log (e.g., "HR did not decrease in the previous cycle, try switching sound types"). These outputs guide immediate intervention adjustments and are also entered into the database through the data archiving module, providing training samples for subsequent model updates.

[0207] 4.3 Data Archiving and Model Updates

[0208] After each intervention, the parameter optimization and self-learning module summarizes and stores the data and effects of the intervention, and uses this data to incrementally update the individual model. This process ensures that the system transforms experience into progress, making it more intelligent in the future. It mainly includes the following steps:

[0209] Intervention Data Archiving: After the intervention end signal is triggered, the system organizes and securely stores all relevant data generated during the intervention. This includes: the entire physiological signal curve (changes in HR, HRV, EDA, etc. over time), the trajectory of sound stimulus parameter changes (when to switch sounds, how to adjust volume, etc.), the utility score U(t) for each cycle, alarm records and abnormal score trends from the panic recognition module, and the user's subjective feedback before and after the intervention (e.g., how much the user's self-reported panic intensity decreased). The module first encrypts and stores this data locally to ensure user privacy. Subsequently, if the user authorizes and the network is available, the module will synchronize the data to the cloud user database via an encrypted channel for long-term storage and in-depth analysis. Each user's data is stored in isolation in the cloud using a unique ID.

[0210] Model Updates and Self-Learning: After data archiving is complete (either in real-time or through scheduled batch processing), the system uses the newly accumulated data to update various models and parameters.

[0211] Gaussian Process Model Update: The parameter combinations and their utility results $(\theta,U)$ tried in this intervention are added as new samples to the training set of the GP model, updating the model's posterior distribution. This improves the GP's understanding of user preference parameters, making the next Bayesian optimization decision more accurate and reliable. In particular, if a new parameter combination that is more effective than before is discovered, the GP will predict higher efficiency in its vicinity, guiding further exploration or utilization in the future.

[0212] Historical Optimal Parameter Database Update: The module extracts the most effective parameter configuration (or several efficient configurations) from the new data during the current intervention and stores it in the user's "optimal parameter database." The next time the user triggers a panic intervention, the system will prioritize using the most effective solution from the historical records as the starting parameters (warm start), rather than conservatively starting from the recommended values ​​in the archive. This reduces parameter trial time. For example, if the user has another episode within 24 hours, the system can directly start the intervention with the most effective sound type and volume from the last intervention, resulting in faster effectiveness and improved immediate and long-term stability of the intervention.

[0213] Panic Recognition Model Update: The panic recognition module model is fine-tuned using newly acquired normal and abnormal data fragments from this episode. For example, calm data after the intervention is merged into the normal dataset, and the boundaries of the One-Class SVM and Isolation Forest models are updated (expanding the normal range). Simultaneously, based on the peak characteristics of this episode, the rationality of the fast threshold pathway setting is examined, and feature thresholds are adjusted if necessary (refer to the aforementioned formula). If a sensor signal is missing during this episode, the corresponding dimensionality reduction recognition model is trained / activated to ensure the system can still function in similar situations in the future.

[0214] Individual acoustic profile updates: If new sound types or different intensities not included in the profile are tried during several pre-treatments and achieve good results, these findings can be fed back into the profile for updates. For example, this could involve increasing the preference score for a particular sound type or adjusting the user's actual response threshold (if a user is found to be more sensitive to a sound than during measurement, the corresponding EMR value can be lowered). The profile's safety parameters can also be gradually optimized based on historical data; for example, if multiple interventions confirm that +5 dB is effective, there's no need to set an upper limit of +9 dB.

[0215] Group-level self-learning: Aggregating multi-user data in the cloud enables group analysis to assist in system optimization. For example, unsupervised clustering can be used to segment user groups with similar response patterns to different sound types, extracting common features from each group (such as a certain type of person being generally sensitive to rain sounds), and optimizing initial parameter suggestions for new users accordingly; or, by statistically analyzing a large amount of test data, the average threshold of various ASMR sounds for the general population can be calculated, serving as a default reference for new users who have not undergone sufficient testing. These group patterns do not directly change individual models, but they are incorporated into the system's default configuration library, improving the overall intelligence level of the system.

[0216] Version management and rollback: The module employs version control for model updates, preserving model parameters from before the update. If the new model experiences performance degradation or anomalies during subsequent use, it can be quickly rolled back to the previous version, ensuring security. Furthermore, strict access permissions are set for archived sensitive physiological data, allowing only the algorithm module to access it, preventing unauthorized personnel from viewing it, and complying with medical data privacy requirements.

[0217] Output: After the model update, the system does not produce directly visible output, but the internal model parameters, thresholds, and files are all updated. The next time the system runs, these updated models and parameters will be automatically loaded, resulting in improved performance. For example, the recognition module will use the updated thresholds and anomaly detection model (reduced false alarm rate), while the control and optimization modules will refer to the updated acoustic files and GP model (better suited to the user's current needs). The entire system continuously optimizes through self-learning. It is important to emphasize that the parameter optimization and self-learning modules only activate after the recognition module confirms the user has entered an episode and triggers intervention; otherwise, they remain in standby mode to conserve resources and avoid overfitting to data from normal states. Through summarizing and analyzing after each round of intervention, the system achieves closed-loop self-improvement, making the intervention increasingly effective and intelligent with each use.

[0218] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A sound stimulus adaptive control system based on abnormal state recognition, characterized in that, It includes a signal acquisition and calculation module, a panic recognition and alarm module, an audio stimulus control module, and a parameter optimization and learning module; The signal acquisition and calculation module acquires various physiological and motion sensing data of the user in real time through wearable sensors, and simultaneously fuses the acquired physiological data to calculate the anxiety index. The panic detection and alarm module first corrects motion state based on motion sensor data, then runs a fast thresholding pathway and an unsupervised model pathway in parallel, and outputs a panic attack alarm signal. and abnormal intensity score Among them, panic attack alarm signals It is a Boolean quantity. A value of 1 indicates that the user is currently experiencing a panic attack. = 0 indicates no seizure alarm or the alarm has been cleared; Abnormal intensity score The risk levels are further categorized into low risk, medium risk, and high risk. The fast thresholding pathway works as follows: Direct threshold judgment rules are set for physiological features extracted from physiological data, for each physiological feature. Set an abnormal threshold Whenever there is a deviation in physiological characteristics Exceeding the abnormal threshold If the direction of change is consistent with the pattern of a panic attack, then this physiological characteristic is marked as a suspicious abnormality at the current moment. ; to include multiple characteristic deviations The combined effect size is obtained by normalizing each effect size to its minimum detectable change value and then weighting and summing them. If the combined effect size Exceeding the set threshold for combined effect size Then the suspicious anomaly will be considered. Perform a count to obtain an anomaly count. Based on joint effect size With joint effect size threshold The relationship between the numbers and the anomaly count The value, the output boolean exception flag. ; The unsupervised model pathway works as follows: Pre-trained single-class support vector machines and isolated forest models are used to detect anomalies in physiological feature patterns; the normalized anomaly probabilities are output as follows: and ; Output panic attack alarm signal and abnormal intensity score The specific process is as follows: Based on the outputs of the fast thresholding pathway and the unsupervised model pathway, a comprehensive anomaly intensity score is calculated. The formula is: ； In the formula, , , Let be the weighting coefficient, satisfying ; The trigger condition is set using m-of-k logic, within a decision window of the most recent k seconds. If at least m moments meet the condition... This will trigger an alarm and output a panic attack alarm signal. ; The preset alarm threshold; The sound stimulation control module selects a sound stimulation mode based on the output of the panic recognition alarm module and the anxiety index, and outputs corresponding sound stimulation according to the user's individual profile under the selected sound stimulation mode. The parameter optimization learning module evaluates the effect of sound stimulation in real time during each stimulation intervention by the sound stimulation control module and optimizes the next stimulation parameters; at the same time, it archives and learns the entire data after the intervention ends.

2. The sound stimulus adaptive control system based on abnormal state recognition according to claim 1, characterized in that, The physiological signals acquired in the signal acquisition and calculation module include: electrocardiogram (ECG) signal, captured heart rate (HR), heart rate variability (HRV), respiratory waveform, and skin conductance (EDA). The collected motion sensing data includes: triaxial accelerometer data and gyroscope data.

3. The sound stimulus adaptive control system based on abnormal state recognition according to claim 1, characterized in that, After acquiring various physiological and motion sensor data, the signal acquisition and calculation module needs to calculate the signal quality index of each channel signal. The formula is as follows: ； In the formula, For signal-to-noise ratio, For data missing rate, The degree of the wake; , , These are the weighting coefficients; The range of values ​​is ,when Once the signal quality is deemed satisfactory, the system's panic detection and alarm module will activate. This is a preset signal quality threshold.

4. The sound stimulus adaptive control system based on abnormal state recognition according to claim 1, characterized in that, The specific process for calculating the anxiety index in the signal acquisition and calculation module is as follows: Physiological signals collected by different sensors are preprocessed to obtain a synchronization signal sequence. Each second in this synchronization signal sequence contains a set of synchronized physiological signal values ​​for feature extraction. With an update step of 1 second, a sliding time window is selected to calculate various physiological characteristics on the synchronization signal sequence; Calculate different weights for each physiological characteristic. Standardized bias of all physiological characteristics According to the corresponding weight Summing yields the overall deviation score. ; The composite deviation score is mapped to an anxiety index in the range of 0–100 using a sigmoid function. ; During mapping, the midpoint θ and slope k of the Sigmoid curve are adjusted according to the individual user situation so that 50 points correspond to the user's resting state; The instantaneous anxiety index calculated every second is smoothed using an exponential moving average, resulting in the final output time. Corresponding smoothed anxiety index for: ； In the formula, This is the smoothing coefficient.

5. The sound stimulus adaptive control system based on abnormal state recognition according to claim 1, characterized in that, The aforementioned panic detection alarm module performs motion state correction based on motion sensor data, specifically including: Calculate the user's average acceleration per unit time based on motion sensor data. and step frequency ; if Exceeding the preset threshold and lasting for more than 5 seconds, or step frequency If the user reaches the brisk walking / running level, the user's exercise status flag is set to TRUE; otherwise, the exercise status flag is set to FALSE. When the motion state flag is set to TRUE, the anomaly detection conditions for the fast thresholding path and the unsupervised model path are corrected; when the internal motion state flag is set to FALSE, no correction is made.

6. The sound stimulus adaptive control system based on abnormal state recognition according to claim 1, characterized in that, The sound stimulation control module includes a rapid intervention mode, a rhythm slowing mode, and a maintenance and consolidation mode. The rapid intervention mode is activated when the user is in a high-risk state, at which point the panic detection alarm module outputs a panic attack alarm signal. = 1 and overall anomaly intensity score Or anxiety index This mode plays fast-paced, relaxing sounds; the duration will not exceed 60 seconds. The condition for entering rhythm-easing mode is: the user's status is in the medium-risk range, at which point... or In this mode, theta rhythm beats are played for 3–5 minutes. The entry condition for maintaining consolidation mode is: the user's status is at a low-risk level. < 0.30 and In this mode, soothing background sounds or soft music play for 3-5 minutes. During the playback of each sound stimulus pattern, the system continuously monitors the real-time changes in the user's physiological indicators and obtains a comprehensive abnormality intensity score. and anxiety index To assess the effectiveness of the intervention, after each sound stimulation pattern ends, the next phase pattern is selected based on the intervention effect, or the sound stimulation is discontinued.

7. The sound stimulus adaptive control system based on abnormal state recognition according to claim 1, characterized in that, A user's personalized profile includes individualized sound stimulation thresholds, sound preferences, and safety restrictions.

Images