Hiv patient privacy preserving data collection system with fused implicit feedback

Abstract

The application discloses a HIV patient privacy protection data acquisition system fusing implicit feedback and relates to the technical field of intelligent medical treatment and health monitoring, and comprises a data acquisition module, which acquires an initial multi-modal data set containing first and second type of sensing data in a first observation time window when detecting a physical disturbance; a feature decoupling module, which decouples the first type of sensing data to extract first and second action features; a dynamic continuation module, which obtains a target micro-motion index based on action feature mapping, and continues to generate a second observation time window and an expanded data set; a coherent noise reduction module, which calculates a coherence coefficient of the two types of data, and when the coherence coefficient is greater than a threshold value, reduces the second type of data based on the first type to obtain a pure state sequence; and a feature generation module, which extracts first and last stable reference values of the pure state sequence, and generates a desensitization state feature code in combination with the target micro-motion index; the application can quantitatively deconstruct the medication psychological game trajectory of a patient, and provides support for long-acting compliance management of chronic infectious diseases.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare and health monitoring technology, specifically to a data collection system for HIV patients that integrates implicit feedback and protects patient privacy. Background Technology

[0002] With the widespread adoption of long-term treatments for chronic infectious diseases, patient medication adherence is directly related to treatment outcomes and the effectiveness of community-based disease control. In order to accurately monitor patients' daily medication behavior while protecting their privacy to the greatest extent, implicit feedback data collection systems based on IoT devices such as smart pillboxes have emerged.

[0003] Existing privacy protection monitoring solutions typically deploy conventional physical sensors such as gravity sensors and opening / closing detection inside the pillbox. Once the pillbox is detected to be touched or the door is opened, the absolute difference in weight change is read within a pre-set fixed time window to determine whether the patient has successfully retrieved the medication.

[0004] However, patients with infectious diseases often experience severe psychological distress and anxiety during long-term treatment due to the stigma associated with the disease or the side effects of medication. This deep-seated psychological state inevitably manifests in the physical operation as involuntary tremors, repetitive rubbing, and extremely hesitant tentative movements of the hands. In this complex micro-interaction scenario, on the one hand, the lengthy operation process caused by the patient's hesitation can easily exceed the rigid observation time set by the system, resulting in abnormal interruption of the monitoring process. On the other hand, the high-frequency micro-tremors of the hands caused by anxiety can be mechanically transmitted along the outer shell of the medicine box, directly causing strong physical coupling interference to the weighing element at the bottom.

[0005] Therefore, how to eliminate the interference caused by involuntary tremors and operational delays resulting from the patient's psychological stress on the underlying physical sensors and restore the true state of medication administration without introducing invasive methods such as video monitoring is a core technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an HIV patient privacy protection data acquisition system that integrates implicit feedback.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, the present invention discloses an HIV patient privacy-preserving data collection system that integrates implicit feedback, comprising: The data acquisition module is used to acquire the initial multimodal dataset of the target object within a preset first observation window when a physical disturbance event that meets the preset activation conditions is detected. The initial multimodal dataset includes a first type of sensor data sequence and a second type of sensor data sequence. The feature decoupling module is used to decouple the features of the first type of sensor data sequence and extract the first action feature under the first preset frequency band and the second action feature under the second preset frequency band. The dynamic extension module is used to perform nonlinear mapping based on the first action feature and the second action feature to obtain the target micro-motion index, dynamically extend the first observation window based on the target micro-motion index to generate the second observation window, and obtain the extended multimodal dataset covering the second observation window. The coherent noise reduction module is used to calculate the coherence coefficient between the first type of sensor data sequence and the second type of sensor data sequence in the extended multimodal dataset within a preset verification frequency band. In response to the coherence coefficient being greater than a preset coherence threshold, adaptive noise reduction is performed on the second type of sensor data sequence using the first type of sensor data sequence as a reference to obtain a clean state data sequence. The feature generation module is used to extract the first stable benchmark value of the pure state data sequence in the preset background period and the second stable benchmark value in the preset period before the end of the second observation window. Based on the difference between the first stable benchmark value and the second stable benchmark value, the module combines the target micro-motion index to generate a desensitized state feature code.

[0008] Secondly, this invention discloses a method for collecting HIV patient privacy-preserving data that incorporates implicit feedback, comprising the following steps: When a physical disturbance event that meets the preset activation conditions is detected, the initial multimodal dataset of the target object within the preset first observation window is obtained. The initial multimodal dataset includes a first type of sensor data sequence and a second type of sensor data sequence. The first type of sensor data sequence is decoupled to extract the first action feature under the first preset frequency band and the second action feature under the second preset frequency band. A nonlinear mapping is performed based on the first action feature and the second action feature to obtain the target micro-motion index. The first observation window is dynamically extended based on the target micro-motion index to generate the second observation window, and an extended multimodal dataset covering the second observation window is obtained. Calculate the coherence coefficient between the first type of sensor data sequence and the second type of sensor data sequence in the extended multimodal dataset within a preset verification frequency band. In response to the coherence coefficient being greater than a preset coherence threshold, perform adaptive noise reduction on the second type of sensor data sequence using the first type of sensor data sequence as a reference benchmark to obtain a clean state data sequence. Extract the first stable benchmark value of the pure state data sequence in the preset background period and the second stable benchmark value in the preset period before the end of the second observation window. Based on the difference between the first stable benchmark value and the second stable benchmark value, combine the target micro-motion index to generate a desensitized state feature code.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The amplitude square coherence of the vertical acceleration component and the original weight sequence is verified by the coherent noise reduction module, and noise reduction is performed by the least mean square adaptive filter. This mechanism accurately removes the physical coupling artifacts caused by the rigid transmission of high-frequency involuntary hand tremors along the hardware shell, restores the real step signal of the tiny pill weight, and solves the problem that the existing technology easily misjudges tremor fluctuations as abnormal drug dispensing. 2. A dynamic extension module was constructed, which calculates the target micro-motion index based on the decoupled first action feature and second action feature, and uses the square term of the index to perform nonlinear multiplication amplification on the base duration of the first observation window; this mechanism adaptively accommodates the repeated probing and prolonged hovering actions caused by the patient's psychological cognitive blockage, and completely avoids the abnormal truncation of the interaction process and loss of monitoring data caused by the traditional rigid time window. 3. The feature decoupling module is used to perform attitude calculation and gravity component stripping on triaxial acceleration and triaxial angular velocity data. Power spectral density integral values ​​representing the intensity of physiological tremor and kinematic parameters representing the degree of cognitive hesitation are extracted in cross-frequency bands. This method externalizes the fuzzy subjective psychological stress into objective physical parameters that can be accurately quantified, providing underlying data support for the system's high-dimensional adaptive modulation and deep attribution. 4. The stable baseline values ​​at the beginning and end of the interaction are extracted by the feature generation module to calculate the actual net weight step change. The thresholded logical result is combined with the target micro-motion index and the hesitation evolution gradient factor and input into the pre-set multi-dimensional joint state coding mapping table to generate an irreversible hexadecimal desensitized state feature code. This architecture completely blocks the physical transmission path of plaintext continuous waveforms without introducing intrusive devices such as video surveillance, and achieves absolute isolation between multimodal behavioral profiling and patient privacy security. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0011] Figure 1 This is an overall block diagram of the system according to Embodiment 1 of the present invention; Figure 2 This is a timing diagram illustrating the dynamic extension of the time window and data truncation in Embodiment 1 of the present invention. Figure 3 This is an overall block diagram of the method in Embodiment 2 of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0013] Application Overview: In the field of long-term treatment and health management of chronic infectious diseases (especially HIV), patient medication adherence is regarded as a key indicator for measuring treatment outcomes and the effectiveness of group prevention and control. The acquisition of such objective adherence data at the Internet of Things (IoT) sensing level is essentially a process of capturing precise physical state transitions. That is, through the weight sensors and inertial measurement units of edge terminals such as smart pillboxes, the patient's single deterministic medication dispensing action is mapped into a pure signal with specific dynamic characteristics and real net weight step in the underlying feature data sequence.

[0014] However, existing technologies lack decoupling and verification mechanisms for cross-modal coupling interference between patients' psychological stress and underlying physical sensing signals. This results in the inability to accurately identify and isolate involuntary tremors and operational hesitation caused by psychological distress during long-term medication use. Involuntary tremors manifest as high-frequency physiological shaking caused by anxiety, which is mechanically transmitted along the hardware casing and severely contaminates the readings of the bottom weighing element. Operational hesitation manifests as repeated probing and hovering caused by cognitive resistance, which greatly delays the interaction process. Consequently, an adaptive purification and extension relationship cannot be established between the artifacts of mechanical transmission and the rigid monitoring window, leading to misjudgment of the actual medication status or abnormal data truncation by the system, thus affecting the objectivity of adherence assessment and the targeting of medical intervention.

[0015] For example, in the scenario of HIV patients taking medication at home, when patients interact with medication collection, traditional monitoring systems can only capture standard short-term interactions through a preset fixed sampling window, but cannot adaptively accommodate actions such as prolonged hovering and sharp-angle reversals caused by extreme hesitation. Furthermore, when patients experience high-frequency micro-vibrations in their hands due to mental tension, the system only records the superficial phenomenon of violently fluctuating weight data, failing to detect the resonance interference of vertical acceleration and the submersion of the actual net weight step signal of the pill. Specifically, the system misjudges prolonged hesitation as medication abandonment and directly issues a missed dose judgment, or incorrectly classifies physical coupling artifacts caused by tremors as abnormal medication dosage. As a result, the patient's actual medication behavior after overcoming psychological resistance is incorrectly recorded, failing to form a desensitized state profile that reflects the actual situation.

[0016] If the above problems are not addressed, the data acquisition system will continue to lose its ability to objectively determine medication status under complex psychological stress. Specifically, the failure to adaptively cancel involuntary tremors will cause the weight signal to be constantly accompanied by mechanical conduction artifacts, resulting in the complete failure of the physical accuracy of the underlying sensors during critical interaction periods, thus masking the true changes in drug weight. At the same time, the failure to dynamically correct operational hesitation will cause the interaction process to be forcibly truncated by a rigid time window, making it impossible for the system to capture the patient's final medication dispensing result. As a result, the distortion of the underlying monitoring data will systematically hinder the cloud-based medical platform from grasping the patient's true medication dosing patterns and psychological evolution, seriously affecting the achievement of long-term HIV chronic disease management goals.

[0017] Example 1: like Figures 1-2 As shown, the HIV patient privacy-preserving data collection system integrating implicit feedback includes: The data acquisition module is used to acquire the initial multimodal dataset of the target object within a preset first observation window when a physical disturbance event that meets the preset activation conditions is detected. The initial multimodal dataset includes a first type of sensor data sequence and a second type of sensor data sequence. In this embodiment, the data acquisition module serves as the sensing entry point for the entire data acquisition system. Its core task is to accurately capture and locate the physical process of patient interaction with the smart medicine box without relying on invasive devices such as cameras or microphones. To achieve this goal, this embodiment constructs a real-time streaming data monitoring and backtracking mechanism based on a synchronous circular buffer.

[0018] Specifically, during system initialization, the edge computing unit (MCU) pre-allocates a fixed-size contiguous storage space in memory to construct a synchronous circular buffer. This buffer is designed as a multi-channel parallel structure to simultaneously receive data streams from the six-axis inertial measurement unit (IMU) and the high-precision weight sensor. To ensure temporal consistency in subsequent signal processing, this embodiment abandons the traditional low-frequency sampling (e.g., 10Hz) of the weight sensor and raises the sampling frequency of the weight sensor to a high-frequency state that is fully synchronized with the IMU (e.g., setting a uniform sampling frequency). This means that each storage unit in the buffer corresponds precisely to a discrete time step t, and that location simultaneously stores triaxial acceleration data. Three-axis angular velocity data and raw weight data The buffer contains seven-dimensional feature vectors. It follows the principle of first-in-first-out (FIFO) and cyclic overwriting, always retaining complete historical data within a certain time window (e.g., 5 seconds) before the current moment, thus providing a data foundation for capturing the background state before a sudden event.

[0019] During system operation, the data acquisition module executes the physical disturbance event monitoring logic. The system reads triaxial acceleration data in real time and calculates the current acceleration vector magnitude. Its calculation formula is expressed as the Euclidean norm: To filter transient noise caused by minor environmental vibrations (such as passing vehicles or slight collisions with the desktop), the system is configured with a silent physical threshold. ) and the number of consecutive trigger frames ( Only if the calculated magnitude of the acceleration vector is... continuous Each sampling period (e.g., 5 consecutive frames, or 50ms) is strictly greater than Only when the conditions are met does the system determine that a substantial physical contact or disturbance has occurred, and marks the start time as the trigger point. This mechanism ensures that the system is only activated when the patient explicitly touches or picks up the medicine box, greatly reducing the false awakening rate.

[0020] Once the trigger time point is locked The system immediately freezes the write pointer of the circular buffer and enters the data backtracking and splicing phase. At this point, in order to construct an initial multimodal dataset containing complete static-dynamic state transitions, the system performs bidirectional data extraction: first, pointer backtracking extracts data from the buffer located at... The previously preset duration (i.e., the preset background time period, for example) The first two seconds of data are recorded. This data sequence records the absolute static state of the medicine box or the ambient noise level before the force is applied, providing a clean reference for subsequent baseline calculations. Next, the system continues with... Starting from a point, the duration of data collected backwards or read from a buffer serves as the basis for the observation window (e.g., ...). The data sequence (last 3 seconds). This part of the data covers the entire interaction process from the patient touching the medicine box, opening the hatch, to the possible tremors.

[0021] Finally, the data acquisition module rigorously aligns and stitches the two data sets along the timeline. The system checks the timestamps of all channels to ensure that the acceleration, angular velocity, and weight data are completely synchronized at the millisecond level, thereby generating an initial multimodal dataset covering the background period and the basic observation window. This dataset is output in matrix form, with rows representing time series and columns representing seven-dimensional sensing features, providing complete and high-precision raw signal support for the subsequent micro-motion extraction and gravity separation in the feature decoupling module.

[0022] The feature decoupling module is used to decouple the features of the first type of sensor data sequence and extract the first action feature under the first preset frequency band and the second action feature under the second preset frequency band. After the system acquires the initial multimodal dataset, the feature decoupling module performs in-depth analysis on the first type of sensor data sequence to separate the two distinct physical representations of physiological tremor and cognitive hesitation.

[0023] Specifically, this module processes the triaxial acceleration and triaxial angular velocity data sequences from the first type of sensor data sequences. In the first core step of feature decoupling, this embodiment pre-constructs a spatial model for attitude calculation and gravity stripping based on multi-sensor fusion. In practical applications, the raw triaxial acceleration data output by the microelectromechanical system (MEMS) is always superimposed with the Earth's gravitational acceleration component. If this component is not removed, the static gravity bias will completely destroy the subsequent extraction of kinematic features. Therefore, the system simultaneously reads the triaxial acceleration and triaxial angular velocity data and uses a pre-set quaternion update algorithm (such as the Mahony complementary filtering algorithm) to perform real-time attitude calculation. The system iteratively calculates and updates the instantaneous attitude quaternion of the smart pillbox in the global navigation coordinate system at a certain period. Subsequently, the system back-projects the standard gravity normalized vector in the global navigation coordinate system to the body coordinate system of the medicine box, and calculates the estimated values ​​of the gravitational acceleration components in the body coordinate system. Its projection transformation formula is as follows: ; in, This is the local gravitational acceleration constant. Therefore, the system compares the original triaxial acceleration vector output by the sensor with the estimated value of this gravitational acceleration component. By performing vector subtraction, the gravitational acceleration components are completely separated, resulting in a pure linear acceleration sequence generated purely by hand manipulation.

[0024] Subsequently, the system performs cross-frequency feature extraction on the pure linear acceleration sequences in parallel using two paths. The first path aims to quantify the patient's involuntary tremors caused by anxiety. To accurately distinguish between the patient's physiological tremors (which typically exhibit significant anisotropy across the three spatial axes) and external environmental impacts (such as table shaking, which usually manifests as a common-mode signal in the same direction across the three axes), the system first calculates the Pearson correlation coefficient between the pure linear acceleration sequences across the three axes. If the correlation coefficient between any two axes is greater than a preset common-mode threshold (e.g., 0.85), the system determines that the current signal originates from environmental interference and directly sets the corresponding tremor feature to zero; otherwise, the system performs subsequent frequency domain analysis.

[0025] The system extracts the target vertical component (i.e., the dynamic acceleration component along the vertical direction of the medicine box) from a purely linear acceleration sequence. The signal is then subjected to a short-time Fourier transform (STFT) using a sliding window to map the time-domain signal to the frequency-domain space, yielding the corresponding discrete power spectral density function. Within this space, the system focuses on the first preset frequency band. In medical dynamics, this frequency band is typically set to 4 to 12 Hz, falling within the frequency range characteristic of intention tremor. The system calculates the power spectral density integral value within this first preset frequency band according to the following formula and determines it as the first action characteristic. : When a patient experiences high-frequency, slight hand tremors, this integral value will increase significantly, objectively and accurately reflecting the energy intensity of intention tremor.

[0026] Meanwhile, the second processing method focuses on capturing the patient's macroscopic hesitation during medication dispensing. The system first rigorously performs bandpass filtering (e.g., setting the passband to 0.1 to 3 Hz) on the pure linear acceleration sequence to remove DC bias drift. After numerical integration of the filtered pure linear acceleration sequence in the discrete-time domain, the system obtains a highly accurate three-dimensional target velocity vector sequence, denoted as... To transform the velocity changes in this three-dimensional space into a quantifiable hesitation index, the system further extracts kinematic hesitation parameters from the target velocity vector sequence.

[0027] Specifically, the system first calculates the Euclidean magnitude of the velocity vector at the current moment in real time. In this embodiment, the preset zero-reaching interval refers to an extremely low velocity envelope range set to filter out dynamic background noise from the sensor and natural physiological tremors of the patient's hand. The system pre-calibrates an upper threshold for the zero-reaching velocity. (For example, if set to 0.05 m / s), then this preset zero-reaching interval is defined as .

[0028] Based on this, the system scans the target velocity vector sequence frame by frame, and the velocity magnitude at a certain sampling time... When this occurs, the system determines that it is currently in a hovering / paused state. The system accumulates the number of sampled frames during which this state lasts and multiplies it by the system's sampling period. (For example, when the sampling frequency is 100Hz,) ), thereby calculating the cumulative duration of pauses (denoted as ). It should be noted that, in order to eliminate the brief pauses that inevitably occur during the normal medication dispensing process, the system has preset a physiological baseline duration. (For example, 0.5 seconds). The system only includes the difference in the valid pause statistics if the duration converted from the cumulative frame count exceeds this baseline. For example, if 150 frames occur during a single medication retrieval action with a speed modulus below 0.05 m / s, then the cumulative pause duration statistics are calculated. The second value objectively reflects the time the patient's hand hovers and hesitates in mid-air.

[0029] On the other hand, in order to characterize the characteristics of the patient's repeated tentative movements, the system performs a detection of acute angle reversal of direction. In physical space, the tentative movement of the patient's hand suddenly retracting due to psychological resistance after extending it will form a V-shaped turning point with an acute angle on the spatial trajectory; mapped to the velocity time series, this is manifested as an obtuse angle reversal of the velocity vector direction between the two effective motion phases (i.e., the angle is greater than 90 degrees).

[0030] In this embodiment, the system employs a fixed-length time observation sliding window strategy. For each determined hovering pause state, the system extracts a first observation sliding window of a preset length (e.g., 0.2 seconds) before its start time and a second observation sliding window of the same length after its end time. The system calculates the arithmetic mean vector of all instantaneous velocity vectors contained within the first observation sliding window of the target velocity vector sequence and defines it as the forward steady-state velocity vector. Similarly, the system calculates the arithmetic mean vector within the second observation sliding window, which is defined as the backward steady-state velocity vector. The system then calculates the spatial dot product of the two vectors and the cosine of their included angle: ; The system sets a reversal threshold (e.g.) When the calculated cosine value When the angle between the two velocity vectors in three-dimensional space exceeds 120 degrees, corresponding to an acute-angle reversal trajectory of less than 60 degrees in physical space, the system determines that an acute-angle reversal event has occurred. The system accumulates the count of this event within a single interactive observation window to obtain the reversal frequency statistics (denoted as ). ).

[0031] Ultimately, the system is based on the first preset weight parameter. (For example, setting it to 0.6 to emphasize punishing repeated trial-and-error behavior) and the second preset weight parameter (For example, set to 0.4, corresponding to hover time), perform a weighted fusion of the above reversal frequency statistics and the cumulative pause duration statistics: ; Through the aforementioned vector dot product and envelope integral operations, the system successfully transforms vague subjective psychological manifestations such as hesitation and repeated probing into expressions derived from... and Supported, quantifiable kinematic hesitation parameters This provides solid underlying data support for the subsequent generation of the target micro-motion index.

[0032] Through the aforementioned physical and mathematical mapping, the system successfully decoupled the first action characteristic (tremor energy) representing physiological stress from massive amounts of raw inertial data. ) and the second motor characteristic representing cognitive impairment (kinematic hesitation parameter) This provides substantial and impeccable data support for subsequent dynamic modulation of time windows.

[0033] The dynamic extension module is used to perform nonlinear mapping based on the first action feature and the second action feature to obtain the target micro-motion index, dynamically extend the first observation window based on the target micro-motion index to generate the second observation window, and obtain the extended multimodal dataset covering the second observation window. This module aims to establish a nonlinear mapping relationship between microscopic movement characteristics at the physical level and stress states at the psychological level, and dynamically adjust the data acquisition time of the system accordingly to prevent data truncation caused by patient hesitation.

[0034] First, the module performs multidimensional feature fusion based on a nonlinear activation function to generate the target micro-motion index. This process is not a simple linear superposition, but rather aims to simulate the stress saturation effect in biopsychology. The system pre-defines a normalized nonlinear mapping model (e.g., a variant based on the Sigmoid function), whose mathematical expression is as follows: ; To ensure that physical characteristics of different dimensions can be effectively substituted into the activation function, the system pre-configures empirical benchmark parameters locally for the smart pillbox hardware and the target object, including the minimum / maximum empirical boundary values ​​of tremor energy. and the minimum / maximum empirical boundary values ​​of the hesitation parameter. . and These are the normalized values ​​of the first action feature and the second action feature after performing a linear mapping based on the above boundary values ​​(the mapping formula is: (and so on), this process ensures that the input variables are strictly constrained. Within the range; and The system is preset with feature sensitivity coefficients (e.g., set to 1.2 and 0.8 respectively, to emphasize tremor weight). To activate the bias term (e.g., set to 5.0 to ensure that the output is forcibly suppressed to an extremely low level when all action features approach zero), this nonlinear formula fuses and maps multidimensional physical features to... The dimensionless interval. When When the value approaches 0, it indicates that the patient's operation is decisive and stable; when When the value approaches 1, it indicates that the patient is in a state of high stress characterized by extreme anxiety and hesitation.

[0035] Subsequently, based on the calculated target micro-motion index The system adaptively and dynamically extends the preset first observation window to generate a second observation window. To address the issue that traditional fixed windows cannot cover long-tail hesitant behavior, the system employs nonlinear multiplication amplification logic to calculate the new cutoff time. Specifically, the system reads the base duration of the first observation window. For example, the system's default interaction window is 3 seconds, and a gain factor is constructed using the squared term of the target micro-motion index to calculate the effective cutoff time of the second observation window. : ; In the formula, This indicates the preset time stretch factor (e.g., set to 2.0). The maximum allowable acquisition time for hardware resources is set for the system (e.g., 10 seconds). The introduction of this threshold can prevent infinite loop acquisition caused by high stress.

[0036] Assuming the system defaults If a patient's mental state is detected to be stable, the following calculations are made. ,but The system maintains almost the original duration; if it detects high levels of anxiety in the patient, it calculates... ,but This mechanism ensures that the system provides more than twice the operating time redundancy for patients with high anxiety.

[0037] Finally, after determining the effective deadline... Subsequently, the dynamic extension module performs underlying data flow control and splicing tasks to obtain the extended multimodal dataset. To clearly define the temporal boundaries of the dataset, this embodiment introduces a background start time. Its definition is: ,in The duration of the preset background period (e.g., 2 seconds). During operation, the system first locks and acquires the data located in the synchronization ring buffer. The dataset consists of a preset background time period within a given time interval. Furthermore, the system does not immediately stop data collection, but rather continues based on the calculated effective cutoff time. The control sensor continuously collects data from the self-triggered time point. Start to Sensor data for the period following the cutoff.

[0038] It is worth noting that the system will eventually be in The dataset of the preset background time period in the interval and the data in the interval The sensor data from subsequent time periods within the interval are time-series stitched together to obtain a complete extended multimodal dataset. This dataset fully covers the second observation window and its preceding background, providing a foundation for subsequent coherent noise reduction modules. Precise processing of the interaction range provides complete signal samples.

[0039] The coherent noise reduction module is used to calculate the coherence coefficient between the first type of sensor data sequence and the second type of sensor data sequence in the extended multimodal dataset within a preset verification frequency band. In response to the coherence coefficient being greater than a preset coherence threshold, adaptive noise reduction is performed on the second type of sensor data sequence using the first type of sensor data sequence as a reference to obtain a clean state data sequence. After the dynamic extension module acquires the extended multimodal dataset covering the second observation window, the coherent noise reduction module further addresses the cross-modal interference problem caused by the mechanical vibration of the patient's high-frequency hand tremors during physical interaction with the smart pillbox. This vibration is rigidly transmitted along the outer shell to the bottom weight sensor, resulting in severe distortion of the weighing data. In practical engineering, relying solely on frequency domain filtering (such as low-pass filtering) often fails to distinguish between weight fluctuations caused by hand tremors and the minute weight steps caused by pill removal, as the two may overlap in the frequency spectrum. To achieve this goal, this embodiment pre-constructs a minimum mean square adaptive filter (LMSFilter) model structure based on digital signal processing before performing specific noise reduction steps. This filter model does not use a fixed cutoff frequency to mechanically filter out a certain frequency band. Instead, it is configured to include a multi-tap finite impulse response (FIR) structure and a weighted adaptive update feedback loop based on the minimum mean square error criterion. This allows it to fit and track the mechanical transfer function of physical vibrations in a specific pillbox material in real time, thereby dynamically stripping coupled noise from the target signal.

[0040] Specifically, the coherent noise reduction module first performs precise data extraction and preprocessing. The system extracts data from the extended multimodal dataset based on timestamp indexes, starting from the moment the physical disturbance event was triggered. Until the effective deadline (Right now The data between these parameters is used to construct a wake-up interaction data subset. To obtain a high-purity reference noise source, the system again performs attitude calculations on the first type of sensor data sequence (acceleration and angular velocity) contained in this subset. Using the aforementioned quaternion gravity stripping algorithm, the system separates the purely linear acceleration sequence and extracts the target-vertical component along the opposite direction of gravity, denoted as... Simultaneously, the system extracts the second type of sensor data sequence (raw weight data) from this subset, denoted as... .

[0041] The system then proceeds to the coherence calculation stage. This step aims to statistically verify whether the weight fluctuations are truly caused by hand tremors. The system uses the Welch mean periodogram method to calculate the vertical component of the target. Compared with the original weight sequence Self-power spectral density , and their cross-power spectral density Based on this, the system calculates the amplitude squared coherence function: ; The system focuses on a preset validation frequency band (i.e., the medically defined intention tremor range, such as 4-12Hz) and calculates the frequency within that band. The mean of the coherence coefficient is obtained. .like If the value is greater than the preset coherence threshold (e.g., 0.6), it indicates that the fluctuation in the weight signal is mainly caused by mechanical transmission and must be canceled; conversely, if the value is less than the threshold, it is considered as no coupling interference, and the system directly outputs the original weight data.

[0042] When the noise reduction condition is triggered, the system activates a least mean square (LMS) adaptive filter. This filter is not a pre-fixed parameter model, but a dynamic system capable of learning and fitting the physical conduction path in real time. In this embodiment, the filter is configured as a transverse FIR structure with an order set to M (e.g., M=32). Specifically, to prevent the weightlessness / weight gain effect (a low-frequency DC component) generated when the patient picks up / puts down the medicine box from being incorrectly filtered out as relevant noise, the desired signal sequence in this embodiment... Before the input filter, an AC fluctuation component is extracted by a high-pass filter (cutoff frequency 0.5Hz) to participate in the noise reduction calculation, while its DC base component is directly bypassed and transmitted. Finally, it is synthesized with the pure AC residual at the output end, so as to perfectly preserve the real tablet weight step signal while eliminating jitter artifacts.

[0043] During the signal input and parameter initialization phase, the system uses the target's vertical component as a reference noise sequence input filter to construct a sliding data window vector. ; and the original weight sequence As the input filter for the desired signal sequence. Specifically, at the initial startup time n=0, the system performs zero-state initialization processing on the filter weight parameter vector configured inside the filter, that is, sets the initial weight vector... This is to establish an unbiased benchmark starting point for adaptive iteration.

[0044] Subsequently, at each discrete time step The system strictly performs the following adaptive iterative calculations: Filter output calculation: using the current filter weight parameter vector Convolve the reference noise sequence to generate a canceled sequence for fitting the interfering noise components. : ; Error signal extraction: Calculate the difference between the desired signal and the canceled sequence. In active noise control logic, this error is actually the true signal after removing jitter artifacts, i.e., the clean-state data sequence. : ; Weight Update: Based on the minimum mean square error criterion, the filter weights are updated in real time using gradient descent to minimize the residual at the next time step. ; in, This is an adaptive variable step size factor. To address the issue of slow convergence or divergence when the patient's grip strength changes dynamically with a fixed step size, this embodiment employs the Normalized Least Mean Square (NLMS) algorithm to adjust the step size in real time based on the energy of the current reference noise vector. ; In the formula, The normalization step size (e.g., 0.1). To prevent the use of tiny constants with a denominator of zero, this mechanism ensures the absolute convergence stability of the algorithm under transient mechanical transmission paths.

[0045] Through the above recursive iteration, the filter weight parameter vector changes over time. It will gradually converge to the optimal estimate of the mechanical conduction function of the pillbox shell, thus accurately subtracting the same-frequency fluctuation component caused by hand tremors from the original weight data. The final output is a pure state data sequence that retains the true DC step (pill weight change) but filters out AC noise. .

[0046] The feature generation module is used to extract the first stable benchmark value of the pure state data sequence in the preset background period and the second stable benchmark value in the preset period before the end of the second observation window. Based on the difference between the first stable benchmark value and the second stable benchmark value, the desensitized state feature code is generated in combination with the target micro-motion index. After the coherent noise reduction module filters out mechanical coupling artifacts and outputs a clean state data sequence, the feature generation module extracts the absolute physical step features from the adaptively purified dynamic time series data, which are not affected by high-frequency vibrations and human hesitation. It then performs a nonlinear spatial mapping with the target micro-motion index, which represents the degree of psychological stress, and finally generates a privacy-preserving state code that cannot be reversed to deduce the original operation details.

[0047] It should be noted that, due to the physical disturbance event triggering at the moment... During the previously preset background period, the target object had not yet made physical contact, so the weight data for this period was in a baseline state unaffected by direct external forces such as hand tremors. Based on this, the system constructs a global clean data source extraction logic: for the baseline state, it directly calls the uncontaminated raw weight data; for the interactive state, it calls the clean state data sequence output by the coherent noise reduction module. .

[0048] Specifically, the feature generation module first performs rigorous time-domain segmented mean calculation. The system extracts data from the acquired raw weight data sequence based on timestamp indices, focusing on the time points triggered by physical disturbance events. The first subset of data within the previously preset background time period. Assume that the discrete sampling point interval covered by this preset background time period is... The system directly samples the raw weight data contained within this interval. The arithmetic mean is calculated using the following formula: ; The system determines the arithmetic mean obtained from the above calculations as the first stable reference value. This value objectively and accurately represents the initial absolute rest weight of the target object before it touches the medicine box.

[0049] Subsequently, the system uses the effective cutoff duration of the second observation window output by the dynamic extension module. The system calculates the boundary index for extracting the terminal steady-state data. It then extracts the cutoff time (i.e., the timestamp) of the second observation window from the pure state data sequence. The second subset of data within a preset time period. The number of sampling points corresponding to this preset time period is set to... (For example, in an actual configuration, it is set to the last 50 sampling periods, corresponding to 0.5 seconds of physical time), at which point the corresponding discrete sampling point interval is: ,in This represents the final sampled sequence points corresponding to the effective cutoff time. The system also calculates the arithmetic mean of the data within this interval: ; The system determines the arithmetic mean as the second stable benchmark value. Through this averaging process, the system effectively eliminates the interference of minor rebound aftershocks after the drug retrieval action is completed, and obtains the final static weight after the interaction is completely finished.

[0050] After obtaining the initial and final physical reference values, the system calculates the difference between the first stable reference value and the second stable reference value. The system performs an absolute difference calculation. This yields the actual net weight step change. To convert continuous physical weight values ​​into discrete objective logical characteristics, the system compares this net weight step change with a preset standard dose range. Perform a rigorous mathematical boundary comparison. If the calculated... If the weight change falls within the closed interval, the system determines that the weight change meets the preset threshold, thus determining that the dosage is up to standard, and outputs the first target logic value (e.g., defined as a binary logic variable R=10); if If the value is outside the interval and greater than the upper limit of the interval, an overdose is determined, and the system outputs a second target logical value (e.g., defined as logical variable R=11); if If the value is outside the interval, less than the lower bound of the interval, and not 0, then the dosage is determined to be insufficient, and the system outputs the third target logic value (e.g., defined as logic variable R=01); otherwise, it is determined that no medication has been taken, and the system outputs the fourth target logic value (e.g., defined as logic variable R=00).

[0051] After completing the calculation of net weight step change and logical threshold comparison at the physical level, in order to further explore the dynamic evolution trend of the patient's psychological state within the extended operation window, the system performs deep attribution analysis based on the evolution of hesitation parameters. This process is no longer limited to static indicators at a single moment, but deconstructs the patient's psychological game trajectory from a time dimension.

[0052] Specifically, after acquiring the extended multimodal dataset covering the second observation window, the system initiates time-series slicing processing. The system divides the time into preset steps. (For example, setting the step size to 0.5 seconds) will expand the corresponding second observation window in the multimodal dataset (i.e., the time from the physical disturbance trigger to the effective cutoff). The complete data segment (the entire process) is divided into K consecutive sub-time windows at equal intervals on the time axis. For the k-th consecutive sub-time window (where... The system extracts the first type of sensor data sequence and reuses the kinematic hesitation parameter calculation logic in the aforementioned feature decoupling module to decouple and calculate the local second action feature corresponding to the sub-time window, denoted as... The system concatenates the calculation results of all sub-time windows in chronological order to construct the second action feature time series. .

[0053] Subsequently, the system calculates the first derivative of the second action feature time series with respect to the time axis to quantify the rate of change of action hesitation. To eliminate random fluctuations in single-point calculations, this embodiment employs the least squares method for the sequence. Perform linear fitting and extract the slope of the fitted line as the hesitant evolution gradient value. The core formula for its calculation is as follows: ; in, The center time point of the k-th sub-window (e.g., if the trigger time is used). If the time is zero, then ), and These are the arithmetic means of the time points and the hesitation characteristic values, respectively. The calculated values ​​are... It accurately characterizes the dynamic change rate of patients' psychological resistance during the operation: positive values ​​represent increased hesitation, and negative values ​​represent decreased hesitation.

[0054] Based on the calculated hesitancy evolution gradient value, the system further generates a compensation factor to correct the final encoding. The system pre-sets a convergence determination threshold. (For example, -0.2 represents a significant trend of hesitation fading) and the divergence judgment threshold. (For example, +0.2 represents a significant trend of hesitation and deterioration).

[0055] Response to hesitant evolution gradient value The system determines that the characteristics of the second action decrease over time, indicating that the patient successfully overcame psychological resistance within the extended time window and became more decisive in their actions. At this point, the system generates a first compensation correction factor (e.g., a marker bit). (This is used to indicate a self-overcoming type of positive compliance state.)

[0056] Response to hesitant evolution gradient value The system determined that the characteristics of the second action increased over time, indicating that the patient's anxiety intensified and the operation became increasingly chaotic over time. At this point, the system generated a second compensatory correction factor (e.g., a marker bit). (This is used to indicate a high-risk state of psychological breakdown.)

[0057] If the gradient value is within the above thresholds, a default factor is generated (e.g., This indicates that the psychological state remains stable.

[0058] Finally, the system combines the generated difference logic comparison results (first / second target logic value R) with the calculated target micro-motion index. And the newly generated compensation correction factor ( This generates the final desensitized state feature code. In this embodiment, to achieve multi-dimensional fine-grained mapping, the joint state coding mapping table pre-configured in local memory is constructed as a three-dimensional index structure. The three input dimensions of this mapping table are: Physical Result Dimension: Binarized Physical Behavior Logical Result ; Psychological intensity dimension: Target micro-movement index The segmented quantification sub-intervals (such as low, medium, and high stress zones); Dynamic evolution dimension: Compensation correction factor (such as self-overcoming, stabilization, psychological breakdown).

[0059] In actual operation, the system inputs the above three parameters as joint query keywords into the mapping table. For example, when the system verifies R=10 (dose target met). (Moderate stress), but When the first compensatory correction factor (representing the resolution of hesitation) is reached, it indicates that although the patient's actual stress level was high, they demonstrated self-regulation ability during the process. The system will output a specific hexadecimal desensitization code (e.g., 0x2A). This code can be parsed in the cloud as a case of self-regulation with high anxiety but good compliance, thus distinguishing it from other cases. This system identifies cases of high anxiety and loss of control. The code extracted from these cases is then used as the final desensitization state feature code. By introducing a hesitation evolution gradient as an additional decision-making dimension, this system not only records the outcome of whether or not medication was taken, but also encodes the dynamic psychological process of how medication was taken under extremely low bandwidth, achieving a deep and comprehensive profile of the patient's behavioral patterns.

[0060] In summary, this embodiment introduces feature decoupling and multimodal coherent noise reduction mechanisms into the underlying physical sensing stage, precisely separating the involuntary tremors caused by psychological stress from the original weight fluctuations, thus completely eliminating the damage to the accuracy of the bottom weighing element caused by high-frequency mechanical conduction. Simultaneously, the system utilizes the decoupled kinematic hesitation parameters to construct a target micro-motion index, thereby adaptively and dynamically extending the observation window, effectively avoiding abnormal process interruptions caused by long-tailed hesitant movements. Based on this, the system uses the purified absolute weight step logic, real-time psychological stress level, and the evolution gradient of movement hesitation as multidimensional decision space variables. Through rigorous nonlinear threshold comparison and joint table lookup, the extremely complex temporal multimodal signal is transformed into a single discrete desensitized state feature code. This closed-loop processing architecture, which integrates underlying hardware timing synchronization with top-level state nonlinear mapping, not only restores the most objective pharmacological results of medication administration without introducing any invasive monitoring equipment, but also deeply quantifies and deconstructs the psychological game trajectory of patients taking medication. This provides a technological foundation for the long-term adherence management of chronic infectious diseases such as HIV, which combines absolute privacy protection with high-dimensional assessment value.

[0061] Example 2: like Figure 3 As shown, the method for collecting HIV patient privacy-preserving data by incorporating implicit feedback includes the following steps: When a physical disturbance event that meets the preset activation conditions is detected, the initial multimodal dataset of the target object within the preset first observation window is obtained. The initial multimodal dataset includes a first type of sensor data sequence and a second type of sensor data sequence. The first type of sensor data sequence is decoupled to extract the first action feature under the first preset frequency band and the second action feature under the second preset frequency band. A nonlinear mapping is performed based on the first action feature and the second action feature to obtain the target micro-motion index. The first observation window is dynamically extended based on the target micro-motion index to generate the second observation window, and an extended multimodal dataset covering the second observation window is obtained. Calculate the coherence coefficient between the first type of sensor data sequence and the second type of sensor data sequence in the extended multimodal dataset within a preset verification frequency band. In response to the coherence coefficient being greater than a preset coherence threshold, perform adaptive noise reduction on the second type of sensor data sequence using the first type of sensor data sequence as a reference benchmark to obtain a clean state data sequence. Extract the first stable benchmark value of the pure state data sequence in the preset background period and the second stable benchmark value in the preset period before the end of the second observation window. Based on the difference between the first stable benchmark value and the second stable benchmark value, combine the target micro-motion index to generate a desensitized state feature code.

[0062] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0063] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0064] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A privacy-preserving data collection system for HIV patients integrating implicit feedback, characterized in that, include: The data acquisition module is used to acquire the initial multimodal dataset of the target object within a preset first observation window when a physical disturbance event that meets the preset activation conditions is detected. The initial multimodal dataset includes a first type of sensor data sequence and a second type of sensor data sequence. The feature decoupling module is used to decouple the features of the first type of sensing data sequence and extract the first action feature under the first preset frequency band and the second action feature under the second preset frequency band. The dynamic extension module is used to perform nonlinear mapping based on the first action feature and the second action feature to obtain the target micro-motion index, dynamically extend the first observation window based on the target micro-motion index to generate the second observation window, and obtain an extended multimodal dataset covering the second observation window. The coherent noise reduction module is used to calculate the coherence coefficient between the first type of sensor data sequence and the second type of sensor data sequence in the extended multimodal dataset within a preset verification frequency band. In response to the coherence coefficient being greater than a preset coherence threshold, adaptive noise reduction is performed on the second type of sensor data sequence with the first type of sensor data sequence as a reference to obtain a clean state data sequence. The feature generation module is used to extract the first stable benchmark value of the pure state data sequence in a preset background period and the second stable benchmark value in a preset period before the end time of the second observation window. Based on the difference between the first stable benchmark value and the second stable benchmark value, and combined with the target micro-motion index, a desensitized state feature code is generated.

2. The HIV patient privacy protection data acquisition system integrating implicit feedback according to claim 1, characterized in that: The first type of sensor data sequence includes a triaxial acceleration data sequence and a triaxial angular velocity data sequence, and the second type of sensor data sequence includes a raw weight data sequence; The process of acquiring the initial multimodal dataset includes: continuously overwriting the triaxial acceleration data sequence, the triaxial angular velocity data sequence, and the original weight data sequence in a preset synchronous circular buffer at the same preset sampling frequency; The acceleration vector magnitude is calculated based on the triaxial acceleration data. In response to the acceleration vector magnitude being greater than the silent physical threshold for a consecutive preset number of frames, the physical disturbance event is determined to be detected, and the trigger time point is recorded. Extract the preset duration data located before the trigger time point from the synchronization ring buffer to construct a dataset for a preset background time period, and extract the basic observation window duration data starting from the trigger time point to construct an interactive behavior dataset; The dataset of the preset background time period is timestamped and concatenated with the dataset of the interaction behavior to generate the initial multimodal dataset.

3. The HIV patient privacy protection data acquisition system integrating implicit feedback according to claim 2, characterized in that: The extraction process of the first action feature and the second action feature includes: Based on the triaxial acceleration data and the triaxial angular velocity data, attitude calculation is performed to separate the gravitational acceleration component and obtain a pure linear acceleration sequence. Extract the target vertical component from the pure linear acceleration sequence and perform a short-time Fourier transform. Calculate the power spectral density integral value within the first preset frequency band and determine the power spectral density integral value as the first action feature, wherein the first preset frequency band is the frequency range characterizing intentional tremor. The pure linear acceleration sequence is subjected to bandpass filtering to remove DC bias drift. The filtered pure linear acceleration sequence is then subjected to time integration to obtain the target velocity vector sequence. The kinematic hesitation parameters of the target velocity vector sequence within the second preset frequency band are extracted as the second action feature, wherein the second preset frequency band is the frequency range characterizing the intended action.

4. The HIV patient privacy-preserving data acquisition system integrating implicit feedback according to claim 3, characterized in that: The calculation process of the kinematic hesitation parameter includes: Statistical values ​​of the frequency of reversals in the target velocity vector sequence that occur at acute angles; The cumulative duration of pauses when the amplitude of the target velocity vector sequence is in a preset zero-approaching interval is statistically analyzed. Based on the first preset weight parameter and the second preset weight parameter, the reverse frequency statistics and the cumulative pause duration statistics are weighted and fused to obtain the second action feature.

5. The HIV patient privacy-preserving data acquisition system integrating implicit feedback according to claim 4, characterized in that: The first observation window is dynamically extended based on the target micro-motion index to generate a second observation window, and an extended multimodal dataset covering the second observation window is obtained, specifically including: Based on the product factor formed by the square of the target micro-motion index and the preset time extension coefficient, the base duration of the first observation window is multiplied and amplified to calculate the effective cutoff duration of the second observation window. The calculation formula is as follows: ； in, Indicates the effective deadline. This indicates the base duration of the first observation window. This represents the preset time stretching factor. Indicates the target micro-motion index. The maximum allowable data collection time set for the system; The dataset of the preset background time period recorded in the synchronization ring buffer is obtained, and the sensor data of the subsequent time period is continuously collected based on the effective cutoff time. The dataset of the preset background time period and the sensor data of the subsequent time period are spliced ​​together in time to obtain the extended multimodal dataset.

6. The HIV patient privacy-preserving data acquisition system integrating implicit feedback according to claim 5, characterized in that: The calculation process of the coherence coefficient includes: From the extended multimodal dataset, a subset of wake-up interaction data is extracted according to timestamps from the time the physical disturbance event was triggered to the effective deadline. The attitude calculation is performed on the first type of sensor data sequence contained in the wake-up interaction data subset to separate the gravitational acceleration component, obtain a pure linear acceleration sequence, and extract the target vertical direction component of the pure linear acceleration sequence. Extract the second type of sensor data sequence contained in the wake-up interaction data subset as the original weight sequence; Calculate the amplitude squared coherence result between the target vertical component and the original weight sequence within the preset verification frequency band, and determine the amplitude squared coherence result as the coherence coefficient.

7. The HIV patient privacy-preserving data acquisition system integrating implicit feedback according to claim 6, characterized in that: The calculation process for the pure state data sequence includes: In response to the coherence coefficient being greater than the preset coherence threshold, it is determined that the original weight sequence contains interference noise components related to the vertical component of the target; The target vertical component is used as a reference noise sequence and input into a preset minimum mean square adaptive filter, and the original weight sequence is used as the desired signal sequence and input into the minimum mean square adaptive filter. The reference noise sequence and the filter weight parameter vector configured in the least mean square adaptive filter are used to perform calculations to generate a cancellation sequence for fitting the interference noise component; Calculate the difference sequence between the desired signal sequence and the cancellation sequence, and output the difference sequence as the pure state data sequence; Based on the difference sequence and the reference noise sequence, the filter weight parameter vector is iteratively updated according to the minimum mean square error criterion.

8. The HIV patient privacy protection data acquisition system integrating implicit feedback according to claim 1, characterized in that: After dynamically extending the first observation window based on the target micro-motion index to generate a second observation window, and obtaining an extended multimodal dataset covering the second observation window, the method further includes: According to a preset time division step size, the complete data segment corresponding to the second observation time window in the extended multimodal dataset is divided into a preset number of continuous sub-time windows; Extract the first type of sensor data sequence within each consecutive sub-time window, decouple and calculate the second action feature corresponding to each consecutive sub-time window, and splice them together in chronological order to construct the second action feature time sequence; Calculate the first derivative of the second action feature time series with respect to the time axis to obtain the hesitation evolution gradient value, which characterizes the dynamic change rate of action hesitation.

9. The HIV patient privacy-preserving data acquisition system integrating implicit feedback according to claim 8, characterized in that: Based on the difference between the first stable benchmark value and the second stable benchmark value, and in conjunction with the target micro-motion index, a desensitized state feature code is generated, including: The hesitant evolution gradient value is compared with a preset convergence threshold and a preset divergence threshold, respectively. In response to the hesitation evolution gradient value being less than the convergence determination threshold, it is determined that the second action feature shows a decreasing trend over time, and a first compensation correction factor is generated; In response to the hesitancy evolution gradient value being greater than the divergence determination threshold, it is determined that the second action feature shows an increasing trend over time, and a second compensation correction factor is generated; Based on the difference between the first stable benchmark value and the second stable benchmark value, the target micro-motion index is combined, and the first compensation correction factor or the second compensation correction factor is used as an additional decision dimension to input into a preset joint state coding mapping table to generate a desensitized state feature code containing dynamic evolution labels.

10. A method for collecting HIV patient privacy-preserving data by incorporating implicit feedback, characterized in that, The HIV patient privacy-preserving data collection system based on the implicit feedback fusion method described in any one of claims 19 performs the following steps: When a physical disturbance event that meets the preset activation conditions is detected, the initial multimodal dataset of the target object within the preset first observation window is obtained. The initial multimodal dataset includes a first type of sensor data sequence and a second type of sensor data sequence. The first type of sensor data sequence is decoupled to extract a first action feature in a first preset frequency band and a second action feature in a second preset frequency band. A nonlinear mapping is performed based on the first action feature and the second action feature to obtain the target micro-motion index. The first observation window is dynamically extended based on the target micro-motion index to generate a second observation window, and an extended multimodal dataset covering the second observation window is obtained. Calculate the coherence coefficient between the first type of sensor data sequence and the second type of sensor data sequence in the extended multimodal dataset within a preset verification frequency band. In response to the coherence coefficient being greater than a preset coherence threshold, perform adaptive noise reduction on the second type of sensor data sequence using the first type of sensor data sequence as a reference benchmark to obtain a clean state data sequence. Extract the first stable benchmark value of the pure state data sequence during a preset background period and the second stable benchmark value during a preset period before the end time of the second observation window. Based on the difference between the first stable benchmark value and the second stable benchmark value, and in combination with the target micro-motion index, generate a desensitized state feature code.

Images