A method and system for evaluating the quality of an intervention based on horticultural therapy

Abstract

The application relates to the technical field of horticultural therapy quality evaluation, and discloses a method and system for evaluating intervention quality based on horticultural therapy, which comprises the following steps: acquiring multi-modal data and synchronously collecting data; establishing an individual baseline database; collecting intervention data of participants during the implementation of horticultural therapy intervention; constructing a multi-modal time series data set from the collected intervention data; extracting physiological process features, behavior process features and experience process features reflecting dynamic characteristics of the intervention process from the multi-modal time series data set; vectorizing the extracted multi-dimensional process features to construct an individual process portrait of each participant; training an efficacy prediction model by using a machine learning algorithm; and generating a multi-dimensional evaluation report of intervention quality based on the trained efficacy prediction model. The application improves the accuracy of quality evaluation.

Description

Technical Field

[0001] This invention relates to the field of horticultural therapy quality assessment technology, and more specifically, to a method and system for assessing the quality of interventions based on horticultural therapy. Background Technology

[0002] Horticultural therapy is a complementary intervention method that uses plants and horticultural activities as a medium. Through participation in gardening operations such as sowing, transplanting, pruning, and harvesting, it promotes individual physical and mental health, cognitive function recovery, and improved social adaptability. In recent years, horticultural therapy has been increasingly widely used in areas such as mental health promotion, geriatric care, rehabilitation medicine, and special education. Numerous studies have shown that horticultural therapy can effectively alleviate mood disorders such as anxiety and depression, reduce perceived stress levels, improve attention and cognitive function, and enhance social connection and quality of life.

[0003] Existing methods for assessing the quality of interventions based on horticultural therapy have a core drawback: they rely solely on discrete measurements before and after the intervention (such as pre- and post-intervention scales and single physiological indicators), treating horticultural therapy as a "black box." This makes it impossible to continuously track and quantify the dynamic characteristics of physiological, behavioral, and experiential aspects during the intervention process, thus failing to reveal the process mechanism of therapeutic efficacy and hindering real-time assessment and personalized feedback of intervention quality.

[0004] Therefore, it is necessary to design an intervention quality assessment method and system based on horticultural therapy to address the problems existing in current technologies. Summary of the Invention

[0005] In view of this, the present invention proposes an intervention quality assessment method and system based on horticultural therapy, aiming to solve the problem of low prediction efficiency in the current system.

[0006] This invention proposes a method and system for assessing the quality of interventions based on horticultural therapy, comprising: Acquire multimodal data synchronously, including wearable physiological monitoring devices, high-definition camera devices, and mobile terminal instantaneous assessment data; Establish an individual baseline database and collect participants' resting physiological data and baseline psychometric data on non-intervention days; During the implementation of horticultural therapy intervention, intervention data was collected from participants, including continuous physiological data, behavioral video data, and instantaneous experience data. The collected intervention data were preprocessed and aligned with the timeline to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential. Physiological process features, behavioral process features, and experiential process features that reflect the dynamic characteristics of the intervention process are extracted from the multimodal time series dataset; The extracted multi-dimensional process features are vectorized to construct an individual process profile for each participant; Using the individual process profile as input features and standardized outcome indicators as prediction targets, a machine learning algorithm is used to train the efficacy prediction model; Based on the trained efficacy prediction model and combined with the current intervention process characteristics, a multidimensional assessment report of intervention quality is generated.

[0007] Furthermore, when establishing an individual baseline database and collecting participants' resting physiological data and baseline psychometric data on non-intervention days, including: Acquire continuous physiological signals from participants in a resting state on non-intervention days, including at least heart rate variability signals, skin conductance signals, and triaxial acceleration signals; The collected raw physiological signals were preprocessed, including: bandpass filtering of the heart rate variability signal and extraction of the RR interval sequence, and removal of abnormal RR intervals; low-pass filtering of the skin conductance signal and removal of motion artifacts; and calculation of the synthetic acceleration from the triaxial acceleration signal. Baseline physiological features are extracted from the preprocessed physiological signals, and the baseline physiological features include at least: Mean RR interval ; RR interval standard deviation ; Root mean square of the difference between adjacent RR intervals ; Low frequency power and high frequency power ; Low-frequency to high-frequency ratio ; Mean skin conductivity ; Number of skin conduction responses That is, the average number of skin conductance responses per minute in a resting state; Average composite acceleration ; Baseline psychometric data of participants were collected, including at least the perceived stress scale score. Anxiety Self-Rating Scale Score Score of the Nature Connection Scale ; The baseline physiological characteristics and baseline psychometric data are combined to form an individual baseline record for each participant. ,in Let N be the RR interval of the i-th normal heartbeat, and N be the total number of valid RR intervals. The baseline mean RR interval is given, and T is the resting state duration. Let t be the skin conductance value at time t. Let x be the acceleration along the x-axis at time t. Let be the Y-axis acceleration at time t. Let be the Z-axis acceleration at time t. Let be the resultant acceleration at time t.

[0008] Furthermore, during the implementation of horticultural therapy interventions, the collection of intervention data from participants includes: Continuously collect participants' heart rate variability signals, skin conductance signals, triaxial acceleration signals, and video behavioral data; Using a mobile instant assessment terminal, based on a combination of preset time triggers and event triggers, the participants' immediate subjective experience is collected through a questionnaire during the intervention process. The questionnaire includes at least assessment items on immediate emotional state, perceptual recovery experience, and social connection. The acquired raw physiological signals were preprocessed, including: extracting the RR interval sequence from the heart rate variability signal; filtering and denoising the skin conductance signal to obtain a clean skin conductance signal; and calculating the synthetic acceleration from the triaxial acceleration signal. The collected behavioral video data is encoded and labeled to generate a sequence of behavioral events. The encoded results are then converted into a time series format, including at least the activity type and task focus at each time point. Organize the instantaneous experience data according to the trigger time point, including at least positive emotion score, negative emotion score, attention recovery level, stress relief and social connection; Using the intervention start time as the origin, all modal data are unified into the same relative time coordinate system to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential.

[0009] Furthermore, when preprocessing and aligning the collected intervention data with the timeline to construct a multimodal time series dataset encompassing physiological, behavioral, and experiential dimensions, the following steps are taken: At the start of intervention Define relative time as the origin. All collected data are unified into the same relative time coordinate system; Extracting RR interval sequences from heart rate variability signals from collected continuous physiological data. ,in , where is the heartbeat sequence number during the intervention process, and M is the total number of effective heartbeats; Interpolation method is used to convert non-equidistant intervals Convert to equal time intervals RR interval estimates The skin conductance signal was filtered and denoised to obtain a clean skin conductance signal. ; The composite acceleration of the triaxial acceleration signal is determined, and the activity intensity is determined based on the composite acceleration. The collected behavioral video data is encoded and labeled to generate a behavioral event sequence containing the start time, end time, activity type, and task focus level of each behavioral event. The behavioral event sequence is then converted into equally spaced time points. Activity type and task focus; The collected instantaneous experience data is used to generate a set of experience events based on the trigger time point, including positive emotion scores, negative emotion scores, attention recovery level, stress relief, and social connection. The preprocessed physiological time series, behavioral time series, and experience event set are aggregated along the time axis to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential.

[0010] Furthermore, when extracting physiological process features, behavioral process features, and experiential process features reflecting the dynamic characteristics of the intervention process from the multimodal time series dataset, the following are included: Physiological process features are extracted from the physiological time series of the multimodal time series dataset. The physiological process features include at least the slope of the stress recovery curve, the area under the stress curve, the physiological recovery efficiency index, and the frequency of skin conductance response. Behavioral process features are extracted from the behavioral time series of the multimodal time series dataset. The behavioral process features include at least the time proportion of various activities, activity transition frequency, average activity segment duration, and comprehensive participation depth index. Experience process features are extracted from the experience event set of the multimodal time series dataset. The experience event set includes the trigger time of each instantaneous evaluation and the scores of each dimension. The experience process features include at least the average positive emotion score, average negative emotion score, average attention recovery level, average stress relief, and average social connection.

[0011] Furthermore, the extracted multi-dimensional process features are vectorized to construct an individual process profile for each participant, including: The extracted physiological process features, behavioral process features, and experiential process features are summarized in a preset order to form an original feature set; Arrange each feature in the original feature set in a fixed order to construct a one-dimensional original feature vector; Each feature in the original feature vector is standardized to convert each feature into a standardized value with a mean of 0 and a standard deviation of 1, thus obtaining a standardized feature vector. The standardized feature vector is used as an individual process profile of the participant; Each participant's individual process profile is linked and stored with their corresponding baseline data, outcome indicators, and covariate information to construct an individual process profile database.

[0012] Furthermore, when training the efficacy prediction model using machine learning algorithms, the following steps are included: Collect sample data from participants who have completed the intervention. The sample data includes an individual process profile vector for each participant and the corresponding standardized outcome indicators. The sample data is divided into a training set and a test set; An ensemble learning model is constructed using random forest, extreme gradient boosting, and lightweight gradient boosting machine as base learners, and linear regression as a meta-learner. Based on the K-fold cross-validation method, each base learner is trained using the training set to obtain the predicted value of each base learner on the training set, and the predicted value is used as a meta-feature. Using the meta-features as input and the standardized outcome index as the target, the meta-learner is trained to obtain the ensemble learning model; The trained ensemble learning model is evaluated using the test set, and the feature importance scores of each base learner are extracted to identify the process features that contribute the most to efficacy prediction.

[0013] Furthermore, when generating a multidimensional assessment report on intervention quality, the following should be included: Obtain the individual process profile vector of the current intervention and input it into the trained efficacy prediction model to obtain the efficacy prediction value; The participants who rank in the top percentage of outcome indicators from the training data are selected, and the average value of each process characteristic is calculated to form the optimal process profile. Calculate the weighted Euclidean distance between the current process profile and the optimal process profile, and calculate the intervention quality index based on the weighted Euclidean distance; Calculate the quality scores of the current intervention in the physiological, behavioral, and experiential dimensions, and compare the scores of each dimension with the baseline data of the intervention and the benchmark data of the same population. Based on the efficacy prediction value, intervention quality index, quality scores of each dimension and comparison results, a multidimensional assessment report of intervention quality is generated, which includes a summary of efficacy prediction, a display of intervention quality index, radar charts of quality scores of each dimension, a list of key process characteristics, and personalized optimization suggestions.

[0014] Furthermore, it also includes: When the intervention quality index is lower than the first threshold and higher than the second threshold, a first-level warning prompt is generated. The first-level warning prompt includes marking the low-scoring dimension and the corresponding general optimization suggestions. When the intervention quality index is lower than or equal to the second threshold, a secondary warning prompt is generated. The secondary warning prompt includes marking low-scoring dimensions, targeted optimization suggestions generated based on feature importance analysis, and suggestions to re-evaluate the intervention plan or increase the number of supplementary interventions. The first threshold is higher than the second threshold.

[0015] Compared with existing technologies, the advantages of this invention are as follows: This invention constructs a three-dimensional time-series dataset integrating continuous physiological signals, behavioral videos, and instantaneous subjective experiences by synchronously acquiring multimodal data and aligning it with the timeline. This enables continuous tracking of the entire horticultural therapy intervention process, improving the dynamism and information completeness of the assessment. Based on this, the extracted physiological, behavioral, and experiential process features can quantify the subtle changes in individual stress recovery, participation quality, and subjective feelings during the intervention process. This elevates the assessment from traditional "comparison of pre- and post-test results" to "process mechanism analysis," providing multi-dimensional process evaluation criteria instead of relying on a single outcome indicator. Simultaneously, by vectorizing multi-dimensional process features to construct individual process profiles, this invention achieves personalized digital modeling of each participant's intervention performance, overcoming the limitation of the group average effect masking individual differences and improving the accuracy and relevance of the assessment. Using individual process profiles as input and standardized outcome indicators as prediction targets, a machine learning algorithm is used to train the efficacy prediction model. This allows for the prediction of possible efficacy levels based on current intervention process characteristics, transforming assessment from post-event summarization to pre-event prediction, reducing feedback lag, and creating conditions for timely adjustment of the intervention plan. The resulting multidimensional intervention quality assessment report not only includes efficacy predictions and intervention quality indices, but also provides quality scores for each dimension and comparisons with baselines and benchmarks for similar populations, as well as personalized optimization suggestions based on feature importance analysis, enriching the interpretability and practicality of the assessment results. The report can also provide tiered warnings based on the degree to which the intervention quality index falls below a preset threshold, indicating specific dimensions of intervention quality deficiencies and providing targeted improvement measures, thus achieving a closed loop from "scoring" to "action guidance." Furthermore, the entire process employs standardized data collection, automated preprocessing, and machine learning model prediction, reducing biases from manual observation and subjective recall, and improving the objectivity, reproducibility, and cross-study comparability of the assessment results. In summary, this invention elevates the assessment of horticultural therapy intervention quality from a static, discrete, population-based, and descriptive paradigm to a dynamic, continuous, individualized, and predictive paradigm, comprehensively enhancing the information dimensions, temporal resolution, personalization level, and clinical practical value of the assessment.

[0016] Secondly, this application also provides an intervention quality assessment system based on horticultural therapy, for applying the aforementioned intervention quality assessment method based on horticultural therapy, including: The data acquisition module is configured to acquire multimodal data synchronously, including wearable physiological monitoring devices, high-definition camera devices, and mobile terminal instantaneous assessment data. The baseline database building module is configured to establish an individual baseline database, collecting participants' resting physiological data and baseline psychometric data on non-intervention days; The intervention data acquisition module is configured to collect intervention data from participants during the implementation of horticultural therapy interventions. The intervention data includes continuous physiological data, behavioral video data, and instantaneous experience data. The preprocessing and alignment module is configured to preprocess and align the collected intervention data along the timeline to construct a multimodal time series dataset that includes physiological, behavioral, and experiential dimensions. The feature extraction module is configured to extract physiological process features, behavioral process features, and experiential process features that reflect the dynamic characteristics of the intervention process from the multimodal time series dataset; The individual profile building module is configured to vectorize the extracted multi-dimensional process features to build an individual process profile for each participant. The model training module is configured to use the individual process profile as input features, standardized outcome indicators as prediction targets, and machine learning algorithms to train the efficacy prediction model. The report generation module is configured to generate a multidimensional assessment report of intervention quality based on a trained efficacy prediction model and the characteristics of the current intervention process. The early warning module is configured to generate corresponding early warning prompts based on the intervention quality index. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating an intervention quality assessment method based on horticultural therapy, provided as an embodiment of the present invention; Figure 2 This is a functional block diagram of an intervention quality assessment system based on horticultural therapy, provided as an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey its scope to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] In some embodiments of this application, see Figure 1 As shown, a method for assessing the quality of interventions based on horticultural therapy includes: Multimodal data is acquired synchronously, including wearable physiological monitoring devices, high-definition camera equipment, and mobile instantaneous assessment data.

[0020] An individual baseline database was established, and participants' resting physiological data and baseline psychometric data were collected on non-intervention days.

[0021] During the implementation of horticultural therapy interventions, intervention data was collected from participants, including continuous physiological data, behavioral video data, and momentary experience data.

[0022] The collected intervention data were preprocessed and aligned with the timeline to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential.

[0023] Extract physiological, behavioral, and experiential features that reflect the dynamic characteristics of the intervention process from multimodal time series datasets.

[0024] The extracted multi-dimensional process features are vectorized to construct an individual process profile for each participant.

[0025] Using individual process profiles as input features and standardized outcome indicators as prediction targets, a machine learning algorithm is used to train an efficacy prediction model.

[0026] Based on the trained efficacy prediction model and combined with the current intervention process characteristics, a multidimensional assessment report of intervention quality is generated.

[0027] Specifically, wearable physiological monitoring devices and high-definition cameras are used to collect behavioral and physiological data of the monitored individuals, as well as physiological data and psychological test data on non-treatment days. An individual baseline database for the monitored individuals is constructed, and various data during the horticultural therapy intervention process are collected and preprocessed. Physiological, behavioral, and experiential characteristics of the dynamic features of the response intervention process after preprocessing are extracted, and the extracted features are vectorized to construct an individual process profile. The individual process profile is used as input features to train the efficacy prediction model, and an evaluation report is generated for the monitored individuals based on the trained efficacy prediction model.

[0028] In some embodiments of this application, when establishing an individual baseline database and collecting participants' resting-state physiological data and baseline psychometric data on non-intervention days, the following are included: Continuous physiological signals of participants at rest were acquired on non-intervention days. These physiological signals included at least heart rate variability, skin conductance, and triaxial acceleration.

[0029] The acquired raw physiological signals were preprocessed, including: bandpass filtering of the heart rate variability signal and extraction of the RR interval sequence, removing abnormal RR intervals; low-pass filtering of the skin conductance signal and removal of motion artifacts; and calculation of the synthetic acceleration from the triaxial acceleration signal.

[0030] Baseline physiological features are extracted from the preprocessed physiological signals. The baseline physiological features include at least the following: Mean RR interval .

[0031] RR interval standard deviation .

[0032] Root mean square of the difference between adjacent RR intervals .

[0033] Low frequency power and high frequency power .

[0034] Low-frequency to high-frequency ratio .

[0035] Mean skin conductivity .

[0036] Number of skin conduction responses This refers to the average number of skin conductance responses per minute in a resting state.

[0037] Average composite acceleration .

[0038] Baseline psychometric data of participants were collected, including at least the perceived stress scale score. Anxiety Self-Rating Scale Score Score of the Nature Connection Scale .

[0039] Baseline physiological characteristics and baseline psychometric data were combined to form an individual baseline record for each participant. ,in Let N be the RR interval of the i-th normal heartbeat, and N be the total number of valid RR intervals. The baseline mean RR interval is given, and T is the resting state duration. Let t be the skin conductance value at time t. Let x be the acceleration along the x-axis at time t. Let be the Y-axis acceleration at time t. Let be the Z-axis acceleration at time t. Let be the resultant acceleration at time t.

[0040] Specifically, under resting conditions, the following three raw signals are continuously recorded: heart rate variability signal: raw photoplethysmography (PPG) or electrocardiogram (ECG) signal; skin conductance signal: raw skin conductance signal; and triaxial acceleration signal: acceleration along three orthogonal axes. The time axis for all signals is in seconds, starting at time 0. During preprocessing, baseline drift and high-frequency noise are removed from the raw PPG signal to obtain a filtered signal. When detecting heartbeat times, the R-wave time sequence is obtained. ,in, , This represents the total number of heartbeats detected. The interval between adjacent heartbeats is calculated. A total of RR interval, Calculate the mean of the original RR intervals. and standard deviation ,like If any of these are found to be abnormal, they are removed. The number of valid RR intervals remaining after removal is denoted as . For the rejected locations, no interpolation is performed. Skin conductance signal preprocessing: High-frequency noise is removed from the original skin conductance signal to obtain a filtered signal. The motion phase is identified using the acceleration signal: The synthetic acceleration is calculated according to the subsequent formula. ,when This is then marked as a motion period. Preprocessing of the triaxial acceleration signal: For each time t, calculate the composite acceleration: ,in The unit is gravitational acceleration. Then, the average resultant acceleration over the entire time period was calculated. Baseline physiological characteristic extraction: Calculation of mean RR interval: Calculate the standard deviation of the RR interval: Calculate the root mean square of the difference between adjacent RR intervals: Heart rate variability frequency domain index: After resampling the RR interval sequence at equal intervals, a fast Fourier transform was used for spectral analysis to obtain the low-frequency power. and high frequency power Calculate the low-frequency to high-frequency ratio: Skin conductivity index: Calculation of average skin conductivity level: ,in For time, The preprocessed skin conductivity signal is averaged from discrete sampling points during actual calculation. Skin conductivity response is detected as follows: the skin conductivity response is defined as a rise in skin conductivity signal exceeding [a certain value] within 1–5 seconds. The event then recovers slowly. The detection algorithm is as follows: Calculate... The first derivative of the integer. Identify events where the derivative exceeds a preset threshold and lasts for 1–5 seconds. Count the total number of events occurring within 30 minutes. Calculate the average number of reactions per minute: Acceleration index, calculate the average composite acceleration: Baseline psychometric data collection: After physiological data collection (to avoid the influence of fatigue), participants were guided by trained researchers to complete the following standardized psychological scale: Perceived Stress Scale (PSS-10): Scores were recorded as follows. (Range 0–40), higher scores indicate greater perceived stress. Self-Rating Anxiety Scale (SAS): Score recorded as... (Range 20–80), higher scores indicate more severe anxiety. Natural Connection Scale (NRS): Scores are recorded as follows: (Range 1–7), higher scores indicate a stronger sense of connection with nature. All scale scores are recorded to one decimal place. The extracted baseline physiological characteristics are combined with the baseline psychometric data to form an individual baseline record B for each participant, where: physiological baseline vector Psychological baseline vector Each record also includes metadata: participant's unique ID, collection date, collection start time, device model, and data quality markers (such as whether there is motion interference).

[0041] Understandably, traditional assessment methods often use the group average as a reference standard, ignoring the natural differences between individuals at rest. This technical approach collects each participant's resting physiological data and baseline psychometric data on non-intervention days, allowing subsequent stress change analysis during the intervention process to be based on the individual, rather than comparing to the group average. This individualized baseline significantly improves the accuracy of assessments of indicators such as stress recovery and autonomic nervous system balance, while reducing the risk of misjudgment due to population heterogeneity—for example, the decrease in heart rate after intervention in individuals with a high baseline heart rate might be misjudged as "ineffective" by traditional methods, while the individualized baseline can accurately identify their true recovery effect. Existing methods often only collect baseline data from a single dimension (such as only scale scores or only heart rate), failing to comprehensively reflect the participant's initial state. This technical solution integrates nine physiological characteristics, including time-domain indicators of heart rate variability (mean RR interval, RR interval standard deviation, and root mean square of the difference between adjacent RR intervals), frequency-domain indicators (low-frequency power, high-frequency power, and low-frequency to high-frequency ratio), skin conductance level, number of skin conductance responses, and synthetic acceleration. It also incorporates scores from three psychological scales to construct an individual baseline record encompassing 12 dimensions. This multidimensional baseline significantly enhances the ability to comprehensively characterize the participant's initial state and reduces the omission of key features due to insufficient information dimensions.

[0042] In some embodiments of this application, the collection of intervention data from participants during the implementation of a horticultural therapy intervention includes: Continuous acquisition of participants' heart rate variability, skin conductance, triaxial acceleration, and video behavioral data.

[0043] Using a mobile instant assessment terminal, based on a combination of preset time triggers and event triggers, participants' immediate subjective experiences are collected through questionnaires during the intervention process. The questionnaires include assessment items of immediate emotional state, perceptual recovery experience, and social connection.

[0044] The acquired raw physiological signals were preprocessed, including: extracting the RR interval sequence from the heart rate variability signal; filtering and denoising the skin conductance signal to obtain a clean skin conductance signal; and calculating the synthetic acceleration from the triaxial acceleration signal.

[0045] The collected behavioral video data is encoded and labeled to generate a sequence of behavioral events. The encoded results are then converted into a time series format, including at least the activity type and task focus at each time point.

[0046] Organize the instantaneous experience data according to the trigger time point, including at least positive emotion scores, negative emotion scores, attention recovery level, stress relief, and social connection.

[0047] Using the intervention start time as the origin, all modal data are unified into the same relative time coordinate system to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential.

[0048] Specifically, participants wore the same wearable physiological monitoring devices (such as medical-grade wristbands) as at baseline, continuously recording the following three signals throughout the horticultural therapy intervention: Heart rate variability signal: raw signal acquired via photoplethysmography or electrocardiogram sensor. Skin conductance signal: raw signal acquired via skin conductance sensor. Triaxial acceleration signals: Acceleration data along three orthogonal axes (X, Y, Z) were collected using accelerometers. High-definition cameras were deployed within the horticultural activity area to record the participants' entire activity from multiple angles. After the intervention, the videos were imported into behavioral coding software for offline annotation. A combination of time-triggered and event-triggered strategies was employed, using a mobile instantaneous assessment terminal (tablet or mobile phone). Short questionnaires were popped up during the intervention to collect participants' immediate subjective experiences. Questionnaires were popped up at random time intervals during the intervention and immediately upon participants completing a pre-set activity type transition. Each questionnaire included the following assessment items, each using a Likert scale of 1 to 5 points: positive emotion score, negative emotion score, attention recovery level, stress relief, and social connection. All trigger times (relative time from the start of the intervention) and their corresponding scores were recorded as a discrete event set for subsequent analysis. The videos were imported into the behavioral coding software and annotated according to a pre-set behavioral coding system. The coding system included at least the following elements: Activity Type: Predefined activity categories, such as sowing, transplanting, pruning, harvesting, resting, etc. Each event records its start relative time, end relative time, and activity type code.

[0049] Task focus: Based on the participant's gaze direction, operational continuity, facial expressions, etc., a continuous score between 0 and 1 is given for each event (or segment within an event), with higher scores indicating greater focus.

[0050] Interaction quality (optional): Record the quality of the participant’s interaction with plants, tools, therapist or partner, on a scale of 1 to 5.

[0051] The final result is a sequence of behavioral events, each event containing a start time, end time, activity type code, task focus score, and interaction quality score.

[0052] Behavioral events were discretized using the same sampling grid as physiological data. For each equally spaced time point, it was first determined which behavioral event it belonged to (i.e., start time ≤ time point ≤ end time), and then the event was assigned a corresponding activity type code, task focus score, and interaction quality score. If a time point did not belong to any event (e.g., activity transition interval), the activity type code was set to empty or a specified rest code, and focus and interaction quality were set to default values. All collected instantaneous assessment data were organized into a discrete event set according to the relative time at which they were triggered. Each event included the following information: relative time, positive emotion score, negative emotion score, degree of attention recovery, sense of stress relief, and sense of social connection. Since these data were only obtained at discrete time points, they were not interpolated when constructing the multimodal time series dataset, but were instead retained as an independent event table and combined with continuous data by associating them with the time axis. Then, the time axis of the multimodal data was aligned: 1. A unified relative time coordinate system was established, with the absolute time at the start of the intervention as the origin, and the relative time was defined as the absolute time minus the intervention start time. The relative time at the end of the intervention was the total intervention duration. Unify the data of all modes into this relative time coordinate system.

[0053] 2. Generate an equally spaced time grid. Assuming a sampling interval of 0.25 seconds (corresponding to 4 Hz), the equally spaced time point sequence will be 0 seconds, 0.25 seconds, 0.5 seconds, ..., until the intervention ends. Each time point represents a data sampling moment.

[0054] 3. Construct a multimodal time series dataset. Align the preprocessed physiological data (RR interval, skin conductance, synthetic acceleration) and behavioral data (activity type, task focus, interaction quality) with the aforementioned equally spaced time points to form a continuous physiological-behavioral time series. Instantaneous experience data is retained as discrete events, and the time point index of each event is recorded for correlation. The final dataset can be organized into a data table, with each row corresponding to an equally spaced time point, containing a timestamp, RR interval value, skin conductance value, synthetic acceleration value, activity type code, task focus value, interaction quality value, and a flag indicating whether instantaneous evaluation data exists at that time point. If so, the scores for each item are stored in an additional column; otherwise, leave it blank.

[0055] In some embodiments of this application, when preprocessing and aligning the collected intervention data with the timeline to construct a multimodal time series dataset containing three dimensions—physiological, behavioral, and experiential—the process includes: At the start of intervention Define relative time as the origin. All collected data are unified into the same relative time coordinate system; Extracting RR interval sequences from heart rate variability signals from collected continuous physiological data. ,in , where is the heartbeat sequence number during the intervention process, and M is the total number of effective heartbeats.

[0056] Interpolation method is used to convert non-equidistant intervals Convert to equal time intervals RR interval estimates The skin conductance signal was filtered and denoised to obtain a clean skin conductance signal. .

[0057] The composite acceleration of the triaxial acceleration signal is determined, and the activity intensity is determined based on the composite acceleration.

[0058] The collected behavioral video data is encoded and labeled to generate a behavioral event sequence containing the start time, end time, activity type, and task focus level of each behavioral event. The behavioral event sequence is then converted into equally spaced time points. The types of activities and task focus on.

[0059] The collected instantaneous experience data is used to generate a set of experience events based on the trigger time point, including positive emotion scores, negative emotion scores, attention recovery level, stress relief, and social connection.

[0060] The preprocessed physiological time series, behavioral time series, and experience event set are aggregated along the time axis to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential.

[0061] Specifically, the absolute moment at which the intervention begins. Define relative time as the origin of time. , ,in The absolute time is denoted as , and the relative time at the end of the intervention is denoted as . The heart rate variability signal was then processed to remove baseline drift and high-frequency noise, and the heartbeat time was detected to obtain the R-wave time sequence. ,in This represents the total number of heartbeats detected. The RR interval is calculated as the interval between adjacent heartbeats. The original RR interval sequence was obtained. Calculate the mean of the original RR intervals. and standard deviation ,like If any of these are found to be abnormal, they are removed. After removal, the valid RR interval sequences are obtained, and their length is denoted as... (Updated value). No interpolation is performed. Because the RR intervals are not uniformly spaced, interpolation to a uniform time grid is required. Set the target sampling interval. Uniform time point: ,in Cubic spline interpolation is used in each... Obtain RR interval estimates The skin conductance signal was then repaired using linear interpolation or wavelet transform and resampled to a uniform time grid with the same RR interval. Subsequently, the triaxial acceleration was analyzed for each time step. Calculate the resultant acceleration: This value reflects the overall intensity of body motion and is resampled to a uniform time grid. To smooth instantaneous fluctuations, the average composite acceleration within a sliding window is calculated. Window length. (correspond (points), activity intensity is defined as: ,in, Within the effective range, fewer points should be used at the edges. The unit remains g. The video will then be imported into encoding software and labeled according to preset rules. Each behavioral event includes: relative start time, relative end time, activity type, and task focus level, resulting in a sequence of behavioral events. For each uniform time point... Find satisfaction If such an event exists, then ,like If it does not belong to any event (such as a transition gap), then (Rest code) .in, For the start relative time, The activity type is defined as follows: (To be continued) Task focus is The mobile device triggers an instantaneous evaluation, and each evaluation records the relative time of triggering. The five scores directly form a discrete set of events, requiring no interpolation. Each event includes: ,in, For positive emotions, score (1–5). Score negative emotions (1–5). The degree of attention recovery (1–5). For a sense of stress relief (1–5). For social connection (1–5). Merge all the above processing results along the timeline to form a unified data structure.

[0062] In some embodiments of this application, extracting physiological process features, behavioral process features, and experiential process features reflecting the dynamic characteristics of the intervention process from a multimodal time series dataset includes: Physiological process features are extracted from the physiological time series of multimodal time series datasets. These physiological process features include at least the slope of the stress recovery curve, the area under the stress curve, the physiological recovery efficiency index, and the frequency of skin conductance response.

[0063] Behavioral process features are extracted from the behavioral time series of multimodal time series datasets. These features include at least the time proportion of each type of activity, the frequency of activity transitions, the average duration of activity segments, and a comprehensive index of participation depth.

[0064] Experience process features are extracted from the experience event set of the multimodal time series dataset. The experience event set includes the trigger time of each instantaneous evaluation and the scores of each dimension. The experience process features include at least the average positive emotion score, average negative emotion score, average attention recovery level, average stress relief, and average social connection.

[0065] Specifically, the physiological process characteristics are based on equally spaced physiological time series (RR intervals, skin conductance levels, activity intensity, etc.). The focus is primarily on the dynamic process of stress recovery. The slope of the stress recovery curve reflects the overall rate of decrease in stress levels (indicated by skin conductance levels) during the intervention. A larger negative slope (i.e., a larger absolute value) indicates faster stress recovery. First, skin conductance levels are obtained at each uniform time point throughout the intervention period, and the overall linear trend slope from the start to the end of the intervention is calculated: the smoothed value at the start of the intervention is denoted as... The smoothing value at the end of the intervention is denoted as The total intervention duration was The slope of the stress recovery curve is: If the result is negative, it indicates a decrease in stress level. The area under the stress curve represents the total stress load during the intervention period, taking into account the deviation of skin conductance level from the individual baseline at each time point. First, the average skin conductance level of the participant is obtained from the baseline database, and the difference is calculated for each time point. The trapezoidal method was used to integrate the entire intervention period: Positive values ​​indicate total stress above baseline, while negative values ​​indicate stress below baseline. The physiological recovery efficiency index comprehensively reflects the stress recovery efficiency per unit "activity dose." A higher index indicates a more significant decrease in stress for a given activity intensity and duration. First, the maximum decrease in skin conductance during the intervention is calculated: That is, the baseline value minus the minimum value during the intervention. If the minimum value is higher than the baseline, The value is negative. Calculate the average activity intensity during the intervention period: ,in This refers to the activity intensity sequence calculated above. The total intervention duration is... The physiological recovery efficiency index is: Skin conductance response frequency is the number of skin conductance responses per minute during the intervention, reflecting the transient fluctuations in sympathetic nerve activity. Using pre-processed clean skin conductance signals, its first derivative (difference) is calculated. Segments with a derivative greater than a preset threshold and a duration between 1 and 5 seconds are identified; each such segment is counted as one skin conductance response. The total number of responses during the entire intervention period is counted. Behavioral process characteristics are based on activity type sequences and task focus sequences on a uniform time grid. They represent the proportion of total time allocated to each horticultural activity (e.g., sowing, transplanting, pruning, harvesting, resting, etc.) during the intervention, reflecting the activity structure. A preset set of activity types is used; for each type, the number of time points in the sequence equal to that type is counted. ,in, For each corresponding type, For statistical sequences, The indicator function (value is 1 if the condition is true, otherwise it is 0), the total number of time points is The time allocation for each type of activity is as follows: ,all The sum is 1. The average activity segment duration is the average duration of performing the same activity type consecutively, reflecting the participant's sustained focus. First, the activity type sequence is divided into consecutive activity segments. Each segment consists of the same activity type, lasting for several consecutive time points. Let there be a total of... The activity segment, the first The duration of each segment is The calculation method is as follows: The average duration of the activity segment: The participation depth comprehensive index comprehensively considers participants' task focus, interaction quality, and social interaction level, quantifies the overall level of participation, and calculates the average task focus: Obtain the average interaction quality score: Calculate the frequency of social interactions The number of verbal / nonverbal interactions between participants and therapists or others was counted from behavioral annotations and divided by the total intervention duration. A weighted summation was used to calculate the comprehensive engagement depth index. ,in, , , These are the weighting coefficients. The experience process features are based on a set of instantaneous evaluation events. Each event records the subjective feeling score at the trigger time point. Let there be a total of... The first effective instantaneous evaluation, the first The score is The average positive emotion score is: The average negative sentiment score is the average of all instantaneous negative sentiment scores obtained from the assessments. The score for negative emotions was: Average negative emotion score: The average degree of attention recovery is the average of the participants' reported sense of attention recovery across all instantaneous assessments, reflecting the intervention's effect on alleviating attentional fatigue. Let the mean degree of attention recovery be... The score for attention recovery is: Average attention recovery level: Average stress relief is the average level of stress relief reported by participants themselves. Let the first part be... The score for the feeling of pressure relief is: Average pressure relief sensation: Average social connection is the average degree of connection participants feel with others or the environment during the intervention. Let the first term be... The score for secondary social connection is Average sense of social connection: After completing the above extraction, physiological process characteristics, behavioral process characteristics, and experiential process characteristics are obtained, which are used to subsequently construct individual process profiles and efficacy prediction models.

[0066] In some embodiments of this application, when constructing an individual process profile for each participant by vectorizing the extracted multi-dimensional process features, the following steps are included: The extracted physiological process features, behavioral process features, and experiential process features are summarized in a preset order to form an original feature set.

[0067] Arrange each feature in the original feature set in a fixed order to construct a one-dimensional original feature vector.

[0068] Each feature in the original feature vector is standardized to convert it into a standardized value with a mean of 0 and a standard deviation of 1, thus obtaining the standardized feature vector.

[0069] The standardized feature vector is used as an individual process profile of the participant.

[0070] Each participant's individual process profile is linked and stored with their corresponding baseline data, outcome indicators, and covariate information to construct an individual process profile database.

[0071] Specifically, firstly, all process features extracted from the multimodal time series dataset are aggregated into a raw feature set in a fixed order. This fixed order ensures consistency in the dimensionality of feature vectors across different participants. The specified feature order is: slope of the stress recovery curve, area under the stress curve, physiological recovery efficiency index, frequency of skin conductance response, time percentage of various activities, activity transition frequency, average activity segment duration, comprehensive participation depth index, average positive emotion score, average negative emotion score, average attention recovery level, average stress release perception, and average social connection perception. All features are arranged in this order to obtain a raw feature vector. If a feature is missing during extraction (e.g., due to data quality issues), its mean among existing participants is used to fill in the gaps. Since different features have different dimensions and numerical ranges, directly using raw values ​​can lead to machine learning models biasing towards features with larger values. Therefore, each feature needs to be standardized, converting it into a standardized value with a mean of 0 and a standard deviation of 1. The standardized feature vector is then... As the first Individual process profiles of each participant, denoted as To create an individual process profile for each participant. The following data are associated and stored to form a structured database: Baseline data: the participant's physiological and psychological baseline vectors extracted from the baseline database; Outcome indicators: changes in psychological scale values ​​before and after the intervention; Covariate information: variables that may affect the efficacy, such as the participant's age, gender, education level, diagnostic information, intervention date, and interventionist number.

[0072] In some embodiments of this application, training the efficacy prediction model using a machine learning algorithm includes: We collected sample data from participants who had completed the intervention. The sample data included individual process profile vectors for each participant and corresponding standardized outcome indicators.

[0073] The sample data is divided into a training set and a test set.

[0074] An ensemble learning model is constructed using random forest, extreme gradient boosting, and lightweight gradient boosting machine as base learners, and linear regression as a meta-learner.

[0075] Based on the K-fold cross-validation method, each base learner is trained separately using the training set to obtain the predicted value of each base learner on the training set, and the predicted value is used as the meta-feature.

[0076] Using meta-features as input and standardized outcome metrics as targets, a meta-learner is trained to obtain an ensemble learning model.

[0077] The trained ensemble learning model was evaluated using a test set, and the feature importance scores of each base learner were extracted to identify the process features that contributed most to efficacy prediction.

[0078] In some embodiments of this application, generating a multidimensional assessment report of intervention quality includes: Obtain the individual process profile vector of the current intervention and input it into the trained efficacy prediction model to obtain the efficacy prediction value.

[0079] The participants who rank in the top percentage of outcome indicators from the training data are selected, and the average value of their process characteristics is calculated to form the optimal process profile.

[0080] Calculate the weighted Euclidean distance between the current process profile and the optimal process profile, and calculate the intervention quality index based on the weighted Euclidean distance.

[0081] The quality scores of the current intervention were calculated in the physiological, behavioral, and experiential dimensions, and the scores of each dimension were compared with the baseline data of the intervention and the benchmark data of the same population.

[0082] Based on the efficacy prediction value, intervention quality index, quality scores of each dimension and comparison results, a multidimensional assessment report of intervention quality is generated, which includes a summary of efficacy prediction, a display of intervention quality index, radar charts of quality scores of each dimension, a list of key process characteristics, and personalized optimization suggestions.

[0083] Specifically, the following two types of data were collected from participants who had completed the intervention: an individual process profile vector for each participant and the corresponding standardized outcome measure. The dataset was randomly divided into a training set and a test set in an 8:2 ratio, and a two-layer ensemble structure was adopted: First layer (base learner): Random Forest (RF), Extreme Gradient Boosting (XGBoost, XGB), Lightweight Gradient Boosting Machine (LightGBM, LGB).

[0084] The second layer (meta-learner) is linear regression with added L2 regularization (ridge regression).

[0085] Hyperparameter presets for base learners: Random Forest: 500 trees, maximum depth 10, minimum number of sample splits 5.

[0086] XGBoost: Learning rate 0.05, maximum depth 6, number of trees 200.

[0087] LightGBM: learning rate 0.05, maximum depth 6, number of leaf nodes 30.

[0088] The base learner is trained using K-fold cross-validation and meta-features are generated. The number of cross-validation folds is set. The training set is randomly divided into equal parts. For each base learner, perform the following operations: arrive One set is used as the validation set, and the rest as the training set. The base learner is trained using the training set. Then, predictions are made for each sample in the validation set, and the predicted values ​​are concatenated in the original order of the samples to form the prediction vector of the base learner across the entire training set. Simultaneously, a final base learner model is trained using all the training data. This process is repeated for all three learners. The random forest is denoted as... XGBoost and LightGBM are respectively and For each sample in the training set The predictions from its three base learners are combined into a single meta-feature vector: The meta-feature matrix is ​​formed by the meta-feature vectors of all samples. Using the meta-feature matrix as input and the true outcome metric as the target, a linear regression model with L2 regularization (ridge regression) is trained. The regularization parameter is selected through cross-validation. After training, the meta-learner model is obtained, which has the following form: ,in, The intercept is... , and The coefficients are used. Three final base learner models are used to predict each sample in the test set, resulting in three predicted values. These predicted values ​​are then combined into a meta-feature vector, which is input into the meta-learner to obtain the final predicted value. Calculate the regression performance metrics: root mean square error (RMSE) and coefficient of determination. , . .in This represents the mean of the true outcomes on the test set. The loss function for Random Forest is to use the mean squared error (MSE) of each tree as the criterion for node splitting (i.e., minimizing the variance of the leaf nodes), and finally take the average of the predictions from all trees. XGBoost and LightGBM, on the other hand, are gradient boosting methods that fit the residual of the previous prediction in each iteration, and their loss function is the squared loss of the MSE. Feature importance scores were extracted from the Random Forest, XGBoost, and LightGBM models respectively. The importance scores of the same feature in the three models were then weighted and averaged (or the maximum value was taken) to obtain a comprehensive importance ranking. The top 5-10 process features that contributed most to predicting treatment efficacy were identified.

[0089] After a new horticultural therapy intervention is completed, an evaluation report containing multi-dimensional quality information is generated based on its process profile and a trained model. The top 20% of participants in terms of outcome metrics (i.e., the group with the best treatment outcomes) are selected from the training data. The average value of these participants for each process feature is calculated to obtain the optimal process profile vector. First, the weighted Euclidean distance between the current profile and the optimal profile is calculated. Let the total number of features be D, and the weight of the d-th feature be... The score is taken as the overall importance score of this feature in the Stacking model. ,in, For Euclidean distance, and Let be the values ​​of the current and optimal portraits on the d-th feature, respectively. An exponential decay function is used to map the distance to 0-100 points. , The closer a value is to 100, the closer the current intervention process is to the historical optimal model. This is based on the following sub-items, each scored using percentile ratings (based on the distribution of similar populations in the training set): The slope of the stress recovery curve (the more negative, the better).

[0090] The area under the stress curve (the smaller the better).

[0091] Physiological recovery efficiency index (the higher the better).

[0092] Skin conductance response frequency (moderate is best, and an ideal range can be set).

[0093] Final physiological dimension score = weighted average of all sub-items.

[0094] 2. Quality Score for Behavioral Dimension Participation depth comprehensive index (the higher the better).

[0095] Activity switching frequency (moderate is best; too high may indicate distraction, while too low indicates monotony).

[0096] Average activity segment duration (too long or too short may be undesirable; take the value that is close to the optimal value).

[0097] Final behavioral dimension score = weighted average.

[0098] 3. Experience Dimension Quality Score Average positive emotion score (the higher the better).

[0099] Average negative emotion score (lower is better, can be converted to a 6-NA score).

[0100] Average attention recovery, stress relief, and sense of social connection (the higher the better).

[0101] Final experience score = average of the above scores.

[0102] 4. Benchmark Comparison Baseline: Compare the improvement of physiological indicators such as skin conductance and heart rate variability relative to baseline during the intervention.

[0103] Benchmark population: A group matching the current participant's baseline characteristics (age, gender, initial symptom severity) is selected from the training data. The mean and standard deviation of this group's scores across each dimension are calculated. The current score is compared to this benchmark, indicating whether it is "better than / on par with / below" the benchmark population.

[0104] 5. Visualize and create a radar chart to display the scores of the current intervention across the physiological, behavioral, and experiential dimensions, as well as the optimal profile and the average score of similar groups. Each axis of the radar chart represents a score from low to high, moving from the inside out.

[0105] VI. Personalized Optimization Suggestions Based on the feature importance ranking and the shortcomings in the current profile, specific suggestions are generated: If the physiological dimension score is low: it is recommended to increase low-intensity, repetitive activities (such as loosening the soil and watering), or to insert 1-2 minutes of mindfulness breathing / plant contact into the activities.

[0106] If the behavioral dimension score is low: it is recommended to break down complex tasks into smaller steps, arrange group collaborative tasks, or increase immediate feedback (such as showing the results of plant growth).

[0107] If the experience dimension score is low: it is recommended to prioritize plants with high sensory stimulation (such as herbs and succulents), or introduce environmental elements such as music and natural sounds.

[0108] If a specific important characteristic (such as the physiological recovery efficiency index) is low: provide targeted enhancement suggestions for that characteristic (such as extending the duration of a single activity or reducing the intensity of the activity).

[0109] At the same time, based on the gap between the predicted efficacy and the historical best sample, suggestions may be made on whether to increase the number of interventions, extend the duration of each intervention, or adjust the activity sequence.

[0110] In some embodiments of this application, it also includes: When the intervention quality index is below the first threshold but above the second threshold, a Level 1 warning is generated. The Level 1 warning includes marking the low-scoring dimensions and corresponding general optimization suggestions.

[0111] When the intervention quality index is lower than or equal to the second threshold, a secondary warning is generated. The secondary warning includes marking low-scoring dimensions, targeted optimization suggestions generated based on feature importance analysis, and suggestions to re-evaluate the intervention plan or increase the number of supplementary interventions. The first threshold is higher than the second threshold.

[0112] In summary, the beneficial effects of this invention are as follows: By synchronously acquiring multimodal data and aligning it with the timeline, this invention constructs a three-dimensional time-series dataset that integrates continuous physiological signals, behavioral videos, and instantaneous subjective experiences. This enables continuous tracking of the entire horticultural therapy intervention process, enhancing the dynamism and information completeness of the assessment. Based on this, the extracted physiological, behavioral, and experiential process features can quantify the subtle changes in individual stress recovery, participation quality, and subjective feelings during the intervention process. This elevates the assessment from traditional "comparison of pre- and post-test results" to "process mechanism analysis," providing multi-dimensional process evaluation criteria instead of relying on a single outcome indicator. Simultaneously, by vectorizing multi-dimensional process features to construct individual process profiles, this invention achieves personalized digital modeling of each participant's intervention performance, overcoming the limitation of the group average effect masking individual differences and improving the accuracy and relevance of the assessment. Using individual process profiles as input and standardized outcome indicators as prediction targets, a machine learning algorithm is used to train the efficacy prediction model. This allows for the prediction of possible efficacy levels based on current intervention process characteristics, transforming the assessment from post-event summarization to pre-event prediction, reducing feedback lag, and creating conditions for timely adjustments to the intervention plan. The resulting multidimensional intervention quality assessment report not only includes efficacy predictions and intervention quality indices, but also provides quality scores for each dimension and comparisons with baselines and benchmarks for similar populations, as well as personalized optimization suggestions based on feature importance analysis, enriching the interpretability and practicality of the assessment results. The report can also provide tiered warnings based on the degree to which the intervention quality index falls below a preset threshold, indicating specific dimensions of intervention quality deficiencies and providing targeted improvement measures, thus achieving a closed loop from "scoring" to "action guidance." Furthermore, the entire process employs standardized data collection, automated preprocessing, and machine learning model prediction, reducing biases from manual observation and subjective recall, and improving the objectivity, reproducibility, and cross-study comparability of the assessment results. In summary, this invention elevates the assessment of horticultural therapy intervention quality from a static, discrete, population-based, and descriptive paradigm to a dynamic, continuous, individualized, and predictive paradigm, comprehensively enhancing the information dimensions, temporal resolution, personalization level, and clinical practical value of the assessment.

[0113] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides an intervention quality assessment system based on horticultural therapy, used to apply the above-mentioned intervention quality assessment method based on horticultural therapy, including: The data acquisition module is configured to acquire multimodal data synchronously, including wearable physiological monitoring devices, high-definition camera equipment, and mobile terminal instantaneous assessment data.

[0114] The baseline database building module is configured to establish an individual baseline database, collecting participants' resting physiological data and baseline psychometric data on non-intervention days.

[0115] The intervention data acquisition module is configured to collect intervention data from participants during the implementation of horticultural therapy interventions. The intervention data includes continuous physiological data, behavioral video data, and momentary experience data.

[0116] The preprocessing and alignment module is configured to preprocess and align the collected intervention data along the timeline to construct a multimodal time series dataset that includes physiological, behavioral, and experiential dimensions.

[0117] The feature extraction module is configured to extract physiological process features, behavioral process features, and experiential process features that reflect the dynamic characteristics of the intervention process from a multimodal time series dataset.

[0118] The individual profile building module is configured to vectorize the extracted multi-dimensional process features to build an individual process profile for each participant.

[0119] The model training module is configured to use individual process profiles as input features, standardized outcome indicators as prediction targets, and machine learning algorithms to train the efficacy prediction model.

[0120] The report generation module is configured to generate a multidimensional assessment report of intervention quality based on a trained efficacy prediction model and the characteristics of the current intervention process.

[0121] The early warning module is configured to generate corresponding early warning prompts based on the intervention quality index.

[0122] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0123] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0124] These computer program instructions may also be stored in a computer-readable storage device that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage device produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for assessing the quality of interventions based on horticultural therapy, characterized in that, include: Acquire multimodal data synchronously, including wearable physiological monitoring devices, high-definition camera devices, and mobile terminal instantaneous assessment data; Establish an individual baseline database and collect participants' resting physiological data and baseline psychometric data on non-intervention days; During the implementation of horticultural therapy intervention, intervention data was collected from participants, including continuous physiological data, behavioral video data, and instantaneous experience data. The collected intervention data were preprocessed and aligned with the timeline to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential. Physiological process features, behavioral process features, and experiential process features that reflect the dynamic characteristics of the intervention process are extracted from the multimodal time series dataset; The extracted multi-dimensional process features are vectorized to construct an individual process profile for each participant; Using the individual process profile as input features and standardized outcome indicators as prediction targets, a machine learning algorithm is used to train the efficacy prediction model; Based on the trained efficacy prediction model and combined with the current intervention process characteristics, a multidimensional assessment report of intervention quality is generated.

2. The intervention quality assessment method based on horticultural therapy according to claim 1, characterized in that, When establishing an individual baseline database, participants' resting physiological data and baseline psychometric data were collected on non-intervention days, including: Acquire continuous physiological signals from participants in a resting state on non-intervention days, including at least heart rate variability signals, skin conductance signals, and triaxial acceleration signals; The collected raw physiological signals were preprocessed, including: bandpass filtering of the heart rate variability signal and extraction of the RR interval sequence, and removal of abnormal RR intervals; low-pass filtering of the skin conductance signal and removal of motion artifacts; and calculation of the synthetic acceleration from the triaxial acceleration signal. Baseline physiological features are extracted from the preprocessed physiological signals, and the baseline physiological features include at least: Mean RR interval ; RR interval standard deviation ; Root mean square of the difference between adjacent RR intervals ; low frequency power and high frequency power ; Low frequency to high frequency ratio ; Mean skin conductivity ; Number of skin conduction responses That is, the average number of skin conductance responses per minute in a resting state; Average composite acceleration ; Baseline psychometric data of participants were collected, including at least the perceived stress scale score. Anxiety Self-Rating Scale Score Score of the Nature Connection Scale ; The baseline physiological characteristics and baseline psychometric data are combined to form an individual baseline record for each participant. ,in Let N be the RR interval of the i-th normal heartbeat, and N be the total number of valid RR intervals. The baseline mean RR interval is given, and T is the resting state duration. The skin conductance value at time t. Let x be the acceleration along the x-axis at time t. Let be the Y-axis acceleration at time t. Let be the Z-axis acceleration at time t. Let be the resultant acceleration at time t.

3. The intervention quality assessment method based on horticultural therapy according to claim 2, characterized in that, When collecting intervention data from participants during the implementation of horticultural therapy interventions, the following should be included: Continuously collect participants' heart rate variability signals, skin conductance signals, triaxial acceleration signals, and video behavioral data; Using a mobile instant assessment terminal, based on a combination of preset time triggers and event triggers, the participants' immediate subjective experience is collected through a questionnaire during the intervention process. The questionnaire includes at least assessment items on immediate emotional state, perceptual recovery experience, and social connection. The acquired raw physiological signals were preprocessed, including: extracting the RR interval sequence from the heart rate variability signal; filtering and denoising the skin conductance signal to obtain a clean skin conductance signal; and calculating the synthetic acceleration from the triaxial acceleration signal. The collected behavioral video data is encoded and labeled to generate a sequence of behavioral events. The encoded results are then converted into a time series format, including at least the activity type and task focus at each time point. Organize the instantaneous experience data according to the trigger time point, including at least positive emotion score, negative emotion score, attention recovery level, stress relief and social connection; Using the intervention start time as the origin, all modal data are unified into the same relative time coordinate system to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential.

4. The intervention quality assessment method based on horticultural therapy according to claim 3, characterized in that, When preprocessing and aligning the collected intervention data with the timeline to construct a multimodal time series dataset encompassing physiological, behavioral, and experiential dimensions, the following steps are included: At the start of intervention Define relative time as the origin. All collected data are unified into the same relative time coordinate system; Extracting RR interval sequences from heart rate variability signals from collected continuous physiological data. ,in , where is the heartbeat sequence number during the intervention process, and M is the total number of effective heartbeats; Interpolation method is used to convert non-equidistant intervals Convert to equal time intervals RR interval estimates The skin conductance signal was filtered and denoised to obtain a clean skin conductance signal. ; The composite acceleration of the triaxial acceleration signal is determined, and the activity intensity is determined based on the composite acceleration. The collected behavioral video data is encoded and labeled to generate a behavioral event sequence containing the start time, end time, activity type, and task focus level of each behavioral event. The behavioral event sequence is then converted into equally spaced time points. Activity type and task focus; The collected instantaneous experience data is used to generate a set of experience events based on the trigger time point, including positive emotion scores, negative emotion scores, attention recovery level, stress relief, and social connection. The preprocessed physiological time series, behavioral time series, and experience event set are aggregated along the time axis to construct a multimodal time series dataset containing three dimensions: physiological, behavioral, and experiential.

5. The intervention quality assessment method based on horticultural therapy according to claim 4, characterized in that, When extracting physiological process features, behavioral process features, and experiential process features reflecting the dynamic characteristics of the intervention process from the multimodal time series dataset, the following are included: Physiological process features are extracted from the physiological time series of the multimodal time series dataset. The physiological process features include at least the slope of the stress recovery curve, the area under the stress curve, the physiological recovery efficiency index, and the frequency of skin conductance response. Behavioral process features are extracted from the behavioral time series of the multimodal time series dataset. The behavioral process features include at least the time proportion of various activities, activity transition frequency, average activity segment duration, and comprehensive participation depth index. Experience process features are extracted from the experience event set of the multimodal time series dataset. The experience event set includes the trigger time of each instantaneous evaluation and the scores of each dimension. The experience process features include at least the average positive emotion score, average negative emotion score, average attention recovery level, average stress relief, and average social connection.

6. The intervention quality assessment method based on horticultural therapy according to claim 5, characterized in that, When constructing an individual process profile for each participant by vectorizing the extracted multi-dimensional process features, the following are included: The extracted physiological process features, behavioral process features, and experiential process features are summarized in a preset order to form an original feature set; Arrange each feature in the original feature set in a fixed order to construct a one-dimensional original feature vector; Each feature in the original feature vector is standardized to convert each feature into a standardized value with a mean of 0 and a standard deviation of 1, thus obtaining a standardized feature vector. The standardized feature vector is used as an individual process profile of the participant; Each participant's individual process profile is linked and stored with their corresponding baseline data, outcome indicators, and covariate information to construct an individual process profile database.

7. The intervention quality assessment method based on horticultural therapy according to claim 6, characterized in that, When training a treatment efficacy prediction model using machine learning algorithms, the following steps are included: Collect sample data from participants who have completed the intervention. The sample data includes an individual process profile vector for each participant and the corresponding standardized outcome indicators. The sample data is divided into a training set and a test set; An ensemble learning model is constructed using random forest, extreme gradient boosting, and lightweight gradient boosting machine as base learners, and linear regression as a meta-learner. Based on the K-fold cross-validation method, each base learner is trained using the training set to obtain the predicted value of each base learner on the training set, and the predicted value is used as a meta-feature. Using the meta-features as input and the standardized outcome index as the target, the meta-learner is trained to obtain the ensemble learning model; The trained ensemble learning model is evaluated using the test set, and the feature importance scores of each base learner are extracted to identify the process features that contribute the most to efficacy prediction.

8. The intervention quality assessment method based on horticultural therapy according to claim 7, characterized in that, When generating a multidimensional assessment report of intervention quality, the following should be included: Obtain the individual process profile vector of the current intervention and input it into the trained efficacy prediction model to obtain the efficacy prediction value; The participants who rank in the top percentage of outcome indicators from the training data are selected, and the average value of each process characteristic is calculated to form the optimal process profile. Calculate the weighted Euclidean distance between the current process profile and the optimal process profile, and calculate the intervention quality index based on the weighted Euclidean distance; Calculate the quality scores of the current intervention in the physiological, behavioral, and experiential dimensions, and compare the scores of each dimension with the baseline data of the intervention and the benchmark data of the same population. Based on the efficacy prediction value, intervention quality index, quality scores of each dimension and comparison results, a multidimensional assessment report of intervention quality is generated, which includes a summary of efficacy prediction, a display of intervention quality index, radar charts of quality scores of each dimension, a list of key process characteristics, and personalized optimization suggestions.

9. The intervention quality assessment method based on horticultural therapy according to claim 8, characterized in that, Also includes: When the intervention quality index is lower than the first threshold and higher than the second threshold, a first-level warning prompt is generated. The first-level warning prompt includes marking the low-scoring dimension and the corresponding general optimization suggestions. When the intervention quality index is lower than or equal to the second threshold, a secondary warning prompt is generated. The secondary warning prompt includes marking low-scoring dimensions, targeted optimization suggestions generated based on feature importance analysis, and suggestions to re-evaluate the intervention plan or increase the number of supplementary interventions. The first threshold is higher than the second threshold.

10. A horticultural therapy-based intervention quality assessment system, applied to the horticultural therapy-based intervention quality assessment method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is configured to acquire multimodal data synchronously, including wearable physiological monitoring devices, high-definition camera devices, and mobile terminal instantaneous assessment data. The baseline database building module is configured to establish an individual baseline database, collecting participants' resting physiological data and baseline psychometric data on non-intervention days; The intervention data acquisition module is configured to collect intervention data from participants during the implementation of horticultural therapy interventions. The intervention data includes continuous physiological data, behavioral video data, and instantaneous experience data. The preprocessing and alignment module is configured to preprocess and align the collected intervention data along the timeline to construct a multimodal time series dataset that includes physiological, behavioral, and experiential dimensions. The feature extraction module is configured to extract physiological process features, behavioral process features, and experiential process features that reflect the dynamic characteristics of the intervention process from the multimodal time series dataset; The individual profile building module is configured to vectorize the extracted multi-dimensional process features to build an individual process profile for each participant. The model training module is configured to use the individual process profile as input features, standardized outcome indicators as prediction targets, and machine learning algorithms to train the efficacy prediction model. The report generation module is configured to generate a multidimensional assessment report of intervention quality based on a trained efficacy prediction model and the characteristics of the current intervention process. The early warning module is configured to generate corresponding early warning prompts based on the intervention quality index.

Images