Manic-depressive state switching recognition system and method based on multi-source feature fusion

Abstract

The application discloses a mania-depression state switching recognition system and method based on multi-source feature fusion, and belongs to the technical field of state recognition. The application is based on collected multi-source time sequence data, quality evaluation and individual dynamic baseline establishment, calculates offset features, extracts cross-source time sequence structures to generate time sequence structure feature vectors, then generates a comprehensive feature vector according to the quality evaluation score, triggers data backfilling when the comprehensive confidence is insufficient, and inputs the comprehensive feature vector into a time sequence probability model to calculate a state switching probability sequence when the comprehensive confidence is sufficient, finally, makes a hierarchical decision according to the probability trend and generates updated collection strategy parameters in reverse, to dynamically schedule multi-source data collection actions in the next time window, improve the timeliness of state switching recognition and the reliability of recognition results, and solve the problems that the existing inpatient state management mode is prone to cause state switching recognition lag and the reliability of analysis results to decrease.

Description

Technical Field

[0001] This invention relates to the field of state recognition technology, and in particular to a system and method for identifying manic-depressive state switching based on multi-source feature fusion. Background Technology

[0002] Bipolar disorder is a mental illness characterized by alternating manic and depressive phases. The transitions between these phases are often gradual and insidious, and patients typically experience changes in various objective aspects of their behavior, including sleep, activity, speech, and mood, during these transitions. For hospitalized patients, timely detection of these early signs of transition is crucial for timely treatment and ensuring patient safety. Therefore, how to continuously and objectively identify these transition trends in clinical management has always been a focus of attention in the field of mental health.

[0003] In current clinical management practices, monitoring changes in hospitalized patients' conditions primarily relies on regular ward rounds and periodic scale assessments by medical staff. Medical staff directly observe and record patients' behavior during ward rounds, and combine this with periodic structured scale evaluations to assess patients' conditions and their trends. With the development of monitoring technology, some management methods have incorporated the collection of physiological and behavioral data such as patient sleep and activity levels. By processing and analyzing this data, medical staff can better understand changes in patients' conditions. These methods typically collect data from multiple sources according to pre-defined procedures, process the data, and generate corresponding analytical results for medical staff's reference.

[0004] However, the existing methods still have certain limitations in practical applications. Firstly, ward rounds and scale assessments are conducted at discrete time points, with long intervals between assessments. Patient status transitions often occur within these intervals, potentially leading to delays in identifying status changes. Secondly, when collecting and processing data from multiple sources, the completeness and reliability of the data vary. If patients fail to cooperate with the data collection process or medical staff fail to enter relevant records in a timely manner, some data sources may be missing or of lower quality. Processing data from different sources in a fixed manner makes it difficult to adjust for these fluctuations in data quality, thus affecting the reliability of the analytical results. Furthermore, existing methods typically involve data collection and observation at a pre-set fixed intensity. This fixed approach cannot promptly increase the monitoring density to accurately grasp changes in the patient's condition, nor can it lead to unnecessary monitoring efforts when the patient's condition is stable, making it difficult to coordinate with limited nursing resources and the patient's need for rest.

[0005] It is evident that existing methods for managing the status of hospitalized patients are inadequate in terms of adaptability to fluctuations in data quality and coordination between monitoring intensity and trends in patient status changes. This can easily lead to delays in status transition identification and a decrease in the reliability of analysis results. Summary of the Invention

[0006] To address the technical problems of existing inpatient state management methods, which easily lead to lag in state transition identification and decreased reliability of analysis results, this invention provides a system and method for identifying manic-depressive state transitions based on multi-source feature fusion. The technical solution is as follows:

[0007] On the one hand, a system for identifying manic-depressive state transitions based on multi-source feature fusion is provided, comprising: a multi-source data initialization module, used to acquire time-series data from various sources based on current acquisition strategy parameters to perform quality assessments and obtain quality assessment scores, and simultaneously acquire ward context labels; an individual adaptive baseline module, used to filter the multi-source time-series data by time period in conjunction with ward context labels, establish an individual dynamic baseline by combining historical filtered data, and calculate the offset feature vector of the multi-source time-series data relative to the individual dynamic baseline to extract cross-source time-series structure and generate a time-series structure feature vector; and a dynamic confidence fusion module, used to perform dynamic confidence weighted fusion of the time-series structure feature vectors of each source based on the quality assessment scores to generate a comprehensive feature vector, and calculate the corresponding comprehensive confidence weight. The system comprises the following modules: a confidence level and a data recovery control module, which generates a data recovery request command and sends it to the nurse station terminal connected to the communication connection when the overall confidence level is detected to be less than a preset threshold. This module obtains the recovered multi-source time-series data and regenerates the overall feature vector based on the recovered multi-source time-series data. A state switching probability calculation module is used to input the overall feature vector into a preset time-series probability model and calculate the state switching probability sequence when the overall confidence level is detected to be not less than a preset threshold. A probability trend classification and scheduling counter-control module is used to perform trend classification judgment based on the state switching probability sequence, generate classification identifiers to generate updated acquisition strategy parameters, and use the updated acquisition strategy parameters as the current acquisition strategy parameters for the next time window to schedule multi-source data acquisition actions.

[0008] On the other hand, a method for identifying manic-depressive state transitions based on multi-source feature fusion is provided. This method includes: acquiring multi-source time-series data based on current acquisition strategy parameters to perform quality assessments and obtain quality assessment scores, while simultaneously acquiring ward context labels; filtering the multi-source time-series data by time period based on the ward context labels, establishing an individual dynamic baseline by combining historical filtered data, and calculating the offset feature vector of the multi-source time-series data relative to the individual dynamic baseline to extract cross-source time-series structure and generate a time-series structure feature vector; and performing dynamic confidence-weighted fusion of the time-series structure feature vectors of each source based on the quality assessment scores to generate a comprehensive feature vector, and calculating the corresponding... The system calculates the overall confidence level. When the overall confidence level is less than a preset threshold, a data recovery request instruction is generated and sent to the nurse station terminal connected to the communication link to obtain the recovered multi-source time-series data. Based on the recovered multi-source time-series data, the overall feature vector is regenerated. When the overall confidence level is not less than a preset threshold, the overall feature vector is input into a preset time-series probability model to calculate the state switching probability sequence. Based on the state switching probability sequence, a trend classification decision is made, and a classification identifier is generated to generate updated acquisition strategy parameters. The updated acquisition strategy parameters are used as the current acquisition strategy parameters for the next time window to schedule multi-source data acquisition actions.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0010] This invention provides a bipolar state transition recognition system based on multi-source feature fusion, targeting the state management scenario of psychiatric inpatients. Through the collaborative processing of modules such as multi-source data initialization, individual adaptive baseline, dynamic confidence fusion, backfilling adjustment, state transition probability calculation, and probability trend grading and scheduling counter-control, the system achieves continuous recognition of state transition trends. Specifically, the system combines ward context labels to filter multi-source time-series data by time period and establishes individual dynamic baselines, effectively eliminating interference from treatment and environmental factors, making the deviation measurement more closely reflect the patient's actual normal state. Furthermore, it dynamically adjusts the fusion weights based on the real-time data quality of each source, and triggers data backfilling to the nurse station terminal when the overall confidence is insufficient, thus ensuring the reliability of the data used for recognition. Simultaneously, it reverse-schedules the intensity of data collection and nursing observation based on the state transition probability, promptly increasing the monitoring density when a transition trend appears and restoring the normal intensity when the state is stable. Therefore, this system, under the condition of limited nursing resources, balances the timeliness of state transition recognition, the reliability of recognition results, and the rational allocation of nursing resources.

[0011] This invention dynamically adjusts fusion weights based on data quality and combines this with confidence-gated data feedback to form a quality closed loop. Unlike conventional fusion methods that use fixed weights, this scheme dynamically determines fusion weights based on the real-time quality assessment scores and historical contributions of each source. When the quality of a source is too low, its weight is forcibly decayed to zero, and the remaining sources are used for normalization compensation. This ensures that the fusion result is not distorted due to the absence or quality decline of individual sources. Furthermore, when the overall confidence level falls below a preset threshold, data feedback is proactively triggered at the nurse station terminal, allowing the supplemented data to re-enter the fusion process. This feature transforms the judgment of insufficient data quality into a reverse control of nursing data entry behavior, ensuring that identification is always based on reliable data.

[0012] This invention employs a probabilistic closed loop, based on the reverse scheduling of data acquisition and nursing observation intensity according to the state switching probability. Unlike methods that rely solely on a single probability value and a fixed threshold, this approach comprehensively considers the current value, upward slope, and duration of the state switching probability to determine a trend classification. Based on the resulting classification identifier, updated acquisition strategy parameters are generated and sent back to the acquisition stage to adjust the sensor sampling frequency, the activated data source, and the frequency of nursing observations. Thus, the monitoring intensity adaptively adjusts with the patient's state switching trend. When a trend emerges, acquisition and observation are intensified to accurately grasp changes; when the state stabilizes, the intensity returns to normal to rationally allocate nursing resources and reduce unnecessary disturbance to the patient. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A schematic diagram of the structure of a manic-depressive state switching recognition system based on multi-source feature fusion provided in an embodiment of this application;

[0015] Figure 2 This is a flowchart illustrating the bipolar state switching identification method based on multi-source feature fusion provided in this application embodiment. Detailed Implementation

[0016] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0017] It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of the embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] like Figure 1 The diagram shown is a structural schematic of a manic-depressive state switching recognition system based on multi-source feature fusion provided in an embodiment of this application. The system includes:

[0020] This embodiment provides a bipolar state transition recognition system based on multi-source feature fusion. This system is designed for state management scenarios of inpatients in psychiatric wards. It includes a multi-source data initialization module, an individual adaptive baseline module, a dynamic confidence fusion module, a backfilling and adjustment module, a state transition probability calculation module, and a probability trend grading and scheduling counter-control module. These modules form both a closed loop around data quality and a closed loop around state transition probability, thus creating a dual-closed-loop control structure. The processing procedures of each module are explained below in conjunction with a practical application scenario in the ward.

[0021] The multi-source data initialization module is used to acquire time-series data from various sources based on the current acquisition strategy parameters, perform quality assessments on each source to obtain quality assessment scores, and simultaneously acquire ward context labels.

[0022] As the data entry point for the entire system, the multi-source data initialization module first acquires multi-source time-series data from multiple data sources based on the current acquisition strategy parameters. Then, it performs quality assessments on the data from each source to obtain a quality assessment score, and simultaneously acquires ward context labels reflecting the patient's environment and treatment situation. The current acquisition strategy parameters specify the acquisition frequency, acquisition duration, and the set of data sources used within the current time window. These values ​​are not fixed but are dynamically adjusted by the probability trend grading and scheduling control module, allowing the acquisition behavior to change dynamically with the patient's condition. In a hospital setting, the multi-source time-series data specifically includes sleep duration and sleep structure data collected by non-contact monitoring devices or wearable devices in the ward; activity level and bed-leaning frequency data collected by activity sensors in the ward; acoustic characteristic data such as speech rate, fundamental frequency, and energy collected by audio pickup devices during daily communication and interviews; and structured scale scores and nursing observation records entered by nursing staff according to nursing procedures.

[0023] To adapt to the actual conditions of different wards, the aforementioned data sources can be added, subtracted, or replaced based on available conditions. For example, non-contact monitoring devices can be used instead of wearable devices to obtain sleep data, or other forms of structured nursing records can be used instead of scale scores; these are all alternative solutions. Furthermore, ward context labels are generated from scheduling records and event registrations in the ward information system. These labels are used to mark environmental events such as visits and outpatient examinations, as well as therapeutic events such as medication adjustments and psychotherapy, so that the individual adaptive baseline module can filter the corresponding time periods when establishing the baseline.

[0024] To ensure that the scheduling intent carried by the acquisition strategy parameters can be executed accurately, this module further analyzes the current acquisition strategy parameters and extracts the sensor sampling frequency, the enabled data source type, and the frequency of nursing data entry suggestions.

[0025] In this process, this module decomposes the acquisition strategy parameters into three specific control information categories, one for objective sensing devices and the other for nursing data entry: the sensor sampling frequency specifies the sampling density of objective sensing devices within the current time window; the enabled data source type indicates which data sources should be used for acquisition within the current time window; and the nursing data entry suggestion frequency provides recommendations on recording frequency to the nursing data entry process. Through this analysis, the abstract scheduling intent is transformed into executable acquisition control criteria, thus laying the foundation for subsequent data acquisition driven by category.

[0026] After parsing the control information above, this module triggers the corresponding objective sensing devices to acquire sleep source data, activity source data, and voice source data based on the sensing sampling frequency and the enabled data source type.

[0027] Specifically, this module determines the objective data sources to be collected within the current time window based on the enabled data source type, and controls the sampling density of the corresponding sensors according to the sensor sampling frequency. The objective sensors then collect sleep source data reflecting the patient's sleep status, activity source data reflecting the patient's activity status, and speech source data reflecting the patient's level of verbal arousal. Since all of the above data is automatically collected by the objective sensors without relying on the patient's active participation, it has good objectivity and continuity. At the same time, the sampling density of the objective sensors can be adjusted according to changes in the current collection strategy parameters, thereby achieving adaptive changes in the objective data collection density as the patient's state changes.

[0028] In conjunction with objective data collection, this module also obtains structured nursing observation data based on the frequency of nursing input suggestions.

[0029] Here, this module obtains structured nursing observation data entered by nursing staff from the nursing data entry process based on the suggested frequency of nursing data entry. This data is a structured record of patient performance observed by nursing staff. This type of data complements objective sensor data: on the one hand, it provides patient performance information that is difficult to directly reflect through objective sensor methods; on the other hand, it can supplement objective sensor data when it is missing or of poor quality. Furthermore, because the recording frequency of nursing observations can be coordinated and adjusted according to the current data acquisition strategy parameters, the intensity of manual observation is further adaptively adjusted according to the changing trends of patient status.

[0030] Considering that the data from the above sources are collected by different objective sensing devices or entered by nursing staff, and that the time and frequency of collection or entry are not consistent, this module further performs timestamp alignment and merging processing on sleep source data, activity source data, voice source data, and structured nursing observation entry data to generate multi-source time-series data.

[0031] In this process, this module aligns data from different sources to a unified time window and merges them based on the timestamps of each data point, thereby generating multi-source time-series data that correspond to each other in time. After timestamp alignment and merging, the time misalignment caused by inconsistent acquisition or entry times of the data from different sources is eliminated, thus providing a reliable data foundation for subsequent quality assessment, baseline establishment, and offset calculation on a unified time dimension. In specific implementation, timestamp alignment can adopt methods such as proximity alignment and interpolation alignment, all of which are replaceable technical means.

[0032] After obtaining the multi-source time series data, in order to measure the reliability of the data from each source, this module obtains the completeness index, timeliness index, and signal-to-noise ratio index of each multi-source time series data.

[0033] The three indicators described above characterize the reliability of data from different perspectives: The completeness indicator reflects the sufficiency of the actual valid samples collected within the current time window relative to the expected sample count for each multi-source time series data point. Its value is the ratio of the actual valid sample count to the expected sample count, indicating whether the source data is missing and the degree of this missing data. The timeliness indicator reflects the time interval between the most recent valid collection or entry of each multi-source time series data point; the shorter the interval, the higher the value, indicating the freshness of the source data in terms of time. The signal-to-noise ratio indicator reflects the proportion of valid signal relative to interference components in each multi-source time series data point, indicating the clarity of the source data itself. All three indicators can be directly calculated based on information such as the number of samples recorded during the collection process, the collection time, and the signal energy, thus ensuring their availability for practical application.

[0034] Based on the above indicators, this module calculates the quality assessment score of the multi-source time series data for each source by weighting and summing the completeness, timeliness and signal-to-noise ratio indicators according to the preset weight system.

[0035] The pre-defined weighting system specifies the proportion of the three indicators in the quality assessment score. The influence of each indicator can be adjusted based on the characteristics of different data sources. For example, for sleep data where continuous data collection is the primary focus, the weight of the completeness indicator can be appropriately increased; for nursing observation data with high timeliness requirements, the weight of the timeliness indicator can be appropriately increased. One possible calculation method is as follows:

[0036] ;

[0037] In the formula, Indicates the first Road data source within the time window Quality assessment score; , , These represent the completeness index, timeliness index, and signal-to-noise ratio index of the source, respectively. , , These represent the weight coefficients of the three corresponding indicators, and they satisfy the normalization condition. The design logic of this formula is to linearly aggregate multiple aspects representing the credibility of data into a single quantitative score according to their respective importance. The effect is to allow for a unified measurement and comparison of the credibility of data from different sources, thus providing a comparable quantitative basis for the subsequent dynamic determination of the fusion weights of each source. In practical implementation, in addition to using the weighted summation method described above, the indicators can also be combined after nonlinear mapping, which is an alternative technical approach.

[0038] The individual adaptive baseline module is used to filter multi-source time series data by combining ward context labels, establish an individual dynamic baseline by combining historical filtered data, and calculate the offset feature vector of multi-source time series data relative to the individual dynamic baseline to extract cross-source time series structure and generate time series structure feature vector.

[0039] The individual adaptive baseline module receives multi-source time-series data and ward context labels from the multi-source data initialization module. It establishes a dynamic individual baseline for each patient that is adaptively updated over time. Based on this baseline, it characterizes the degree of deviation of the patient's current state from their normal state and the temporal correlation structure between data from various sources. Considering the significant individual differences in patients' normal levels of sleep, activity, and speech, a uniform fixed baseline would be insufficient to accurately reflect the true changes in a specific patient. Therefore, this module replaces the fixed baseline with an individual dynamic baseline and combines it with context labels for time-period filtering. This ensures that subsequent identification more closely reflects the patient's actual state changes and effectively reduces interference introduced by individual differences and contextual factors.

[0040] To this end, this module first parses the ward context labels to identify environmental event time periods and therapeutic event time periods.

[0041] In this process, this module categorizes the events marked in the ward context labels into two types: environmental event time periods corresponding to time periods caused by changes in the patient's external environment, such as visits or outings for examinations; and therapeutic event time periods corresponding to time periods caused by treatment measures applied to the patient, such as medication adjustments or psychotherapy. By distinguishing between these two types of event time periods, it becomes clear which time periods of multi-source time-series data may be affected by treatment or environmental factors, thus providing a basis for subsequent time period filtering.

[0042] After identifying the aforementioned event periods, this module removes data segments corresponding to environmental event periods and therapeutic event periods from the multi-source time-series data to generate clean sliding statistics.

[0043] The reason for this exclusion is that multi-source time-series data occurring during environmental or therapeutic events are often subject to short-term influences from treatment measures or changes in the external environment, and therefore do not accurately reflect the patient's normal level. Thus, this module retains only data unaffected by the corresponding events to form clean sliding statistics, thereby preventing short-term fluctuations caused by treatment or environmental factors from being mistakenly included in the baseline, and providing a reliable data foundation for establishing accurate individual dynamic baselines.

[0044] After obtaining the clean sliding statistics, this module updates and establishes the individual dynamic baseline based on the clean sliding statistics. It should be noted that the clean sliding statistics here are the historical filtered data that has been continuously accumulated over time after the above-mentioned time period filtering process. The two refer to the same object, and the individual dynamic baseline is established based on this historical filtered data and updated as it is updated.

[0045] Because a patient's normal state may change slowly as the disease progresses and treatment advances, a fixed baseline is difficult to consistently reflect their true normal state. Therefore, this module updates the individual dynamic baseline accordingly with the continuous updating of pure sliding statistical data, thus dynamically following changes in the patient's normal state. This established individual dynamic baseline provides a personalized and time-adaptive benchmark for subsequent deviation measurements, ensuring that deviation measurements are always referenced to a benchmark that closely reflects the patient's current normal state. In practical implementation, the updating and establishment of the individual dynamic baseline can employ methods such as sliding window statistics and exponentially weighted moving averages, all of which are replaceable technical approaches.

[0046] After establishing the individual dynamic baseline, this module calculates the difference between the multi-source time series data and the individual dynamic baseline, extracts the offset, rate of change and trend slope relative to the individual dynamic baseline, and combines them to generate an offset feature vector.

[0047] Among them, the offset is used to characterize the degree of deviation of the current multi-source time-series data from the individual dynamic baseline; the rate of change is used to characterize how quickly the deviation changes over time; and the trend slope is used to characterize the overall trend of the deviation over a period of time. These three factors jointly depict the deviation of the patient's state from their normal state and its development trend from different perspectives. One possible method for calculating the offset is as follows:

[0048] ;

[0049] In the formula, Indicates the first Road data source within the time window Offset relative to the individual dynamic baseline; This indicates the current value of the source within this time window; This represents the individual dynamic baseline value corresponding to the source. The formula uses a relative offset to characterize the degree of deviation. Its design logic is to eliminate the differences in dimensions and benchmark levels between different patients and different sources. Its effect is to make the degree of deviation from each source comparable, thereby comprehensively depicting the deviation of the patient's state from multiple perspectives such as value deviation, rate of change, and overall trend under an individualized benchmark.

[0050] After obtaining the offset feature vectors of each source, in order to further utilize the various temporal relationships, this module calculates the lead delay and lag delay between the offset feature vectors of different sources.

[0051] Specifically, this module analyzes the temporal sequence of offset features from different sources, calculating the lead time (earlier occurrence) and lag time (later occurrence) of an offset from one source relative to an offset from another. This approach is necessary because various manifestations during patient state transitions often do not occur simultaneously but rather in a specific sequence; for example, decreased sleep often precedes increased activity. Therefore, quantitatively characterizing this sequence helps to grasp the evolutionary pattern of state transitions. In practice, the lead and lag times can be obtained by performing cross-correlation analysis on offset sequences from different sources and taking the time shift corresponding to the maximum cross-correlation value. Alternatively, other time delay estimation methods can be used; all are replaceable techniques.

[0052] After obtaining the time delay relationship, this module extracts the precursor chain pattern data by matching the preset cascading trigger pattern library based on the lead delay and lag delay.

[0053] The pre-stored cascading trigger pattern library contains several cascading trigger patterns reflecting the sequential triggering relationship of various source offsets in time. These patterns describe a typical evolutionary pattern where one type of source offset appears first, subsequently triggering another type of source offset. This pattern library is pre-established based on statistical analysis of historical multi-source time-series data and can be continuously updated as new data accumulates. This module matches the current sequential relationship of each source offset, characterized by lead and lag delays, with the patterns stored in this library, extracting precursor chain pattern data that matches the current time-series relationship. The extracted precursor chain pattern data reflects the sequential evolutionary chain relationship of various manifestations in the current multi-source time-series data that may indicate state transitions. Its effect is to capture precursory evolutionary signs in the early stages of a transition trend. In specific implementation, matching can employ pattern comparison, similarity calculation, etc., all of which are replaceable technical means.

[0054] Based on this, this module performs time-series encoding and splicing on the leading delay, lagging delay, and precursor chain pattern data to generate a time-series structure feature vector containing time-series causal relationships.

[0055] Among them, the precursor chain pattern data can be represented as data consisting of the pattern identifier of the matched cascaded triggering pattern and its corresponding matching degree. The pattern identifier is used to indicate the matched evolution pattern, and the matching degree is used to characterize the degree of consistency between the current temporal relationship and the pattern. The temporal coding and splicing process specifically involves splicing the offset features of each source, the encoded lead delay and lag delay, and the precursor chain pattern data into a one-dimensional feature vector in a preset dimensional order, and arranging the feature components of each source in a unified order and dimension to generate a temporal structure feature vector.

[0056] In this process, this module characterizes the lead and lag delays of the sequential relationships between each source offset, as well as the precursor chain pattern data reflecting the sequential evolution of each source offset. These are then concatenated into a unified feature representation after temporal encoding. The resulting temporal structure feature vector not only contains information about each source offset itself, but also information about the temporal sequence and causal relationships between them. This allows subsequent fusion and recognition to utilize the temporal causal relationships between the source offsets, capturing signs of state transition trends in their early stages based on the evolutionary order of various manifestations, thereby improving the timeliness of recognition. In practical implementation, temporal encoding can employ positional encoding, temporal embedding, etc., all of which are replaceable techniques.

[0057] The dynamic confidence fusion module is used to perform dynamic confidence weighted fusion of the temporal structure feature vectors of each source based on the quality assessment score, generate a comprehensive feature vector, and calculate the corresponding comprehensive confidence.

[0058] The dynamic confidence fusion module takes the time-series structural feature vectors generated by the individual adaptive baseline module and the quality assessment scores obtained by the multi-source data initialization module. It weights and fuses the features from multiple sources according to the current confidence level of each source to obtain a comprehensive feature vector, and simultaneously provides the comprehensive confidence level of the fusion result. Unlike fusion methods that use fixed weights, this module dynamically determines the fusion weights of each source based on the quality assessment scores, ensuring that higher-quality sources occupy a larger proportion in the fusion result. This maintains the stability and reliability of the fusion result even under fluctuating data quality conditions.

[0059] To ensure the historical reliability of each source, this module first obtains preset historical contribution data, which is used to characterize the effective weight of the corresponding source in historical state classification decisions.

[0060] Historical contribution data reflects the extent to which each source has contributed to obtaining correct results in past status identification processes; that is, the more reliably a source has historically reflected changes in patient status, the higher its value. In terms of acquisition method, this data is pre-determined based on statistical analysis of the consistency between the characteristics of each source during historical identification processes and the final results confirmed by medical staff. It can be updated as the identification process progresses, thus allowing the historical reliability of each source to be incorporated into the determination of subsequent weights.

[0061] Specifically, the historical contribution data is updated based on the verification and confirmation results of the identification results by medical staff. That is, after the patient's state has indeed changed, the consistency between the features provided by each source in this identification and the confirmation result is statistically analyzed, and the historical contribution data of the corresponding source is adjusted accordingly. This allows the source that has been more reliable in reflecting the real state change in the past to obtain a higher value, thus forming a closed loop of verification results flowing back to correct the reliability of each source.

[0062] After obtaining historical contribution data, this module performs mapping calculations based on the quality assessment scores and historical contribution data, and normalizes the calculation results to obtain the dynamic fusion weight of each source.

[0063] In this process, this module combines the quality assessment score reflecting real-time quality with historical contribution data reflecting historical reliability, giving greater weight to sources with higher real-time quality and historical reliability. Then, normalization is applied to sum the weights of all sources to one. The following calculation methods can be used:

[0064] ;

[0065] In the formula, Indicates the first Road data source within the time window Dynamic fusion weights, This indicates the quality assessment score of the source. This represents the historical contribution data of the source, with the denominator being the sum of the products of the quality assessment scores of each source and the historical contribution data, used to normalize the weights. The design logic of this formula is to simultaneously consider the real-time quality and historical reliability of each source in a product form. Its effect is to make the fusion result more reliably based on currently trustworthy and historically valid sources, thus making the fusion result more reasonable and reliable when the reliability of different sources is inconsistent. In practical implementation, in addition to the above multiplication method, other mapping methods that can simultaneously reflect real-time quality and historical reliability can also be used for mapping calculation, making them alternative technical means.

[0066] Furthermore, to address situations where the quality of individual sources is too low or missing, this module, upon detecting that the quality assessment score of any multi-source time series data is lower than a preset lower limit, forcibly decays its corresponding dynamic fusion weight to zero, and then uses normalization processing to dynamically compensate for the missing weights from other sources.

[0067] A preset lower limit value is used to define the minimum level of credibility at which source data can participate in fusion. The value directly affects the strictness of source rejection: a higher value indicates a stricter requirement for source credibility, increasing the likelihood of sources being forcibly decayed to zero; a lower value indicates a higher tolerance, reducing the likelihood of rejection. When a source is decayed to zero, since the sum of the weights remains one after normalization, the decayed weights are distributed among the remaining sources after normalization, thus achieving dynamic compensation for missing weights.

[0068] In hospital settings, when a shift's nursing observation record has not yet been entered, or when a patient leaves the hospital causing a missing objective source of data, this process ensures that the fusion results will not be distorted due to the low quality or absence of individual sources. Even when some sources of data are missing, stable and reliable fusion results can still be obtained based on other credible sources.

[0069] After determining the weights of each source, this module performs a weighted summation based on the temporal structure feature vectors of each source and their corresponding dynamic fusion weights to generate a comprehensive feature vector. Then, it uses the sum of the products of the quality assessment scores of each source participating in the fusion and the dynamic fusion weights to generate a comprehensive confidence score.

[0070] The generation of the comprehensive feature vector involves weighting and superimposing the temporal structure feature vectors from each source according to their respective dynamic fusion weights, so that sources with higher credibility occupy a larger proportion in the comprehensive feature vector. The calculation method can be as follows:

[0071] ;

[0072] In the formula, Indicates time window The comprehensive feature vector, Indicates the first Dynamic fusion weights of road data sources; Indicates the first The temporal structure feature vector of the road data source.

[0073] It should be noted that before participating in the weighted summation, the temporal structure feature vectors from each source are first processed by a unified feature representation to ensure that the temporal structure feature vectors from each source have a consistent dimension. For example, the same type and number of feature components are extracted from each source, or they are aligned to a unified dimension through mapping transformation, thereby ensuring that the above weighted summation operation can be performed according to the corresponding components.

[0074] Based on the generated comprehensive feature vector, this module further generates a comprehensive confidence score by summing the products of the quality assessment scores of each source involved in the fusion and the dynamic fusion weights. The calculation method can be as follows:

[0075] ;

[0076] In the formula, Indicates time window The overall confidence level, Indicates the first Dynamic fusion weights of road data sources This represents the quality assessment score of the source. The design logic of this formula is to sum the quality assessment scores of each source according to its proportion in the fusion, thereby aggregating the quality of each source participating in the fusion into a unified measure of the overall credibility of the fusion result. Its effect is to enable the generated comprehensive confidence level to objectively reflect the overall credibility of this fusion result, and thus serve as the basis for subsequent judgment on whether data backfilling is needed.

[0077] The data recovery and control module is used to generate a data recovery request command and send it to the nurse station terminal connected to the communication connection when the overall confidence level is detected to be less than a preset threshold, so as to obtain the recovered multi-source time series data and regenerate the overall feature vector based on the recovered multi-source time series data.

[0078] The data recovery and control module gates the overall reliability of the fusion results and proactively triggers data recovery when the reliability is insufficient, thus forming a closed-loop control around data quality. Specifically, this module compares the overall confidence level with a preset threshold. When the overall confidence level is detected to be less than the preset threshold, it indicates that the overall quality of the multi-source time-series data currently involved in the fusion is insufficient to support reliable identification. At this time, this module does not force the subsequent state switching probability calculation based on low-confidence data. Instead, it generates a data recovery request instruction and sends it to the nurse station terminal connected to the communication link to prompt nursing staff to supplement the entry of relevant scale scores or nursing observation records, check the patient's data acquisition device wearing status, and then obtain the recovered multi-source time-series data. Based on the recovered multi-source time-series data, it regenerates the comprehensive feature vector until the overall confidence level reaches the required level.

[0079] To avoid prolonged processing delays due to nursing data entry congestion, the data recovery control module includes a recovery termination condition. Specifically, when the cumulative number of data recovery attempts reaches a preset limit, or when the waiting time since the triggering of the data recovery request reaches a preset timeout period, the module stops waiting for data recovery and instead generates a comprehensive feature vector based on currently available multi-source time-series data in a downgraded manner. The output result is then marked with a flag indicating insufficient data reliability for later reference. This ensures both the required data reliability and prevents the processing from entering an infinite waiting state.

[0080] The preset threshold here serves as a quality gate, and its value directly determines the system's requirements for data reliability: the higher the threshold, the stricter the requirements for the quality of the data involved in the identification, and the more likely the data will be triggered to be backfilled; the lower the threshold, the higher the tolerance, and the less likely the data will be backfilled. In practical applications, the threshold can be set according to the trade-off between the reliability of identification and the workload of nursing staff.

[0081] By transforming the judgment of insufficient data quality into a data replenishment request for the nurse station terminal, and allowing the replenished data to re-enter the fusion process, this module forms a closed loop in which the data processing stage reversely controls the nursing data entry behavior. Its effect is that it can proactively replenish data when it is missing or of poor quality, thereby ensuring that the comprehensive feature vector entering the subsequent identification stage is always based on a sufficiently reliable data foundation.

[0082] In practice, the data recovery request command can be sent to terminals other than the nurse station terminal, or other terminals in the ward that perform data entry or equipment management functions, making it a replaceable technical means.

[0083] The state transition probability calculation module is used to calculate the state transition probability sequence by inputting the comprehensive feature vector into the preset time series probability model when the comprehensive confidence level is not less than the preset threshold.

[0084] The state switching probability calculation module corresponds to the backfilling and control module. The two modules form a complete branching process based on the comparison between the comprehensive confidence level and a preset threshold: when the comprehensive confidence level is detected to be no less than the preset threshold, it indicates that the current comprehensive feature vector has been established on a sufficiently reliable data foundation. At this point, this module inputs the comprehensive feature vector into a preset time-series probability model to calculate the state switching probability sequence. Since this calculation is only performed when the comprehensive confidence level meets the requirements, the results are always based on reliable data, thereby effectively improving the reliability of the recognition results.

[0085] To enable the model to utilize the continuous changes in the comprehensive features over time, this module first parses the comprehensive feature vector into a time series feature matrix according to the time window.

[0086] In this process, this module organizes the comprehensive feature vectors corresponding to different time windows according to their chronological order, and parses them into a time series feature matrix arranged sequentially in the time dimension, so that the comprehensive features of each time window can be presented in an orderly manner, thereby providing an orderly input for the subsequent time series modeling.

[0087] After obtaining the time series feature matrix, this module inputs the time series feature matrix into a pre-trained hidden Markov model or a neural network model with a temporal structure.

[0088] Hidden Markov Models (HMMs) can characterize the migration patterns of patient states between different categories through state transitions, while neural network models with temporal structures can capture the evolution of comprehensive features over time through their ability to model temporal dependencies. Both can take into account temporal dependencies when processing time-series feature matrices. Regarding training methods, the models are pre-trained using time-series feature matrices obtained from processed historical multi-source time-series data labeled with state categories. The HMM estimates parameters such as state transition probabilities and observation probabilities through statistical analysis of the labeled data, while the neural network models with temporal structures iteratively optimize and adjust their network parameters based on the labeled data. This allows the trained models to characterize patient state migrations based on the time-series feature matrix. These two models are interchangeable techniques; in practical applications, one can be chosen based on data characteristics and recognition needs, or other probabilistic models with temporal modeling capabilities can be used as alternatives.

[0089] Based on the model processing, this module calculates and outputs the continuous probability distribution sequence of the current state category to the target state category through a hidden Markov model or a neural network model.

[0090] Specifically, based on the input time-series feature matrix, the model calculates the probability of the patient transitioning from their current state to the target state, and outputs this probability as a continuously ordered probability distribution sequence over time. This continuous probability distribution sequence is the state transition probability sequence. The resulting probability sequence not only reflects the likelihood of the patient transitioning to their current state but also reflects the continuous change of that likelihood over time. This allows subsequent decisions to be made based on this trend rather than on the value at a single isolated moment, thus providing a more robust grasp of the development of the state transition.

[0091] The probability trend classification and scheduling reverse control module is used to make trend classification decisions based on the state switching probability sequence, generate classification identifiers to generate updated acquisition strategy parameters, and use the updated acquisition strategy parameters as the current acquisition strategy parameters for the next time window to schedule multi-source data acquisition actions.

[0092] The probability trend classification and scheduling counter-control module classifies and determines the development trend of state transitions, and adjusts data collection behavior accordingly, thus forming another closed-loop control around the state transition probability. Unlike methods that judge solely based on whether a single probability value exceeds a fixed threshold, this module comprehensively considers factors such as the current value of the state transition probability, its upward trend over time, and the duration of its sustained high level to make trend classification judgments. This more robustly reflects the development degree of the transition trend and reduces repeated judgments caused by short-term probability fluctuations.

[0093] The closed loop implemented in this module, which uses the reverse scheduling of data acquisition and nursing observation intensity based on state switching probability, works in conjunction with the closed loop implemented by the supplementary control module, which focuses on data quality. Together, they form the dual closed-loop control structure of this system, thereby coordinating the monitoring intensity with the patient's state switching trend under the condition of limited nursing resources.

[0094] To this end, this module first extracts the current probability value, the probability rise slope, and the duration of exceeding the preset probability threshold of the state switching probability sequence.

[0095] The three aspects of information described above depict different facets of the switching trend: the current probability value reflects the likelihood of the patient's current state transition; the probability rise slope reflects how quickly this likelihood increases over time; and the duration exceeding the preset probability threshold reflects how long the state switching probability remains at a high level. The preset probability threshold defines whether the state switching probability is already at a high level; the longer the duration exceeding this threshold, the stronger the persistence of the switching trend. By simultaneously extracting information from these three aspects, this module can comprehensively depict the development degree of the state switching trend from multiple perspectives, such as the value level, the rate of rise, and the duration, thus avoiding the one-sidedness caused by judging solely based on the probability value at a single isolated moment.

[0096] The current probability value is taken from the probability value corresponding to the current time window in the state transition probability sequence; the probability rise slope can be obtained by linearly fitting the probability values ​​of the most recent time windows in the state transition probability sequence and using the slope of the fitted line as the probability rise slope, or by using the average of the probability differences between adjacent time windows as the probability rise slope, all of which are alternative technical means; the duration of exceeding the preset probability threshold is the number of time windows in the state transition probability sequence where the probability value is continuously not lower than the preset probability threshold.

[0097] After extracting the above information, this module inputs the current probability value, probability rise slope, and duration into a preset hierarchical decision logic tree for joint decision processing.

[0098] The preset hierarchical judgment logic tree contains pre-defined logical rules for hierarchical judgment based on the above three aspects of information. These rules are pre-defined based on historical data analysis and clinical management experience and can be adjusted according to actual application.

[0099] As an illustrative judgment logic, the graded judgment logic tree can be set as follows: when the current probability value is lower than the preset probability threshold and the probability rise slope is not positive, it is judged as normal; when the current probability value reaches the preset probability threshold, but the probability rise slope is small and the duration is short, it is judged as attention level; when the current probability value reaches the preset probability threshold and the probability rise slope is large or the duration is long, it is judged as warning level; when the current probability value is significantly higher than the preset probability threshold, and the probability rise slope is large and the duration is further prolonged, it is judged as emergency level. The comparison benchmarks for the above criteria can be preset and adjusted based on historical data analysis and clinical management experience. Those skilled in the art can use this to achieve a joint judgment from three criteria to four levels.

[0100] This module takes the current probability value, the probability rise rate, and the duration as input, and makes a joint judgment based on the rules set by the logic tree, rather than making a judgment based on any single piece of information in isolation. This approach allows the hierarchical decision-making to comprehensively consider the value, rate of rise, and duration of the state transition probability, thus preventing repetitive judgments due to instantaneous changes in a single piece of information when the probability fluctuates briefly. The effect is to improve the robustness of the hierarchical decision-making and reduce misjudgments. In practical implementation, in addition to using a hierarchical decision-making logic tree, rule-based scoring, fuzzy decision-making, and other methods can also be used to jointly determine the above three aspects of information; all are alternative technical means.

[0101] After completing the joint judgment, this module outputs a judgment result, including normal level, attention level, warning level or emergency level, based on the judgment processing result, and encapsulates the judgment result into a classification identifier.

[0102] In this process, this module categorizes the development level of the state transition trend into one of four levels—Normal, Attention, Warning, or Emergency—based on the results of the joint judgment. Each level reflects the increasing degree of development of the state transition trend. The judgment result is then encapsulated as a hierarchical identifier. This clearly defines the development level of the state transition trend in a hierarchical manner, enabling subsequent scheduling of data collection and observation intensity based on different levels. In practice, the number of levels assigned by the judgment result can be increased or decreased according to actual management needs, making it a replaceable technical approach.

[0103] After generating the grading identifier, this module generates corresponding scheduling instructions for different grades. Specifically, when the grading identifier indicates a normal grade and is not a decline from a higher grade, the probability trend grading and scheduling counter-control module maintains the basic sensor sampling frequency and the frequency of routine nursing observation recommendations unchanged. That is, it continues to schedule multi-source data acquisition actions for the next time window using the basic acquisition strategy parameters, thus ensuring that all four grades—normal, attention, warning, and emergency—have corresponding acquisition and observation scheduling actions. When the grading identifier indicates a attention grade, this module generates a first scheduling instruction to increase the sensor sampling frequency and the frequency of nursing observation recommendations.

[0104] This indicates that the patient's condition has begun to show signs of improvement but has not yet reached a level requiring warning. Therefore, this module increases the sampling density of objective sensors and suggests increasing the frequency of observation and recording in the nursing data entry process. This allows for the confirmation of the condition change trend through more intensive objective data and manual observation. The effect of this approach is that the monitoring density can be increased appropriately and promptly when signs of a trend first appear to accurately grasp its subsequent changes, while avoiding premature and excessively intensive monitoring.

[0105] Furthermore, when the classification indicator is identified as a warning level or an emergency level, this module generates an encrypted acquisition activation command to enable all preset data sources and outputs a priority review signal to the medical terminal.

[0106] At this point, the patient's condition shift trend has become quite evident. Therefore, this module generates an encrypted data acquisition activation command to enable all preset data sources. This mobilizes all data sources and collects data at a density higher than the level of concern, thus providing sufficient and reliable confirmation of the condition shift trend. Simultaneously, a priority review signal is output to the medical terminal, ensuring that the corresponding identification results are prioritized for review by medical staff. The effect of this approach is to capture changes with the most intensive monitoring when the trend is significant, and to ensure that identification results reach medical staff promptly, thereby shortening the lag in detecting condition shifts. It should be noted that the priority review signal is output to the medical terminal as a data signal for medical staff reference; the final judgment and action are made by the medical staff.

[0107] Encrypted acquisition here refers to increasing the acquisition density, that is, further increasing the sensor sampling frequency and enabling all preset data sources on the basis of the increased acquisition frequency at the level of interest. Its acquisition intensity is higher than that corresponding to the acquisition intensity at the level of interest, and the two are progressive in acquisition intensity; encrypted acquisition is unrelated to data encryption.

[0108] In contrast, when the classification flag is detected to have fallen back to the normal level, this module generates a second scheduling instruction to restore the basic sensing frequency.

[0109] This indicates that the patient's condition has stabilized, so this module restores the sampling density of objective sensors and the corresponding observation recording intensity to normal levels. This process corresponds to the aforementioned process of increasing monitoring intensity at higher levels, and together they embody the scheduling logic of monitoring intensity rising and falling with the grading indicator. The effect is to reduce monitoring intensity when the condition is stable, thereby avoiding unnecessary monitoring input, rationally allocating nursing resources, and reducing unnecessary disturbance to the patient's rest.

[0110] Finally, this module performs parameter conversion processing on the first scheduling instruction, the encryption activation instruction, or the second scheduling instruction to generate updated acquisition strategy parameters.

[0111] In this process, this module converts the various scheduling instructions expressed in the form of scheduling intentions into acquisition strategy parameters that can be directly executed by the multi-source data initialization module. Specifically, it converts the adjustment intentions regarding sensor sampling frequency, enabled data source type, and nursing data entry suggestion frequency into corresponding specific values ​​in the acquisition strategy parameters. The generated updated acquisition strategy parameters are then sent back to the multi-source data initialization module as the current acquisition strategy parameters for the next time window. This ensures that the scheduling intentions expressed in the various scheduling instructions are implemented as actual acquisition actions, forming a closed loop where the recognition results control data acquisition in reverse. The effect is that the intensity of acquisition and observation can be adaptively adjusted according to the changes in the grading indicators as the patient's condition changes.

[0112] like Figure 2 The diagram shown is a flowchart illustrating the bipolar state switching identification method based on multi-source feature fusion provided in this application embodiment, including:

[0113] Based on the current acquisition strategy parameters, multi-source time-series data are acquired to perform quality assessments and obtain quality assessment scores, while ward context labels are acquired simultaneously.

[0114] By combining ward context labels, multi-source time series data are filtered for different time periods. Individual dynamic baselines are established by combining historical filtered data. The offset feature vectors of multi-source time series data relative to the individual dynamic baselines are calculated to extract cross-source time series structure and generate time series structure feature vectors.

[0115] Based on the quality assessment scores, the temporal structure feature vectors of each source are dynamically weighted and fused to generate a comprehensive feature vector, and the corresponding comprehensive confidence is calculated.

[0116] When the overall confidence level is detected to be less than a preset threshold, a data recovery request instruction is generated and sent to the nurse station terminal connected to the communication to obtain the recovered multi-source time series data, and the overall feature vector is regenerated based on the recovered multi-source time series data.

[0117] When the overall confidence level is detected to be no less than a preset threshold, the overall feature vector is input into a preset time-series probability model to calculate the state switching probability sequence.

[0118] Trend classification is determined based on the state switching probability sequence, and a classification identifier is generated to generate updated acquisition strategy parameters. The updated acquisition strategy parameters are then used as the current acquisition strategy parameters for the next time window to schedule multi-source data acquisition actions.

[0119] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above functions can be divided into different functional modules to complete all or part of the functions described above.

[0120] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0121] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units, located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0122] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0123] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the solution, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A manic-depressive state switching recognition system based on multi-source feature fusion, characterized in that, include: The multi-source data initialization module is used to acquire time-series data from various sources based on the current acquisition strategy parameters, perform quality assessments on each source to obtain quality assessment scores, and simultaneously acquire ward context labels. The individual adaptive baseline module is used to filter the multi-source time series data by combining the ward context labels, establish an individual dynamic baseline by combining historical filtered data, and calculate the offset feature vector of the multi-source time series data relative to the individual dynamic baseline to extract the cross-source time series structure and generate a time series structure feature vector. The dynamic confidence fusion module is used to perform dynamic confidence weighted fusion of the temporal structure feature vectors of each source based on the quality assessment scores, generate a comprehensive feature vector, and calculate the corresponding comprehensive confidence. The data recovery and control module is used to generate a data recovery request instruction and send it to the nurse station terminal connected to the communication when the overall confidence level is detected to be less than a preset threshold, so as to obtain the recovered multi-source time series data and regenerate the overall feature vector based on the recovered multi-source time series data. The state transition probability calculation module is used to input the comprehensive feature vector into a preset time-series probability model when the comprehensive confidence level is detected to be not less than the preset threshold, and calculate the state transition probability sequence. The probability trend classification and scheduling reverse control module is used to make trend classification decisions based on the state switching probability sequence, generate classification identifiers to generate updated acquisition strategy parameters in reverse, and use the updated acquisition strategy parameters as the current acquisition strategy parameters for the next time window to schedule multi-source data acquisition actions.

2. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 1, characterized in that: When the multi-source data initialization module acquires time-series data from various sources based on the current acquisition strategy parameters, it is specifically used for: The current acquisition strategy parameters are analyzed to extract the sensor sampling frequency, the type of data source enabled, and the frequency of nursing data entry suggestions. Based on the sensor sampling frequency and the enabled data source type, the corresponding objective sensor device is triggered to acquire sleep source data, activity source data and voice source data; Based on the frequency of the nursing entry suggestions, structured nursing observation entry data is obtained; The sleep source data, activity source data, voice source data, and structured nursing observation data are timestamped and merged to generate the multi-source time-series data.

3. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 1, characterized in that: When the multi-source data initialization module performs quality assessments to obtain quality assessment scores, it is specifically used for: Obtain the completeness index, timeliness index, and signal-to-noise ratio index of each of the multi-source time-series data; Based on a preset weighting system, the completeness index, the timeliness index, and the signal-to-noise ratio index are weighted and summed to calculate the quality assessment score of the multi-source time series data corresponding to each source.

4. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 1, characterized in that: The individual adaptive baseline module, when filtering the multi-source time-series data by combining the ward context labels, establishing an individual dynamic baseline by combining historical filtered data, and calculating the offset feature vector of the multi-source time-series data relative to the individual dynamic baseline, is specifically used for: The ward context labels were analyzed to identify environmental event time periods and therapeutic event time periods; Data segments corresponding to the environmental event time period data and the therapeutic event time period data are removed from the multi-source time series data to generate clean sliding statistics data; The individual dynamic baseline is updated and established based on the pure sliding statistics. The difference between the multi-source time-series data and the individual dynamic baseline is calculated, and the offset, rate of change and trend slope relative to the individual dynamic baseline are extracted and combined to generate the offset feature vector.

5. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 1, characterized in that: The individual adaptive baseline module, when extracting cross-source temporal structure and generating temporal structure feature vectors, is specifically used for: Calculate the lead time and lag time between the offset feature vectors of different sources; Based on the lead delay and the lag delay, a preset cascaded triggering mode library is matched to extract the predecessor chain mode data; The leading delay, the lagging delay, and the precursor chain pattern data are time-series encoded and concatenated to generate the time-series structure feature vector containing time-series causal relationships.

6. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 1, characterized in that: The dynamic confidence fusion module is specifically used for: Obtain preset historical contribution data, which is used to characterize the effective proportion of the corresponding source in historical state classification decisions; Mapping calculations are performed based on the quality assessment scores and the historical contribution data, and the calculation results are normalized to obtain the dynamic fusion weight of each source. If the quality assessment score of any of the multi-source time series data is detected to be lower than the preset lower limit value, the corresponding dynamic fusion weight is forcibly decayed to zero, and the missing weight is dynamically compensated by other sources through the normalization process. The comprehensive feature vector is generated by weighted summation of the temporal structure feature vectors of each source and their corresponding dynamic fusion weights, and the comprehensive confidence score is generated by summing the products of the quality assessment scores of each source participating in the fusion and the dynamic fusion weights.

7. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 1, characterized in that: The state transition probability calculation module is specifically used for: The comprehensive feature vector is parsed into a time series feature matrix according to the time window; The time series feature matrix is ​​input into a pre-trained hidden Markov model or a neural network model with a temporal structure; The continuous probability distribution sequence of the current state category to the target state category is calculated using the Hidden Markov Model or the Neural Network Model and then output.

8. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 1, characterized in that: When the probability trend classification and scheduling counter-control module performs trend classification judgment based on the state switching probability sequence and generates classification identifiers, it is specifically used for: Extract the current probability value, probability rise slope, and duration exceeding the preset probability threshold of the state transition probability sequence; The current probability value, the probability rise slope, and the duration are input into a preset hierarchical decision logic tree for joint determination processing; Based on the judgment and processing results, output a judgment result including normal level, attention level, warning level or emergency level, and encapsulate the judgment result as the classification identifier.

9. The manic-depressive state switching recognition system based on multi-source feature fusion as described in claim 8, characterized in that: The probability trend classification and scheduling reverse control module, when generating updated acquisition strategy parameters in reverse, is specifically used for: When the classification indicator is identified as a level of concern, a first scheduling instruction is generated to increase the sensor sampling frequency and the frequency of nursing observation recommendations. When the classification indicator is identified as a warning level or an emergency level, an encrypted collection activation command is generated to enable all preset data sources, and a priority review signal is output to the medical terminal. When the classification flag is detected to have fallen back to the normal level, a second scheduling instruction to restore the basic sensing frequency is generated; The first scheduling instruction, the encryption activation instruction, or the second scheduling instruction are processed by parameter conversion to generate the updated acquisition strategy parameters.

10. A method for identifying manic-depressive state switching based on multi-source feature fusion, characterized in that, Includes the following steps: Based on the current acquisition strategy parameters, acquire multi-source time-series data to perform quality assessments and obtain quality assessment scores, and simultaneously acquire ward context labels; The multi-source time series data is filtered by time period based on the ward context labels, and an individual dynamic baseline is established by combining historical filtered data. The offset feature vector of the multi-source time series data relative to the individual dynamic baseline is calculated to extract the cross-source time series structure and generate a time series structure feature vector. Based on the quality assessment scores, the temporal structure feature vectors of each source are dynamically weighted and fused with confidence to generate a comprehensive feature vector, and the corresponding comprehensive confidence is calculated. When the overall confidence level is detected to be less than a preset threshold, a data recovery request instruction is generated and sent to the nurse station terminal with the communication connection to obtain the recovered multi-source time series data, and the overall feature vector is regenerated based on the recovered multi-source time series data. When the overall confidence level is detected to be not less than the preset threshold, the overall feature vector is input into the preset time-series probability model to calculate the state switching probability sequence. Based on the state switching probability sequence, a trend classification decision is made, and a classification identifier is generated to generate updated acquisition strategy parameters in reverse. The updated acquisition strategy parameters are used as the current acquisition strategy parameters for the next time window to schedule multi-source data acquisition actions.

Images