An oral clinic nursing adverse event early warning method and system based on an LSTM time sequence model

Abstract

The application discloses an oral clinic nursing adverse event early warning method and system based on an LSTM time sequence model, relates to the technical field of intelligent nursing early warning systems, and comprises the following steps: collecting continuous multi-modal physiological signals output by a wearable device during the diagnosis and treatment of oral clinic patients, time-stamped operation events recorded by a clinical terminal, and preoperative static clinical characteristics, and forming a multi-source heterogeneous original data stream; a dynamic risk assessment system is constructed by fusing the multi-modal physiological signals, the operation events and the static clinical characteristics, taking the individual baseline of the patient as a reference; sample entropy is used to capture early physiological instability signals, the current state deviation degree is quantified by means of Mahalanobis distance, and a double-channel LSTM model is used to realize collaborative modeling of group rules and individual differences, thereby providing high-sensitivity and high-specificity graded early warning before the occurrence of adverse events; and an attention mechanism is introduced to locate the risk source, and an interpretable and executable intervention instruction is generated. The application can early warn oral clinic nursing adverse events.

Description

Technical Field

[0001] This invention relates to the field of intelligent nursing early warning system technology, and in particular to a method and system for early warning of adverse events in dental clinic nursing based on LSTM time series model. Background Technology

[0002] Intelligent nursing early warning system technology is a technical system that integrates multi-source physiological monitoring data, clinical operation information, and individual patient characteristics, and uses artificial intelligence algorithms to analyze, dynamically assess, and provide graded early warnings for potential nursing risks in real time. This technology constructs a patient-centered, continuous perception-intelligent judgment-proactive intervention closed loop, enabling it to identify early signs before adverse events occur, provide interpretable risk alerts, and automatically trigger standardized response measures in conjunction with medical terminals or medical devices. This improves nursing safety and clinical response efficiency, and is widely used in high-risk medical scenarios such as operating rooms, emergency rooms, intensive care units, and specialist outpatient clinics.

[0003] Traditional dental clinic care relies on manual observation or alarms based on single vital sign thresholds. Often, these alarms are only detected after patients have developed obvious symptoms. This lacks the ability to sensitively capture early physiological instability, making it difficult to achieve true pre-warning. Furthermore, existing warning models are mostly based on population statistical thresholds, ignoring the differences in patients' individual physiological baselines. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an early warning method for adverse events in dental clinic care based on an LSTM time series model. This solves the problem that traditional dental clinic care relies on manual observation or single vital sign threshold alarms, which often only detect adverse events after patients have developed obvious symptoms. This lack of sensitivity in capturing early physiological instability makes it difficult to achieve true pre-event warning. Furthermore, existing warning models are mostly based on population statistical thresholds, ignoring differences in patients' individual physiological baselines. This invention also provides an early warning system for adverse events in dental clinic care based on an LSTM time series model, using the aforementioned method.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for early warning of adverse events in dental outpatient care based on an LSTM time series model, comprising:

[0008] The system collects continuous multimodal physiological signals output by wearable devices during the diagnosis and treatment of patients in the dental clinic, timestamped operation events recorded by clinical terminals, and static clinical features entered before the operation, forming a multi-source heterogeneous raw data stream.

[0009] A unified timeline is constructed based on the time when the patient enters the consultation room. Cross-modal temporal alignment processing is performed on the multi-source heterogeneous raw data stream. Sliding window segmentation of physiological signals and event-driven interpolation operation are adopted. Missing mask is generated to mark unreliable signal segments, and a temporal feature matrix is ​​output.

[0010] Based on the physiological signal segmentation in the time-series feature matrix, the physiological signal complexity index under the sliding time window is calculated, and the physiological signal complexity index is fused with the original physiological signal, operation event and static clinical features to form a fused feature vector.

[0011] The fused feature vector is input into the dual-channel LSTM risk prediction model. The absolute risk probability is output through the general LSTM sub-model. The dynamic deviation of the current state from the individual baseline is calculated by combining the patient's personal health baseline file. The two are then fused to obtain a personalized dynamic risk score.

[0012] The risk warning level is determined based on the personalized dynamic risk score and its change slope. When the preset intervention threshold is reached, the key risk-causing feature combination is located through the attention mechanism, and the warning information is generated and pushed to the nursing terminal.

[0013] As a preferred embodiment of the LSTM time-series model-based early warning method for adverse events in dental outpatient care according to the present invention, the specific steps for acquiring continuous multimodal physiological signals, operational events, and static clinical features are as follows:

[0014] Wearable devices are used to collect patients' heart rate, respiratory rate, blood oxygen saturation, skin conductance, and body movement intensity signals in real time during the diagnosis and treatment process, forming the first signal set;

[0015] The second event sequence is formed by recording the start time of anesthetic injection, the activation time of high-speed turbine handpiece, the start and stop time of saliva suction device, and the time of patient position change through dental chair integrated system or nurse handheld terminal;

[0016] The electronic medical record system is used to obtain the patient's age, gender, history of syncope, history of allergies, and number of underlying diseases before surgery, forming a third feature vector.

[0017] The first signal set, the second event sequence, and the third feature vector are bound together according to the patient's identity to form a multi-source heterogeneous raw data stream.

[0018] As a preferred embodiment of the LSTM-based time-series model-based early warning method for adverse events in dental clinics according to the present invention, the cross-modal time-series alignment processing specifically includes the following steps:

[0019] A high-resolution time axis with a uniform sampling frequency is established, with the moment the patient enters the examination room as the zero point.

[0020] Each physiological signal in the first signal set is segmented using a sliding window of fixed length to obtain a discrete physiological segment sequence;

[0021] For each event timestamp in the second event sequence, a preset physiological response duration template is matched according to the event type to generate a unit pulse extension signal covering the corresponding time period, and interpolated as a pseudo-continuous event signal on a unified time axis;

[0022] For invalid segments in physiological signals caused by motion artifacts or device detachment, generate a binary mask sequence of the same dimension;

[0023] All segmented physiological signals, interpolated event signals, and mask sequences are concatenated along the time axis to output a time-series feature matrix.

[0024] As a preferred embodiment of the LSTM time-series model-based early warning method for adverse events in dental clinics according to the present invention, the calculation of the physiological signal complexity index specifically includes the following steps:

[0025] For any physiological signal segment in the time-series feature matrix, construct an embedding dimension of... The sequence of embedded vectors;

[0026] Set similarity tolerance This is a fixed proportion of the standard deviation of the signal segments;

[0027] Calculate the Chebyshev distance between any two distinct embedding vectors;

[0028] Statistical satisfaction of distance less than The number of vector pairs, denoted as And calculate their logarithmic mean to obtain the sample entropy, which is expressed as:

[0029] ;

[0030] in, This represents the embedding dimension, which takes a positive integer value and is used to control the dimension of the reconstructed phase space. Indicates similarity tolerance; This represents the total number of sampling points in the signal segment; This indicates that in the embedding dimension is At that time, with the first A vector originating from a point and other vectors within the tolerance range Matching ratio within; The smaller the output value, the more regular the signal, the weaker the system's regulatory ability, and the more likely it is to be in a state of stress or pre-pathological condition.

[0031] The entropy value of the sample is used as an index of the physiological signal complexity under this window, and together with the original signal, event signal and static features, it forms a fused feature vector.

[0032] As a preferred embodiment of the LSTM time-series model-based early warning method for adverse events in dental clinics according to the present invention, the calculation of the dynamic deviation degree specifically includes the following steps:

[0033] Retrieve historical steady-state distribution parameters of the patient's corresponding physiological signals from the individual's health baseline profile, including the mean vector. With covariance matrix ;

[0034] The current fused feature vector With mean vector Align to the same feature space;

[0035] The Mahalanobis distance is calculated using the following expression:

[0036] ;

[0037] in, This represents the fused feature vector at the current moment, with the same dimensions as the baseline features; This represents the characteristic mean vector formed by the patient's multiple visits to the clinic under non-stress conditions; This represents the corresponding covariance matrix, reflecting the correlation and fluctuation range among the features; It is the inverse of the covariance matrix, used to eliminate the influence of dimensions and correlations between features; The distance is dimensionless;

[0038] right Perform min-max normalization to obtain the dynamic deviation. ;

[0039] The absolute risk probability output by the general LSTM sub-model and Personalized dynamic risk scores are synthesized according to weighted fusion rules, where the weight coefficients are preset constants.

[0040] As a preferred embodiment of the LSTM time-series model-based early warning method for adverse events in dental outpatient care according to the present invention, the attention mechanism for locating key risk-causing feature combinations comprises the following steps:

[0041] Receive the hidden states output by the final layer of the dual-channel LSTM model at each time step;

[0042] Attention scores are calculated using learnable weight vectors and biases;

[0043] Apply the Softmax function to the score sequence to obtain the normalized attention weights;

[0044] Select the continuous time window with the largest weight, and within this time window, extract the top three feature dimensions that contribute the most to the hidden state from the original input features, as the core risk factors;

[0045] The contribution is determined by calculating the gradient magnitude or perturbation sensitivity between the input features and the hidden state, ensuring that the selected features have causal explanatory power rather than just statistical correlation.

[0046] As a preferred embodiment of the LSTM time-series model-based early warning method for adverse events in dental clinics according to the present invention, the generation and push of the early warning information specifically includes the following steps:

[0047] The warning level is determined based on the personalized dynamic risk score and its slope of change within the most recent time window, including:

[0048] If the risk score is lower than the first threshold and the slope is less than the first slope threshold, no warning will be issued;

[0049] If the risk score is between the first threshold and the second threshold, or the slope is greater than or equal to the first slope threshold, a level two warning is triggered.

[0050] If the risk score is greater than or equal to the second threshold, a level 3 warning will be triggered;

[0051] Based on the warning level, a pre-set intervention plan library is matched to generate warning information containing standardized instructions such as pausing operation, adjusting to a supine position, preparing for oxygen inhalation, or calling for emergency medical assistance.

[0052] The warning information is pushed to the nursing terminal through the hospital's internal communication protocol, and a lock command is sent to the dental chair control system at the same time to prohibit non-emergency operations.

[0053] The first threshold, the second threshold, and the first slope threshold are all configurable parameters, which are jointly determined by the medical institution based on the historical incidence of adverse events and clinical tolerance.

[0054] After receiving the warning information, the nursing terminal automatically highlights the core risk factors and corresponding intervention steps, and records the nurse's confirmation and execution time.

[0055] Secondly, this invention provides an early warning system for adverse events in dental outpatient care based on an LSTM time series model, comprising:

[0056] The module includes a source data acquisition module, a time-series alignment processing module, a complexity fusion module, a dual-channel risk prediction module, and an intelligent early warning push module.

[0057] The multi-source data acquisition module is used to collect continuous multimodal physiological signals output by wearable devices from dental clinic patients during the diagnosis and treatment process, time-stamped operation events recorded by clinical terminals, and static clinical features entered before the operation, forming a multi-source heterogeneous raw data stream.

[0058] The time-series alignment processing module is used to construct a unified timeline based on the time when the patient enters the examination room, perform cross-modal time-series alignment processing on multi-source heterogeneous raw data streams, segment physiological signals using a sliding window, perform event-driven interpolation operations on events, generate missing mask markers to mark unreliable signal segments, and output a time-series feature matrix.

[0059] The complexity fusion module is used to calculate the physiological signal complexity index under the sliding time window based on the physiological signal segmentation in the time-series feature matrix, and to fuse the physiological signal complexity index with the original physiological signal, operation event and static clinical features to form a fused feature vector.

[0060] The dual-channel risk prediction module is used to input the fused feature vector into the dual-channel LSTM risk prediction model, output the absolute risk probability through the general LSTM sub-model, and calculate the dynamic deviation of the current state from the individual baseline by combining the patient's personal health baseline file, and fuse the two to obtain a personalized dynamic risk score.

[0061] The intelligent early warning push module is used to determine the risk warning level based on the personalized dynamic risk score and its change slope. When the preset intervention threshold is reached, the module uses an attention mechanism to locate the key risk-causing feature combination, generate a warning message containing standardized instructions such as pausing operation, adjusting to a supine position, preparing for oxygen inhalation, or calling for emergency help, and push it to the nursing terminal. At the same time, it sends a locking command to the dental chair control system to prohibit the execution of non-emergency operations.

[0062] The beneficial effects of this invention are as follows: By integrating multimodal physiological signals, operational events, and static clinical characteristics, a dynamic risk assessment system with the patient's individual baseline as a reference is constructed. Sample entropy is used to capture early physiological instability signals, Mahalanobis distance is combined to quantify the degree of deviation from the current state, and a dual-channel LSTM model is used to achieve collaborative modeling of group patterns and individual differences. This provides highly sensitive and specific graded early warnings before adverse events occur. At the same time, an attention mechanism is introduced to locate the root cause of the risk, generate interpretable and executable intervention instructions, and link the nursing terminal and the dental chair control system to form a closed-loop response. This improves the safety level of dental outpatient nursing and reduces the incidence and severity of adverse events such as syncope, allergies, and cardiovascular and cerebrovascular accidents. Attached Figure Description

[0063] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of an early warning method for adverse events in dental clinic nursing based on an LSTM time series model, as shown in the embodiment.

[0065] Figure 2 This is a schematic diagram of an early warning system for adverse events in dental clinic nursing based on an LSTM time series model, as shown in the embodiment. Detailed Implementation

[0066] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0067] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0068] Reference Figure 1 and Figure 2 This embodiment provides a method for early warning of adverse events in dental outpatient care based on an LSTM time series model, including the following steps:

[0069] S1. Collect continuous multimodal physiological signals output by wearable devices from patients in the dental clinic during the diagnosis and treatment process, timestamped operation events recorded by clinical terminals, and static clinical features entered before the operation to form a multi-source heterogeneous raw data stream.

[0070] Furthermore, wearable devices are used to collect the patient's heart rate, respiratory rate, blood oxygen saturation, skin conductance, and body movement intensity signals in real time during the diagnosis and treatment process, forming a first signal set; the dental chair integrated system or nurse handheld terminal records the start time of anesthetic injection, the activation time of the high-speed turbine handpiece, the start and stop time of the saliva suction device, and the time of patient position change, forming a second event sequence; the electronic medical record system obtains the patient's age, gender, history of syncope, allergy history, and number of underlying diseases before the operation, forming a third feature vector; the first signal set, the second event sequence, and the third feature vector are bound according to the patient's identity to form a multi-source heterogeneous raw data stream.

[0071] It should be noted that by synchronously integrating real-time physiological signals collected by wearable devices, timestamp records of clinical operation events, and preoperative static medical history information, and binding them with the patient's identity as the unique index, a multi-dimensional data foundation covering the entire diagnosis and treatment process can be constructed, effectively avoiding information silos and improving the comprehensiveness and individual adaptability of subsequent risk identification.

[0072] In particular, the multimodal data collected in the steps is not simply piled up, but is specifically selected around the stress scenarios unique to oral diagnosis and treatment. For example, skin conductance and body movement intensity can sensitively reflect patients' anxiety or discomfort, while operational events such as the activation of high-speed turbine handpieces are directly related to the intensity of local stimulation. The data composition method allows the system to focus on key factors that may induce adverse events from the source, rather than generalized monitoring, thereby laying a high-quality input foundation for subsequent accurate early warning without increasing the burden of redundancy.

[0073] S2. Construct a unified timeline based on the time the patient enters the consultation room, perform cross-modal time alignment processing on the multi-source heterogeneous raw data stream, use a sliding window to segment physiological signals, perform event-driven interpolation operations on events, generate a missing mask to mark unreliable signal segments, and output a time series feature matrix.

[0074] Furthermore, a high-resolution time axis with a uniform sampling frequency is established, with the moment the patient enters the examination room as the zero point. Each physiological signal in the first signal set is segmented using a sliding window of fixed length to obtain a discrete physiological segment sequence. For each event timestamp in the second event sequence, a preset physiological response duration template is matched according to the event type to generate a unit pulse extension signal covering the corresponding time period, which is then interpolated as a pseudo-continuous event signal on the unified time axis. For invalid segments in the physiological signals caused by motion artifacts or equipment detachment, a binary mask sequence of the same dimension is generated. All segmented physiological signals, interpolated event signals, and mask sequences are concatenated along the time axis to output a temporal feature matrix.

[0075] It should be noted that by taking the patient's entry into the consultation room as the unified zero point of time, high-resolution temporal alignment of heterogeneous data is performed. By employing strategies such as sliding window segmentation, event-driven interpolation, and missing data masking, not only are practical problems such as inconsistent sampling frequencies of multi-source signals, event sparsity, and data missingness solved, but the temporal causal relationship between key operations and physiological responses is also preserved, providing the model with a clear and semantically complete input representation.

[0076] In particular, the construction of a unified timeline is not merely for formal alignment, but rather to simulate the natural cognitive logic of clinical observation, starting with the patient's admission to the room; it transforms discrete events into pseudo-continuous signals with persistent physiological responses, essentially restoring the causal chain of operation-response at the data level; and the introduction of a missing mask allows the model to actively identify and avoid the influence of unreliable data segments; thus truly transforming the original heterogeneous data into a structured representation that conforms to clinical temporal logic and can be effectively understood by deep models.

[0077] S3. Based on the physiological signal segmentation in the time-series feature matrix, calculate the physiological signal complexity index under the sliding time window, and fuse the physiological signal complexity index with the original physiological signal, operation event and static clinical features to form a fused feature vector.

[0078] Furthermore, for any physiological signal segment in the time-series feature matrix, an embedding dimension of [missing value] is constructed. Embedded vector sequence; setting similarity tolerance The standard deviation of the signal segment is a fixed proportion; the Chebyshev distance between any two different embedding vectors is calculated; statistics are performed to show that the distance is less than a certain value. The number of vector pairs, denoted as And calculate their logarithmic mean to obtain the sample entropy, which is expressed as:

[0079] ;

[0080] in, This represents the embedding dimension, which takes a positive integer value and is used to control the dimension of the reconstructed phase space. Indicates similarity tolerance; This represents the total number of sampling points in the signal segment; This indicates that in the embedding dimension is At that time, with the first A vector originating from a point and other vectors within the tolerance range Matching ratio within; The smaller the output value, the more regular the signal and the weaker the system's regulatory ability, and the more likely it is to be in a state of stress or pre-pathology. The entropy value of this sample is used as an indicator of the physiological signal complexity under this window, and together with the original signal, event signal and static features, they form a fusion feature vector.

[0081] It should be noted that the introduction of a physiological signal complexity index based on phase space reconstruction can effectively capture changes in the system's regulatory capacity that cannot be reflected by traditional mean or variance, identify potential unstable states before patients show obvious clinical symptoms, thereby shifting the warning window forward and enhancing sensitivity to early stress responses or precursors of adverse events.

[0082] In particular, the introduction of complexity indicators breaks away from the limitations of traditional vital signs that only focus on numerical values, and instead captures the dynamic changes in the physiological system's internal regulatory capacity. In short-term, high-stimulation scenarios such as dental clinics, patients may not yet show obvious abnormalities in heart rate or blood oxygenation, but their autonomic nervous system regulation has already begun to become unbalanced. By quantifying the regularity of signals, the system can identify potential risks at an earlier stage. The judgment logic based on system stability is more forward-looking and adaptable than simply relying on threshold alarms.

[0083] S4. Input the fused feature vector into the dual-channel LSTM risk prediction model, output the absolute risk probability through the general LSTM sub-model, and calculate the dynamic deviation of the current state from the individual baseline by combining the patient's personal health baseline file. Combine the two to obtain a personalized dynamic risk score.

[0084] Furthermore, historical steady-state distribution parameters of the patient's corresponding physiological signals, including the mean vector, can be retrieved from the individual's health baseline file. With covariance matrix .

[0085] The individual health baseline profile is compiled from the patient's previous oral treatment records under conditions of no stress and stable physiological state. Each time a patient completes a treatment process and it is clinically confirmed that no adverse stress response occurred, the system automatically extracts and verifies the fusion feature vector of the corresponding steady-state period during that treatment and securely incorporates it into the patient's individual baseline database. Through a data accumulation mechanism, the system can continuously update and optimize the distribution parameters (mean vector) representing the patient's individualized steady-state physiological characteristics. With covariance matrix Therefore, during the risk assessment phase of the current diagnosis and treatment, when calculating the individual dynamic deviation, the optimized steady-state distribution parameters can be directly retrieved from the pre-built and maintained personal health baseline file, thereby avoiding the computational overhead and delay of temporary data collection and online modeling, and effectively ensuring the real-time nature and individual specificity of risk warning.

[0086] It should be noted that for new patients visiting for the first time or those without sufficient steady-state records, the system will automatically use a population reference baseline as a temporary alternative. This population reference baseline is statistically derived from the historical steady-state characteristics of a large number of similar individuals who have completed treatment and have been confirmed to have no stress response. Its mean vector and covariance matrix have been pre-trained and fixed in the system's model library. During the initial treatment, the system uses this population baseline for preliminary risk assessment and collects the patient's physiological and behavioral characteristics in real time during the initial stable phase of treatment (5 minutes of rest before surgery) to initialize their personal baseline profile. After the treatment is completed, if no adverse events are confirmed, this effective steady-state data will be stored as the patient's first baseline record in their personal health record and will be iteratively optimized in subsequent treatments to achieve a smooth transition from the population baseline to the individualized baseline.

[0087] The current fused feature vector With mean vector Align to the same feature space; calculate the Mahalanobis distance, expressed as:

[0088] ;

[0089] in, This represents the fused feature vector at the current moment, with the same dimensions as the baseline features; This represents the characteristic mean vector formed by the patient's multiple visits to the clinic under non-stress conditions; This represents the corresponding covariance matrix, reflecting the correlation and fluctuation range among the features; It is the inverse of the covariance matrix, used to eliminate the influence of dimensions and correlations between features; The distance is dimensionless.

[0090] right Perform min-max normalization to obtain the dynamic deviation. The absolute risk probability output by the general LSTM sub-model. and Personalized dynamic risk scores are synthesized according to weighted fusion rules, where the weight coefficients are preset constants.

[0091] The dual-channel LSTM risk prediction model consists of a general LSTM sub-model and an individual dynamic deviation calculation module. The general LSTM sub-model is a single-input, single-output temporal neural network that receives a fused feature vector sequence segmented by a sliding window and aligned to the time axis as input. Its network structure includes at least one LSTM unit and a fully connected output layer. The output layer uses the Sigmoid activation function to generate the absolute risk probability. ;

[0092] The individual dynamic deviation calculation module is non-parametric; it retrieves historical steady-state distribution parameters of corresponding physiological signals, including the mean vector, from the patient's individual health baseline file. With covariance matrix ; Fuse the feature vector at the current time step and After aligning to the same feature space, calculate the Mahalanobis distance. ; then Perform min-max normalization to obtain the dynamic deviation. ; Absolute risk probability With dynamic deviation Personalized dynamic risk scores are synthesized according to weighted fusion rules. The expression is:

[0093] ;

[0094] in The weight coefficients are preset constants; during the training phase, the dual-channel LSTM risk prediction model only updates the trainable parameters of the general LSTM sub-model, while the individual dynamic deviation calculation module remains fixed.

[0095] It should be noted that by integrating the general risk prediction results of the group with the dynamic deviation of the individual's historical health baseline, the risk score is both universal and personalized, making it more universal and avoids a one-size-fits-all approach, thus effectively improving the accuracy and clinical reliability of the early warning.

[0096] In particular, the core of the dual-channel design lies in distinguishing between general risks and individual abnormalities; the general LSTM learns the common response patterns of a large number of patients under similar operations, while the dynamic deviation characterizes the difference between the current patient and their own normal state; the fusion of the two means that even if a patient's baseline vital signs are always high or low, as long as their current state deviates significantly from their personal baseline, the system can still effectively identify risks; avoiding missed or false alarms due to individual differences, making the warning truly consistent with the physiological reality of each patient.

[0097] S5. Based on the personalized dynamic risk score and its change slope, the risk warning level is determined. When the preset intervention threshold is reached, the key risk-causing feature combination is located through the attention mechanism, and the warning information is generated and pushed to the nursing terminal.

[0098] Furthermore, the hidden states output by the final layer of the dual-channel LSTM model at each time step are received; attention scores are calculated using learnable weight vectors and biases; the Softmax function is applied to the score sequence to obtain normalized attention weights; a continuous time window with the largest weight is selected, and within this time window, the top three feature dimensions that contribute the most to the hidden state from the original input features are extracted as core risk factors; the contribution is determined by calculating the gradient magnitude or perturbation sensitivity between the input features and the hidden state, ensuring that the selected features have causal interpretability rather than just statistical correlation.

[0099] The warning level is determined based on the personalized dynamic risk score and its slope within the most recent time window. If the risk score is below the first threshold and the slope is less than the first slope threshold, no warning is issued; if the risk score is between the first and second thresholds, or the slope is greater than or equal to the first slope threshold, a level two warning is triggered; if the risk score is greater than or equal to the second threshold, a level three warning is triggered.

[0100] Based on the warning level, a pre-set intervention plan library is matched to generate warning information containing standardized instructions such as pausing operation, adjusting to a supine position, preparing for oxygen inhalation, or calling for emergency help. The warning information is pushed to the nursing terminal through the hospital's internal communication protocol, and a locking command is sent to the dental chair control system at the same time to prohibit the execution of non-emergency operations.

[0101] The first threshold, the second threshold, and the first slope threshold are all configurable parameters, which are jointly calibrated by medical institutions based on the historical incidence of adverse events and clinical tolerance. After receiving the warning information, the nursing terminal automatically highlights the core risk factors and corresponding intervention steps, and records the nurse's confirmation and execution time.

[0102] Among them, attention score The calculation expression is:

[0103] ;

[0104] In the formula, For the LSTM model at time step The hidden state, The weight matrix is ​​a learnable matrix. For learnable context weight vectors, For learnable biases; contribution The calculation uses the gradient magnitude method, and its expression is:

[0105] ;

[0106] In the formula, This is the final output of the model (personalized dynamic risk score). The first feature vector in the original input feature vector One dimension, This represents taking the absolute value, used for quantifying features. For output Local sensitivity.

[0107] It should be noted that by combining the absolute value of the risk score and its changing trend to conduct graded early warnings, and by automatically locating the root cause of the risk through the attention mechanism, not only can the interpretability of the risk be improved from whether it is dangerous to why it is dangerous, but it can also link the nursing terminal and the dental chair control system to generate standardized and executable intervention instructions, forming a closed loop and improving the timeliness and standardization of emergency response in dental clinics.

[0108] In particular, the warning system not only focuses on whether there is danger, but also emphasizes why there is danger and how to deal with it. By retracing the risk-causing characteristics through attention mechanisms, the system can point out specific combinations such as anesthesia operation combined with low blood oxygen, or changes in body position causing sudden changes in heart rate, giving nurses a clear direction for intervention. At the same time, the warning information is linked with the dental chair control, directly translating decisions into action constraints to prevent continued operation in high-risk conditions. This seamless connection from identification to execution allows the technology to be truly embedded in the clinical workflow, rather than just serving as a bystander reminder.

[0109] This embodiment also provides an early warning system for adverse events in dental outpatient care based on an LSTM time series model, including:

[0110] The system includes a source data acquisition module, a time-series alignment processing module, a complexity fusion module, a dual-channel risk prediction module, and an intelligent early warning push module.

[0111] The multi-source data acquisition module is used to collect continuous multimodal physiological signals output by wearable devices from dental clinic patients during the diagnosis and treatment process, time-stamped operation events recorded by clinical terminals, and static clinical features entered before the operation, forming a multi-source heterogeneous raw data stream.

[0112] The time alignment processing module is used to construct a unified time axis based on the time when the patient enters the clinic, perform cross-modal time alignment processing on multi-source heterogeneous raw data streams, segment physiological signals using a sliding window, perform event-driven interpolation operations on events, generate missing mask markers to mark unreliable signal segments, and output a time sequence feature matrix.

[0113] The complexity fusion module is used to calculate the physiological signal complexity index under the sliding time window based on the physiological signal segmentation in the time-series feature matrix, and to fuse the physiological signal complexity index with the original physiological signal, operation event and static clinical features to form a fused feature vector.

[0114] The dual-channel risk prediction module is used to input the fused feature vector into the dual-channel LSTM risk prediction model, output the absolute risk probability through the general LSTM sub-model, and calculate the dynamic deviation of the current state from the individual baseline by combining the patient's personal health baseline profile. The two are then fused to obtain a personalized dynamic risk score.

[0115] The intelligent early warning push module is used to determine the risk warning level based on the personalized dynamic risk score and its change slope. When the preset intervention threshold is reached, the module uses an attention mechanism to locate the key risk-causing feature combination, generate a warning message containing standardized instructions such as pausing operation, adjusting to a supine position, preparing for oxygen inhalation, or calling for emergency help, and push it to the nursing terminal. At the same time, it sends a lock command to the dental chair control system to prohibit the execution of non-emergency operations.

[0116] In summary, this invention integrates multimodal physiological signals, operational events, and static clinical characteristics to construct a dynamic risk assessment system based on the patient's individual baseline. It utilizes sample entropy to capture early physiological instability signals, combines Mahalanobis distance to quantify the degree of deviation from the current state, and achieves collaborative modeling of group patterns and individual differences through a dual-channel LSTM model. This provides highly sensitive and specific graded early warnings before adverse events occur. Simultaneously, it introduces an attention mechanism to locate the root cause of the risk, generates interpretable and executable intervention instructions, and links the nursing terminal and dental chair control system to form a closed-loop response. This improves the safety level of dental outpatient nursing and reduces the incidence and severity of adverse events such as syncope, allergies, and cardiovascular and cerebrovascular accidents.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for early warning of adverse events in dental clinic nursing based on an LSTM time series model, characterized in that: include: The system collects continuous multimodal physiological signals output by wearable devices during the diagnosis and treatment of patients in the dental clinic, timestamped operation events recorded by clinical terminals, and static clinical features entered before the operation, forming a multi-source heterogeneous raw data stream. A unified timeline is constructed based on the time when the patient enters the consultation room. Cross-modal temporal alignment processing is performed on the multi-source heterogeneous raw data stream. Sliding window segmentation of physiological signals and event-driven interpolation operation are adopted. Missing mask is generated to mark unreliable signal segments, and a temporal feature matrix is ​​output. Based on the physiological signal segmentation in the time-series feature matrix, the physiological signal complexity index under the sliding time window is calculated, and the physiological signal complexity index is fused with the original physiological signal, operation event and static clinical features to form a fused feature vector. The fused feature vector is input into the dual-channel LSTM risk prediction model. The absolute risk probability is output through the general LSTM sub-model. The dynamic deviation of the current state from the individual baseline is calculated by combining the patient's personal health baseline file. The two are then fused to obtain a personalized dynamic risk score. The risk warning level is determined based on the personalized dynamic risk score and its change slope. When the preset intervention threshold is reached, the key risk-causing feature combination is located through the attention mechanism, and the warning information is generated and pushed to the nursing terminal.

2. The method for early warning of adverse events in dental clinic nursing based on LSTM time series model as described in claim 1, characterized in that: The specific steps for acquiring the continuous multimodal physiological signals, operational events, and static clinical features are as follows: Wearable devices are used to collect the patient's heart rate, respiratory rate, blood oxygen saturation, skin conductance, and body movement intensity signals in real time during the diagnosis and treatment process, forming the first signal set; The second event sequence is formed by recording the start time of anesthetic injection, the activation time of high-speed turbine handpiece, the start and stop time of saliva suction device, and the time of patient position change through dental chair integrated system or nurse handheld terminal; The electronic medical record system is used to obtain the patient's age, gender, history of syncope, history of allergies, and number of underlying diseases before surgery, forming a third feature vector. The first signal set, the second event sequence, and the third feature vector are bound together according to the patient's identity to form a multi-source heterogeneous raw data stream.

3. The method for early warning of adverse events in dental clinic nursing based on LSTM time series model as described in claim 2, characterized in that: The cross-modal timing alignment process includes the following steps: A high-resolution time axis with a uniform sampling frequency is established, with the moment the patient enters the examination room as the zero point. Each physiological signal in the first signal set is segmented using a sliding window of fixed length to obtain a discrete physiological segment sequence; For each event timestamp in the second event sequence, a preset physiological response duration template is matched according to the event type to generate a unit pulse extension signal covering the corresponding time period, and interpolated as a pseudo-continuous event signal on a unified time axis; For invalid segments in physiological signals caused by motion artifacts or device detachment, generate a binary mask sequence of the same dimension; All segmented physiological signals, interpolated event signals, and mask sequences are concatenated along the time axis to output a time-series feature matrix.

4. The method for early warning of adverse events in dental clinic nursing based on LSTM time series model as described in claim 3, characterized in that: The calculation of the physiological signal complexity index involves the following steps: For any physiological signal segment in the time-series feature matrix, construct an embedding dimension of... The sequence of embedded vectors; Set similarity tolerance This is a fixed proportion of the standard deviation of the signal segments; Calculate the Chebyshev distance between any two distinct embedding vectors; Statistical satisfaction of distance less than The number of vector pairs, denoted as And calculate their logarithmic mean to obtain the sample entropy, which is expressed as: ; in, This represents the embedding dimension, which takes a positive integer value and is used to control the dimension of the reconstructed phase space. Indicates similarity tolerance; This represents the total number of sampling points in the signal segment; This indicates that the embedding dimension is At that time, with the first A vector originating from a point and other vectors within the tolerance range Matching ratio within; The smaller the output value, the more regular the signal, the weaker the system's regulatory ability, and the more likely it is to be in a state of stress or pre-pathological condition. The entropy value of the sample is used as an index of the physiological signal complexity under this window, and together with the original signal, event signal and static features, it forms a fused feature vector.

5. The method for early warning of adverse events in dental clinic nursing based on LSTM time series model as described in claim 4, characterized in that: The calculation of the dynamic deviation is performed using the following steps: Retrieve historical steady-state distribution parameters of the patient's corresponding physiological signals from the individual's health baseline profile, including the mean vector. With covariance matrix ; The current fused feature vector With mean vector Align to the same feature space; The Mahalanobis distance is calculated using the following expression: ; in, This represents the fused feature vector at the current moment, with the same dimensions as the baseline features; This represents the characteristic mean vector formed by the patient's multiple visits to the clinic under non-stress conditions; This represents the corresponding covariance matrix, reflecting the correlation and fluctuation range among the features; It is the inverse of the covariance matrix, used to eliminate the influence of dimensions and correlations between features; The distance is dimensionless; right Perform min-max normalization to obtain the dynamic deviation. ; The absolute risk probability output by the general LSTM sub-model and Personalized dynamic risk scores are synthesized according to weighted fusion rules, where the weight coefficients are preset constants.

6. The method for early warning of adverse events in dental clinic nursing based on LSTM time series model as described in claim 5, characterized in that: The attention mechanism locates key risk-causing feature combinations, and the specific steps are as follows: Receive the hidden states output by the final layer of the dual-channel LSTM model at each time step; Attention scores are calculated using learnable weight vectors and biases; Apply the Softmax function to the score sequence to obtain the normalized attention weights; Select the continuous time window with the largest weight, and within this time window, extract the top three feature dimensions that contribute the most to the hidden state from the original input features, as the core risk factors; The contribution is determined by calculating the gradient magnitude or perturbation sensitivity between the input features and the hidden state, ensuring that the selected features have causal explanatory power rather than just statistical correlation.

7. The method for early warning of adverse events in dental clinic nursing based on LSTM time series model as described in claim 6, characterized in that: The specific steps for generating and pushing the warning information are as follows: The warning level is determined based on the personalized dynamic risk score and its slope of change within the most recent time window, including: If the risk score is lower than the first threshold and the slope is less than the first slope threshold, no warning will be issued; If the risk score is between the first threshold and the second threshold, or the slope is greater than or equal to the first slope threshold, a level two warning is triggered. If the risk score is greater than or equal to the second threshold, a level 3 warning will be triggered; Based on the warning level, a pre-set intervention plan library is matched to generate warning information containing standardized instructions such as pausing operation, adjusting to a supine position, preparing for oxygen inhalation, or calling for emergency medical assistance. The warning information is pushed to the nursing terminal through the hospital's internal communication protocol, and a lock command is sent to the dental chair control system at the same time to prohibit non-emergency operations. The first threshold, the second threshold, and the first slope threshold are all configurable parameters, which are jointly determined by the medical institution based on the historical incidence of adverse events and clinical tolerance. After receiving the warning information, the nursing terminal automatically highlights the core risk factors and corresponding intervention steps, and records the nurse's confirmation and execution time.

8. A pre-warning system for adverse events in dental clinic nursing based on an LSTM time series model, based on the pre-warning method for adverse events in dental clinic nursing based on an LSTM time series model as described in any one of claims 1-7, characterized in that: include: The module includes a source data acquisition module, a time-series alignment processing module, a complexity fusion module, a dual-channel risk prediction module, and an intelligent early warning push module. The multi-source data acquisition module is used to collect continuous multimodal physiological signals output by wearable devices from dental clinic patients during the diagnosis and treatment process, time-stamped operation events recorded by clinical terminals, and static clinical features entered before the operation, forming a multi-source heterogeneous raw data stream. The time-series alignment processing module is used to construct a unified timeline based on the time when the patient enters the examination room, perform cross-modal time-series alignment processing on multi-source heterogeneous raw data streams, segment physiological signals using a sliding window, perform event-driven interpolation operations on events, generate missing mask markers to mark unreliable signal segments, and output a time-series feature matrix. The complexity fusion module is used to calculate the physiological signal complexity index under the sliding time window based on the physiological signal segmentation in the time-series feature matrix, and to fuse the physiological signal complexity index with the original physiological signal, operation event and static clinical features to form a fused feature vector. The dual-channel risk prediction module is used to input the fused feature vector into the dual-channel LSTM risk prediction model, output the absolute risk probability through the general LSTM sub-model, and calculate the dynamic deviation of the current state from the individual baseline by combining the patient's personal health baseline file, and fuse the two to obtain a personalized dynamic risk score. The intelligent early warning push module is used to determine the risk warning level based on the personalized dynamic risk score and its change slope. When the preset intervention threshold is reached, the module uses an attention mechanism to locate the key risk-causing feature combination, generate a warning message containing standardized instructions such as pausing operation, adjusting to a supine position, preparing for oxygen inhalation, or calling for emergency help, and push it to the nursing terminal. At the same time, it sends a locking command to the dental chair control system to prohibit the execution of non-emergency operations.

Images