An intelligent monitoring and early warning system for infusion liquid abnormalities

Abstract

The application discloses an infusion liquid abnormality intelligent monitoring and early warning system, which is characterized in that a multispectral sensing module, an environment self-adaptive compensation module, a patient data linkage module, a data fusion processing module, a hierarchical early warning module and a medical data interaction module are communicatively connected, can accurately identify and individually and hierarchically early warn three abnormal types of sedimentation, discoloration and flocculation, and through linkage of a HIS system and a quality control platform, can realize monitoring, early warning, disposal data tracing, a traceable infusion safety quality control file and improved infusion process safety.

Description

Technical Field

[0001] This invention relates to the field of infusion abnormality monitoring and early warning technology, specifically to an intelligent monitoring and early warning system for infusion fluid abnormalities. Background Technology

[0002] Intravenous infusion is a common clinical procedure, but during infusion, the fluid is prone to abnormalities such as precipitation, discoloration, flocculent formation, and crystallization due to improper drug compatibility, abnormal storage, or prolonged infusion time. These abnormalities can easily lead to medical risks such as allergies and embolisms. Some drugs may only cause transient discoloration or turbidity, which can be easily overlooked if not carefully observed. However, when two drugs are mixed and become turbid, their chemical composition has changed. When the mixture clears, it is no longer the original composition. Injecting such a drug into the body not only fails to achieve the therapeutic effect but may even be harmful, leading to adverse consequences. Sometimes, the drugs themselves do not exhibit crystallization during preparation, but during infusion, the two drug solutions come into contact and react, easily precipitating crystals or flocculent substances. Once injected into the body, these substances can easily cause adverse reactions. Furthermore, the warning standards for abnormal situations should differ for different populations; otherwise, over-warning or false alarms may occur, increasing the workload of medical staff. Therefore, it is necessary to strictly monitor the quality of infusion fluids and provide dynamic warnings for potential abnormalities.

[0003] In summary, existing infusion monitoring technologies have the following problems: Limited monitoring dimensions: Most technologies only monitor single abnormalities such as precipitation or discoloration, failing to simultaneously identify precipitation, discoloration, and flocculent abnormalities, resulting in low accuracy; Insufficient identification precision: Relying on a single optical sensor or simple threshold judgment, they are greatly affected by ambient light and container material, exhibiting low recognition rates for slight discoloration and trace precipitation; Lack of personalized early warning: Using fixed thresholds for "one-size-fits-all" early warnings without considering individual characteristics such as patient age, disease type, and treatment stage, easily leading to false alarms in special populations such as pediatricians and those undergoing chemotherapy. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by providing an intelligent monitoring and early warning system for abnormal infusion fluids. This system can monitor and identify three types of abnormalities in infusion fluids in real time and with high accuracy: sedimentation, discoloration, and flocculent matter. By combining individual patient characteristics with clinical diagnosis and treatment data, it can achieve personalized and graded early warning.

[0005] To achieve the above objectives, the present invention provides an intelligent monitoring and early warning system for abnormal infusion fluids, comprising: Multispectral sensing module: used to attach to the outer wall of infusion container and Mofe's dropper to collect multispectral signals of infusion fluid; Environmental adaptive compensation module: It is communicatively connected to the multispectral sensing module, and has built-in ambient light sensor and temperature sensor to collect environmental parameters and perform interference correction on the collected multispectral signals; Patient data linkage module: It is equipped with a standardized interface for interfacing with the hospital's HIS system to automatically acquire individualized patient diagnosis and treatment data, including age, disease type, treatment stage data, and information on the current infusion solution; Data fusion processing module: used to receive corrected multispectral signals and individualized diagnosis and treatment data, and perform the following operations: based on the multispectral signals, through a pre-trained fusion recognition model, through feature extraction and fusion recognition, determine whether there is at least one abnormality among precipitation, discoloration and flocculent matter in the infusion fluid, and output the abnormality recognition result and comprehensive confidence level; The graded early warning module communicates with the data fusion processing module and is used to dynamically calculate personalized early warning thresholds based on the identification results output by the data fusion processing module, combined with the current infusion fluid information and individualized diagnosis and treatment data; it compares the model identification results with the personalized early warning thresholds, and judges the abnormal type of infusion fluid and the corresponding early warning level based on the comprehensive confidence level, triggering graded early warning actions corresponding to the early warning level; Medical data interaction module: It communicates with the graded early warning module, the hospital HIS system, and the hospital quality control platform to upload monitoring and early warning data.

[0006] To optimize the above technical solution, the specific measures also include: Furthermore, the multispectral sensing module includes at least four spectral sensors: a first spectral sensor with an operating wavelength of 450±10nm, used to detect scattering signals from sediment and flocculent matter; a second spectral sensor with an operating wavelength of 550±10nm, used to detect transmission or reflection signals of the liquid in the yellow spectral region; a third spectral sensor with an operating wavelength of 650±10nm, used to detect transmission or reflection signals of the liquid in the red spectral region; and a fourth spectral sensor with an operating wavelength of 850±20nm, used to acquire the intensity characteristics of the scattering signal from flocculent matter.

[0007] Furthermore, the data fusion processing module includes a lightweight multi-branch neural network model for feature screening and classification of the corrected multispectral signal; the lightweight multi-branch neural network model includes three feature extraction branches corresponding to the identification of sedimentation, discoloration and flocculent matter, respectively, and a fusion decision layer, which integrates the output of each branch, the current drug solution type and individualized diagnosis and treatment data, and outputs the final anomaly identification result and comprehensive confidence level. The anomaly identification results include the anomaly type, feature value, and confidence level for each anomaly type branch; the overall confidence level = precipitation confidence level × precipitation weight + color change confidence level × color change weight + flocculent confidence level × flocculent weight; the precipitation weight, color change weight, and flocculent weight are all determined based on their historical infusion data, drug characteristics, and patient-specific diagnosis and treatment data.

[0008] Furthermore, the precipitation feature extraction branch is used to receive signals from the first spectral sensor and extract the scattering coefficient, the proportion of abnormal duration of the scattering signal, and the time-series change features; the color change feature extraction branch is used to receive signals from the second and third spectral sensors and calculate the color difference value and determine the color change tendency; the flocculent feature extraction branch is used to receive signals from the first and fourth spectral sensors and fuse and extract the scattering signal fluctuation features of flocculents, the time-series decay trend features of signal intensity, and the ratio of scattering coefficients at 450nm to 850nm.

[0009] Furthermore, the graded early warning module is preset with a basic early warning threshold based on the type of medication; the personalized early warning threshold is: Personalized early warning threshold = Basic early warning threshold × (1 + α × A + β × D + γ × S), where α is the age weight coefficient; β is the disease weight coefficient; γ is the treatment stage weight coefficient; α, β, and γ are fixed coefficients preset based on clinical sample statistical analysis and medical expert consensus, and their values ​​range from 0 to 1; A is the risk coefficient mapped according to the patient's actual age; D is the risk coefficient mapped according to the patient's disease type; S is the risk coefficient mapped according to the patient's treatment stage, where A = 0.3 for ages ≤ 14 years or ≥ 65 years, A = 0.1 for ages 15-64 years; D = 0.4 for severe cases, D = 0.1 for common diseases, D = 0.2 for postoperative diseases; S = 0.3 for acute phase, S = 0.1 for recovery phase, and S = 0.05 for maintenance phase.

[0010] Furthermore, the warning levels in the graded warning module include at least the following: Level 1 warning: triggered when the basic warning threshold < the identified abnormal feature value < the corresponding personalized warning threshold, and the confidence level of the abnormal type ≥ the basic confidence level; Level 2 warning: triggered when the corresponding personalized warning threshold ≤ the identified abnormal feature value < the fixed threshold, and the confidence level of the abnormal type ≥ the basic confidence level; Level 3 warning: triggered when the identified abnormal feature value ≥ the fixed threshold, or when there are multiple types of abnormalities, each abnormal feature value ≥ its corresponding personalized warning threshold, and the confidence level of each abnormal type is greater than the fixed confidence level, while the overall confidence level ≥ the preset overall confidence level; wherein, the basic confidence level is obtained by training based on historical clinical data, the fixed threshold is the preset highest-level feature value red line, the fixed confidence level is the preset highest-level confidence level red line, and the preset overall confidence level is the preset highest-level overall confidence level.

[0011] Furthermore, when a Level 2 warning is triggered, if the overall confidence level is higher than the preset overall confidence level, the system outputs a confirmation of an anomaly and an immediate medication change instruction; if the overall confidence level is lower than the preset overall confidence level but higher than the baseline overall confidence level, the system outputs a suspected anomaly and an immediate medication verification instruction; when a Level 1 warning is triggered, if the overall confidence level is greater than the observation threshold, the system automatically upgrades the monitoring level and outputs a low-priority prompt; where the observation threshold is a preset observation level standard; and the baseline overall confidence level is a comprehensive risk probability standard for the presence of an anomaly in the infusion fluid, based solely on the general characteristics of the medication and clinical risks.

[0012] Furthermore, the medical data interaction module is used to bind the identification results, warning level, trigger time, and subsequent medical staff treatment records of the data fusion processing module with the patient identifier and timestamp, and then upload them to the hospital quality control platform to generate a traceable infusion safety quality control file.

[0013] This invention also provides a method for intelligent monitoring and early warning of abnormal infusion fluids, comprising the following steps: Step 1: System initialization and benchmark calibration: The multispectral sensing module is attached to the lower side of the infusion container and the lower surface of the Mofe's dropper to collect multi-channel optical benchmark values ​​and environmental benchmark parameters of the clear infusion liquid. At the same time, the individualized diagnosis and treatment data and treatment stage data of the target patient are obtained through the hospital HIS system interface. Step 2: Multispectral signal acquisition and environmental calibration: The optical signal of the infusion fluid is acquired using a multispectral sensing module, and the optical signal is calibrated by environmental adaptive compensation in combination with the current environmental parameters; Step 3: Anomaly feature extraction: Based on the calibrated optical signal, extract the features of sedimentation, discoloration and flocculent matter respectively; among which, the features of flocculent matter include the spatial proportion of filamentous morphological regions based on the near-infrared channel signal analysis, and / or the dynamic sedimentation trajectory of the abnormal region in the continuous frame signal analysis. Step 4: Abnormal Feature Fusion and Identification: Input the extracted feature data into the pre-trained fusion and identification model to identify and output whether there are abnormalities such as precipitation, discoloration or flocculent matter, and output the identification results and comprehensive confidence level. Step 5: Personalized warning threshold calculation and level determination: Determine the basic warning threshold based on the current infusion information, and calculate the personalized warning threshold for the current patient by combining the individualized diagnosis and treatment data obtained in Step 1; compare the identification results of Step 4 with the personalized warning threshold, and determine the warning level by combining the comprehensive confidence level and trigger the corresponding graded warning action; Step 6: Upload the entire process data of each monitoring, identification, early warning and follow-up treatment to the hospital quality control platform.

[0014] Furthermore, the fusion recognition model adopts a lightweight multi-branch neural network model, LSTM-Transformer, including: sedimentation, discoloration and flocculent recognition branches; the recognition method is: feature screening and classification of the corrected multispectral signal; each branch outputs the feature value and confidence level of sedimentation, discoloration and flocculent abnormalities according to its corresponding features. The results output from each branch enter the fusion decision layer. Combining individualized diagnosis and treatment data with information on the current infusion solution, personalized early warning thresholds for abnormal precipitation, discoloration, and flocculent matter are calculated: Personalized early warning threshold = Basic early warning threshold × (1 + α × A + β × D + γ × S), where α is the age weight coefficient; β is the disease weight coefficient; γ is the treatment stage weight coefficient; α, β, and γ are fixed coefficients preset based on clinical sample statistical analysis and medical expert consensus; A is the risk coefficient mapped according to the patient's actual age; D is the risk coefficient mapped according to the patient's disease type; S is the risk coefficient mapped according to the patient's treatment stage, where A = 0.3 for ages ≤ 14 years or ≥ 65 years; A = 0.1 for ages 15-64 years; D = 0.4 for severe cases, D = 0.1 for common diseases, and D = 0.2 for postoperative diseases; S = 0.3 for the acute phase, S = 0.1 for the recovery phase, and S = 0.05 for the maintenance phase.

[0015] The beneficial effects of this invention are: by providing personalized graded early warning, combining the patient's age, disease type, and treatment stage to dynamically calculate the early warning threshold, and setting three levels of early warning and corresponding treatment instructions, this invention can accurately monitor and warn different patients; by linking the HIS system and the quality control platform, it can achieve traceability of monitoring, early warning, and treatment data, thereby improving the safety of the infusion process. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the intelligent monitoring and early warning system for abnormal infusion fluids of the present invention; Figure 2 This is a flowchart illustrating the intelligent monitoring and early warning method for abnormal infusion fluids according to the present invention. Detailed Implementation

[0017] The invention will now be described in further detail with reference to the accompanying drawings.

[0018] The embodiments described in this invention are merely some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0019] like Figure 1As shown, this invention discloses an intelligent monitoring and early warning system for abnormal infusion fluids, comprising: a multispectral sensing module, an environmental adaptive compensation module, a patient data linkage module, a data fusion processing module, a graded early warning module, a medical data interaction module, and communication connections between the various modules.

[0020] Multispectral sensing module: This module is attached to the outer wall of the infusion container and the Mofee drip chamber to collect multispectral signals of the infusion fluid. In one implementation, the multispectral sensing module includes a flexible substrate and four spectral sensors mounted on the substrate. The flexible substrate uses medical-grade silicone as the substrate, which can fit tightly with round or square infusion bottles and bags to reduce interference when collecting spectral data. Specifically, since infusion fluids are prone to precipitation and flocculent formation near the bottom of the container, and where two liquids meet, the flexible substrate is attached to the bottom of the infusion container where there is liquid. At the Mofee drip chamber, the sensors are attached to the outer wall of the lower section of the drip chamber, i.e., the section below the drip chamber where the liquid is completely filled and continuously flowing. This region is fundamentally different from the dripping area at the top of the dropper (the interface between the gas chamber and the liquid surface): the liquid in the lower liquid column region is continuous and the flow is stable, unaffected by the liquid surface fluctuations and gas-liquid interface instability caused by the periodic dripping of the droplets above, thus enabling the establishment of a stable and reliable optical reference signal.

[0021] In one implementation, four spectral sensors are configured as follows: a first spectral sensor (SGL-450) with an operating wavelength of 450±10nm, used to detect scattering signals from sediment and flocculent matter; a second spectral sensor (SGL-550) with an operating wavelength of 550±10nm, used to detect transmission or reflection signals of the liquid in the yellow spectral region; a third spectral sensor (SGL-650) with an operating wavelength of 650±10nm, used to detect transmission or reflection signals of the liquid in the red spectral region; and a fourth spectral sensor (SGL-850) operating in near-infrared light at a wavelength of 850±20nm, used to acquire the intensity characteristics of scattering signals from flocculent matter. The sampling frequency of each sensor is set to 10Hz to capture changes in the liquid signal in real time.

[0022] In one implementation, the system also includes an environmental adaptive compensation module: it is communicatively connected to the multispectral sensing module and has a built-in ambient light sensor and temperature sensor to collect ambient light intensity and infusion ambient temperature, and to perform interference correction on the collected multispectral signals (for example: when the ambient light intensity exceeds 500 lux, the collected signals of the blue light, yellow light and red light sensors are linearly attenuated and corrected; when the temperature deviates from the 25℃ reference temperature, the near-infrared sensor signal is compensated according to the temperature coefficient, and the correction formula is: corrected signal = original signal × (1 + k × (T - 25)), where k is the temperature compensation coefficient, which is 0.005, and T is the temperature).

[0023] In one implementation, the patient data linkage module is equipped with a standardized interface for interfacing with the hospital's HIS system. This interface is used to automatically acquire personalized medical data such as the patient's age, disease type (e.g., cardiovascular disease, diabetes, infectious disease), and treatment stage (e.g., acute phase, recovery phase, maintenance phase), as well as information on the currently infused medication (name, concentration, compatibility information, etc.). The data is updated once per minute to ensure real-time data availability.

[0024] In one implementation, the data fusion processing module receives the corrected multispectral signal and individualized diagnosis and treatment data, and performs the following operations: based on the multispectral signal, through a pre-trained fusion recognition model, through feature extraction and fusion recognition, it determines whether there is at least one abnormality among precipitation, discoloration and flocculent matter in the infusion fluid, and outputs the abnormality recognition result and comprehensive confidence level.

[0025] In one implementation, the data fusion processing module uses an edge computing gateway (model EC-600S) as its hardware platform. It incorporates a pre-trained LSTM (Long Short-Term Memory) and a lightweight multi-branch neural network model with a Transformer architecture for feature filtering and classification of the corrected multispectral signals. This model is trained on a large number of normal and abnormal infusion fluid samples (including multispectral signal samples of different types of sediment, varying degrees of discoloration, and different forms of flocculent matter). The training process uses the Adam optimizer with a learning rate of 0.001 and 100 iterations.

[0026] The model comprises three feature extraction branches corresponding to sedimentation, discoloration, and flocculent identification, respectively. The sedimentation feature extraction branch receives correction signals from the first spectral sensor, extracts the scattering coefficient, the proportion of abnormal scattering signal duration, and time-series variation features. The discoloration feature extraction branch receives correction signals from the second and third spectral sensors, converts the signals to RGB color values, calculates the color difference with the reference color value, and determines the discoloration tendency (e.g., yellowing, reddening, or turbidity) through a fully connected layer. The flocculent feature extraction branch receives correction signals from the first and fourth spectral sensors, fuses and extracts the morphological features of the flocculents (e.g., length, width, area, and proportion of filamentous grayscale areas) using an edge detection algorithm, analyzes dynamic sedimentation trajectory features, and calculates the ratio of the scattering coefficients at 450nm to 850nm.

[0027] In one implementation, the aforementioned lightweight multi-branch neural network model further includes a fusion decision layer. This layer employs an attention mechanism to weightedly fuse the output feature values ​​of the three branches. Combining the current drug solution type and individualized treatment data, it outputs anomaly identification results (types and feature values ​​of precipitation, discoloration, and flocculents) and the confidence score (range 0-1) for each anomaly type. The overall confidence score is the weighted average of the confidence scores for each anomaly type: Overall Confidence Score = Precipitation Confidence Score × Precipitation Weight + Discoloration Confidence Score × Discoloration Weight + Flocculent Confidence Score × Flocculent Weight. The precipitation weight, discoloration weight, and flocculent weight are all determined based on a "basic weight - dynamic weight - personalized weight" system. The basic weights are based on clinical infusion big data statistics: Precipitation: 0.4; Discoloration: 0.3; Flocculent: 0.3. The dynamic weights are adjusted dynamically based on the physicochemical properties of the drug solution, increasing the weight of common abnormalities and decreasing the weight of less common abnormalities. For example, for antibiotics, the weights are +0.1 for flocculent matter, -0.05 for precipitation, and -0.05 for discoloration; for chemotherapy, the weights are +0.1 for precipitation, -0.05 for flocculent matter, and -0.05 for discoloration. Personalized weights are further fine-tuned based on the dynamic weights, taking into account individual patient characteristics (age, disease, treatment stage), with the adjustment range controlled within ±0.01 to ±0.03.

[0028] In one implementation, the tiered early warning module is communicatively connected to the data fusion processing module. The tiered early warning module includes an early warning output unit, a wireless linkage unit, and an infusion pump linkage interface. The early warning output unit includes an LED light and a buzzer. The wireless linkage unit can synchronize early warning information to the nurse station terminal and the smart wristband of medical staff. When the infusion pump triggers a level 3 early warning, it can automatically pause the infusion and push alarm information including the abnormality type and emergency treatment suggestions to the mobile terminal of medical staff.

[0029] In one implementation, the graded early warning module presets a basic early warning threshold based on the type of drug solution (determined experimentally, for example: the abnormal duration of precipitation scattering signal accounts for 5% of the basic threshold, the basic threshold for discoloration is 3ΔE color difference value, and the basic threshold for flocculent area accounts for 3%). Based on the characteristic values ​​of precipitation, discoloration, and flocculents output by the data fusion processing module, combined with the current infusion drug solution information and individualized diagnosis and treatment data, a personalized early warning threshold is dynamically calculated. The personalized early warning threshold = basic early warning threshold × (1 + α × A + β × D + γ × S), where α is the age weight coefficient; β is the disease weight coefficient; γ is the treatment stage weight coefficient; α, β, and γ are fixed coefficients preset based on clinical sample statistical analysis and medical expert consensus, and their values ​​range from 0 to 1. A represents the risk coefficient mapped based on the patient's actual age; D represents the risk coefficient mapped based on the patient's disease type; and S represents the risk coefficient mapped based on the patient's treatment stage. Specifically, A=0.3 for patients ≤14 years or ≥65 years old; A=0.1 for patients 15-64 years old; D=0.4 for severe cases, D=0.1 for common diseases, and D=0.2 for postoperative diseases; S=0.3 for the acute phase, S=0.1 for the recovery phase, and S=0.05 for the maintenance phase. The feature values ​​of precipitation, discoloration, and flocculent matter in the model identification results are compared with their corresponding personalized warning thresholds. Combined with the comprehensive confidence level, the warning level corresponding to the abnormal infusion fluid is determined, triggering the graded warning action corresponding to the warning level.

[0030] In one implementation, the warning level and corresponding action settings are as follows: The warning levels include: Level 1 warning, where the system compares the real-time characteristic values ​​of sediment, flocculent matter, and discoloration with the corresponding personalized warning thresholds for sedimentation, flocculent matter, and discoloration, respectively. A warning is triggered when the characteristic value of the identified abnormal type is higher than the basic warning threshold but lower than its corresponding personalized warning threshold; the LCD screen displays a yellow warning message, and a buzzer emits a low-frequency warning sound. After triggering Level 1 warning, if the overall confidence level is greater than the preset observation threshold (the observation threshold is the standard for the preset observation level), the system automatically upgrades the monitoring level and outputs a low-priority prompt (such as "It is recommended to strengthen patrols"); if the observation threshold is not exceeded, the system continues monitoring by default.

[0031] A Level 2 alert is triggered when the characteristic value of the identified abnormal type reaches or exceeds its corresponding personalized alert threshold, but is below a fixed threshold (the fixed threshold is a preset red line for the highest level characteristic value; once exceeded, a Level 3 alert is immediately triggered), and the confidence level of the abnormal type is greater than or equal to the baseline confidence level. The LCD screen displays an orange alert, and a buzzer emits a mid-frequency alert tone. After triggering a Level 2 alert, if the overall confidence level is higher than the preset overall confidence level (the preset overall confidence level is the highest level overall confidence level, derived from historical clinical abnormality experience; once exceeded, an abnormality is highly likely), the system outputs a "Confirmed abnormality, change medication immediately" command and pushes it to the medical staff's mobile terminal; if the overall confidence level is lower than the preset overall confidence level but higher than the baseline overall confidence level, the system outputs a "Suspected abnormality, check medication immediately" command.

[0032] The Level 3 warning includes two scenarios: Scenario 1: It is triggered directly when the identified abnormal feature value reaches or exceeds a fixed threshold, without considering the confidence level; Scenario 2: When multiple types of abnormalities exist simultaneously, each abnormal feature value is greater than or equal to its corresponding personalized warning threshold, and the confidence level of each abnormality is greater than the fixed confidence level (the fixed confidence level is the preset highest level confidence red line, obtained based on historical abnormality experience; once this confidence threshold is exceeded, it indicates that the confidence levels of multiple abnormal types are relatively high), and the overall confidence level is greater than or equal to the preset overall confidence level, it is triggered. The LCD screen displays a red warning prompt, the buzzer emits a high-frequency continuous alarm sound, and the power supply to the infusion pump in the infusion tubing is automatically cut off. An emergency treatment command is generated and pushed to the mobile terminal of medical staff and the monitoring platform of the nurse station.

[0033] In one implementation, the medical data interaction module uses a 5G module to communicate with the tiered early warning module, the hospital's HIS system, and the hospital's quality control platform. It uploads the identification results (anomaly type, feature value, confidence level), early warning level, early warning trigger time, environmental parameters, patient information, and subsequent medical staff handling records (such as verification results, dressing change time, patient reaction, etc.) from the data fusion processing module, bound to the patient's unique identifier (hospitalization number) and timestamp, to the hospital's quality control platform. The platform generates a traceable infusion safety quality control file based on the uploaded data, supporting query and statistical analysis by patient, department, time, and other dimensions.

[0034] This invention also discloses a method for intelligent monitoring and early warning of abnormal infusion fluids, which includes the following steps in one implementation: Step 1: System initialization and benchmark calibration: Medical staff attach the multispectral sensing module to the lower side of the infusion container and the lower surface of the Mofe's dropper, turn on the system power, and the system automatically completes the initialization.

[0035] Multi-channel optical reference values ​​of clear infusion fluid are collected. In this embodiment, each sensor collects 10 sets of data and averages them. Environmental reference parameters (light intensity, temperature) are collected and stored. Simultaneously, individualized diagnosis and treatment data such as the target patient's age, disease type, treatment stage, and current infusion solution are obtained through the hospital's HIS system interface.

[0036] Step 2: Multispectral signal acquisition and environmental calibration: The optical signal (frequency 10Hz) of the infusion fluid is acquired using a multispectral sensing module, and the optical signal is calibrated and corrected by environmental adaptive compensation in combination with the current environmental parameters. Step 3: Anomaly Feature Extraction: The data fusion processing module receives the calibrated multispectral signal and extracts features of precipitation, discoloration, and flocculent matter respectively; the precipitation feature extraction branch receives the signal from the first spectral sensor and extracts the scattering coefficient, the proportion of abnormal duration of the scattering signal, and the time sequence change features; the discoloration feature extraction branch receives the signals from the second and third spectral sensors, calculates the color difference value, and judges the discoloration tendency; the flocculent matter feature extraction branch receives the signals from the first and fourth spectral sensors, extracts the proportion of filamentous grayscale regions, dynamic sedimentation trajectory, and the ratio of scattering coefficients at 450nm to 850nm.

[0037] Step 4: Abnormal Feature Fusion and Recognition: Input the extracted feature data into the pre-trained LSTM-Transformer model for recognition. Each branch outputs whether there is precipitation, discoloration or flocculent abnormality based on its corresponding features, outputs the weight and confidence of each type of abnormality, and calculates the comprehensive confidence: Comprehensive confidence = precipitation confidence × precipitation weight + discoloration confidence × discoloration weight + flocculent confidence × flocculent weight.

[0038] The method for identifying flocculent matter is as follows: Connected components are extracted from the 850nm near-infrared light channel image, and their aspect ratio and texture are analyzed to screen candidate regions with filamentous or mesh-like morphology. The motion trajectory of these candidate regions is tracked in consecutive frame images, and their settling velocity is calculated. Trajectories conforming to slow settling patterns are selected, and the ratio of the scattered signal in the 450nm and 850nm channels for these candidate regions is calculated. Based on a pre-trained fusion classifier, the morphological features, motion trajectory features, and spectral ratio features are fused and judged, and the probability and confidence level of being a flocculent anomaly are output. In one implementation, the time interval between consecutive frame images is set to 50ms, and the settling trajectory must meet the requirements of continuity, directional consistency, and velocity stability. The sedimentation trajectory of the candidate region must continuously cover ≥8 frames (corresponding to a time length of 400ms), without any breaks, disappearances or reappearances, and the angle between the trajectory direction of the consecutive 8 frames and the vertical downward direction of gravity must be <15°. The sedimentation velocity of the consecutive 8 frames must fall within the corrected floc velocity threshold range, which is set at 0.1−0.5mm / s (based on actual measurements of clinical infusion samples).

[0039] Step 5: Personalized Early Warning Threshold Calculation and Level Determination: The results output from each branch enter the fusion decision layer. Based on the current infusion drug information, a basic early warning threshold is determined. Combined with the individualized diagnosis and treatment data obtained in Step 1, a personalized early warning threshold for the current patient's abnormal precipitation, discoloration, and flocculent matter is calculated. Personalized early warning threshold = basic early warning threshold × (1 + α × A + β × D + γ × S). Where α is the age weight coefficient; β is the disease weight coefficient; and γ is the treatment stage weight coefficient. These three weight coefficients are fixed coefficients preset based on statistical analysis of a large number of clinical samples and consensus among medical experts. A represents the risk coefficient mapped based on the patient's actual age. Based on the varying tolerance to infusion abnormalities among different age groups, the system presets: A=0.3 for patients ≤14 years or ≥65 years old; A=0.1 for patients 15-64 years old. D represents the risk coefficient mapped based on the patient's disease type. Based on the varying tolerance to infusion abnormalities and complication risks among patients with different diseases, the system presets: D=0.4 for severe illnesses such as immunodeficiency diseases; D=0.1 for common diseases; D=0.2 for postoperative diseases. S represents the risk coefficient mapped based on the patient's treatment stage. Based on the varying physical condition and sensitivity to infusion abnormalities among patients at different treatment stages, the system presets: S=0.3 for the acute phase; S=0.1 for the recovery phase; S=0.05 for the maintenance phase.

[0040] The identification results from step 4 are compared with the personalized early warning threshold, and the overall confidence level is used to determine the early warning level and trigger the corresponding graded early warning action. The determination of the early warning level and the graded early warning action are the same as in the intelligent monitoring and early warning system for abnormal infusion fluids, and will not be described again.

[0041] Step 6: Medical Data Interaction: The system will bind the identification results (slight precipitation, slight discoloration, etc.), warning level, trigger time, environmental parameters, patient information and subsequent medical staff treatment records to the hospital quality control platform after binding the patient identifier and timestamp, and generate an infusion safety quality control file.

[0042] The following example illustrates the abnormal flocculent matter in an adult patient undergoing intravenous infusion. Patient Wang (52 years old) is recovering from gastrointestinal surgery. The current infusion is "ceftazidime injection," and the infusion environment is a temporary infusion point in a corridor (light intensity 520 lux, temperature 25.5℃). Due to the movement of people, the ambient light fluctuates, and postoperative patients have low tolerance for abnormal medication, requiring precise monitoring of flocculent matter and other abnormalities.

[0043] (1) System initialization and benchmark calibration: Medical staff attach the multispectral sensing module to the lower side of the infusion container and the outer wall of the liquid column area of ​​the lower section of the Mofe's drip tube (i.e., the section of the tube below the drip tube where the liquid is continuously filled, the liquid flow in this area is stable and not affected by the droplets falling from above and the fluctuation of the gas-liquid interface), and start the system; the multispectral sensing module collects the multi-channel benchmark values ​​of clear ceftazidime injection: 450nm channel 0.93, 550nm channel 0.89, 650nm channel 0.86, 850nm channel 0.91 (unit: absorbance), and stores them as the benchmark reference for model recognition; the environmental adaptive compensation module collects the environmental benchmark parameters: light intensity 520 lux, temperature 25.5℃, and generates the environmental compensation coefficient simultaneously; the patient data linkage module connects to the HIS system to obtain individualized data: age 52 years old, disease: post-gastrointestinal surgery (recovery period), current drug information (ceftazidime, incompatibility reminder: avoid mixing with calcium-containing solutions), and the data is synchronized to the LSTM-Transformer model of the data fusion processing module.

[0044] (2) The multispectral sensing module collects real-time optical signals at a frequency of 10Hz. The real-time signal is collected after 15 minutes of infusion. The environmental adaptive compensation module detects the current environmental parameters: illumination 530 lux (exceeding the 500 lux threshold) and temperature 25.6℃. It then initiates environmental interference correction: ① Linear attenuation correction is performed on the 450nm (precipitate / flocculent scattering channel), 550nm, and 650nm visible light channel signals, with a correction coefficient of 0.98 (derived from the light intensity fitting). ② The temperature deviates from the baseline (25℃) by 0.6℃, and compensation is performed on the 850nm near-infrared channel signal. After calibration, the final signal group is synchronized to the data fusion processing module.

[0045] (3) The flocculent identification branch of the LSTM-Transformer model performs feature extraction on the 450nm scattering signal and the 850nm near-infrared signal. The steps are as follows: Step 1: Signal Preprocessing: The model normalizes the input calibration signal to obtain a normalized signal; simultaneously, it calls the reference signal from the initialization phase to calculate the signal difference, highlighting anomalous signal components. Bandpass filtering (passband frequency 0.1Hz to 4Hz) is applied to the original timing signals of each channel to eliminate regular noise introduced by the periodic dripping of the Moffield dropper, while retaining non-periodic signal changes caused by anomalous events.

[0046] Step 2: Static Feature Extraction: ① Based on the 450nm scattering signal time series, the scattering coefficient (0.72) is extracted through a one-dimensional temporal convolutional layer (with a kernel size of 3, i.e., a weighted moving average of the absorbance values ​​of 3 consecutive sampling points). Then, the proportion of the duration of abnormal scattering signals is calculated through an adaptive threshold segmentation algorithm (threshold 0.65) (i.e., the ratio of the number of sampling points with scattering coefficients exceeding the threshold to the total number of sampling points within the sampling window; this ratio is the core static feature of the flocculent); ② Based on the 850nm near-infrared time series signal, the scattering features of the flocculent are extracted through signal gradient abrupt change detection: signal rising slope, number of consecutive sampling frames of abnormal signals, and deviation of the peak scattering intensity from the baseline value; ③ The ratio of the scattering coefficients at 450nm to those at 850nm is calculated.

[0047] Step 3: Dynamic Feature Extraction: The model calls the LSTM layer to perform time-series analysis on the 450nm and 850nm calibration signals of 10 consecutive frames to capture the time-series attenuation trend of the flocculent signal: the attenuation rate of the scattered signal intensity between consecutive frames is calculated by the signal differential analysis method, and it is determined that the signal has a "monotonically attenuating" trend (which is consistent with the law that the scattered signal in the monitoring area weakens frame by frame due to the settling of flocculents under the action of gravity, and excludes the interference of signal enhancement frame by frame caused by the rising of bubbles), and generates a time-series feature vector.

[0048] (4) LSTM-Transformer model fusion recognition: The static and dynamic feature vectors of the flocculent matter identification branch, along with the features of the sedimentation and discoloration branches (sedimentation: no abnormality; discoloration: no abnormality), are input into the model fusion decision layer to perform the following steps: Step 1: Feature weighted encoding: The Transformer layer uses an attention mechanism to combine individualized patient data (postoperative recovery period, low tolerance to flocculent material, weight coefficient increased to 0.2) and current drug information (ceftazidime easily produces flocculent material, weight coefficient 0.3) to weight the flocculent material feature vector, thus obtaining a weighted feature vector; Step 2: Model Inference and Confidence Calculation: The weighted feature vector is input into the fully connected layer and matched with the pre-trained sample library (containing 1000+ ceftazidime flocculent samples). The output recognition result is: anomaly type + feature value + confidence (e.g., "flocculent anomaly exists"; filamentous region accounts for 4.8%, confidence: 0.92). Step 3: Output of comprehensive results: The final identification result of the model is "only flocculent abnormality exists", with a comprehensive confidence level of 0.92.

[0049] (5) The graded early warning module presets a basic early warning threshold for ceftazidime flocculent material (3% of the filamentous area and scattering coefficient of 0.68); and calculates a personalized early warning threshold: α=0.2, β=0.3, γ=0.1; A=0.1 (15-64 years old), D=0.2 (postoperative recovery period, disease weight), S=0.1 (recovery period), personalized early warning threshold = 3%÷(1+0.2×0.1+0.3×0.2+0.1×0.1)=3%÷1.09≈2.75%. This threshold is 3% lower than the basic threshold, indicating that the early warning sensitivity of patients in the postoperative recovery period is higher than that of ordinary patients. The early warning is triggered when the proportion of abnormal scattering signal duration reaches 2.75%, realizing earlier and more sensitive abnormal detection for high-risk groups.

[0050] Comparison and judgment: Personalized threshold 2.75% < percentage of abnormal scattering signal duration 4.8% < 10% (fixed threshold), confidence level 0.92 ≥ 0.8 (basic confidence level), triggering a level 2 warning; Since there is no precipitation or discoloration abnormality, the comprehensive confidence level 0.92 ≥ 0.9 (preset comprehensive confidence level), generating the instruction "Confirm the existence of abnormality, change the medicine immediately".

[0051] The medical data interaction module binds the hospital number and timestamp to "abnormal flocculent matter, 4.8% proportion of filamentous area, scattering coefficient 0.72, identification confidence level 0.92, level 2 warning, trigger time, environmental parameters, and postoperative patient information" and uploads them to the HIS system and quality control platform. When medical staff arrive at the scene, they confirm that the medication has produced flocculent matter due to prolonged storage after mixing. They then change the medication and record the treatment process. The treatment data is archived simultaneously, generating a traceable infusion safety quality control file.

[0052] The present invention also discloses an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the above-mentioned intelligent monitoring and early warning method for abnormal infusion fluids. The memory is used to store the intelligent monitoring and early warning system for abnormal infusion fluids.

[0053] The present invention also discloses a computer-readable storage medium comprising a stored computer program, wherein the computer program, when running, controls the device containing the computer-readable storage medium to execute the above-mentioned intelligent monitoring and early warning method for abnormal infusion fluids.

[0054] This invention, by establishing communication connections between a multispectral sensing module, an environmental adaptive compensation module, a patient data linkage module, a data fusion processing module, a graded early warning module, and a medical data interaction module, can accurately identify and provide personalized graded early warnings for three abnormal types: precipitation, discoloration, and flocculent matter. This ensures the personalization (thresholds vary from person to person) and accuracy (false signals are filtered out through confidence levels), thereby improving the safety of the infusion process. Furthermore, by linking the HIS system and the quality control platform, it enables data traceability for monitoring, early warning, and treatment, creating a traceable infusion safety quality control record, which greatly assists in medical quality control management.

[0055] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. An intelligent monitoring and early warning system for abnormal infusion fluids, characterized in that, The system includes: Multispectral sensing module: used to attach to the outer wall of the infusion container and the outer wall of the lower liquid column area of ​​the Mofe's dropper to collect multispectral signals of the infusion liquid; Environmental adaptive compensation module: It is communicatively connected to the multispectral sensing module, and has built-in ambient light sensor and temperature sensor to collect environmental parameters and perform interference correction on the collected multispectral signals; Patient data linkage module: It is equipped with a standardized interface for interfacing with the hospital's HIS system to automatically acquire individualized patient diagnosis and treatment data, including age, disease type, treatment stage data, and information on the current infusion solution; Data fusion processing module: used to receive corrected multispectral signals and individualized diagnosis and treatment data, and perform the following operations: based on the multispectral signals, through a pre-trained fusion recognition model, through feature extraction and fusion recognition, determine whether there is at least one abnormality among precipitation, discoloration and flocculent matter in the infusion fluid, and output the abnormality recognition result and comprehensive confidence level; The graded early warning module communicates with the data fusion processing module and is used to dynamically calculate personalized early warning thresholds based on the identification results output by the data fusion processing module, combined with the current infusion fluid information and individualized diagnosis and treatment data; it compares the model identification results with the personalized early warning thresholds, and judges the abnormal type of infusion fluid and the corresponding early warning level based on the comprehensive confidence level, triggering graded early warning actions corresponding to the early warning level; Medical data interaction module: It communicates with the graded early warning module, the hospital HIS system, and the hospital quality control platform to upload monitoring and early warning data.

2. The intelligent monitoring and early warning system for abnormal infusion fluids according to claim 1, characterized in that, The multispectral sensing module includes at least four spectral sensors: a first spectral sensor with an operating wavelength of 450±10nm, used to detect scattering signals from sediments and flocculents; a second spectral sensor with an operating wavelength of 550±10nm, used to detect transmission or reflection signals of the liquid in the yellow spectral region; a third spectral sensor with an operating wavelength of 650±10nm, used to detect transmission or reflection signals of the liquid in the red spectral region; and a fourth spectral sensor with an operating wavelength of 850±20nm, used to acquire the intensity characteristics of the scattering signals from flocculents.

3. The intelligent monitoring and early warning system for abnormal infusion fluids according to claim 2, characterized in that, The data fusion processing module includes a lightweight multi-branch neural network model for feature screening and classification of the corrected multispectral time-series signal; the lightweight multi-branch neural network model includes three feature extraction branches corresponding to the identification of precipitation, discoloration and flocculent matter, respectively, and a fusion decision layer, which integrates the output of each branch, the current drug solution type and individualized diagnosis and treatment data, and outputs the final anomaly identification result and comprehensive confidence level. The anomaly identification results include the anomaly type, feature value, and confidence level for each anomaly type branch; the overall confidence level = precipitation confidence level × precipitation weight + color change confidence level × color change weight + flocculent confidence level × flocculent weight; the precipitation weight, color change weight, and flocculent weight are all determined based on their historical infusion data, drug characteristics, and patient-specific diagnosis and treatment data.

4. The intelligent monitoring and early warning system for abnormal infusion fluids according to claim 3, characterized in that, The precipitation feature extraction branch is used to receive the time-series signal from the first spectral sensor and extract the scattering coefficient, the proportion of abnormal duration of the scattering signal, and time-series variation features. The color change feature extraction branch is used to receive signals from the second and third spectral sensors, calculate the color difference value, and determine the color change tendency; the flocculent feature extraction branch is used to receive time-series signals from the first and fourth spectral sensors, and fuse and extract the flocculent scattering signal fluctuation characteristics, signal intensity time-series attenuation trend characteristics, and the ratio of scattering coefficients at 450nm to 850nm.

5. The intelligent monitoring and early warning system for abnormal infusion fluids according to claim 4, characterized in that, The graded early warning module has a preset basic early warning threshold based on the type of medication. The personalized early warning threshold is: Personalized early warning threshold = Basic early warning threshold ÷ (1 + α × A + β × D + γ × S), where α is the age weight coefficient; β is the disease weight coefficient; γ is the treatment stage weight coefficient; α, β, and γ are fixed coefficients preset based on clinical sample statistical analysis and medical expert consensus, and their values ​​range from 0 to 1; A is the risk coefficient mapped according to the patient's actual age; D is the risk coefficient mapped according to the patient's disease type; S is the risk coefficient mapped according to the patient's treatment stage, where A = 0.3 for ages ≤ 14 years or ≥ 65 years, and A = 0.1 for ages 15-64 years; D = 0.4 for severe cases, D = 0.1 for common diseases, and D = 0.2 for postoperative diseases; S = 0.3 for the acute phase, S = 0.1 for the recovery phase, and S = 0.05 for the maintenance phase. Therefore, the higher the patient's risk, the larger A, D, and S are, the lower the personalized early warning threshold is, and the higher the system sensitivity is, enabling high-risk individuals to trigger early warnings even in the slightest abnormal state.

6. The intelligent monitoring and early warning system for abnormal infusion fluids according to claim 5, characterized in that, The tiered early warning module includes at least the following warning levels: Level 1 warning: triggered when the basic warning threshold < the identified abnormal feature value < the corresponding personalized warning threshold, and the confidence level of the abnormal type ≥ the basic confidence level; Level 2 warning: triggered when the corresponding personalized warning threshold ≤ the identified abnormal feature value < the fixed threshold, and the confidence level of the abnormal type ≥ the basic confidence level; Level 3 warning: triggered when the identified abnormal feature value ≥ the fixed threshold, or when multiple types of abnormalities exist, each abnormal feature value ≥ its corresponding personalized warning threshold, and the confidence level of each abnormal type is greater than the fixed confidence level, while the overall confidence level ≥ the preset overall confidence level; wherein, the basic confidence level is obtained by training based on historical clinical data, the fixed threshold is the preset highest-level feature value red line, the fixed confidence level is the preset highest-level confidence level red line, and the preset overall confidence level is the preset highest-level overall confidence level.

7. The intelligent monitoring and early warning system for abnormal infusion fluids according to claim 6, characterized in that, When a Level 2 warning is triggered, if the overall confidence level is higher than the preset overall confidence level, the system outputs a confirmation of an anomaly and an immediate medication change instruction; if the overall confidence level is lower than the preset overall confidence level but higher than the baseline overall confidence level, the system outputs a suspected anomaly and an immediate medication check instruction; when a Level 1 warning is triggered, if the overall confidence level is greater than the observation threshold, the system automatically upgrades the monitoring level and outputs a low-priority prompt; where the observation threshold is a preset observation level standard; and the baseline overall confidence level is a comprehensive risk probability standard for the presence of an anomaly in the infusion fluid, based solely on the general characteristics of the medication and clinical risks.

8. The intelligent monitoring and early warning system for abnormal infusion fluids according to claim 1, characterized in that, The medical data interaction module is used to bind the identification results, warning level, trigger time and subsequent medical staff treatment records of the data fusion processing module with the patient identifier and timestamp, and then upload them to the hospital quality control platform to generate a traceable infusion safety quality control file.

9. A method for intelligent monitoring and early warning of abnormal infusion fluids based on the intelligent monitoring and early warning system according to any one of claims 1-8, characterized in that, The method includes the following steps: Step 1: System initialization and benchmark calibration: The multispectral sensing module is attached to the lower side of the infusion container and the outer wall of the lower section of the liquid column area of ​​the Mofe's drip tube to collect multi-channel optical benchmark values ​​and environmental benchmark parameters of the clear infusion liquid. At the same time, the individualized diagnosis and treatment data and treatment stage data of the target patient are obtained through the hospital HIS system interface. Step 2: Multispectral signal acquisition and environmental calibration: The absorbance time-series signal of each channel of the infusion fluid is acquired at a fixed sampling frequency using a multispectral sensing module, and the noise interference introduced by the periodic fluctuation of the droplets is eliminated by bandpass filtering. The optical signal is calibrated by environmental adaptive compensation in combination with the current environmental parameters. Step 3: Anomaly feature extraction: Based on the calibrated multi-channel absorbance time-series signal, extract the features of precipitation, discoloration and flocculent matter respectively; among which, the flocculent matter features include the proportion of the duration of abnormal signals in the fluctuation features of the scattering signal based on the near-infrared channel time-series signal, and / or the analysis of the temporal decay trend of the intensity of abnormal signals in the continuous sampling frame signal; Step 4: Abnormal Feature Fusion and Identification: Input the extracted feature data into the pre-trained fusion and identification model to identify and output whether there are abnormalities such as precipitation, discoloration or flocculent matter, and output the identification results and comprehensive confidence level. Step 5: Personalized early warning threshold calculation and level determination: Determine the basic early warning threshold based on the current infusion information. Combined with the individualized diagnosis and treatment data obtained in Step 1, calculate the personalized early warning threshold for the current patient using the formula: Personalized early warning threshold = Basic early warning threshold ÷ (1 + α × A + β × D + γ × S). Compare the identification results from Step 4 with the personalized early warning threshold, determine the early warning level based on the comprehensive confidence level, and trigger the corresponding graded early warning action. Step 6: Upload the entire process data of each monitoring, identification, early warning and follow-up treatment to the hospital quality control platform.

10. The method according to claim 9, characterized in that, The fusion recognition model adopts a lightweight multi-branch neural network model, LSTM-Transformer, which includes: sedimentation, discoloration and flocculent recognition branches; the recognition method is: feature screening and classification of the corrected multispectral time series signal; each branch outputs the feature value and confidence level of sedimentation, discoloration and flocculent anomalies according to its corresponding features; The results output from each branch enter the fusion decision layer. Combining individualized diagnosis and treatment data with information on the current infused medication, personalized early warning thresholds for abnormal precipitation, discoloration, and flocculent matter are calculated: Personalized early warning threshold = Basic early warning threshold ÷ (1 + α × A + β × D + γ × S), where α is the age weight coefficient; β is the disease weight coefficient; γ is the treatment stage weight coefficient; α, β, and γ are fixed coefficients preset based on clinical sample statistical analysis and medical expert consensus; A is the risk coefficient mapped according to the patient's actual age; D is the risk coefficient mapped according to the patient's disease type; S is the risk coefficient mapped according to the patient's treatment stage, where A = 0.3 for ages ≤ 14 years or ≥ 65 years; A = 0.1 for ages 15-64 years; D = 0.4 for severe cases, D = 0.1 for common diseases, and D = 0.2 for postoperative diseases; S = 0.3 for the acute phase, S = 0.1 for the recovery phase, and S = 0.05 for the maintenance phase.

Images