Real-time big data integration and analysis method for medical internet of things

Abstract

The present application belongs to the technical field of medical internet of things and real-time big data analysis, and particularly relates to a real-time big data integration analysis method for medical internet of things. By continuously obtaining physiological signal data stream from medical internet of things equipment, an adaptive data analysis window mechanism is adopted, that is, a dynamic comparison and data accumulation are carried out based on a real-time extracted feature set and a preset composite analysis condition until the condition is met or the maximum window is reached, and then a high-precision abnormal state analysis is triggered. The present application realizes earlier and more accurate detection of abnormal physiological state of patients, and through correlation of analysis results, intervention measures and response data, an evolvable clinical evidence chain is constructed, and finally a systematized knowledge capable of continuous optimization is formed, thereby providing a high-quality data processing and analysis basis for personalized clinical intervention and intelligent management of hospitals.

Description

Technical Field

[0001] This invention belongs to the field of medical Internet of Things and real-time big data analysis technology, specifically relating to a real-time big data integration and analysis method for medical Internet of Things. Background Technology

[0002] In the fields of medical IoT and big data analytics, real-time processing of continuous physiological data streams generated by medical IoT terminals such as bedside monitoring devices and wearable sensors is crucial for achieving intelligent early warning and early clinical intervention for high-risk patients. This type of data is characterized by high frequency, continuity, and multi-dimensionality. Its effective integration and real-time analysis aim to capture clinically significant abnormal patterns from massive amounts of fluctuating information, thereby providing decision support for the early identification and intervention of emergencies. This is expected to reduce the rate of emergency surgeries due to delayed diagnosis and accumulate high-quality clinical diagnostic evidence.

[0003] However, existing medical IoT data analysis methods face a series of prominent contradictions when addressing the aforementioned demands for real-time performance, accuracy, and systematization. First, regarding real-time performance, traditional methods often employ fixed time windows for data slicing and analysis, making it difficult to adapt to the inherent physiological rhythms of the data and the speed of abnormal signal evolution, easily leading to delayed warnings or insufficient sensitivity. Second, in terms of analytical accuracy and reliability, simple threshold alarms are susceptible to transient noise interference, while complex analytical models struggle to balance computational efficiency with robustness to data quality (such as transient signal loss and artifacts) in real-time streaming processing. Furthermore, the system lacks an effective knowledge evolution mechanism; once deployed, the analytical logic and model parameters become relatively fixed, unable to continuously and securely self-optimize and update based on constantly emerging clinical intervention results and new data patterns, making it difficult to form a closed loop of transformation from data to clinical evidence and then to systematic knowledge. In addition, at the data processing level, ensuring the secure and reliable transmission of real-time data streams from the terminal to the analysis engine while meeting low latency requirements also presents technical challenges.

[0004] Existing technologies often struggle to simultaneously and satisfactorily resolve the contradictions between real-time response and high-precision analysis, secure data transmission and low-latency processing, and static analysis models and the demands of dynamically evolving clinical knowledge in practical applications of real-time big data analytics in the medical IoT. Therefore, there is an urgent need for a novel data integration and analysis method that can balance real-time performance, security, analytical accuracy, and the system's sustainable evolution capabilities, thereby providing a more solid technological foundation for personalized medicine and hospital management. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution: A real-time big data integration and analysis method for the Internet of Things in healthcare, characterized by comprising: The system continuously acquires real-time data streams representing the patient's physiological state from medical IoT devices, performs real-time feature extraction to obtain a first feature set, and determines whether the first feature set meets preset real-time analysis conditions. The real-time analysis conditions are comprehensive conditions describing the abnormal significance and analysis reliability of the first feature set. If the first feature set does not meet the instant analysis conditions, then data for a subsequent set duration is obtained from the real-time data stream, and the subsequent data is spliced ​​with the current segment of the real-time data stream to form an extended data segment. Feature extraction is performed on the extended data segment to obtain a second feature set, and it is determined whether the second feature set meets the instant analysis conditions; If the second feature set does not meet the real-time analysis conditions and the cumulative duration of the extended data segment is less than the preset maximum analysis window duration, then return to the step of obtaining data for a subsequent set duration from the real-time data stream and concatenating the subsequent data with the current extended data segment to form a new extended data segment; If the second feature set meets the real-time analysis conditions, or the cumulative duration of the extended data segment reaches the maximum analysis window duration, then based on the current first or second feature set that meets the conditions, or based on the feature set corresponding to the extended data segment that has reached the maximum analysis window duration, high-precision anomaly analysis is performed to generate analysis results containing anomaly type, severity level and predicted trajectory. The analysis results are pushed to the clinical intervention system in real time.

[0006] Further, after performing real-time feature extraction on the acquired current segment of real-time data stream to obtain a first feature set, and determining whether the first feature set meets preset real-time analysis conditions, the process also includes: If the first feature set meets the real-time analysis conditions, then the high-precision abnormal state analysis is performed based on the first feature set to generate a first analysis result, and the first analysis result is pushed to the clinical intervention system in real time.

[0007] Furthermore, the process of determining whether the first feature set or the second feature set meets the preset real-time analysis conditions includes: Extract the dynamic trend features and stability quantification features of core physiological indicators from the first feature set or the second feature set; The dynamic trend feature is compared with a preset abnormal trend threshold, and the stability quantification feature is compared with a preset stability threshold. When the dynamic trend feature exceeds the abnormal trend threshold and the stability quantification feature is lower than the stability threshold, it is determined that the current feature set meets the instant analysis conditions. When the dynamic trend feature does not exceed the abnormal trend threshold, or the stability quantification feature is not lower than the stability threshold, it is determined that the current feature set does not meet the instant analysis conditions. The core physiological indicators include at least heart rate variability, rate of decrease in blood oxygen saturation, and respiratory rhythm disorder index.

[0008] Furthermore, before the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices, it also includes: Based on the pathophysiological characteristics of the target monitored diseases, determine the length of the basic time window used for initial feature analysis; Based on the basic time window length and the data sampling frequency of the medical IoT device, the initial data volume corresponding to the current segment of real-time data stream is calculated; Specifically, determining the basic time window length includes: obtaining the typical physiological parameter fluctuation cycle of the target monitored disease in the early compensatory stage, and setting an integer multiple of the fluctuation cycle as the basic time window length.

[0009] Furthermore, before the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices, it also includes: Based on the clinical risk stratification of the target patient group, set personalized analysis trigger sensitivity levels; The preset maximum analysis window duration is dynamically adjusted based on the analysis trigger sensitivity level. The higher the analysis trigger sensitivity level, the shorter the corresponding maximum analysis window duration. The adjustment of the maximum analysis window duration is based on the principle that a shorter window is used for high-risk patient groups to pursue analysis timeliness, while a longer window is used for low-risk patient groups to obtain more sufficient feature information for discrimination.

[0010] Furthermore, based on the current first or second feature set that meets the conditions, or based on the feature set corresponding to the extended data segment that has reached the maximum analysis window duration, the process of performing high-precision anomaly state analysis includes: The feature set is input into a pre-constructed anomaly analysis network; Through multi-level abstraction processing of the anomaly analysis network, anomaly probability vectors and state evolution vectors are output. By fusing the anomaly probability vector and the state evolution vector, the analysis results containing the anomaly type, severity level, and predicted trajectory are generated.

[0011] Furthermore, after pushing the analysis results to the clinical intervention system in real time, the process also includes: Continuously monitor the patient's physiological state in response to the intervention measures corresponding to the pushed analysis results; The analysis results, the intervention measures, and the response data are stored together to form a complete clinical decision evidence unit. The stored clinical decision evidence units are periodically aggregated and pattern-mined to generate systematic knowledge for updating the real-time analysis conditions or the high-precision abnormal state analysis parameters.

[0012] Furthermore, the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices specifically includes: At the medical IoT device end, the raw physiological sensor signals are subjected to localized preliminary filtering and analog-to-digital conversion to generate time-series data packets in a standard format; The time-series data packets are sent from the device to the edge computing node at a first preset frequency through a dedicated secure data transmission channel. At the edge computing node, multiple received time-series data packets are buffered and aligned at a second preset frequency, where the second preset frequency is higher than the first preset frequency. When the required amount of data for the current segment of real-time data stream is met, valid data segments are extracted from the verified time-series data packets and assembled into the real-time data stream for subsequent processing. The proprietary secure data transmission channel employs an end-to-end encryption mechanism based on temporary session keys, and the edge computing node destroys the temporary session key immediately after completing data assembly.

[0013] Furthermore, the construction and updating process of the pre-built anomaly analysis network is independent of the execution process of the real-time big data integration and analysis method, specifically including: The anomaly analysis network was initially trained offline based on historical desensitized clinical data. During the operation of the real-time big data integrated analysis method, the structure and parameters of the anomaly analysis network remain frozen; Based on the generated systematic knowledge, incremental updates and evaluations of the anomaly analysis network are periodically triggered; The updated anomaly analysis network version will only be switched to the online service version if the incremental update evaluation passes the preset verification criteria.

[0014] Furthermore, the method also includes: A unique process traceability identifier is generated for each execution of the high-precision anomaly analysis process; Record the following metadata associated with the process traceability identifier: The snapshot of the feature set that triggered the analysis, the cumulative duration of the extended data segment used, the version of the core parameters called during the analysis, and the final analysis results pushed out; All process traceability identifiers and their associated metadata are stored in an immutable audit log for quality review and algorithm performance backtracking.

[0015] This invention belongs to the field of medical IoT and real-time big data analytics, specifically relating to a real-time big data integration and analysis method for medical IoT. It continuously acquires physiological signal data streams from medical IoT devices and employs an adaptive data analysis window mechanism. This mechanism dynamically compares and accumulates data based on real-time extracted feature sets and preset composite analysis conditions until the conditions are met or the maximum window is reached, thereby triggering high-precision abnormal state analysis. This invention enables earlier and more accurate detection of abnormal physiological states in patients. By correlating analysis results, intervention measures, and response data to construct an evolutionary chain of clinical evidence, it ultimately forms a continuously optimized systematic knowledge base, providing a high-quality data processing and analysis foundation for personalized clinical intervention and intelligent hospital management. Attached Figure Description

[0016] Figure 1 The present invention requests protection for a flowchart of a real-time big data integration and analysis method for the medical Internet of Things; Figure 2 The present invention requests protection for a second flowchart of a real-time big data integration and analysis method for the medical Internet of Things; Figure 3 The present invention requests protection for a third flowchart of a real-time big data integration and analysis method for the medical Internet of Things. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a real-time big data integration and analysis method for the medical Internet of Things, comprising: The system continuously acquires real-time data streams representing the patient's physiological state from medical IoT devices, performs real-time feature extraction to obtain a first feature set, and determines whether the first feature set meets preset real-time analysis conditions. The real-time analysis conditions are comprehensive conditions describing the abnormal significance and analysis reliability of the first feature set. If the first feature set does not meet the conditions for real-time analysis, then data for the subsequent set duration is obtained from the real-time data stream, and the subsequent data is spliced ​​with the current segment of the real-time data stream to form an extended data segment. Feature extraction is performed on the extended data segment to obtain a second feature set, and it is determined whether the second feature set meets the conditions for real-time analysis. If the second feature set does not meet the conditions for real-time analysis and the cumulative duration of the extended data segment is less than the preset maximum analysis window duration, then return to the step of obtaining data for the subsequent set duration from the real-time data stream and concatenating the subsequent data with the current extended data segment to form a new extended data segment. If the second feature set meets the conditions for immediate analysis, or the cumulative duration of the extended data segment reaches the maximum analysis window duration, then based on the current first or second feature set that meets the conditions, or based on the feature set corresponding to the extended data segment that has reached the maximum analysis window duration, high-precision anomaly analysis is performed to generate analysis results containing anomaly type, severity level and predicted trajectory. The analysis results are pushed to the clinical intervention system in real time.

[0021] In this embodiment, raw physiological signals are continuously and synchronously acquired from one or more medical IoT sensing devices deployed on the patient's body; the raw physiological signals include at least electrocardiogram waveforms, photoplethysmography pulse waves, and impedance respiratory waves; while acquiring the raw physiological signals, anti-power frequency interference filtering and baseline drift correction are performed on the raw physiological signals to generate a pre-purified continuous digital signal sequence. Using a configurable initial time window as a unit, the signal data of the current segment is extracted from the continuous digital signal sequence. For the current segment of signal data, a real-time feature extraction operation is performed: First, all complete cardiac cycles in the signal segment are identified, and the standard deviation of adjacent cardiac intervals is calculated as the first feature. Second, in the photoplethysmography (PPG) signal segment, the trough of each pulse wave is located, and the average slope of amplitude decrease between consecutive troughs is calculated as the second feature. Finally, spectral analysis is performed on the impedance respiratory wave signal segment, and the ratio of the dominant frequency band energy to the total energy is extracted as the third feature. The first, second, and third features are combined to form the first feature set. The first feature set is input into a preset composite logic judge for real-time analysis condition compliance judgment: the composite logic judge has a built-in first threshold list and a second threshold list; during the judgment, the first feature is compared with the heart rate variability abnormality threshold in the first threshold list, the second feature is compared with the blood oxygen decline rate threshold in the first threshold list, and the third feature is compared with the respiratory rhythm stability threshold in the second threshold list; when the first feature exceeds its corresponding threshold, the second feature exceeds its corresponding threshold, and the third feature is lower than its corresponding threshold, it is determined that the real-time analysis condition is met; if the above three conditions are not met at the same time, it is determined that the condition is not met. If the system determines that the conditions for real-time analysis are not met, it immediately starts the data accumulation mode: it continues to extract a segment of subsequent signal data of a fixed duration from the continuous digital signal sequence; it then aligns and splices the subsequent signal data with the previously accumulated current segment of signal data on the time axis to form a new, longer signal data segment as an extended data segment; and it re-executes the same real-time feature extraction operation as described above on the extended data segment to obtain a second feature set covering the same type of features. The second feature set is input again into the preset composite logic judge, and the judgment is made according to the same threshold list and logic rules; If the second feature set is still determined not to meet the conditions for real-time analysis, the system checks the total duration of the currently accumulated extended data segments; compares the total duration with a predefined maximum allowable analysis duration; if the total duration is less than the maximum allowable analysis duration, the system automatically returns to the execution data accumulation mode, that is, it again extracts and splices subsequent signal data of a fixed duration, extracts features from the new extended data segments and makes judgments, and this process is repeated. If, during the loop, the feature set extracted in a certain round is determined by the composite logic judge to meet the immediate analysis conditions, the loop will immediately exit; or, if the accumulated total time reaches or exceeds the maximum allowed analysis time during the loop, the loop will be forcibly terminated; when one of the above two situations occurs, the system enters the high-precision abnormal state analysis stage. In the high-precision abnormal state analysis phase, the system locks the final feature set that triggered this phase, or, when triggered by reaching the maximum duration, locks the final feature set extracted from the extended data segment corresponding to that duration. The system dynamically compares this final feature set with the patient's historical feature baseline to calculate the deviation. Simultaneously, it matches this final feature set with a pre-stored library of typical feature templates corresponding to different clinical emergency patterns, calculating a similarity score. By combining the deviation and multiple similarity scores, the system infers the most likely type of abnormal physiological state, assesses the quantitative level of its severity, and extrapolates its short-term evolution trajectory based on the direction of change of this feature set in the most recent cycles. Finally, a structured analysis result report is packaged and generated. The system uses a dedicated message middleware on the hospital's intranet to push the analysis results messages to the designated clinical intervention system workstation in real time with the highest priority; at the same time, a delivery record with a timestamp and a unique identifier for this analysis process is generated in the system log.

[0022] Further, after performing real-time feature extraction on the acquired current segment of real-time data stream to obtain a first feature set, and determining whether the first feature set meets preset real-time analysis conditions, the process also includes: If the first feature set meets the real-time analysis conditions, then the high-precision abnormal state analysis is performed based on the first feature set to generate a first analysis result, and the first analysis result is pushed to the clinical intervention system in real time.

[0023] In this embodiment, if the first feature set is determined by the composite logic judge to meet the instant analysis conditions during the first judgment, the system will omit all data accumulation and loop judgment steps and directly enter the high-precision abnormal state analysis stage. Under this direct path, the system takes the first feature set as the final feature set and immediately executes the same analysis process as the high-precision abnormal state analysis stage: that is, it performs dynamic comparison with the patient's individual historical baseline to calculate the deviation, performs matching with the typical feature template library to calculate the similarity score, and comprehensively infers the abnormality type, severity level and evolution trajectory. After completing the analysis, the system generates a structured analysis result message in the same manner; Finally, the system also uses a dedicated message middleware to push the generated analysis result message to the clinical intervention system workstation in real time with the highest priority, and generates corresponding delivery log records.

[0024] Furthermore, the process of determining whether the first feature set or the second feature set meets the preset real-time analysis conditions includes: Extract the dynamic trend features and stability quantification features of core physiological indicators from the first feature set or the second feature set; The dynamic trend feature is compared with a preset abnormal trend threshold, and the stability quantification feature is compared with a preset stability threshold. When the dynamic trend feature exceeds the abnormal trend threshold and the stability quantification feature is lower than the stability threshold, it is determined that the current feature set meets the instant analysis conditions. When the dynamic trend feature does not exceed the abnormal trend threshold, or the stability quantification feature is not lower than the stability threshold, it is determined that the current feature set does not meet the instant analysis conditions. The core physiological indicators include at least heart rate variability, rate of decrease in blood oxygen saturation, and respiratory rhythm disorder index.

[0025] In this embodiment, the operation of extracting dynamic trend features and stability quantification features from the feature set has different specific implementation methods for different core physiological indicators: For the heart rate variability index, its dynamic trend characteristic is specifically defined as: the slope value of the standard deviation sequence of adjacent heart rate intervals calculated within the current analysis time window; the slope value is calculated by linearly fitting multiple adjacent heart rate interval standard deviation data points sampled at equal intervals within the time window, and the slope of the resulting straight line is the dynamic trend characteristic; its stability quantification characteristic is specifically defined as: the root mean square error of the adjacent heart rate interval standard deviation data points relative to the fitted straight line; For the blood oxygen saturation decline rate index, its dynamic trend characteristics are specifically defined as: the average decline rate of blood oxygen saturation estimates continuously calculated from the photoplethysmography signal within the current analysis time window; this average decline rate is obtained by calculating the linear regression slope of the blood oxygen saturation estimate sequence within this time window; its stability quantification characteristics are specifically defined as: the sample entropy of the blood oxygen saturation estimate sequence, used to quantify the unpredictability and complexity of the sequence pattern; For the respiratory rhythm disorder index, its dynamic trend characteristics are specifically defined as: the direction of change of the moving average of the proportion of respiratory dominant frequency energy within the current analysis time window, which is quantified as a discrete state value representing rising, falling or stable; its stability quantification characteristics are specifically defined as: the decay rate of the autocorrelation function of the respiratory dominant frequency energy proportion sequence after the first positive zero crossing point. The preset abnormal trend thresholds are positive for the dynamic trend feature of heart rate variability, representing the maximum allowable upward slope; negative for the dynamic trend feature of blood oxygen saturation decline rate, representing the absolute value of the maximum allowable decline rate; and a discrete set of states for the dynamic trend feature of respiratory rhythm disorder index. When the feature value falls into the set of states representing "rising", it is considered to exceed the threshold. The preset stability thresholds are: for the heart rate variability stability quantification feature, an upper limit threshold representing the maximum allowable root mean square error; for the blood oxygen saturation decline rate stability quantification feature, a lower limit threshold representing the minimum allowable sample entropy value, below which the sequence is considered too regular and may indicate an abnormality; and for the respiratory rhythm disorder index stability quantification feature, an upper limit threshold representing the maximum allowable decay rate. The comparison and judgment logic is as follows: For the three indicators extracted from the current feature set, their dynamic trend characteristics are compared with the corresponding abnormal trend thresholds, and their stability quantification characteristics are compared with the corresponding stability thresholds. For each indicator, it is only marked as "triggered state" when its dynamic trend characteristics exceed (or fall into) its abnormal trend threshold, and its stability quantification characteristics are lower than (or higher than, depending on the threshold type) its stability threshold. Finally, the composite logic judge determines that the current feature set meets the instant analysis conditions if and only if all three core physiological indicators are marked as "triggered state". If one or more indicators are not in the "triggered state", it is determined that it does not meet the conditions.

[0026] Furthermore, before the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices, it also includes: Based on the pathophysiological characteristics of the target monitored diseases, determine the length of the basic time window used for initial feature analysis; Based on the basic time window length and the data sampling frequency of the medical IoT device, the initial data volume corresponding to the current segment of real-time data stream is calculated; Specifically, determining the basic time window length includes: obtaining the typical physiological parameter fluctuation cycle of the target monitored disease in the early compensatory stage, and setting an integer multiple of the fluctuation cycle as the basic time window length.

[0027] In this embodiment, determining the length of the base time window for initial feature analysis based on the pathophysiological characteristics of the target monitored disease is a configuration process that includes clinical knowledge input: the system maintains a disease-physiological cycle mapping table, which is initialized by clinical expert knowledge; when configuring a monitoring plan for a target patient, the system first queries the mapping table for the key physiological parameters corresponding to the target monitored disease, as well as the typical fluctuation range of these parameters in the early compensatory stage of the disease; for example, for postoperative atrial fibrillation monitoring, the key physiological parameter is atrial electrical activity instability, which in its early stage manifests as irregular fluctuations in the atrial phase, with a typical cycle range of 2 to 5 minutes; the system takes the median of this range as the reference cycle; The base time window length is set as an integer multiple of the fluctuation period. The specific rules are as follows: the system sets a base multiple N, where N is an integer not less than 1, with a default value of 2; the product of the reference period and the base multiple N is calculated, and the result is used as the suggested base time window length; at the same time, the system provides an adjustment interface, allowing clinical engineers to fine-tune the suggested value within a certain range (e.g., ±20%) to adapt to different ward environmental noise levels or patient activity characteristics; the finally confirmed value is saved as the initial time window parameter for that patient analysis instance. The initial data volume is calculated based on the basic time window length and the data sampling frequency of the medical IoT device. The specific process is as follows: After the basic time window length is finally confirmed, the system reads the fixed data sampling frequency of the connected medical IoT device, which is in Hertz; the basic time window length (in seconds) is multiplied by the sampling frequency (in samples per second), and the product is the number of original data points to be collected when extracting the current segment of signal data from the continuous digital signal sequence. This number is the initial data volume. When the system starts the real-time data stream processing thread, it first loads the initial data volume parameters of the patient instance and uses them as the initial basis for buffer size allocation to ensure that the first segment of signal data used for feature extraction can be fully accommodated.

[0028] Furthermore, before the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices, it also includes: Based on the clinical risk stratification of the target patient group, set personalized analysis trigger sensitivity levels; The preset maximum analysis window duration is dynamically adjusted based on the analysis trigger sensitivity level. The higher the analysis trigger sensitivity level, the shorter the corresponding maximum analysis window duration. The adjustment of the maximum analysis window duration is based on the principle that a shorter window is used for high-risk patient groups to pursue analysis timeliness, while a longer window is used for low-risk patient groups to obtain more sufficient feature information for discrimination.

[0029] In this embodiment, personalized analysis trigger sensitivity levels are set based on the clinical risk stratification of the target patient group. Specifically, the system connects to the hospital's electronic medical record system and, after authorization, automatically reads the patient's admission diagnosis, surgical records, key laboratory test results, and past medical history. According to a preset rule engine, this information is parsed; for example, conditions such as "within 24 hours after coronary artery bypass surgery" and "age greater than 75 years and elevated serum creatinine" are marked as high-risk factors. The system accumulates the number of high-risk factors and classifies patients into "very high risk," "high risk," "intermediate risk," or "low risk" risk levels based on the accumulated number. Each risk level is pre-mapped to an analysis trigger sensitivity level, ranging from "very high sensitivity" to "low sensitivity." Based on the sensitivity level triggered by the analysis, the preset maximum analysis window duration is dynamically adjusted. The mapping relationship is stored in the form of a configuration table: for the "extremely high sensitivity" level, the mapped maximum analysis window duration is T1 seconds; for the "high risk" level, the mapped duration is T2 seconds, and T2>T1; for the "intermediate risk" level, the mapped duration is T3 seconds, and T3>T2; for the "low risk" level, the mapped duration is T4 seconds, and T4>T3; the specific values ​​of T1 to T4 are set by the system administrator based on clinical feedback. The logic behind using a shorter window to pursue timely analysis for high-risk patient groups is as follows: when a patient is classified as "extremely high-risk" or "high-risk", the system uses T1 or T2 as the maximum allowable analysis time. This means that the total time allowed for the system to accumulate signal data in the data accumulation cycle is relatively short. Once this time has been accumulated, regardless of whether the features meet the conditions, the system will force entry into the high-precision anomaly analysis stage to avoid delays in judging potential anomalies in high-risk patients due to waiting for more "ideal" features. The processing logic for using a longer window to obtain more comprehensive feature information for low-risk patient groups is as follows: when a patient is classified into the "intermediate risk" or "low risk" level, the system uses T3 or T4 as the maximum allowable analysis time. This gives the system a longer observation period, allowing for more rounds of data accumulation and feature extraction cycles, in order to obtain a more stable and discriminative feature set, thereby improving the certainty of the analysis conclusions and reducing the possibility of false alarms for low-risk patients.

[0030] Furthermore, referring to Figure 2 The process of performing high-precision anomaly analysis based on the current first or second feature set that meets the conditions, or based on the feature set corresponding to the extended data segment that has reached the maximum analysis window duration, includes: The feature set is input into a pre-constructed anomaly analysis network; Through multi-level abstraction processing of the anomaly analysis network, anomaly probability vectors and state evolution vectors are output. By fusing the anomaly probability vector and the state evolution vector, the analysis results containing the anomaly type, severity level, and predicted trajectory are generated.

[0031] In this embodiment, the pre-built anomaly analysis network is a deep neural network model that has been trained and deployed, and its input layer dimension matches the dimension of the final feature set. The operation of inputting the feature set into the pre-built anomaly analysis network is to fill the final feature set values ​​into each neuron of the input layer according to the order defined in the network input layer. The anomaly analysis network employs multi-level abstraction processing, specifically referring to the forward propagation of data within the network: the input feature values ​​undergo linear transformation and non-linear activation function processing in the first fully connected layer to generate the first layer of abstract features; these abstract features are then passed to the second fully connected layer for further transformation and abstraction, generating deeper feature representations; this process proceeds sequentially through all hidden layers until the output layer. The output layer of the anomaly analysis network is designed to output at least one anomaly probability vector and one state evolution vector. The output layer is divided into two parallel sub-output layers. The first sub-output layer has K neurons, where K is equal to the number of anomaly types defined by the system. Each neuron is processed by the Softmax function and outputs a probability value. These K probability values ​​together constitute the anomaly probability vector, representing the probability that the current feature set belongs to each anomaly type. The second sub-output layer has M neurons and outputs unnormalized real values, which constitute the state evolution vector, used to describe the coordinates of the current state in multiple preset physiological state dimensions. The subsequent decision-making logic involves fusing the anomaly probability vector and the state evolution vector to generate analysis results: First, the system identifies the anomaly type with the highest probability from the anomaly probability vector as the primary suspected type. Then, the state evolution vector is input into a predefined state-severity mapping matrix, and a severity quantification value is obtained through matrix multiplication. Simultaneously, the system queries the historical state evolution vectors of the same patient saved in the most recent analysis cycles, calculates the difference vector between the current vector and the historical vector, and extrapolates the possible trajectory of the state in the next short cycle based on the direction and magnitude of this difference vector. Finally, the primary suspected type, severity quantification value, and extrapolated trajectory described in text are encapsulated in the specified fields of the analysis result message.

[0032] Furthermore, referring to Figure 3 After pushing the analysis results to the clinical intervention system in real time, the system also includes: Continuously monitor the patient's physiological state in response to the intervention measures corresponding to the pushed analysis results; The analysis results, the intervention measures, and the response data are stored together to form a complete clinical decision evidence unit. The stored clinical decision evidence units are periodically aggregated and pattern-mined to generate systematic knowledge for updating the real-time analysis conditions or the high-precision abnormal state analysis parameters.

[0033] In this embodiment, the system continuously monitors the patient's physiological response to the intervention measures corresponding to the pushed analysis results. Specifically, after the analysis results are pushed, the system initiates a follow-up observation window lasting L minutes. During this period, the system continues to acquire physiological signals from the same medical IoT device and extracts a set of concise physiological indicator snapshots at short fixed intervals (e.g., 1 minute), including average heart rate, blood oxygen saturation, and respiratory rate. These snapshot data are labeled as "post-intervention response data". The analysis results, intervention measures, and response data are stored together in a linked manner. Specifically, the system asynchronously reads a summary of the specific intervention measures taken in response to the analysis results from the feedback interface of the clinical intervention system (if any) or the nurses manually enter the summary from the workstation. Examples of such measures include "administer oxygen via face mask at 3L / min" and "administer 20mg of diuretic via intravenous bolus." Subsequently, the system creates a new data structure called the Clinical Decision Evidence Unit. This unit contains the following associated fields: a process traceability identifier for this analysis, a complete copy of the analysis result message, a summary of the recorded intervention measures, and a snapshot of "post-intervention response data" arranged in time series. The system periodically aggregates and mines patterns from multiple stored clinical decision evidence units. Specifically, a scheduled task is set up to run every 24 hours. Once started, the task scans all clinical decision evidence units generated in the past 24 hours. First, these units are grouped according to their abnormality type. For each group, the task performs aggregation to calculate the frequency of various interventions under that type of abnormality, as well as the average improvement rate of key indicators (such as blood oxygen saturation) in the response data within the observation window after taking various measures. Then, the task performs pattern mining, using association rule analysis to find frequently occurring combination rules among "specific initial feature patterns," "specific interventions," and "good physiological responses." The system generates a structured knowledge base for updating real-time analysis conditions or high-precision anomaly analysis parameters. Specifically, it encodes the aggregated statistical results (such as frequency and improvement rate) and the mined association rules in a predefined format to generate a knowledge update report. This report is then sent to a separate "knowledge management and evaluation module." Based on the strong association rules indicated in the report, this module suggests fine-tuning a threshold (such as the blood oxygen decline rate threshold) in the composite logic judge or proposes sample feature patterns that should be focused on when retraining the anomaly analysis network.

[0034] Furthermore, the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices specifically includes: At the medical IoT device end, the raw physiological sensor signals are subjected to localized preliminary filtering and analog-to-digital conversion to generate time-series data packets in a standard format; The time-series data packets are sent from the device to the edge computing node at a first preset frequency through a dedicated secure data transmission channel. At the edge computing node, multiple received time-series data packets are buffered and aligned at a second preset frequency, where the second preset frequency is higher than the first preset frequency. When the required amount of data for the current segment of real-time data stream is met, valid data segments are extracted from the verified time-series data packets and assembled into the real-time data stream for subsequent processing. The proprietary secure data transmission channel employs an end-to-end encryption mechanism based on temporary session keys, and the edge computing node destroys the temporary session key immediately after completing data assembly.

[0035] In this embodiment, at the medical IoT device, the raw physiological sensing signal undergoes localized preliminary filtering and analog-to-digital conversion. Specifically, the microcontroller within the sensing device reads the analog voltage signal generated by the electrodes or optical sensors in real time. First, this signal passes through a hardware analog filter to remove 50Hz / 60Hz power frequency interference and its main harmonics. Then, the filtered analog signal is fed into a high-resolution analog-to-digital converter and digitized according to a preset fixed sampling frequency to generate the original discrete digital signal sequence. Finally, the device's firmware runs a lightweight software digital filter to perform a moving average processing on the original digital signal sequence to further suppress baseline drift, ultimately generating a standard-format time-series data packet containing a header, device ID, timestamp, and payload data segment. Through a dedicated secure data transmission channel, time-series data packets are sent at a first preset frequency. Specifically, during the communication initialization phase, the device and the edge computing node negotiate and generate a temporary symmetric key for this session using a key exchange protocol based on elliptic curve cryptography. For each time-series data packet to be sent, the device uses this temporary symmetric key and the AES-GCM encryption algorithm to encrypt and perform integrity authentication calculations on the payload portion of the data packet, generating ciphertext and an authentication tag. The encrypted data packets are then sent out via Wi-Fi or Bluetooth Low Energy (BLE) protocols at a first preset frequency (e.g., 10 times per second). The principle for setting the first preset frequency is that it is higher than twice the highest frequency component of the key physiological signal to satisfy the Nyquist sampling theorem. At the edge computing node, multiple received time-series data packets are buffered and aligned at a second preset frequency. The operational details are as follows: the service process on the edge computing node polls the network interface at a second preset frequency (e.g., 100 times per second), which is much higher than the first preset frequency, to receive data packets; each received data packet is immediately decrypted and its integrity is verified using the same temporary symmetric key stored during the session; after successful verification, the device ID, timestamp, and valid data are extracted; the service process maintains a first-in-first-out data buffer for each device ID and inserts the valid data into the corresponding position in the buffer according to the order of the timestamps; alignment verification refers to checking whether the timestamps of consecutive data packets in the buffer are consecutive. If a timestamp jump exceeds a certain threshold, the data packet is considered lost and recorded in the log. When the required data volume for the current segment's real-time data stream is met, valid data segments are extracted from the verified time-series data packets and assembled into a real-time data stream. The specific logic is as follows: The processing thread of the real-time data integration and analysis method requests data from the data service process of the edge computing node; the request specifies the required device ID and initial data volume; after receiving the request, the data service process checks the buffer corresponding to the device ID; only when the number of valid data points arranged in chronological order in the buffer reaches or exceeds the requested initial data volume will the service process extract a precise number of data points sequentially from the head of the buffer, assemble them into a continuous data array, and return it to the processing thread as the "signal data of the current segment"; this operation ensures that even with minor network jitter, a continuous and uninterrupted real-time data stream can be provided for the analysis method. The immediate destruction of the temporary session key by the edge computing node after data assembly means that in each session (usually defined as a connection from the start to the end of patient monitoring), when the real-time data analysis method processing thread explicitly notifies the data service process that the current monitoring and analysis task has been terminated, or when the network connection is abnormally disconnected, the data service process will actively erase the temporary symmetric key corresponding to the device ID from memory. Before the device reconnects and establishes a new session, the old key no longer exists and cannot be used to decrypt any historical communication data.

[0036] Furthermore, the construction and updating process of the pre-built anomaly analysis network is independent of the execution process of the real-time big data integration and analysis method, specifically including: The anomaly analysis network was initially trained offline based on historical desensitized clinical data. During the operation of the real-time big data integrated analysis method, the structure and parameters of the anomaly analysis network remain frozen; Based on the generated systematic knowledge, incremental updates and evaluations of the anomaly analysis network are periodically triggered; The updated anomaly analysis network version will only be switched to the online service version if the incremental update evaluation passes the preset verification criteria.

[0037] In this embodiment, the anomaly analysis network is initially trained offline based on historical desensitized clinical data. The specific process is completed in an isolated development and training environment: the training data comes from desensitized historical medical IoT data accumulated over the past few years and corresponding clinical outcomes annotated by experts; in this environment, data scientists use a deep learning framework to design the network structure, using the feature set from the historical data as input and the annotated anomaly types as supervision signals, and iteratively adjust the network parameters through the backpropagation algorithm until the model reaches the preset accuracy index on an independent validation set; the trained model parameter file is then exported. During the operation of the real-time big data integrated analysis method, the structure and parameters of the anomaly analysis network remain frozen. This means that the anomaly analysis network deployed in the production system is a read-only file with pre-loaded trained parameters. When processing real-time data, the system only performs forward propagation calculations of the network to obtain the output, and never modifies any weights or bias parameters of the network during this process. The network's computing resources are pre-allocated and isolated to ensure the stable performance of real-time analysis. Furthermore, the method also includes: A unique process traceability identifier is generated for each execution of the high-precision anomaly analysis process; Record the following metadata associated with the process traceability identifier: The snapshot of the feature set that triggered the analysis, the cumulative duration of the extended data segment used, the version of the core parameters called during the analysis, and the final analysis results pushed out; All process traceability identifiers and their associated metadata are stored in an immutable audit log for quality review and algorithm performance backtracking.

[0038] In this embodiment, after generating a knowledge update report, if the report contains suggestions for retraining the network, the "knowledge management and evaluation module" will send an evaluation task request to the model management service. Upon receiving the request, the model management service will not immediately update the production model, but will first use a batch of recently accumulated desensitized clinical data (containing new feature patterns) as a test set in the training environment to evaluate the performance indicators of the current production version model on this new data, such as precision and recall. The updated anomaly analysis network version is switched to the online service version only when the incremental update evaluation passes the preset validation criteria. The process is as follows: If the evaluation results show that the current model performance has significantly decreased (below the preset threshold), the model management service initiates a retraining process, using an augmented dataset containing new data to incrementally train the network and obtain a candidate new model. This candidate new model must outperform the current production version model in all aspects on an independent, more recent validation dataset, and pass a regression test for historical data stability to ensure that there will be no large number of misjudgments for cases that were correctly judged in the past. Only when all these validation criteria are passed will the candidate new model be approved for release. During release, the system gradually switches online real-time analysis requests to the new version model through blue-green deployment or canary release strategies, and closely monitors the quality of analysis results during the switch.

[0039] A unique process traceability identifier is generated for each high-precision anomaly state analysis process. The specific rule is that the identifier is composed of the following parts: analysis initiation timestamp (accurate to milliseconds), unique hardware ID of the source medical IoT device, and a 4-digit serial number generated by a local atomic counter. This combination ensures the global uniqueness of the identifier throughout the entire system. The metadata associated with the process traceability identifier is recorded, specifically in the following ways: When the analysis process is triggered, the system immediately assigns a traceability identifier and creates a temporary metadata record object; at different stages of the analysis process, data is populated into this object: after feature extraction is completed, a numerical snapshot of the current final feature set is saved; before entering high-precision analysis, the actual cumulative duration of the extended data segment used in this analysis is recorded; when calling the anomaly analysis network, the file hash value of the network model is recorded as its version identifier; after the analysis results are generated, a complete copy of the result message is linked to this record; all these operations are completed in memory until the analysis results are successfully pushed. All process traceability identifiers and their associated metadata are stored in an immutable audit log. Specifically, the system uses a dedicated log file or database table with append-only write capabilities to store the audit log. Once all operations of an analysis process are completed and the results have been pushed, the system serializes the metadata record object in memory, along with its process traceability identifier, into a structured log entry. The system process has only append-only write permissions to this log file or database table, and no modification or deletion permissions. When a log entry is written, the system's current timestamp and the writing process's digital signature are automatically appended. This is for quality review and algorithm performance backtesting. Specific application scenarios include: when a clinical team has doubts about the accuracy of an alarm, they can retrieve complete metadata from the audit log using the process traceability identifier (usually obtained from the intervention system log); the reviewer can verify whether the feature snapshot values ​​that triggered the analysis, the data accumulation time, and the model version used are reasonable, thereby determining whether it is a real abnormality in the patient's physiological state or a false alarm caused by algorithm parameters or data quality issues; in addition, algorithm engineers can regularly analyze the logs to calculate the detection rate and false positive rate of various abnormalities under different model versions, quantitatively evaluating the evolution of algorithm performance.

[0040] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0041] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0042] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A real-time big data integration and analysis method for the Internet of Things in healthcare, characterized in that, include: The system continuously acquires real-time data streams representing the patient's physiological state from medical IoT devices, performs real-time feature extraction to obtain a first feature set, and determines whether the first feature set meets preset real-time analysis conditions. The real-time analysis conditions are comprehensive conditions describing the abnormal significance and analysis reliability of the first feature set. If the first feature set does not meet the instant analysis conditions, then data for a subsequent set duration is obtained from the real-time data stream, and the subsequent data is spliced ​​with the current segment of the real-time data stream to form an extended data segment. Feature extraction is performed on the extended data segment to obtain a second feature set, and it is determined whether the second feature set meets the instant analysis conditions; If the second feature set does not meet the real-time analysis conditions and the cumulative duration of the extended data segment is less than the preset maximum analysis window duration, then return to the step of obtaining data for a subsequent set duration from the real-time data stream and concatenating the subsequent data with the current extended data segment to form a new extended data segment; If the second feature set meets the real-time analysis conditions, or the cumulative duration of the extended data segment reaches the maximum analysis window duration, then based on the current first or second feature set that meets the conditions, or based on the feature set corresponding to the extended data segment that has reached the maximum analysis window duration, high-precision anomaly analysis is performed to generate analysis results containing anomaly type, severity level and predicted trajectory. The analysis results are pushed to the clinical intervention system in real time.

2. The real-time big data integration and analysis method for the medical Internet of Things according to claim 1, characterized in that, After performing real-time feature extraction on the acquired current segment of real-time data stream to obtain a first feature set, and determining whether the first feature set meets preset real-time analysis conditions, the process further includes: If the first feature set meets the real-time analysis conditions, then the high-precision abnormal state analysis is performed based on the first feature set to generate a first analysis result, and the first analysis result is pushed to the clinical intervention system in real time.

3. The real-time big data integration and analysis method for the medical Internet of Things according to claim 1, characterized in that, The process of determining whether the first feature set or the second feature set meets the preset real-time analysis conditions includes: Extract the dynamic trend features and stability quantification features of core physiological indicators from the first feature set or the second feature set; The dynamic trend feature is compared with a preset abnormal trend threshold, and the stability quantification feature is compared with a preset stability threshold. When the dynamic trend feature exceeds the abnormal trend threshold and the stability quantification feature is lower than the stability threshold, it is determined that the current feature set meets the instant analysis conditions. When the dynamic trend feature does not exceed the abnormal trend threshold, or the stability quantification feature is not lower than the stability threshold, it is determined that the current feature set does not meet the instant analysis conditions. The core physiological indicators include at least heart rate variability, rate of decrease in blood oxygen saturation, and respiratory rhythm disorder index.

4. The real-time big data integration and analysis method for the medical Internet of Things according to claim 1, characterized in that, Before the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices, it also includes: Based on the pathophysiological characteristics of the target monitored diseases, the length of the basic time window used for initial feature analysis is determined; Based on the basic time window length and the data sampling frequency of the medical IoT device, the initial data volume corresponding to the current segment of real-time data stream is calculated; Specifically, determining the basic time window length includes: obtaining the typical physiological parameter fluctuation cycle of the target monitored disease in the early compensatory stage, and setting an integer multiple of the fluctuation cycle as the basic time window length.

5. The real-time big data integration and analysis method for the medical Internet of Things according to claim 1, characterized in that, Before the process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices, it also includes: Based on the clinical risk stratification of the target patient group, set personalized analysis trigger sensitivity levels; The preset maximum analysis window duration is dynamically adjusted based on the analysis trigger sensitivity level. The higher the analysis trigger sensitivity level, the shorter the corresponding maximum analysis window duration. The adjustment of the maximum analysis window duration is based on the principle that a shorter window is used for high-risk patient groups to pursue analysis timeliness, while a longer window is used for low-risk patient groups to obtain more sufficient feature information for discrimination.

6. The real-time big data integration and analysis method for the medical Internet of Things according to claim 1, characterized in that, The process of performing high-precision anomaly analysis based on the current first or second feature set that meets the conditions, or based on the feature set corresponding to the extended data segment that has reached the maximum analysis window duration, includes: The feature set is input into a pre-constructed anomaly analysis network; Through multi-level abstraction processing of the anomaly analysis network, anomaly probability vectors and state evolution vectors are output. By fusing the anomaly probability vector and the state evolution vector, the analysis results containing the anomaly type, severity level, and predicted trajectory are generated.

7. The real-time big data integration and analysis method for the medical Internet of Things according to claim 1, characterized in that, After the analysis results are pushed to the clinical intervention system in real time, the system also includes: Continuously monitor the patient's physiological state in response to the intervention measures corresponding to the pushed analysis results; The analysis results, the intervention measures, and the response data are stored together to form a complete clinical decision evidence unit. The stored clinical decision evidence units are periodically aggregated and pattern-mined to generate systematic knowledge for updating the real-time analysis conditions or the high-precision abnormal state analysis parameters.

8. The real-time big data integration and analysis method for the medical Internet of Things according to claim 1, characterized in that, The process of continuously acquiring real-time data streams characterizing a patient's physiological state from medical IoT devices specifically includes: At the medical IoT device end, the raw physiological sensor signals are subjected to localized preliminary filtering and analog-to-digital conversion to generate time-series data packets in a standard format; The time-series data packets are sent from the device to the edge computing node at a first preset frequency through a dedicated secure data transmission channel. At the edge computing node, multiple received time-series data packets are buffered and aligned at a second preset frequency, where the second preset frequency is higher than the first preset frequency. When the required amount of data for the current segment of real-time data stream is met, valid data segments are extracted from the verified time-series data packets and assembled into the real-time data stream for subsequent processing. The proprietary secure data transmission channel employs an end-to-end encryption mechanism based on temporary session keys, and the edge computing node destroys the temporary session key immediately after completing data assembly.

9. The real-time big data integration and analysis method for the medical Internet of Things according to claim 6, characterized in that, The construction and updating process of the pre-built anomaly analysis network is independent of the execution process of the real-time big data integration and analysis method, and specifically includes: The anomaly analysis network was initially trained offline based on historical desensitized clinical data. During the operation of the real-time big data integrated analysis method, the structure and parameters of the anomaly analysis network remain frozen; Based on the generated systematic knowledge, incremental updates and evaluations of the anomaly analysis network are periodically triggered; The updated anomaly analysis network version will only be switched to the online service version if the incremental update evaluation passes the preset verification criteria.

10. The real-time big data integration and analysis method for the medical Internet of Things according to any one of claims 1 to 5 and 7 to 9, characterized in that, The method further includes: A unique process traceability identifier is generated for each execution of the high-precision anomaly analysis process; Record the following metadata associated with the process traceability identifier: The snapshot of the feature set that triggered the analysis, the cumulative duration of the extended data segment used, the version of the core parameters called during the analysis, and the final analysis results pushed out; All process traceability identifiers and their associated metadata are stored in an immutable audit log for quality review and algorithm performance backtracking.

Images