Medical data risk early warning system and method based on AI and big data

Abstract

The invention discloses a medical data risk early warning system and method based on AI and big data, and belongs to the technical field of medical information processing of the big data technology, and the system comprises an intelligent denoising processing unit which is used for obtaining historical medical data and analyzing the noise features of the historical medical data to generate denoised data; the data structure analysis unit is used for analyzing the denoised data to extract data structure features; the self-adaptive algorithm selection unit is used for selecting an early warning algorithm combination and configuring algorithm parameters according to the data structure characteristics so as to generate an algorithm selection result; the multi-algorithm fusion detection unit is used for processing the de-noised data by adopting an algorithm selection result so as to identify abnormity and generate an abnormity detection signal; and the early warning result explanation unit is used for carrying out judgment standard tracking on the abnormal detection signal to determine a judgment basis and generating early warning explanation information in combination with the judgment basis, and the stability and robustness of the whole early warning system are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information processing of big data technology, and particularly to a medical data risk early warning system and method based on AI and big data. BACKGROUND

[0002] In the field of medical data monitoring and analysis, traditional risk early warning technology mainly relies on single anomaly detection algorithm and manual experience for model construction and parameter configuration. These methods usually expose problems such as single detection method, poor model adaptability, and difficult-to-explain early warning results when dealing with complex and variable medical data. This situation makes it difficult for medical managers and clinical experts to efficiently and accurately identify potential risks, affecting the scientificity of early intervention and decision-making.

[0003] Due to the inherent limitations of existing technology, a single detection algorithm can only identify specific types of anomalies, such as global outliers, and is not sensitive to other risk patterns such as local fluctuations or trend deviations, resulting in monitoring blind spots. The selection of algorithms and the adjustment of parameters are highly dependent on the prior knowledge of domain experts, which is not only inefficient, but also causes the model performance to decrease significantly when the data characteristics change, resulting in poor stability. Most importantly, many existing AI models are like black boxes, and the early warning signals generated by them lack transparent and traceable judgment basis, making it difficult for users to trust and adopt them.

[0004] Finally, when potential risks occur in medical data, managers and doctors cannot fully and accurately capture all abnormal signals, nor can they understand the specific reasons behind the early warning, which greatly limits the application value and credibility of intelligent early warning systems in clinical and management practice. SUMMARY

[0005] The present application aims to provide a medical data risk early warning system and method based on AI and big data to solve the problems raised in the background.

[0006] The technical solution of the present application is as follows: An intelligent denoising processing unit is used to obtain historical medical data and analyze the noise characteristics of the historical medical data to generate denoised data. A data structure analysis unit is used to analyze the denoised data to extract data structure features. An adaptive algorithm selection unit is used to select a combination of early warning algorithms and configure algorithm parameters based on the data structure features to generate an algorithm selection result. A multi-algorithm fusion detection unit is used to process the denoised data using the algorithm selection result to identify anomalies and generate an anomaly detection signal. The early warning result interpretation unit is configured to track a judgment standard of the anomaly detection signal to determine a judgment basis, and generate early warning interpretation information in combination with the judgment basis.

[0007] Preferably, the intelligent denoising processing unit is configured to: identify spikes and high-frequency fluctuations in the historical medical data, and compare the spikes and high-frequency fluctuations with a preset noise threshold; when the spikes or high-frequency fluctuations exceed the preset noise threshold, perform smoothing or denoising processing on the historical medical data to generate denoised data; when the spikes or high-frequency fluctuations do not exceed the preset noise threshold, directly use the historical medical data as the denoised data.

[0008] Preferably, the data structure features include data distribution patterns, periodicity and trend intensity, noise levels, and fluctuation frequencies.

[0009] Preferably, the data distribution patterns include skewness and heavy-tailed features.

[0010] Preferably, the anomalies include global outliers, local fluctuation anomalies, and trend deviation anomalies.

[0011] Preferably, the adaptive algorithm selection unit is specifically configured to: automatically match and select a combination of early warning algorithms from a density-first strategy, a distribution-first strategy, and a trend-first strategy according to the data structure features.

[0012] Preferably, the algorithm parameters are parameters selected from a preset parameter set including IQR multiples, DBSCAN radii, LOF neighbor numbers, SARIMA order numbers, and Prophet holiday impact intensities; the adaptive algorithm selection unit is further configured to dynamically adjust the algorithm parameters according to the data structure features.

[0013] Preferably, the combination of early warning algorithms includes prediction algorithms and unsupervised detection algorithms. The judgment basis includes trigger rules, detection algorithms, and key parameters.

[0014] Preferably, the trigger rules are that the standard deviation of more than the mean ± a preset multiple or the local density is lower than a neighborhood threshold; and the early warning interpretation information further includes visual threshold and prediction residual display.

[0015] A medical data risk early warning method based on AI and big data, comprising the following steps: obtaining historical medical data, and analyzing noise features of the historical medical data to generate denoised data; analyzing the denoised data to extract data structure features; According to the data structure characteristics, a combination of early warning algorithms is selected and algorithm parameters are configured to generate an algorithm selection result; The algorithm selection result is used to process the de-noised data to identify abnormalities and generate an abnormality detection signal; The abnormality detection signal is subjected to judgment standard tracking to determine a judgment basis, and early warning interpretation information is generated in combination with the judgment basis.

[0016] The present application provides an AI and big data based medical data risk early warning system and method, which has the following improvements and advantages compared with the prior art: 1. The present application sets up an intelligent de-noising processing unit to establish a judgment mechanism based on a preset noise threshold; this mechanism can identify spikes and high-frequency fluctuations in historical medical data, and only performs targeted smoothing or de-noising when these noise characteristics exceed the threshold; this design avoids information loss that may be caused by indiscriminate processing of all data, realizes accurate filtering of strong interference signals, and maximizes the retention of potential real dramatic fluctuations in the data, enhancing the stability and robustness of the entire early warning system; 2. The data structure analysis unit of the system can analyze the de-noised data to extract a set of quantitative data structure characteristics, including data distribution pattern, periodicity and trend intensity, noise level and fluctuation frequency; this converts the construction of the early warning model from relying on prior knowledge and manual selection to an automatic process driven by the inherent properties of the data; the adaptive algorithm selection unit automatically matches and selects the optimal combination of early warning algorithms from density-first, distribution-first and trend-first strategies according to these specific characteristics, and dynamically adjusts the algorithm parameters; this data-driven adaptive modeling process significantly reduces the dependence on the experience of domain experts, ensures that the model can always maintain near-optimal performance when dealing with diverse data streams, and greatly improves the accuracy and generalization ability of the early warning; 3. The present application can comprehensively identify multiple risk patterns including global outlier, local fluctuation anomaly and trend deviation anomaly by fusing prediction algorithms and unsupervised detection algorithms; this design overcomes the limitations of traditional single algorithms that can only identify specific types of anomalies, realizes broad-spectrum coverage of risk signals, significantly improves the comprehensiveness and accuracy of anomaly detection, and avoids the existence of monitoring blind spots; 4. The early warning result interpretation unit of the system provides a solution to the black box problem commonly existing in existing AI models; this unit can track the judgment criteria of the abnormal detection signal to determine the judgment basis composed of trigger rules, detection algorithms and key parameters, and generate early warning interpretation information combined with visual threshold and prediction residual display; this mechanism converts the judgment process of AI into a completely transparent and traceable process, enabling medical experts not only to know the early warning, but also to clearly understand the internal logic and quantitative standards of the early warning trigger, greatly enhancing the trust between man and machine and improving the practical value and adoption willingness of AI early warning results in medical management and decision-making practice. BRIEF DESCRIPTION OF DRAWINGS

[0017] The application will be further explained in conjunction with the accompanying drawings and embodiments: Figure 1 is a flowchart of a medical data risk early warning system based on AI and big data. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in conjunction with specific embodiments.

[0019] Example 1 Please refer to Figure 1 The present application provides a medical data risk early warning system based on AI and big data, comprising: An intelligent denoising processing unit is used to obtain historical medical data and analyze the noise characteristics of the historical medical data to generate denoised data; A data structure analysis unit is used to analyze the denoised data to extract data structure features; An adaptive algorithm selection unit is used to select early warning algorithm combinations and configure algorithm parameters according to the data structure features to generate algorithm selection results; A multi-algorithm fusion detection unit is used to process the denoised data using the algorithm selection results to identify abnormalities and generate abnormal detection signals; An early warning result interpretation unit is used to track the judgment criteria of the abnormal detection signal to determine the judgment basis, and generate early warning interpretation information combined with the judgment basis; The embodiment of the application provides a medical data risk early warning system and method based on AI and big data; the system aims to solve the problems of single abnormality detection means, dependence on artificial experience, poor model stability and difficult-to-explain results in the prior art; the system realizes efficient, accurate and interpretable automatic early warning of medical data risk through a data-driven and multi-stage fusion processing technical path; the system is constructed as an integrated information processing system containing several functional units, and the units work cooperatively to complete the complete process from raw data processing to interpretable early warning output; The intelligent denoising processing unit aims to improve the quality of input data and lay a foundation for the accuracy of subsequent analysis; in the embodiment, the unit is configured to obtain historical medical data, which can be time series format patient monitoring indicators, outpatient flow or drug usage, etc.; the analysis module built in the unit identifies and quantifies the noise characteristics of the data, and then generates a set of denoised data based on the analysis results; the operation of the unit effectively avoids the influence of random interference in the raw data on the subsequent algorithm determination, thereby enhancing the stability and robustness of the entire early warning system; The data structure analysis unit aims to deeply understand the inherent characteristics of the data and provide decision basis for the adaptive selection of subsequent algorithms; the unit receives the denoised data generated by the intelligent denoising processing unit and performs multi-dimensional statistical analysis on the data to extract a set of data structure characteristics that can comprehensively describe the characteristics of the data; this design changes the construction of the early warning model from artificial selection relying on prior knowledge to a data-driven automatic process, which is a key step to realize the adaptive ability of the system; The adaptive algorithm selection unit aims to automatically construct the optimal early warning algorithm model according to the actual characteristics of the data; the unit selects the most suitable early warning algorithm combination for the current data scenario from the preset algorithm strategy library according to the data structure characteristics extracted by the data structure analysis unit, and automatically configures and optimizes the key parameters of the selected algorithm to generate an algorithm selection result; the unit enables the system to flexibly cope with the monitoring needs of different types of medical data, and significantly improves the generalization ability and application range of the model; The multi-algorithm fusion detection unit aims to perform specific abnormality detection tasks to identify potential risk points in the data; the unit receives the denoised data and the algorithm selection result generated by the adaptive algorithm selection unit, and strictly follows the algorithms and parameter configurations specified in the result to process and analyze the data; by executing the algorithm, the unit can identify abnormal points in the data sequence and generate corresponding abnormality detection signals, which mark the time point and specific value of the abnormality occurrence; The early warning result interpretation unit aims to improve the transparency and credibility of the early warning result and solve the black box problem of the existing AI model. The unit starts a judgment standard tracking mechanism in response to receiving the abnormality detection signal generated by the multi-algorithm fusion detection unit. The mechanism determines the specific judgment basis for the abnormal point by backtracking the detection process of the abnormal point. The unit combines the judgment basis with the abnormal signal to generate early warning interpretation information that is easy for human experts, such as doctors or hospital managers, to understand. The above-mentioned units work together to form a complete technical closed loop from data input, denoising, feature analysis, algorithm adaptive selection, fusion detection to result interpretation. The technical effect is that the system realizes the full-process automation of medical data risk early warning, significantly improves the accuracy and comprehensiveness of abnormality detection, and greatly enhances the user's trust and willingness to adopt the AI early warning result by providing transparent decision basis, thereby improving the intelligent level of medical management and decision-making. To further improve the accuracy and practicality of early warning, the system can also be configured with a context information fusion interface, not shown. The interface is used to receive and integrate external events or state information synchronized with the historical medical data time point, such as the patient's electronic medical record, such as medication changes, surgery records, hospital management logs, such as equipment maintenance plans or public health event reports. In the multi-algorithm fusion detection unit, these context information can be used as additional features or conditional variables to adjust the sensitivity of the early warning model or perform secondary verification on the detected abnormal signals, thereby effectively distinguishing between normal fluctuations caused by specific known events and true unknown risks, and avoiding oversimplification of complex real-world problems. To ensure that the model can accurately depict the dose-effect relationship of key medical indicators, the system can also introduce a configurable domain knowledge base in the adaptive algorithm selection unit. The knowledge base pre-stores risk weights, normal value ranges, and clinically significant fluctuation thresholds defined by medical experts for different types of medical data, such as heart rate, blood pressure, and drug usage. When configuring algorithm parameters, the system not only refers to data structure characteristics, but also queries the knowledge base to weight or fine-tune key parameters such as IQR multiple or preset multiples in decision rules , thereby balancing the flexibility of data-driven and the fidelity of knowledge-driven.

[0020] The intelligent denoising processing unit is used to: identify spikes and high-frequency fluctuations in historical medical data and compare spikes and high-frequency fluctuations with a pre-set noise threshold; When the spike or high-frequency fluctuation exceeds the pre-set noise threshold, perform smoothing or denoising processing on the historical medical data to generate denoised data. When the spike or high-frequency fluctuation does not exceed the preset noise threshold, the historical medical data is directly taken as the denoised data; In the embodiment, the working mode of the intelligent denoising processing unit is further limited; the unit executes its denoising function through a threshold-based judgment mechanism; The intelligent denoising processing unit is used to identify spikes and high-frequency fluctuations in the historical medical data; a spike refers to an isolated value that appears instantaneously and deviates significantly from adjacent data points, which may be caused by device acquisition errors or transient interference; a high-frequency fluctuation refers to rapid and frequent reciprocating changes in data values within a very short time window, which may be caused by noise introduced during signal transmission; To quantify the above noise characteristics, the system introduces a preset noise threshold; the preset noise threshold The setting principle is that, for a representative benchmark historical data set collected under normal operating conditions and confirmed by domain experts, the absolute value distribution of the deviation between the data points and the local moving average is calculated, and the 99th percentile of the distribution is taken to ensure that only statistically extreme changes are considered as potential noise; the unit compares the absolute value of the identified spike or the amplitude of the high-frequency fluctuation with the preset noise threshold; When the value of the spike or high-frequency fluctuation exceeds the preset noise threshold, it is determined that the data point or data segment is significantly contaminated by noise, at which time the unit will automatically perform smoothing or denoising processing on the historical medical data; to deal with different noise types, when a spike is identified, median filtering can be used for processing; when a high-frequency fluctuation is identified, wavelet transform is used for denoising, and a processed denoised data is generated; To further improve the intelligent level of denoising, the window size of the median filter and the decomposition level of the wavelet transform are not fixed values, but can be adaptively adjusted according to the fluctuation frequency of the data; for example, when the dominant fluctuation frequency calculated by the data structure analysis unit, which will be described in detail later, is low, it means that the data changes are relatively smooth, at which time the window size of the median filter can be appropriately increased to obtain better smoothing effect; otherwise, the window size is reduced; similarly, for the wavelet transform, the decomposition level may also be inversely related to the fluctuation frequency Not a fixed value, but can be adaptive adjustment according to the fluctuation frequency F of data, for example, can be established as follows exemplary function relationship: ; Wherein, is the base window size, for example, 5, is the adjustment coefficient, the dimension is time, for example, second, is the dominant fluctuation frequency of the current data, is the reference frequency; the formula makes the window size increase when the fluctuation frequency F is low, to obtain better smoothing effect; in order to ensure that the window size is positive odd number, the calculation result also needs to be constrained; Similarly, for wavelet transform, the decomposition layer number can be negatively related to the fluctuation frequency F, to ensure that while filtering out high frequency noise, the details of the effective signal are preserved to the greatest extent; When the value of the peak or high frequency fluctuation does not exceed the preset noise threshold, it is considered that the fluctuation is within the normal range, which may be an effective physiological signal or business fluctuation; at this time, the unit will directly output the original historical medical data as the denoised data, in order to retain all potential effective information in the data; The gain technical effect brought by the embodiment is that: by introducing the intelligent judgment mechanism based on threshold, the intelligent denoising processing unit avoids the one-size-fits-all smoothing processing of all data, realizes the accurate identification and targeted processing of noise; this can not only effectively filter out strong interference signal, but also maximize the retention of real but violent fluctuation information that may exist in the data, so as to improve the stability of the model while avoiding information loss caused by excessive processing, and improve the accuracy of subsequent anomaly detection.

[0021] Embodiment 2 Data structure features include data distribution pattern, periodicity and trend intensity, noise level and fluctuation frequency; Data distribution pattern includes skewness and heavy tail features; In this embodiment, the data structure features extracted by the data structure analysis unit are described; these features provide quantitative basis for the decision of the subsequent adaptive algorithm selection unit; the data structure features include data distribution pattern, periodicity and trend intensity, noise level and fluctuation frequency; Data distribution pattern, the function is to reveal the overall statistical distribution characteristics of the data; in this embodiment, the feature is further embodied as including skewness and heavy tail features; skewness refers to the asymmetry of data distribution, which is quantified by calculating skewness coefficient ; heavy tail feature refers to the tail of the distribution is thicker than the normal distribution, which means that the probability of extreme value is higher, which is quantified by calculating kurtosis coefficient The technical significance of identifying these patterns is that drug usage data with significant right skewness and heavy tails suggest that robust statistical methods that are insensitive to outliers should be used when selecting algorithms, rather than methods based on the normal distribution assumption. The function of periodicity and trend strength is to identify the regularity of data in the time dimension; this unit calculates the periodicity strength score of data through methods such as autocorrelation function analysis or spectral analysis. and Trend Strength Score ; Periodicity refers to the pattern of data repeating at regular time intervals, and trend refers to the long-term upward or downward tendency of data over time; Noise level, which quantifies the intensity of random interference in the data; this unit calculates the signal-to-noise ratio of the data sequence or calculates the energy ratio after high-pass filtering the data to obtain the noise level index ; Fluctuation frequency is used to measure the speed of data change. This unit obtains the dominant fluctuation frequency by calculating the zero-crossing rate or short-time Fourier transform of the sequence. ; The technical gain effect of this embodiment is that: by extracting the above-mentioned set of specific and quantitative data structure features, the data structure analysis unit provides a solid data foundation for the system's adaptive ability; the algorithm selection no longer relies on vague human judgment, but is based on a deep insight into the intrinsic properties of the data; for example, when a strong periodicity is detected and low noise When the system can confidently select the trend prediction algorithm; when the heavy tail feature is detected and high fluctuation frequency When the local anomaly detection algorithm based on density is selected first, the algorithm selection is made intelligent, precise and optimized.

[0022] Anomalies include global outlier anomalies, local fluctuation anomalies, and trend deviation anomalies; In this embodiment, the anomaly types that can be identified by the multi-algorithm fusion detection unit are classified; this classification ensures that the system can comprehensively capture risk signals of different natures, including global outlier anomalies, local fluctuation anomalies, and trend deviation anomalies; Global outliers are defined as data points that appear unusually isolated within the entire dataset. These anomalies typically represent independent, rare events, such as a single erroneous device reading or a single large-scale public health emergency. Detecting these anomalies typically relies on assessing how far a data point deviates from the overall data distribution. Local fluctuation anomaly, defined as the data showing a significantly different behavior pattern from its neighboring data points within a certain local time window, e.g. a sharp shock or a sudden change in density within a short time; such anomaly may not be prominent in the global range, but is crucial for monitoring short-term burst risks, such as a sharp deterioration of a patient's vital signs after surgery; Trend deviation anomaly, defined as the development trend of the data points significantly deviates from the normal track predicted based on historical data; this anomaly focuses on the dynamic behavior pattern of the data, rather than the size of a single value; for example, during the flu season, the outpatient volume is already at a high level, but if its growth rate is much higher than the prediction model of the same period in history, it constitutes a trend deviation anomaly; The technical effect of the gain of the embodiment is that: by classifying the anomalies, the system can use different algorithm strategies for detection, achieving broad-spectrum coverage of risk signals; traditional single algorithm can only identify one type of anomaly, but the system can capture global, local and trend-based risk patterns by integrating multiple detection mechanisms, significantly improving the comprehensiveness and accuracy of anomaly detection, and avoiding the existence of monitoring blind area.

[0023] Embodiment 3 The adaptive algorithm selection unit is specifically configured to: automatically match and select a combination of early warning algorithms from a density-first strategy, a distribution-first strategy, and a trend-first strategy according to data structure characteristics; The algorithm parameters are parameters selected from a preset parameter set including IQR multiple, DBSCAN radius, LOF neighbor number, SARIMA order, and Prophet holiday impact strength; the adaptive algorithm selection unit is further configured to dynamically adjust the algorithm parameters according to the data structure characteristics; The combination of early warning algorithms includes prediction algorithms and unsupervised detection algorithms; The determination basis includes trigger rules, detection algorithms, and key parameters; The trigger rule is that the standard deviation of more than the mean ± a preset multiple or the local density is lower than the neighborhood threshold; the early warning explanation information further includes a visualization threshold and a prediction residual display; In the embodiment, the working mechanism of the adaptive algorithm selection unit, the algorithm parameters involved, and the composition of the combination of early warning algorithms are described in detail; these features together constitute a data-driven adaptive modeling and parameter adjustment process; The adaptive algorithm selection unit works in a way that: according to the data structure characteristics output by the data structure analysis unit, it automatically matches and selects a combination of early warning algorithms from a preset strategy library; the strategies include: Density-first strategy: when it is detected that the data noise level is high and periodic When the probability of anomaly is weak, the system prefers this strategy; this strategy tends to use unsupervised detection algorithms that are insensitive to noise, such as density-based DBSCAN or isolation-based isolation forest, to effectively identify global outlier anomalies and local fluctuation anomalies; Distribution-first strategy: When the data exhibits a stable distribution, e.g., skewness and kurtosis This strategy is used when the data is within a stable range but has no obvious trend. This strategy selects a detection method based on statistical distribution, such as the IQR method, and identifies anomalies by setting a statistical threshold. Trend-first strategy: When the data has obvious periodicity or trend This strategy is activated when ; it will give priority to forecasting algorithms, such as SARIMA or Prophet, and identify trend deviation anomalies by comparing the residuals between the actual value and the model prediction value; Therefore, the early warning algorithm combination is structurally designed to include prediction algorithms and unsupervised detection algorithms. This combination takes advantage of the complementary strengths of different algorithms: prediction algorithms are good at processing time-dependent data, while unsupervised detection algorithms can effectively discover unknown and sudden abnormal patterns. After selecting the algorithm combination, the adaptive algorithm selection unit is also used to dynamically adjust the algorithm parameters according to the data structure characteristics; the algorithm parameters are selected from a preset parameter set, each parameter in the set has a clear engineering practice significance; the set includes: IQR multiples : Used for distribution-based anomaly detection, the value will be based on the heavy tail characteristics of the data distribution Make adjustments; if A higher value indicates that there are more extreme values. It will be appropriately increased to reduce false positives; This dynamic adjustment can be exemplarily defined by a specific functional relationship; for example, a basic multiple can be set , usually 1.5, and based on the kurtosis coefficient The adjustment is made by the difference in kurtosis from the normal distribution, typically 3; an exemplary adjustment formula is: ; in, : IQR, a multiple of the interquartile range method, used to adjust the edge of anomaly detection; is the kurtosis value of the normal distribution, for example 3, is a preset sensitivity coefficient, such as 0.5, used to control The magnitude increases with kurtosis; this formula ensures that as the tail of the data distribution becomes heavier, The larger it is, the looser the boundary for judging anomalies is, thus improving the algorithm's tolerance for extreme values ​​inherent in the data; : function, returns the maximum value of two parameters; DBSCAN radius and the number of LOF neighbors : used for density-based detection, their values ​​will be based on the frequency of fluctuations in the data and noise levels Make adjustments; the internal logic is that in a high noise environment, will be increased to obtain more robust local density estimates; Number of LOF neighbors The value can be compared with the noise level indicator Establish a positive correlation functional relationship, for example; ; in, : LOF, the number of neighbors of the local outlier factor algorithm; : Floor rounding function; is the number of base neighbors, for example 20, is the scaling factor; to ensure the robustness of the model under extreme inputs, and Function pair The value range of , such as 5 and the maximum value , such as 10% of the total number of data points or a fixed upper limit, such as 200; this can prevent the extreme value of the noise level N from causing The parameter values ​​are unreasonable, thus ensuring the stability and effectiveness of the algorithm; for the core parameter radius of the DBSCAN algorithm , can be automatically determined by analyzing the k-distance graph of the data; the method is to calculate the distance from each point in the data set to its first The distances to the nearest neighbors are sorted in descending order and plotted into a graph. The distance value corresponding to the inflection point with the most significant slope change in the graph is selected as ; Here The value is the noise level mentioned above The number of neighbors calculated adaptively achieves the linked adaptive configuration of the two key parameters; SARIMA order (p, d, q) (P, D, Q) s: for trend prediction, the order is automatically determined based on the analysis of the autocorrelation function and partial autocorrelation function of the denoised data; to realize automatic determination, the system uses an automatic model selection algorithm, for example, by performing grid search in the preset p, d, q, P, D, Q order range, and taking Akaike information criterion or Bayesian information criterion minimization as the objective function, the optimal order combination is automatically optimized and determined; wherein, : non-seasonal autoregressive order; : non-seasonal difference order; : non-seasonal moving average order; : seasonal autoregressive order; : seasonal difference order; : seasonal moving average order; : seasonal cycle length; Adjust the hyperparameters such as adjustment coefficient γ, sensitivity coefficient α, and scaling coefficient β, the values of which can be determined by grid search on a standard validation dataset or using other optimization algorithms such as Bayesian optimization to achieve the best early warning performance; Prophet holiday impact strength : for processing predictions containing holiday effects, the strength is adaptively set by maximum likelihood estimation according to the fluctuation amplitude of the data during the holidays marked in the historical data; The gain technical effect of the embodiment is that by establishing a dynamic mapping relationship between data characteristics and algorithm strategies and parameters, the system realizes the transition from manual parameter adjustment to data-driven adaptive modeling; this mechanism not only significantly reduces the dependence on the experience of domain experts, greatly improves the modeling efficiency, but also ensures that the model can always maintain near-optimal performance when facing diversified and dynamically changing data streams, thereby greatly improving the accuracy, generalization ability and automation level of the early warning system.

[0024] In this embodiment, the determination basis and the content of the early warning explanation information generated by the early warning result explanation unit are described in detail to enhance the transparency and credibility of the early warning results; The determination basis is structured to include three core elements: trigger rules, detection algorithms, and key parameters; when an abnormal point is detected, the early warning result explanation unit will backtrack its detection link and automatically generate these three pieces of information; Detection algorithm: clearly indicates which algorithm or combination of algorithms ultimately identified the anomaly; for example, SARIMA prediction model or DBSCAN clustering algorithm; Key parameters: list the core parameter configurations used when executing the detection algorithm; for example, SARIMA order = (1, 1, 1) (1, 1, 0) 12 or DBSCAN radius = 0.5, minimum point number = 5 Triggering rule: describe the specific conditions for determining an anomaly in a clear and quantitative manner; in this embodiment, the rule is embodied in two typical forms: exceeding the mean ± a preset multiple of the standard deviation or the local density being lower than the neighborhood threshold; The first rule is mathematically expressed as wherein, is the denoised data point at the current time, and are the mean and standard deviation calculated according to the historical denoised data within a preset time window, is a preset multiple, the value of which is determined based on the receiver operating characteristic curve analysis, aiming to optimize the sensitivity and specificity of the early warning, and is commonly used in statistical distribution-based detection; The determination process is as follows: on a validation dataset with true labels, normal / abnormal, by constantly changing the value of , for example, from 1.5 to 5.0 with a step size of 0.1, a set of true positive rates, TPR, i.e. sensitivity, and false positive rates, FPR, i.e. 1-specificity, can be calculated for each value. Plotting these ( , ) pairs in a two-dimensional coordinate system forms the receiver operating characteristic curve; the final value can be selected based on the maximization principle of the Youden index, and the calculation formula is: sensitivity specificity

[0025] wherein, : Youden index, used to find the best threshold value in receiver operating characteristic curve analysis; sensitivity : true positive rate, the ability of the system to correctly identify anomalies; specificity : the ability of the system to correctly identify normal data; : true positive rate; : false positive rate; select the value of the point that maximizes , i.e. the threshold multiple that performs best on the validation set, which achieves the best balance between sensitivity and specificity; The second rule, the principle of which is that the number of samples in the neighborhood of the data point is lower than a certain threshold, is commonly used in density algorithms such as DBSCAN; ​For a more intuitive explanation, the early warning explanation information further includes a visualization of the threshold and the prediction residual display; To achieve this purpose, the system is configured to automatically generate a visualization chart on which: The original data curve and the prediction curve of the model are plotted, if a prediction class algorithm is used; The dynamic or static abnormality determination threshold line is clearly marked, for example, the upper and lower tracks of the mean ± kσ; The degree of deviation of the data from the normal trend is displayed in the form of a residual plot, i.e. the difference between the actual value and the predicted value; Highlight the abnormal data points that triggered the early warning; The technical effect of the embodiment is that by providing structured, multi-dimensional explainability information, the system converts the black box AI decision-making process into a completely transparent, traceable and auditable process; The doctor or manager not only knows when the early warning occurs, but also clearly understands why the early warning is triggered, the specific algorithm, parameters and quantitative rules, and intuitively perceives the severity of the anomaly through the visualization chart; This greatly enhances the trust between man and machine, reduces the adoption threshold of AI early warning results in clinical and management practice, thereby improving the practical value and credibility of the entire system.

[0026] Embodiment 4 An AI and big data-based medical data risk early warning method, comprising the following steps: Obtain historical medical data and analyze the noise characteristics of the historical medical data to generate denoised data; Analyze the denoised data to extract data structure features; According to the data structure features, select a combination of early warning algorithms and configure algorithm parameters to generate an algorithm selection result; Process the denoised data using the algorithm selection result to identify abnormalities and generate an abnormality detection signal; Determine the determination basis by tracking the determination criteria of the abnormality detection signal, and generate early warning explanation information in combination with the determination basis; The embodiment of the present application provides an AI and big data-based medical data risk early warning method, which realizes efficient, accurate and explainable automatic early warning of medical data risk through a data-driven, multi-stage fusion processing technical path; The method constitutes a complete technical closed loop from data input, denoising, feature analysis, algorithm adaptive selection, fusion detection to result explanation, including the following steps: Perform intelligent denoising processing to improve the quality of input data. This step obtains historical medical data, such as time series format patient monitoring indicators, outpatient flow or drug usage, etc. Then, identify the spikes and high-frequency fluctuations in the data. Spikes refer to isolated values that appear instantaneously and significantly deviate from adjacent data points. High-frequency fluctuations refer to rapid reciprocating changes in data within a very short period of time. Compare the identified spike or high-frequency fluctuation value with the preset noise threshold. When the value exceeds the threshold, it is determined that the data is significantly contaminated by noise. Perform targeted processing according to the noise type: use median filtering for spikes, and use wavelet transform for high-frequency fluctuations to generate a set of denoised data. If the threshold is not exceeded, it is considered that the fluctuation is within the normal range, and the original historical medical data is directly output as the denoised data.

[0027] Perform data structure analysis to deeply understand the intrinsic characteristics of the data. This step receives the denoised data generated by the previous step and performs multi-dimensional statistical analysis to extract a set of data structure features that can comprehensively describe the characteristics of the data. These data structure features include: Data distribution pattern: Calculate skewness coefficient S and kurtosis coefficient K to quantify the skewness and heavy-tailed characteristics of data distribution; Periodicity and trend intensity: Calculate the period intensity score P and trend intensity score T through autocorrelation function analysis or spectral analysis; Noise level: Calculate the signal-to-noise ratio or high-pass filtered energy ratio of the data sequence to obtain the noise level indicator N; Fluctuation frequency: Calculate the zero-crossing rate or short-time Fourier transform of the sequence to obtain the dominant fluctuation frequency F; Perform adaptive algorithm selection to automatically build the optimal early warning algorithm model. Based on the data structure features extracted in the previous step, select the most suitable early warning algorithm combination from the pre-set algorithm strategy library. Strategies include: Density-first strategy: preferentially selected when detecting high data noise level N and weak periodicity P, tend to use unsupervised detection algorithms such as DBSCAN or Isolation Forest; Distribution-first strategy: used when data shows stable distribution pattern but no obvious trend, will select detection methods based on IQR statistical distribution; Trend-first strategy: activated when data has obvious periodicity P or trend T, preferentially selects prediction algorithms such as SARIMA or Prophet; After selecting the algorithm combination, the key parameters of the algorithm will also be automatically configured and optimized according to the data structure features, such as adjusting the IQR multiple ki according to the heavy-tailed feature K, or adjusting the DBSCAN radius according to the fluctuation frequency F and noise level N and the number of LOF neighbors kl, finally generate an algorithm selection result; Perform multi-algorithm fusion detection to identify potential risk points in the data. This step receives the denoised data and the algorithm selection result, and strictly follows the algorithm and parameter configuration specified in the result to process and analyze the data. By executing a combination of prediction algorithms and unsupervised detection algorithms, it can identify abnormal points in the data sequence, including global outliers, local volatility anomalies, and trend deviation anomalies. After identifying the anomalies, generate corresponding anomaly detection signals that mark the time points and specific values of the anomalies; Perform early warning result interpretation to solve the black box problem of AI models. This step starts the judgment standard tracking mechanism after receiving the anomaly detection signal. By tracing back the detection process of the abnormal point, determine the specific judgment basis for its generation. This judgment basis is structured to include three core elements: trigger rule, detection algorithm, and key parameters. The trigger rule is specifically defined as a condition that is clear and quantifiable, such as exceeding the mean ± a preset multiple of the standard deviation or having a local density lower than the neighborhood threshold. Combine this judgment basis with the anomaly signal and generate a pre-warning interpretation information that includes visual threshold and prediction residual display for human experts to understand. This information is presented in the form of charts, which plot the original data and model prediction curve, mark the abnormal judgment threshold line, and highlight the abnormal data points that triggered the warning, making the judgment process completely transparent and traceable.

[0028] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application.

Claims

1. A medical data risk early warning system based on AI and big data, characterized by: include: An intelligent denoising processing unit, used to obtain historical medical data and analyze the noise characteristics of the historical medical data to generate denoised data; A data structure analysis unit, used to analyze the denoised data to extract data structure features; An adaptive algorithm selection unit is used to select a warning algorithm combination and configure algorithm parameters based on data structure characteristics to generate an algorithm selection result; A multi-algorithm fusion detection unit is used to process the denoised data using the algorithm selection results to identify anomalies and generate anomaly detection signals; The warning result interpretation unit is used to track the judgment criteria of the abnormal detection signal to determine the judgment basis, and generate warning explanation information based on the judgment basis.

2. The medical data risk early warning system based on AI and big data according to claim 1 is characterized in that: The intelligent denoising processing unit is used for: Identify spikes and high-frequency fluctuations in historical medical data and compare them with a preset noise threshold; When a spike or high-frequency fluctuation exceeds a preset noise threshold, smoothing or noise reduction processing is performed on the historical medical data to generate denoised data; When the spike or high-frequency fluctuation does not exceed the preset noise threshold, the historical medical data is directly used as the denoised data.

3. The medical data risk early warning system based on AI and big data according to claim 1 is characterized in that: The data structure characteristics include data distribution form, periodicity and trend strength, noise level and fluctuation frequency.

4. The medical data risk early warning system based on AI and big data according to claim 3 is characterized in that: The data distribution includes skewness and heavy-tail characteristics.

5. The medical data risk early warning system based on AI and big data according to claim 1 is characterized in that: The anomalies include global outlier anomalies, local fluctuation anomalies and trend deviation anomalies.

6. The medical data risk early warning system based on AI and big data according to claim 1 is characterized in that: The adaptive algorithm selection unit is specifically used to automatically match and select an early warning algorithm combination from a density priority strategy, a distribution priority strategy, and a trend priority strategy according to data structure characteristics.

7. The medical data risk early warning system based on AI and big data according to claim 1 is characterized in that: The algorithm parameters are parameters selected from a preset parameter set including IQR multiple, DBSCAN radius, number of LOF neighbors, SARIMA order, and Prophet holiday impact intensity; The adaptive algorithm selection unit is also used to dynamically adjust algorithm parameters according to data structure characteristics.

8. The medical data risk early warning system based on AI and big data according to claim 1 is characterized in that: The early warning algorithm combination includes a prediction algorithm and an unsupervised detection algorithm; The determination basis includes triggering rules, detection algorithms and key parameters.

9. The medical data risk early warning system based on AI and big data according to claim 8, characterized in that: The triggering rule is that the standard deviation exceeds the mean ± a preset multiple or the local density is lower than the neighborhood threshold; the warning explanation information also includes a visualization threshold and a prediction residual display.

10. A medical data risk warning method based on AI and big data, applied to a medical data risk warning system based on AI and big data as claimed in any one of claims 1 to 8, characterized in that: The following steps are involved: Acquire historical medical data and analyze the noise characteristics of the historical medical data to generate denoised data; Analyzing the denoised data to extract data structure features; According to the data structure characteristics, select a warning algorithm combination and configure algorithm parameters to generate an algorithm selection result; Processing the denoised data using the algorithm selection result to identify anomalies and generate an anomaly detection signal; The abnormal detection signal is tracked according to the judgment criteria to determine the judgment basis, and early warning explanation information is generated in combination with the judgment basis.

Images