Medical AI data risk early warning system based on transparency and interpretability

Abstract

The invention relates to the technical field of medical data processing and artificial intelligence, in particular to a medical AI data risk early warning system based on transparency and interpretability. The system comprises a data acquisition module, an intelligent denoising module, a structure perception and configuration module, a fusion detection module and an interpretability enhancement module. The system generates preprocessed data based on noise distribution of historical time series data, and automatically calculates an optimal algorithm combination to perform anomaly recognition; the core of the method is that a judgment standard tracking mechanism is used for backtracking a mathematical judgment boundary of an abnormal result and converting the mathematical judgment boundary into a natural language triggering rule; according to the invention, the conversion from single alarm to logic transparency is realized, a user is enabled to clearly know violation reasons, the technical barrier is broken, and the clinical credibility of the system is greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of medical data processing and artificial intelligence technology, specifically a medical AI data risk early warning system based on transparency and interpretability. Background Technology

[0002] With the deepening of medical informatization, hospital information systems and material management systems have accumulated massive amounts of historical time-series operational data, which cover complex business scenarios such as outpatient traffic and consumable usage. Currently, monitoring of such data mainly relies on manual experience to set fixed thresholds, or on traditional data cleaning and single monitoring models for anomaly screening. However, traditional monitoring methods have many limitations. On the one hand, conventional data preprocessing often adopts a one-size-fits-all noise reduction strategy, which is prone to accidentally deleting real public health emergency signals. Moreover, relying on manually specified algorithms and fixed parameters makes it difficult to adapt to medical business forms with different periodicity and noise levels, resulting in weak model generalization ability and insufficient detection accuracy. On the other hand, existing AI early warning models are often black boxes, unable to explain the judgment boundaries and logical sources behind the alarms to medical management personnel, resulting in low credibility of early warning results and difficulty in meeting the actual clinical needs. Therefore, how to build a medical data risk early warning system that can adapt to complex business characteristics and has clear interpretability has become an urgent problem to be solved in this field. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a medical AI data risk warning system based on transparency and interpretability. Specifically, the technical solution of this invention includes: The data acquisition module is used to acquire historical time-series operational data in medical business scenarios, including outpatient traffic data and consumable usage data. The intelligent noise reduction module is used to determine whether to perform smooth noise reduction processing based on the noise distribution characteristics in the historical time-series running data, so as to generate normalized preprocessed data. The structure perception and configuration module is used to extract statistical features from the preprocessed data and automatically calculate the optimal early warning algorithm combination and the corresponding model parameter configuration based on the extracted data structure indicators. The fusion detection module is used to call the optimal early warning algorithm combination and the model parameter configuration to perform anomaly identification on the preprocessed data and generate anomaly detection results including global outliers, local fluctuations and trend deviations; The interpretability enhancement module is used to respond to the anomaly detection result, and to reverse-engineer the judgment logic of the anomaly detection result through the judgment standard tracking mechanism, and generate risk warning explanation information including trigger rules, algorithm source and parameter settings.

[0004] Preferably, determining whether to perform smoothing and noise reduction processing based on the noise distribution characteristics in the historical time-series operational data includes: Retrieve the historical time-series execution data; Analyze the peak characteristics and high-frequency fluctuation characteristics in the historical time-series data; When the detected spike feature or the high-frequency fluctuation feature exceeds the preset interference threshold, the smoothing and noise reduction processing logic is triggered to denoise the historical time-series running data and output the preprocessed data.

[0005] Preferably, statistical features are extracted from the preprocessed data, and the optimal combination of early warning algorithms is automatically calculated based on the extracted data structure indicators, including: Call the preprocessed data; The data distribution pattern, periodicity, trend intensity, noise level, and fluctuation frequency of the preprocessed data are extracted as data structure indicators. Based on the data structure indicators, a prediction algorithm or unsupervised detection algorithm that is suitable for the current data characteristics is matched in a preset algorithm library to determine the optimal combination of early warning algorithms.

[0006] Preferably, based on the data structure indicators, a prediction algorithm or unsupervised detection algorithm that is suitable for the current data characteristics is matched from a preset algorithm library, including: When the data structure indicator indicates that the data is high-noise and short-period, a density-first algorithm is selected as the optimal combination of early warning algorithms. When the data structure indicator shows that the data has a steady-state distribution characteristic, the distribution-first algorithm is selected as the optimal early warning algorithm combination. When the data structure indicators indicate that the data has obvious periodic or seasonal trend characteristics, a trend-first algorithm is selected as the optimal early warning algorithm combination.

[0007] Preferably, based on the extracted data structure indicators, the corresponding model parameter configuration is automatically calculated, including: Invoke the data structure indicators and the determined optimal early warning algorithm combination; Based on the fluctuation frequency and distribution pattern in the data structure indicators, the interquartile range multiple, density cluster radius, or number of neighbors of local outliers are dynamically calculated as model parameter configurations for unsupervised detection. Based on the periodicity and trend intensity in the data structure indicators, the order of the differential autoregressive moving average model or the influence intensity coefficient of holidays is dynamically calculated as the parameter configuration for the prediction model.

[0008] Preferably, the optimal early warning algorithm combination and the model parameter configuration are invoked to perform anomaly identification on the preprocessed data, generating anomaly detection results including global outliers, local fluctuations, and trend deviations, including: Using the distribution detection logic in the optimal early warning algorithm combination, the overall outlier behavior of the preprocessed data in the long-term pattern is identified; Using the density detection logic in the optimal early warning algorithm combination, local density changes and short-term drastic fluctuations in the preprocessed data are identified. By utilizing the trend prediction logic in the optimal early warning algorithm combination, the trend deviation behavior of the preprocessed data relative to the prediction baseline is identified; The anomaly detection results are generated by aggregating the identification results of the overall outlier behavior, the local density changes and short-term drastic fluctuations, and the trend deviation behavior through a unified scheduling interface.

[0009] Preferably, the judgment logic of the anomaly detection result is inferred through the judgment standard tracking mechanism to generate risk warning explanation information including trigger rules, algorithm source, and parameter settings, including: For each data point marked as an anomaly, trace back the mathematical decision boundary upon which it was identified; The mathematical decision boundary is transformed into triggering rules described in natural language. The triggering rules include exceeding the mean standard deviation range or local density being lower than the neighborhood threshold. Extract the name of the algorithm used to generate the data points marked as anomalies and their key parameter values; The risk warning explanation information is generated by combining the triggering rule, the algorithm name, and the key parameter values.

[0010] Preferably, the risk warning explanation information further includes visual auxiliary charts, and the interpretability enhancement module is further configured as follows: Based on the output of the optimal early warning algorithm combination, plot the prediction trend curve and the confidence interval threshold channel; Calculate the residual between the actual and predicted values ​​of the preprocessed data, and plot the residual distribution. The predicted trend curve, the confidence interval threshold channel, and the residual distribution map are used as the visualization auxiliary charts and output synchronously with the risk warning interpretation information.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This system utilizes a judgment standard tracking mechanism to trace back the mathematical judgment boundaries behind abnormal detection results and transform them into natural language trigger rules that medical administrators can understand, such as values ​​exceeding the mean standard deviation range. Simultaneously, it uses visual aids to display predicted trends and confidence intervals. This combination of logical deduction and graphical display breaks down the barriers between technology and business, allowing users not only to know the alarm results but also to clearly understand the alarm reasons and violation logic, thereby significantly improving the system's credibility and decision-making support value in clinical applications. 2. Through the structure perception and configuration module, this system can deeply analyze the statistical characteristics of data such as distribution patterns, periodicity, and trends, and automatically select the optimal strategy from the algorithm library accordingly: density-first algorithms are used for high-noise, short-period data, distribution-first algorithms are used for steady-state data, and trend-first algorithms are used for strong-period data. This mechanism of matching one optimal strategy to one type of data significantly enhances the system's generalization ability and detection accuracy for different business forms such as outpatient traffic and consumable usage. 3. Unlike the traditional one-size-fits-all smoothing process, this system's intelligent noise reduction module only intervenes when it detects spikes or high-frequency fluctuations that exceed the interference threshold. This mechanism removes system input errors or non-business random noise while preserving, to the greatest extent possible, the true dynamic characteristics such as the surge in traffic at the beginning of a public health emergency, thus reducing the risk of missed reports caused by the loss of true signals due to excessive cleaning. 4. This system establishes a mapping relationship between data structure indicators and model hyperparameters, and can dynamically calculate key parameters such as density cluster radius, interquartile range multiple, or seasonality coefficient based on the fluctuation frequency and distribution pattern of the data. This gives the system a biological adaptive adjustment capability, which can automatically relax the threshold when the data fluctuates drastically and automatically tighten the threshold when the data is stable, thus maintaining the best balance between sensitivity and specificity without the need for repeated manual adjustments. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1: Please see Figure 1 A medical AI data risk early warning system based on transparency and explainability includes: The data acquisition module is used to acquire historical time-series operational data in medical business scenarios. The historical time-series operational data includes outpatient traffic data and consumable usage data. The intelligent denoising module is used to determine whether to perform smoothing and denoising processing based on the noise distribution characteristics in historical time-series running data, so as to generate normalized preprocessed data. The structure perception and configuration module is used to extract statistical features from preprocessed data and automatically calculate the optimal combination of early warning algorithms and the corresponding model parameter configuration based on the extracted data structure indicators. The fusion detection module is used to call the optimal combination of early warning algorithms and model parameter configurations to perform anomaly identification on preprocessed data and generate anomaly detection results including global outliers, local fluctuations and trend deviations. The interpretability enhancement module is used to respond to anomaly detection results by inferring the judgment logic of the anomaly detection results through the judgment standard tracking mechanism, and generate risk warning explanation information including trigger rules, algorithm source and parameter settings.

[0015] This embodiment details the architecture logic and interaction details of the aforementioned medical AI data risk warning system based on transparency and interpretability. This system aims to solve the black-box problem in existing medical data monitoring, namely, the inability to explain why an alarm is triggered, and the algorithm's inability to adapt to the ever-changing characteristics of medical business data. The data acquisition module, as the system's sensing front end, extracts raw business flows from underlying databases such as the Hospital Information System (HIS) and the Materials Management System (HRP) through a pre-defined ETL interface. This module acquires historical time-series operational data at predetermined time intervals. This historical time-series operational data originates from the HIS / HRP database and physically represents a continuous numerical sequence generated over time in medical business scenarios, measured in times / day or items / day. Specifically, it covers outpatient traffic data such as emergency room registrations and the number of patients seen at fever clinics, as well as consumable usage data such as the quantity of high-value consumables issued and the dosage of antibiotics. The intelligent denoising module acts as a gatekeeper for data quality. It receives raw data and analyzes its noise distribution characteristics. This module does not blindly denoise, but intelligently determines whether smoothing and noise reduction processing is needed based on the feature analysis results. In response to the determination that processing is needed, the system generates regularized preprocessed data to remove random interference that is not essential to the business. In response to the determination that the signal-to-noise ratio of the data itself is good, the original data is directly transmitted. The structure perception and configuration module, as the core brain of the system, is responsible for transforming configuration from manual experience to data-driven configuration. This module performs in-depth statistical analysis on the input preprocessed data and extracts statistical features. Based on the extracted data structure indicators, such as the skewness coefficient and periodicity intensity value, the module automatically matches them in the built-in algorithm library to calculate the optimal combination of early warning algorithms for the current data scenario and simultaneously calculates the appropriate model parameter configuration. On this basis, the fusion detection module, as the executor for anomaly identification, calls the determined optimal combination of early warning algorithms and model parameter configuration through a unified scheduling interface to perform point-by-point or sliding window scanning on the preprocessed data and perform anomaly identification. The anomaly detection results output by this module are multi-dimensional, covering global outliers where the overall data value deviates from the normal distribution range, local fluctuations with drastic changes in a short period of time, and trend deviations where the actual trend does not match the predicted baseline trend. In response to the generation of anomaly detection results, the interpretability enhancement module immediately triggers the judgment standard tracking mechanism. This mechanism backtracks the decision boundary of the algorithm to deduce the judgment logic of the anomaly detection results and finally generates risk warning explanation information for users. This information clearly includes the reason for the violation described in natural language, the specific algorithm name for judging the anomaly, and the key threshold parameters used in the judgment. This embodiment constructs an adaptive intelligent monitoring closed loop through the cascading collaboration of five core modules. When facing complex business scenarios in hospitals, the system can not only adapt to different data types with huge differences in the distribution of outpatient volume and consumable volume, and automatically select the optimal algorithm to improve detection accuracy, but more importantly, it breaks the black box state of traditional AI early warning. Through the logical reverse reasoning of the interpretability enhancement module, medical management personnel can clearly understand the mathematical logic and business meaning behind the alarm, thereby greatly improving the credibility of the system and the feasibility of clinical implementation.

[0016] Example 2: Based on the noise distribution characteristics in historical time-series operational data, determine whether to perform smoothing and noise reduction processing, including: Retrieve historical time-series execution data; Analyze the peak characteristics and high-frequency fluctuation characteristics in historical time-series operational data; When a spike feature or high-frequency fluctuation feature is detected that exceeds the preset interference threshold, the smoothing and noise reduction processing logic is triggered to denoise the historical time-series running data and output the preprocessed data.

[0017] This embodiment details the specific execution logic of the intelligent denoising module. This logic aims to accurately remove non-business noise caused by input errors or occasional system malfunctions without damaging the true business characteristics. The system reads the data sequence within the time window to be analyzed, i.e., it calls historical time-series running data. The intelligent denoising module uses time-domain statistical analysis combined with frequency-domain signal decomposition technology to analyze peak characteristics and high-frequency fluctuation characteristics respectively. Peak characteristics: derived from sliding window statistical analysis, physically meaning the numerical change amplitude at a single point or within a very short time in the data sequence exceeds the neighborhood mean by a multiple, dimensionless; the specific calculation formula is:

[0018] in, This is the value at the current moment. and They are respectively of length ,like Select time window Historical data within the current point are excluded. The mean and standard deviation within the sliding window formed by the interference. To prevent the minimum value where the denominator is zero, it has the same characteristics as... Same dimensions; High-frequency fluctuation characteristics: derived from Fourier transform or wavelet decomposition, physically representing the energy percentage within a preset high-frequency band; specifically, through discrete wavelet transform, using the Daubechies4 (db4) wavelet basis, the signal is decomposed into approximate coefficients. and detail coefficient The calculation formula is:

[0019] in, The first layer detail coefficients represent the highest frequency layer. This represents the total energy after signal decomposition, i.e., the sum of the squares of all approximation coefficients and detail coefficients; the system performs interference detection and smoothing; the system sets a preset interference threshold, for example... The threshold is 3.5. The threshold is 15%. In response to the detection of spikes or high-frequency energy proportions in the data exceeding this interference threshold, the system determines that the current data quality is impaired and triggers smoothing and noise reduction processing logic. During this process, either a moving average filter or a Savitsky-Gore filter is used, where the window length of the Savitsky-Gore filter is... Set to 7, order of polynomial fitting. Set to 2, the system denoises the historical time-series data and outputs the smoothed sequence as preprocessed data; if the threshold is not exceeded, the system retains the original data characteristics to prevent excessive smoothing from masking the true signals of public health emergencies. This embodiment adopts an adaptive judgment mechanism based on noise distribution characteristics, which effectively avoids the drawbacks of the one-size-fits-all preprocessing in traditional data cleaning. In the medical data scenario, this mechanism only intervenes when the data is indeed disturbed, which not only ensures the purity of the downstream model input, but also preserves the real dynamic characteristics such as the surge in traffic at the beginning of a flu outbreak to the greatest extent, thereby reducing the false alarm rate caused by data distortion at the source.

[0020] Example 3: Statistical features are extracted from the preprocessed data, and based on the extracted data structure indicators, the optimal combination of early warning algorithms is automatically calculated, including: Recall preprocessed data; Extract the data distribution pattern, periodicity, trend intensity, noise level, and fluctuation frequency of the preprocessed data as data structure indicators; Based on data structure indicators, predictive or unsupervised detection algorithms that are suitable for the current data characteristics are matched from a pre-defined algorithm library to determine the optimal combination of early warning algorithms.

[0021] This embodiment details the process of automatically calculating the optimal early warning algorithm combination in the structure perception and configuration module, aiming to achieve adaptive matching of one type of data with one optimal strategy. The structure perception and configuration module calls preprocessed data and calculates a set of multi-dimensional feature vectors as data structure indicators. Data distribution pattern: derived from statistical calculation, its physical meaning is the data pattern quantified by skewness and kurtosis, dimensionless; periodicity intensity: derived from the autocorrelation function (ACF), its physical meaning is the correlation coefficient judgment value under a specific lag order, dimensionless; trend intensity: derived from the Mann-Kendall trend test, its physical meaning is the monotonicity of the sequence over time, dimensionless; noise level: derived from residual sequence analysis, its physical meaning is the variance of the residual sequence, with units consistent with the original data; fluctuation frequency: derived from the statistical calculation of the coefficient of variation, its physical meaning is the ratio of the sequence standard deviation to the mean, used to quantify the relative dispersion of data over time, dimensionless. The system performs mapping and matching based on the above indicators in a pre-set algorithm library. This algorithm library includes prediction algorithms suitable for regular sequences, such as ARIMA, Prophet, and LSTM, as well as unsupervised detection algorithms suitable for unlabeled and disordered sequences, such as Isolation Forest, DBSCAN, and LOF. The optimal combination of early warning algorithms is determined based on data structure indicators. For example, in response to indicators showing that the data has strong periodicity, the system prioritizes matching prediction algorithms; in response to the data exhibiting random distribution, the system matches unsupervised detection algorithms. This embodiment introduces multi-dimensional data structure indicators to automate and intelligently select algorithms. Compared with the traditional method of manually specifying algorithms, this solution can deeply analyze the inherent usage characteristics of medical data and dynamically adjust the detection strategy according to physical attributes such as periodicity and dispersion. This significantly improves the model's generalization ability to different business scenarios and solves the problem that a single model cannot take into account multiple medical business forms.

[0022] Example 4: Based on data structure indicators, predictive or unsupervised detection algorithms that are suitable for the current data characteristics are matched from a pre-defined algorithm library, including: When the data structure indicators show that the data is high-noise and has short-period characteristics, density-first algorithms are selected as the optimal combination of early warning algorithms. When the data structure indicators suggest that the data has a steady-state distribution, the distribution-first algorithm is selected as the optimal combination of early warning algorithms. When data structure indicators show that the data has obvious periodic or seasonal trends, trend-first algorithms are selected as the optimal combination of early warning algorithms.

[0023] This embodiment further refines the logical rules for selecting specific algorithm types based on data structure indicators. For data with high noise and short-period characteristics, when the data structure indicators show a high coefficient of variation and a significant autocorrelation coefficient at short lag orders, the system identifies the data as non-stationary discrete data; that is, the coefficient of variation in the data structure indicators... Furthermore, the autocorrelation coefficient has a significant lag order. In response, the system selects density-first algorithms, such as DBSCAN or LOF, as the optimal combination of early warning algorithms. This is because such algorithms are not sensitive to noise and can identify anomalies based on the local density of sample points, making them suitable for handling irregular data such as sudden surges in emergency room traffic. For data exhibiting steady-state distribution characteristics, when indicators show that the data skewness is close to 0 and the trend is weak, it indicates that the data is in a statistical steady state; that is, the absolute value of skewness in the data structure indicators. And the Mann-Kendall trend test statistic In response to this, the system selects distribution-first algorithms, such as the 3-Sigma principle and box plot detection, as the optimal combination of early warning algorithms. These algorithms are based on statistical assumptions, have extremely high computational efficiency, and are suitable for handling stable business flows such as routine consumable requisition. For data with obvious periodic or seasonal trends, when the indicators show a significant long-term trend or obvious weekly or monthly cycles, the system selects trend-first algorithms, such as SARIMA, Prophet, and Holt-Winters, as the optimal combination of early warning algorithms. These algorithms can model the time dependence of time series and identify deviations by predicting the baseline, making them suitable for handling data with obvious time patterns, such as outpatient volume and hospitalization numbers. This embodiment constructs a refined classification and matching mechanism to accurately align the mathematical characteristics of the algorithm with the physical features of the data. This mechanism avoids model mismatch situations, such as using linear prediction algorithms to fit noisy nonlinear data, thereby significantly improving the accuracy of anomaly identification while ensuring detection coverage and ensuring the robustness of the early warning system in complex medical data environments.

[0024] Example 5: Based on the extracted data structure metrics, the corresponding model parameter configurations are automatically calculated, including: Invoke the data structure indicators and the determined optimal early warning algorithm combination; Based on the fluctuation frequency and distribution pattern in the data structure indicators, the interquartile range multiple, density cluster radius, or number of neighbors of local outlier factors are dynamically calculated as model parameter configurations for unsupervised detection classes. Based on the periodicity and trend strength in the data structure indicators, the order of the differential autoregressive moving average model or the influence strength coefficient of holidays is dynamically calculated as the parameter configuration for predictive models.

[0025] This embodiment details the process of automatically calculating the corresponding model parameter configuration, aiming to abandon fixed parameters and adopt a dynamic calculation method to optimize sensitivity; the system calls data structure indicators and the determined optimal early warning algorithm combination; for unsupervised detection algorithms, the system performs calculations based on fluctuation frequency and distribution pattern; specifically, it dynamically calculates the interquartile range multiple. The system was built Mapping function between values ​​and skewness coefficient and kurtosis:

[0026] in, As an adjustment factor, it is preferably set to 0.5. Responding to the data having a heavy-tailed distribution (i.e., high kurtosis), the system automatically increases the k value using this formula to reduce false alarms. Simultaneously, regarding the density clustering radius epsilon, the system constructs a K-distance map using the data distribution morphology characteristics. Specifically, it calculates the distance from each sample point to its k-th ... The distance between the nearest neighbors, here The default value is a preset lower limit of 5. Distance curves are generated in descending order and Gaussian smoothing is applied to these curves to eliminate minor noise interference. The sensitivity of curve inflection points is corrected by combining fluctuation frequency. The curve is automatically determined by finding the point of maximum second derivative. This is to achieve adaptive radius to data density; the specific calculation formula is as follows:

[0027] in, This represents the distance value corresponding to the maximum value of the second derivative of the K-distance graph. The preset sensitivity gain coefficient is preferably set to 0.2. The fluctuation frequency; Adjusting the scaling factor The system performs dynamic mapping based on the fluctuation frequency index, constructing the following linear equation:

[0028] in, The base factor is preferably set to 8. The gain factor is preferably set to 15. The coefficient of variation, extracted from the data structure metrics, is the fluctuation frequency. This equation ensures that as the fluctuation frequency of the data increases, The value automatically increases linearly, and will Set as The floor value, i.e.:

[0029] Thus by increasing To smooth out the effects of high-frequency noise; the calculated value, and This corresponds to the minimum number of points in density clustering or the number of neighbors of local outliers, and is used as the model parameter configuration for unsupervised detection. For prediction algorithms, the system calculates based on periodicity and trend strength. Specifically, the system uses an automated order determination algorithm to determine the ARIMA model parameters: iterative ADF unit root tests are performed to determine the difference order. That is, in a non-stationary sequence, the statistical probability value The time loop performs difference operations and accumulates the results. Until the sequence becomes stationary or reaches a preset depth; determine the autoregression order based on ACF / PACF statistical properties. and moving average order The model order can be determined by either using the critical order at which the absolute value of the correlation coefficient decays to within the confidence interval, or by directly using the AIC criterion to minimize and search for the optimal model. Combination; for models such as Prophet, the system calculates based on the ratio of historical holiday data amplitude to weekday amplitude, i.e.:

[0030] in, and These are the average values ​​for holidays and weekdays in historical data, respectively; the system maps this ratio to the `holidays_prior_scale` parameter of the Prophet model: if The parameter value is set to 10.0 to enhance the model's ability to fit sudden changes during holidays; otherwise, it is set to 0.05 to prevent overfitting, thus serving as the parameter configuration for the prediction class model. This embodiment achieves customized model parameters. By establishing a functional mapping relationship between data features and model hyperparameters, the system can automatically relax the threshold when data volatility increases and automatically tighten the threshold when data is stable. This dynamic adjustment mechanism achieves the best balance between sensitivity and specificity, eliminating the need for repeated manual optimization, greatly reducing operation and maintenance costs and improving adaptability to unexpected situations.

[0031] Example 6: The optimal combination of early warning algorithms and model parameter configurations are invoked to perform anomaly identification on preprocessed data, generating anomaly detection results including global outliers, local fluctuations, and trend deviations, including: By utilizing the distribution detection logic in the optimal early warning algorithm combination, we can identify the overall outlier behavior of preprocessed data in long-term patterns. By utilizing the density detection logic in the optimal early warning algorithm combination, local density changes and short-term drastic fluctuations in preprocessed data are identified. By utilizing the trend prediction logic in the optimal early warning algorithm combination, the trend deviation behavior of preprocessed data relative to the prediction baseline is identified; The anomaly detection results are generated by aggregating the identification results of the overall outlier behavior, the local density changes and short-term drastic fluctuations, and the trend deviation behavior through a unified scheduling interface.

[0032] This embodiment details the execution logic of the fusion detection module. This module does not execute in isolation, but rather calls multiple logics in parallel or sequentially to comprehensively capture anomalies. The system utilizes the distribution detection logic in the optimal early warning algorithm combination to calculate the probability density of data points over a long historical period. When the probability of a data point appearing is below a very low threshold, it is identified as a global outlier, such as identifying a department's consumable usage suddenly reaching a multiple of its historical peak. Simultaneously, the system uses density detection logic to calculate the distance or density reachability of a data point to its nearest neighbors through clustering or nearest neighbor algorithms. This is in response to a data point appearing within a local time window... The system identifies isolated outliers as local density changes and short-term drastic fluctuations, such as sudden, brief traffic spikes during low-flow periods at night. Furthermore, it utilizes trend prediction logic to generate a prediction baseline and confidence interval through a time-series prediction model, calculates the residual between the actual value and the prediction baseline, and identifies trend deviation behavior when the actual value falls outside the confidence interval. The system aggregates the identification results of overall outlier behavior, local density changes and short-term drastic fluctuations, and trend deviation behavior through a unified scheduling interface, and uses a logical OR operation; that is, if any of the above detection logic outputs an anomaly, the final anomaly detection result is generated. This embodiment employs a multi-view fusion detection mechanism, which effectively solves the blind spot problem of a single algorithm. This mechanism can not only detect obvious total excesses, but also keenly capture local abnormal patterns or trend deterioration lurking under normal totals. By integrating the detection capabilities of distribution, density and trend, it significantly improves the comprehensiveness and penetration of risk warning.

[0033] Example 7: By inferring the judgment logic of anomaly detection results through the standard tracking mechanism, risk warning explanation information is generated, including trigger rules, algorithm source, and parameter settings, including: For each data point marked as an anomaly, trace back the mathematical decision boundary upon which it was identified; The mathematical decision boundary is transformed into triggering rules described in natural language. The triggering rules include exceeding the mean standard deviation range or local density being lower than the neighborhood threshold. Extract the name of the algorithm used to generate the data points marked as anomalies and their key parameter values; The triggering rule, algorithm name, and key parameter values ​​are combined to generate risk warning explanation information.

[0034] This embodiment details the interpretability enhancement module and its execution judgment standard tracking mechanism, aiming to help doctors understand why the AI ​​triggers an alarm. For each data point marked as abnormal, the system traces back the mathematical judgment boundary it touched during detection. For complex prediction algorithms, this mathematical judgment boundary specifically refers to the upper and lower limits of the confidence interval calculated based on the statistical distribution of the prediction residuals. The mathematical judgment boundary originates from the internal state of the algorithm model and has the physical meaning of the critical mathematical expression for the algorithm to determine an anomaly, such as a value greater than the third quartile plus 1.5 times the interquartile range. The system calls the built-in rule translation template and the preset mapping template library to convert the mathematical judgment boundary into trigger rules described in natural language. For example, it converts mathematical inequalities into descriptions that business personnel can understand, such as values ​​exceeding the mean standard deviation range or local density being lower than the neighborhood threshold. The system extracts the algorithm name and key parameter values ​​used when generating the data points marked as abnormal from the model configuration. The system combines the trigger rules, algorithm names, and key parameter values ​​to generate structured risk warning explanation information, such as clearly indicating that the alarm reason is local fluctuation anomaly, and the trigger rule is that the actual value deviates from the prediction baseline by more than 95% confidence interval. This embodiment eliminates the gap between the technical black box and business understanding by translating the complex underlying mathematical decision-making logic into rule descriptions that business personnel can read. This enables medical managers not only to know what is happening, but also why it is happening, allowing them to quickly locate the source of risk, greatly assisting subsequent investigation and decision-making, and enhancing the depth of human-machine collaboration.

[0035] Example 8: The risk warning explanation information also includes visual auxiliary charts, and the interpretability enhancement module is also configured as follows: Based on the output of the optimal combination of early warning algorithms, plot the prediction trend curve and the confidence interval threshold channel; Calculate the residuals between the actual and predicted values ​​of the preprocessed data, and plot the residual distribution. The predicted trend curve, confidence interval threshold channel, and residual distribution map are used as auxiliary visualization charts and output synchronously with the risk warning interpretation information.

[0036] This embodiment further refines the generation logic of the visualization auxiliary charts; the interpretability enhancement module, based on the output of the optimal early warning algorithm combination, draws a predicted trend curve on the time axis, which represents the theoretically normal trend, and draws a confidence interval threshold channel that surrounds the curve, usually a semi-transparent band, representing the allowable range of normal fluctuations; outliers are explicitly marked outside the channel to visually show the degree of deviation; the system calculates the difference between the actual value and the predicted value of the preprocessed data, i.e., the residual, and draws a residual distribution chart accordingly, which is used to show the statistical characteristics of data deviation; the system uses the predicted trend curve, confidence interval threshold channel, and residual distribution chart as visualization auxiliary charts, and outputs them synchronously on the user interface along with the risk warning explanation information in text form; This embodiment leverages the advantages of human visual perception, making complex temporal anomalies readily apparent through a combination of text and images. Doctors can quickly verify the rationality of the system's judgment by observing trend curves and confidence channels. This intuitive display of the evidence chain not only improves information transmission efficiency but also further enhances users' trust in AI decision-making.

[0037] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A medical AI data risk early warning system based on transparency and interpretability, characterized in that, include: The data acquisition module is used to acquire historical time-series operational data in medical business scenarios, including outpatient traffic data and consumable usage data. The intelligent noise reduction module is used to determine whether to perform smooth noise reduction processing based on the noise distribution characteristics in the historical time-series running data, so as to generate normalized preprocessed data. The structure perception and configuration module is used to extract statistical features from the preprocessed data and automatically calculate the optimal early warning algorithm combination and the corresponding model parameter configuration based on the extracted data structure indicators. The fusion detection module is used to call the optimal early warning algorithm combination and the model parameter configuration to perform anomaly identification on the preprocessed data and generate anomaly detection results including global outliers, local fluctuations and trend deviations; The interpretability enhancement module is used to respond to the anomaly detection result by inferring the judgment logic of the anomaly detection result through the judgment standard tracking mechanism, and generate risk warning explanation information including trigger rules, algorithm source and parameter settings.

2. The medical AI data risk early warning system based on transparency and interpretability according to claim 1, characterized in that, Based on the noise distribution characteristics in the historical time-series operational data, determine whether to perform smoothing and noise reduction processing, including: Retrieve the historical time-series execution data; Analyze the peak characteristics and high-frequency fluctuation characteristics in the historical time-series data; When the peak feature or the high-frequency fluctuation feature is detected to exceed the preset interference threshold, the smoothing and noise reduction processing logic is triggered to denoise the historical time-series running data and output the preprocessed data.

3. The medical AI data risk early warning system based on transparency and interpretability according to claim 1, characterized in that, Statistical features are extracted from the preprocessed data, and based on the extracted data structure indicators, the optimal combination of early warning algorithms is automatically calculated, including: Call the preprocessed data; The data distribution pattern, periodicity, trend intensity, noise level, and fluctuation frequency of the preprocessed data are extracted as data structure indicators. Based on the data structure indicators, a prediction algorithm or unsupervised detection algorithm that is suitable for the current data characteristics is matched in a preset algorithm library to determine the optimal combination of early warning algorithms.

4. A medical AI data risk early warning system based on transparency and interpretability according to claim 3, characterized in that, Based on the aforementioned data structure indicators, a prediction algorithm or unsupervised detection algorithm that is suitable for the current data characteristics is matched from a preset algorithm library, including: When the data structure indicator indicates that the data is high-noise and short-period, a density-first algorithm is selected as the optimal combination of early warning algorithms. When the data structure indicator shows that the data has a steady-state distribution characteristic, the distribution-first algorithm is selected as the optimal early warning algorithm combination. When the data structure indicators indicate that the data has obvious periodic or seasonal trend characteristics, a trend-first algorithm is selected as the optimal early warning algorithm combination.

5. A medical AI data risk early warning system based on transparency and interpretability according to claim 1, characterized in that, Based on the extracted data structure metrics, the corresponding model parameter configurations are automatically calculated, including: Invoke the data structure indicators and the determined optimal early warning algorithm combination; Based on the fluctuation frequency and distribution pattern in the data structure indicators, the interquartile range multiple, density cluster radius, or number of neighbors of local outliers are dynamically calculated as model parameter configurations for unsupervised detection. Based on the periodicity and trend intensity in the data structure indicators, the order of the differential autoregressive moving average model or the influence intensity coefficient of holidays is dynamically calculated as the parameter configuration for the prediction model.

6. A medical AI data risk early warning system based on transparency and interpretability according to claim 1, characterized in that, The optimal early warning algorithm combination and the model parameter configuration are invoked to perform anomaly identification on the preprocessed data, generating anomaly detection results including global outliers, local fluctuations, and trend deviations, including: Using the distribution detection logic in the optimal early warning algorithm combination, the overall outlier behavior of the preprocessed data in the long-term pattern is identified; Using the density detection logic in the optimal early warning algorithm combination, local density changes and short-term drastic fluctuations in the preprocessed data are identified. By utilizing the trend prediction logic in the optimal early warning algorithm combination, the trend deviation behavior of the preprocessed data relative to the prediction baseline is identified; The anomaly detection results are generated by aggregating the identification results of the overall outlier behavior, the local density changes and short-term drastic fluctuations, and the trend deviation behavior through a unified scheduling interface.

7. A medical AI data risk early warning system based on transparency and interpretability according to claim 1, characterized in that, By using a standard tracking mechanism to deduce the judgment logic of the anomaly detection results, risk warning explanation information is generated, including trigger rules, algorithm source, and parameter settings, including: For each data point marked as an anomaly, trace back the mathematical decision boundary upon which it was identified; The mathematical decision boundary is transformed into triggering rules described in natural language. The triggering rules include exceeding the mean standard deviation range or local density being lower than the neighborhood threshold. Extract the name of the algorithm used to generate the data points marked as anomalies and their key parameter values; The risk warning explanation information is generated by combining the triggering rule, the algorithm name, and the key parameter values.

8. A medical AI data risk early warning system based on transparency and interpretability according to claim 7, characterized in that, The risk warning explanation information also includes visual auxiliary charts, and the interpretability enhancement module is further configured as follows: Based on the output of the optimal early warning algorithm combination, plot the prediction trend curve and the confidence interval threshold channel; Calculate the residual between the actual and predicted values ​​of the preprocessed data, and plot the residual distribution. The predicted trend curve, the confidence interval threshold channel, and the residual distribution map are used as the visualization auxiliary charts and output synchronously with the risk warning interpretation information.

Images