A method for outlier detection in medical data based on the CMOD algorithm and its application
By constructing a weighted minimum spanning tree and a two-dimensional scatter plot using the CMOD algorithm, the problems of inaccurate feature recognition and poor algorithm adaptability in outlier detection of medical data are solved. This enables efficient and accurate outlier sample detection and anomaly cause analysis, supporting medical insurance reimbursement review and hospital management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-05-13
- Publication Date
- 2026-07-31
AI Technical Summary
Existing outlier detection methods for medical data suffer from low feature recognition accuracy, poor algorithm adaptability, and difficulty in handling high-dimensional medical data, resulting in large errors in medical insurance reimbursement review and an inability to effectively identify hidden violation patterns in medical insurance reimbursement.
We construct a weighted minimum spanning tree based on the CMOD algorithm, calculate edge scaling weights using Euclidean distance, segment the spanning tree into multiple clusters, identify abnormal and normal clusters, and display feature differences using a two-dimensional scatter plot to pinpoint the cause of the anomaly.
It enables efficient and accurate outlier sample detection, reduces waste of medical insurance funds, provides quality monitoring and audit support, and improves the adaptability and interpretability of the test.
Smart Images
Figure CN122494286A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical data processing and anomaly detection, specifically relating to a method for detecting outliers in medical data based on the CMOD algorithm and its application. Background Technology
[0002] Current technologies for outlier detection in medical data mainly include traditional outlier techniques based on distance, density, and clustering, as well as methods combining machine learning. These methods generally suffer from the following shortcomings: First, low feature recognition accuracy. Existing methods mostly rely on manually preset features, failing to automatically distinguish between valid numerical features and non-analytical fields (such as hospital numbers and medical insurance codes) in medical data. Furthermore, they are weak in handling features with special formats (such as percentages), while medical insurance reimbursement audits require accurate extraction of core features such as costs and treatment items; feature recognition errors directly lead to auditing errors. Second, poor adaptability of outlier detection algorithms. Traditional algorithms (such as statistical Z-score and clustering-based K-means) are insufficient for processing high-dimensional medical data, struggling to capture complex relationships between features. They also depend on data quality and are sensitive to data sparsity. Detection accuracy drops significantly when the amount of surgical data is small or the feature dimensionality is high, making it difficult to effectively identify hidden violations in medical insurance reimbursement (such as cost splitting and overtreatment).
[0003] Therefore, there is an urgent need for an outlier detection method tailored to the characteristics of medical data, which can solve problems such as inaccurate feature recognition and poor algorithm adaptability, and achieve efficient and accurate outlier sample detection, providing reliable technical support for clinical diagnosis and treatment, hospital management and medical insurance reimbursement review. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention proposes a method for outlier detection in medical data based on the CMOD algorithm, specifically including the following steps:
[0005] Acquire users' medical data and filter the data categories that need to be detected as outliers, treating each user's medical data as a sample;
[0006] Construct a weighted minimum spanning tree based on the Euclidean equations between samples;
[0007] The constructed weighted minimum spanning tree is divided into multiple clusters, and the number of samples in each cluster is no greater than a set threshold.
[0008] Clusters with more than a set threshold of samples within a cluster are classified as abnormal clusters, while other clusters are classified as normal clusters. Samples within normal clusters whose distance from the cluster center exceeds a set threshold are classified as abnormal samples. All samples within abnormal clusters and abnormal samples within normal clusters are output as all abnormal users.
[0009] Furthermore, constructing a weighted minimum spanning tree includes the following steps:
[0010] 101. Randomly select a user from the set of users to be inspected as the root node, and initialize the set of visited nodes and the set of unvisited nodes, including adding the root node to the set of visited nodes and setting the set of unvisited nodes to empty;
[0011] 102. Find the node with the smallest scaling weight among the nodes in the visited set and remove it from the unvisited set and add it to the visited set.
[0012] 103. Select the node with the smallest scaling weight between the visited node and the newly added node as the parent node of the newly added node;
[0013] 104. Repeat steps 102 to 103 until all nodes are added to the visited set, and finally form a weighted minimum spanning tree covering all samples.
[0014] Furthermore, the calculation of node scaling weights includes:
[0015] ;
[0016] in, Let x represent the i-th sample. i With the p-th sample x p The edge scaling weights, if the p-th sample x p If it is the root node, then it is forced to be associated with the p-th sample x. p The scaling weight of the edges between them is set to 1.0, which serves as the initial baseline value for the entire spanning tree; Let x represent the i-th sample. i With the i-th sample x p The Euclidean distance between them; For the p-th sample x p With its parent node x parent The Euclidean distance between them; This represents a minimal correction constant used to prevent the denominator from being zero.
[0017] Furthermore, the constructed weighted minimum spanning tree is divided into multiple clusters, specifically including the following steps:
[0018] Sort all edges in the spanning tree in descending order of scaling weight, find the edge with the largest scaling weight, delete the edge to get two subtrees;
[0019] Determine if the size of the subtree has reached the set threshold. If not, continue to delete the edge with the largest scaling weight in the subtree and split the subtree until the size of the subtree reaches the set threshold.
[0020] Furthermore, if the number of nodes in the subtree is greater than n-min normal The segmentation stops when n is the total number of samples and min is the minimum value. normal This represents the minimum normal cluster size.
[0021] Furthermore, if the final cluster size is smaller than the minimum normal cluster size, then the cluster is a normal cluster; otherwise, the cluster is an outlier cluster, meaning that the samples within the cluster are an abnormal set of samples.
[0022] Furthermore, the minimum normal cluster size min normal Represented as: d represents the number of metrics that users participate in outlier detection.
[0023] This invention also proposes an application of a medical data outlier detection method based on the CMOD algorithm. Specifically, it addresses outlier clusters detected by a CMOD-based medical data outlier detection method, and identifies the causes of anomalies in the abnormal sample set. The specific steps include:
[0024] Once an abnormal cause is confirmed, the sample characteristics that directly or indirectly cause the abnormal cause are used as the abnormal factors of the abnormal cause.
[0025] For anomalies with an outlier factor greater than or equal to 2, select any two features as the center points of a two-dimensional coordinate system. Take the mean of these two features of the samples in the normal cluster as the cluster center. Calculate the distance between each sample in the outlier cluster and the cluster center in this coordinate system. If this distance is greater than 1.5 times the average distance of each sample in the outlier cluster to the cluster center, then the current anomaly is the cause of the current sample being out of the group.
[0026] Furthermore, the causes of abnormalities include at least the following: treatment efficiency, cost structure, cost control, and service consumption. Abnormal factors for treatment efficiency include: patient age, length of hospital stay, and nursing fees; abnormal factors for cost structure include: the proportion of medication costs, total Diagnosis Related Groups (DRG) costs, the proportion of examination and testing costs, and examination fees; abnormal factors for cost control include: material costs, total DRG costs, and surgical fees; and abnormal factors for service consumption include: nursing fees, bed fees, consultation fees, and length of hospital stay. Specifically, the cost proportions mentioned in this invention refer to the percentage of the total costs incurred.
[0027] Furthermore, when determining whether there are abnormalities in treatment efficiency, a two-dimensional coordinate system can be constructed using patient age and length of hospital stay, or length of hospital stay and nursing care costs; when determining whether there are abnormalities in cost structure, a two-dimensional coordinate system can be constructed using the proportion of medication costs and total DRG costs, or the proportion of examination and testing costs and examination fees; when determining whether there are abnormalities in cost control, a two-dimensional coordinate system can be constructed using material costs and total DRG costs, or surgical fees and total DRG costs; when determining whether there are abnormalities in service consumption, a two-dimensional coordinate system can be constructed using nursing care costs and bed fees, or consultation fees and length of hospital stay.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] 1. This invention can identify outliers in surgical data (such as cases with abnormal costs, cases with special conditions, cases with abnormal treatment processes, and cases suspected of violating medical insurance reimbursement regulations) in clinical practice, medical management, and medical insurance review. These outliers may reflect special clinical situations, potential medical risks, data recording errors, or unreasonable expenditures in medical insurance reimbursement. Timely discovery and analysis of these samples can provide targeted diagnostic and treatment references for medical staff, provide quality monitoring basis for hospital management departments, and provide support for accurate review by medical insurance institutions, thereby reducing the waste of medical insurance funds.
[0030] 2. This invention uses the CMOD algorithm to construct a weighted minimum spanning tree and recursively cut it, which can effectively capture the complex relationships between samples in high-dimensional medical data; by dynamically calculating the clustering scale threshold, it adapts to scenarios with different sample sizes and feature dimensions, solving the problem of poor adaptability of traditional algorithms.
[0031] 3. This invention presents the results in an intuitive and easy-to-understand interface. Through feature comparison statistical tables and pairwise feature scatter plots, the differences between normal samples and outlier samples are clearly displayed, allowing medical personnel to quickly locate the causes of outliers and improving the practicality of the technological achievement. Attached Figure Description
[0032] Figure 1 This is a flowchart of a medical data outlier detection method based on the CMOD algorithm according to the present invention. Detailed Implementation
[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0034] This invention proposes a method for outlier detection in medical data based on the CMOD algorithm, specifically including the following steps:
[0035] Acquire users' medical data and filter the data categories that need to be detected as outliers, treating each user's medical data as a sample;
[0036] Construct a weighted minimum spanning tree based on the Euclidean equations between samples;
[0037] The constructed weighted minimum spanning tree is divided into multiple clusters, and the number of samples in each cluster is no greater than a set threshold.
[0038] Clusters with more than a set threshold of samples within a cluster are classified as abnormal clusters, while other clusters are classified as normal clusters. Samples within normal clusters whose distance from the cluster center exceeds a set threshold are classified as abnormal samples. All samples within abnormal clusters and abnormal samples within normal clusters are output as all abnormal users.
[0039] In this embodiment, the raw data includes non-numerical information such as the patient's hospital number and primary diagnosis, as well as numerical information such as examination indicators, length of hospital stay, and medical expenses. The data format is an Excel file (.xlsx or .xls).
[0040] As an optional implementation method, effective feature identification and standardized preprocessing are performed on the patient's raw data, specifically including:
[0041] First, define the set of non-analysis fields and the range of analysis fields;
[0042] These effective numerical features are then extracted to form the original feature matrix. Finally, the standardized feature matrix is combined with key information such as hospital number and main diagnosis to form a feature set for modeling.
[0043] As a specific implementation method, this embodiment uses a partial medical cost dataset of spine-related surgeries from a certain hospital. The dataset contains 342 patient records, stored as an Excel file, and includes 18 fields such as hospital number, length of stay, total DRG medical cost, and drug ratio.
[0044] Data cleaning is performed, specifically including:
[0045] First, verify the existence and format of the file to confirm that it is in .xlsx or .xls format.
[0046] The data content is then structurally validated to ensure that it includes the core fields;
[0047] During the data cleaning phase, the system employs an intelligent strategy for handling duplicate records: for multiple medical records of the same hospital number, numerical features are summed and aggregated, while the first recorded value is retained for non-numerical features.
[0048] Effective feature identification and standardized preprocessing of user data specifically includes:
[0049] 1. Define the set of non-analysis fields and the range of analysis fields, extract effective numerical features from the raw data, and construct a feature set focused on analysis by excluding non-analysis fields such as "hospitalization number".
[0050] 2. Standardize the original feature matrix, for example, by using Z-score standardization.
[0051] This embodiment constructs a weighted minimum spanning tree (Weighted MST), using any sample as the root node, and calculates the Euclidean distance from the root node to all other samples as the initial edge distance, including:
[0052]
[0053] in, Let x represent the i-th sample. i With the j-th sample x j Euclidean distance; x i,k Let x represent the i-th sample. i The value of the k-th feature; Let x represent the j-th sample. j The value of the k-th feature; d is the dimension of the sample feature.
[0054] In this embodiment, a "scaling weight" is introduced to replace the original Euclidean distance as the edge weight, including:
[0055]
[0056] in, Let x represent the i-th sample. i With the p-th sample x p Edge scaling weights; Indicates the first Sample With the i-th sample x i With the p-th sample x p The Euclidean distance between them; For the p-th sample x p With its parent node x parent The Euclidean distance between them; This represents a minimal correction constant used to prevent the denominator from being zero.
[0057] The constructed weighted minimum spanning tree is partitioned to obtain multiple clusters, specifically including the following steps:
[0058] Sort all edges in the spanning tree in descending order of scaling weight, find the edge with the largest scaling weight, delete the edge to get two subtrees;
[0059] Determine if the size of the subtree is less than n-min. normal If the target is not reached, continue deleting the edge with the largest scaling weight from the subtree and splitting the subtree until the subtree size is less than n-min. normal .
[0060] Clustering is performed based on spanning trees. To distinguish between "normal clusters" and "outlier clusters," two core thresholds are set, including:
[0061]
[0062] Where, min normal denoted by , where n is the total number of samples and d represents the number of metrics that users participate in outlier detection.
[0063] max cut =n-min normal
[0064] Where, max cut This indicates the maximum size of the class that can be divided.
[0065] As an optional implementation, outlier clusters are directly assigned a fixed, extremely high score, while for normal clusters, the "Euclidean distance from the sample to the cluster center" is used as the outlier score—the greater the distance, the further the sample deviates from the normal level within the cluster, and the higher the degree of outlier.
[0066] As an optional implementation of this embodiment, samples whose distance from the cluster center to the normal cluster center exceeds a set threshold are considered abnormal samples. This embodiment calculates the cluster center for each feature dimension of each sample, as follows:
[0067]
[0068] in, This represents the center value of the t-th normal cluster at the k-th eigenvalue. This represents the size of the t-th normal cluster. Let represent the k-th feature value of the i-th sample within the t-th normal cluster.
[0069] Next, Euclidean distance is used to calculate the outlier score of the sample, including:
[0070]
[0071] in, Let represent the outlier score of the i-th sample.
[0072] This embodiment also proposes an application of a medical data outlier detection method based on the CMOD algorithm. The method is characterized by detecting the cause of the outlier clusters detected by the medical data outlier detection method based on the CMOD algorithm, specifically including the following steps: confirming the existence of the cause of the outlier, and using the sample features that directly or indirectly cause the cause of the outlier as the outlier factor of the cause of the outlier.
[0073] For anomalies with an outlier factor greater than or equal to 2, select any two features as the center points of a two-dimensional coordinate system. Take the mean of these two features of the samples in the normal cluster as the cluster center. Calculate the distance between each sample in the outlier cluster and the cluster center in this coordinate system. If this distance is greater than 1.5 times the average distance of each sample in the outlier cluster to the cluster center, then the current anomaly is the cause of the current sample being out of the group.
[0074] Specifically, the outlier score calculated by this invention is used to quantify the degree to which a sample deviates from the normal distribution of medical data. The higher the score, the higher the degree of sample abnormality. In the application process, abnormal samples are first screened based on the outlier score, and then the feature distribution of abnormal samples is compared with that of normal cluster samples to finally locate the specific dimension and violation type of the abnormality.
[0075] Following the principles of "strong business correlation, interpretable anomalies, and verifiable violations," feature pairs are formed by combining two features with a direct causal relationship and reflecting the same type of anomaly pattern. Multiple feature pairs can be created within the same dimension. Compared to one-dimensional distribution plots, two-dimensional scatter plots can better illustrate the constraints and combined anomalies between two features, avoiding hidden violations that cannot be identified by single-dimensional methods. Compared to three-dimensional and higher-dimensional visualizations, two-dimensional scatter plots have no projection distortion or curse of dimensionality, offer clear distribution, and are highly readable. They achieve an optimal balance between anomaly identification and business interpretability, making them more suitable for medical review and actual clinical scenarios. The X and Y axes of the scatter plot are directly mapped to preprocessed standardized numerical features. The purpose of pairing features is to project high-dimensional medical data onto a two-dimensional plane, visually displaying the sample distribution while preserving business interpretability.
[0076] As an optional implementation, in the two-dimensional feature scatter plot, the arithmetic mean of the two-dimensional standardized features of all samples in the current normal cluster is taken according to the dimension, and the center coordinates in the X-axis direction and the Y-axis direction are obtained respectively. Together, they constitute the cluster center of the normal cluster in the two-dimensional scatter plot. Based on the cluster center of the normal cluster, the two-dimensional Euclidean distance from the sample to the cluster center is calculated. If the distance is greater than 1.5 times the average distance in the cluster, it is determined to be a dense cluster that significantly deviates from the normal sample.
[0077] This embodiment takes the determination of whether there is an abnormality in the treatment efficiency as an example. It selects the combination of features such as age and length of hospital stay, and length of hospital stay and nursing cost to draw a pairwise comparison scatter plot. Based on the above quantitative rules, it identifies high abnormality samples to verify the consistency of the treatment process and identify unreasonable treatment costs.
[0078] This embodiment takes the determination of whether there is an anomaly in the cost structure as an example. It selects a combination of features such as the proportion of medication costs and total DRG costs, the proportion of examination and testing costs and examination fees, etc., to carry out comparative analysis. It focuses on the rationality of the cost structure of high-abnormal samples, checks whether it complies with medical standards and DRG payment requirements, and identifies anomalies such as cost structure imbalance.
[0079] This embodiment takes determining whether there are abnormalities in cost control as an example. It selects characteristics such as material costs and total DRG costs, surgical fees and core costs of total DRG costs for comparison, conducts special analysis on high-abnormal samples, focuses on core surgical cost items, and accurately locates the root cause of cost out-of-control.
[0080] This embodiment takes determining whether service consumption is abnormal as an example. It selects service-related feature combinations such as nursing fees and bed fees, and consultation fees and length of hospital stay for comparison. High-abnormal samples are identified based on quantified distance, verifying the matching between service consumption and billing, and identifying anomalies such as inflated or omitted services. Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for outlier detection in medical data based on the CMOD algorithm, characterized in that, Specifically, the following steps are included: Acquire users' medical data and filter the data categories that need to be detected as outliers, treating each user's medical data as a sample; Construct a weighted minimum spanning tree based on the Euclidean equations between samples; The constructed weighted minimum spanning tree is divided into multiple clusters, and the number of samples in each cluster is no greater than a set threshold. Clusters with more than a set threshold of samples within a cluster are classified as abnormal clusters, while other clusters are classified as normal clusters. Samples within normal clusters whose distance from the cluster center exceeds a set threshold are classified as abnormal samples. All samples within abnormal clusters and abnormal samples within normal clusters are output as all abnormal users.
2. The method for outlier detection in medical data based on the CMOD algorithm according to claim 1, characterized in that, Constructing a weighted minimum spanning tree involves the following steps:
101. Randomly select a user from the set of users to be inspected as the root node, and initialize the set of visited nodes and the set of unvisited nodes, including adding the root node to the set of visited nodes and setting the set of unvisited nodes to empty; 102. Find the node with the smallest scaling weight among the nodes in the visited set and remove it from the unvisited set and add it to the visited set.
103. Select the node with the smallest scaling weight between the visited node and the newly added node as the parent node of the newly added node; 104. Repeat steps 102 to 103 until all nodes are added to the visited set, and finally form a weighted minimum spanning tree covering all samples.
3. The method for outlier detection in medical data based on the CMOD algorithm according to claim 2, characterized in that, The calculation of node scaling weights includes: ； in, Let x represent the i-th sample. i With the p-th sample x p Edge scaling weights; Let x represent the i-th sample. i With the i-th sample x p The Euclidean distance between them; For the p-th sample x p With its parent node x parent The Euclidean distance between them; This represents a minimal correction constant used to prevent the denominator from being zero.
4. The method for outlier detection in medical data based on the CMOD algorithm according to claim 1, characterized in that, The constructed weighted minimum spanning tree is partitioned to obtain multiple clusters, specifically including the following steps: Sort all edges in the spanning tree in descending order of scaling weight, find the edge with the largest scaling weight, delete the edge to get two subtrees; Determine if the size of the subtree has reached the set threshold. If not, continue to delete the edge with the largest scaling weight in the subtree and split the subtree until the size of the subtree reaches the set threshold.
5. The method for outlier detection in medical data based on the CMOD algorithm according to claim 4, characterized in that, If the number of nodes in the sub-tree is greater than n-min normal then stop splitting, where n is the total number of samples and min normal is the minimum normal cluster size.
6. The method for outlier detection in medical data based on the CMOD algorithm according to claim 4, characterized in that, If the final cluster size is smaller than the minimum normal cluster size, then the cluster is a normal cluster; otherwise, the cluster is an outlier cluster, meaning that the samples within the cluster are an abnormal set of samples.
7. A method for outlier detection in medical data based on the CMOD algorithm according to claim 5 or 6, characterized in that, Minimum normal cluster size min normal Represented as: , where n is the total number of samples and d represents the number of metrics that users participate in outlier detection.
8. An application of a medical data outlier detection method based on the CMOD algorithm, characterized in that, For outlier clusters detected by the medical data outlier detection method based on the CMOD algorithm as described in claim 1, the specific steps for detecting the reasons for the anomalies in the abnormal sample set include: Once an abnormal cause is confirmed, the sample characteristics that directly or indirectly cause the abnormal cause are used as the abnormal factors of the abnormal cause. For anomalies with an outlier factor greater than or equal to 2, select any two features as the center points of a two-dimensional coordinate system. Take the mean of these two features of the samples in the normal cluster as the cluster center. Calculate the distance between each sample in the outlier cluster and the cluster center in this coordinate system. If this distance is greater than 1.5 times the average distance of each sample in the outlier cluster to the cluster center, then the current anomaly is the cause of the current sample being out of the group.
9. The application of the medical data outlier detection method based on the CMOD algorithm according to claim 8, characterized in that, Abnormal factors include at least the following: treatment efficiency, cost structure, cost control, and service consumption. Abnormal factors for treatment efficiency include: patient age, length of hospital stay, and nursing fees. Abnormal factors for cost structure include: the proportion of medication costs, the total cost of disease diagnosis-related groups, the proportion of examination and testing costs, and examination fees. Abnormal factors for cost control include: material costs, the total cost of disease diagnosis-related groups, and surgical fees. Abnormal factors for service consumption include: nursing fees, bed fees, consultation fees, and length of hospital stay.
10. The application of the medical data outlier detection method based on the CMOD algorithm according to claim 9, characterized in that, When assessing whether there are abnormalities in treatment efficiency, a two-dimensional coordinate system can be constructed using patient age and length of hospital stay, or length of hospital stay and nursing care costs. When assessing whether there are abnormalities in cost structure, a two-dimensional coordinate system can be constructed using the proportion of medication costs and the total cost of the disease diagnosis-related group, or the proportion of examination and testing costs and the examination fees. When assessing whether there are abnormalities in cost control, a two-dimensional coordinate system can be constructed using material costs and the total cost of the disease diagnosis-related group, or the surgical fees and the total cost of the disease diagnosis-related group. When assessing whether there are abnormalities in service consumption, a two-dimensional coordinate system can be constructed using nursing care costs and bed fees, or consultation fees and length of hospital stay.