A clinical pathway establishment and optimization method satisfying differential privacy constraints
By applying the differential privacy noise mechanism and the co-occurrence frequency pruning strategy of treatment projects during the clinical path establishment process, the problem of low accuracy in clinical path generation is solved, and an efficient generation of accurate clinical paths under privacy protection is achieved.
Patent Information
- Application Number
- CN202210290953.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-23
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2042-03-23
AI Technical Summary
Existing clinical pathway establishment methods have problems with low accuracy in the privacy protection process, especially when excessive noise is added during sequence pattern mining, resulting in a decrease in the accuracy of clinical pathway generation, and untrustworthy researchers infer the increased risk of patient sensitive data by analyzing statistical values.
The differential privacy noise mechanism is used to perform noise perturbation processing on the clinical path establishment process, including adding Laplace noise to atomic expansion projects, sequence expansion projects, vertical lists and candidate path sets, combined with the co-existing frequency pruning strategy of treatment projects, reducing the amount of noise and filtering infrequent modes, and improving generation efficiency and accuracy.
It achieves the accuracy and efficiency of clinical path generation under the satisfaction of differential privacy constraints, avoids untrustworthy researchers infer patient sensitive data through statistical analysis, and ensures the security of data privacy.
Smart Images
Figure CN114664457B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical data processing, and in particular to a method for establishing and optimizing a clinical pathway that satisfies differential privacy constraints. Background Art
[0002] A clinical pathway is a standardized treatment model and procedure established for a specific disease. It divides the diagnosis and treatment process into several stages and specifies the diagnostic and treatment items to be used in each stage. It is an effective way to improve the quality of clinical care, enhance medical quality, and control medical costs. Traditional clinical pathways are mostly developed through expert discussions. From the perspective of evidence-based medicine, they lack data support. With the rapid development of medical information technology, a large amount of medical data has been accumulated across the country. Data mining technology can fully utilize the information in this existing data, providing reference and guidance for the development of clinical pathways, making them more scientific and reasonable. Therefore, the use of data mining technology to generate clinical pathways from massive amounts of medical data has attracted increasing attention.
[0003] Although there are a lot of technologies and methods for establishing clinical pathways based on data mining, these methods still have the following problems: they do not pay attention to the privacy leakage of patient medical data. Figure 1 As shown, a trusted data administrator preprocesses real patient data to create a patient sequence dataset. They then perform sequence pattern mining on the sequence dataset and provide the mining results (frequent sequence patterns) to third-party researchers through a data query interface. Untrusted researchers then evaluate and select frequent sequence patterns to establish clinical pathways. In this scenario, because the medical data shared by patients is sensitive personal information, even if the medical institution releases statistical information of real data to third-party research institutions, malicious attackers can combine existing background knowledge with re-identification attacks, link attacks, and other means to infer a patient's medical information by leveraging changes in the statistical analysis values of anonymous data. Therefore, there is an urgent need to protect user data privacy in the establishment of clinical pathways.
[0004] Differential privacy is a widely adopted data privacy model in academia and industry. Compared to recent privacy-preserving techniques such as k-anonymity, l-diversity, and t-closeness, which rely on specialized attack assumptions and background knowledge, differential privacy has garnered significant attention and extensive research due to its ability to prevent attacks even with arbitrary background knowledge and provide strong privacy protection. The key principle of differential privacy is to ensure that the addition or deletion of any piece of data from a dataset does not significantly affect the algorithm's computational results, thereby protecting individual privacy. The primary implementation mechanism for differential privacy is noise, which includes Laplace noise, Gaussian noise, and exponential noise. To achieve differential privacy for a specific data analysis task, noise with a Laplace or Gaussian distribution is added, depending on the sensitivity of the task and the privacy budget.
[0005] A clinical pathway is essentially a chronological sequence of examinations, treatments, and nursing procedures for a specific disease. Therefore, the problem of clinical pathway mining can be transformed into the problem of mining frequent sequential patterns. The key issue in establishing clinical pathways based on differential privacy is adding noise that satisfies differential privacy constraints during the sequential pattern mining process. According to the principle of differential privacy, random noise perturbation must be added to the sequential pattern mining process. However, this presents the following challenges: First, some sequential pattern mining algorithms increase the risk of privacy leakage due to multiple scans of the original sensitive dataset. This requires adding more noise to achieve the same level of privacy protection, reducing the accuracy of the clinical pathway. Second, the number of candidate clinical pathways generated during the sequential pattern mining process is too large. The arbitrary addition or deletion of a patient's treatment sequence data significantly changes the number of candidate clinical pathways. To mask this change and protect user data privacy, excessive noise must be added, severely reducing the accuracy of the generated clinical pathways. Summary of the Invention
[0006] To address the low accuracy of privacy-based clinical pathways in existing methods, the present invention provides a technical solution for establishing and optimizing clinical pathways that meets differential privacy constraints, including the following steps:
[0007] Obtain hospitalization treatment items and time information for patients with the target disease, and establish an anonymous treatment item sequence dataset corresponding to the target disease;
[0008] Counting the co-occurrence frequencies between different treatment items in the anonymous treatment item sequence dataset, obtaining an atomic expansion item and a sequence expansion item for each treatment item in the anonymous treatment item sequence dataset according to the co-occurrence frequencies, and adding a first differential privacy noise to each atomic expansion item and each sequence expansion item;
[0009] Scan the anonymous treatment item sequence dataset, establish a 1-sequence vertical list corresponding to the anonymous treatment item sequence dataset, and add a second differential privacy noise to the 1-sequence vertical list; based on the 1-sequence vertical list after adding the second differential privacy noise, find a 1-sequence pattern whose noise support is greater than a support threshold and record it as a frequent 1-sequence pattern;
[0010] Based on frequent 1-sequence patterns, the candidate sequence pattern generation process is recursively called to generate a set of candidate paths. The generated candidate path set is pruned using the atomic expansion project and sequence expansion project after adding the first differential privacy noise, and the third differential privacy noise is added to the pruned candidate path set. Based on the candidate path set after adding the third differential privacy noise, the candidate path set with noise support greater than the support threshold is found as a frequent k-sequence pattern, k≥2; the frequent n-sequence pattern is provided to untrusted clinical pathway researchers through the query interface, n≥1.
[0011] Furthermore, the first differential privacy noise is Laplace(b), Laplace(b) represents the Laplace noise with noise scale parameter b, ω2 represents the maximum number of candidate 2-sequences contained in the longest sequence of the anonymous treatment project sequence dataset, |C2| represents the number of true candidate 2-sequences in mining, ε represents the set privacy protection strength, α is the privacy budget allocation ratio, 0<α<1.
[0012] Furthermore, the second differential privacy noise is m represents the maximum length of a frequent sequence; Δf represents the global sensitivity of calculating candidate 1-sequences, which is the maximum change in the number of candidate 1-sequences after adding or deleting any sequence; ε represents the set privacy protection strength; α is the privacy budget allocation ratio, 0<α<1.
[0013] Furthermore, the third differential privacy noise is m represents the maximum length of a frequent sequence; Δf represents the global sensitivity of calculating the candidate k-sequence, which is the maximum change in the number of candidate k-sequences after adding or deleting any sequence; ε represents the set privacy protection strength; α is the privacy budget allocation ratio, 0<α<1.
[0014] Furthermore, adding a second differential privacy noise to the 1-sequence vertical list includes:
[0015] Positive noise perturbation processing: Based on the characteristics of the 1-sequence vertical list structure, positive Laplace noise is added to the real count value of the header node. At the same time, dummy nodes are added to the 1-sequence vertical list items according to the noise amount. Each dummy node item contains a dummy SID and a dummy TID.
[0016] Negative noise disturbance processing: According to the structural characteristics of the 1-sequence vertical list, negative Laplace noise is added to the real count value of the header node. At the same time, the node records in the 1-sequence vertical list are deleted according to the amount of noise, and a deletion item list is designed to record the deleted records in each pattern IDList.
[0017] Furthermore, based on the frequent 1-sequence pattern, the candidate sequence pattern generation process is recursively called to generate a candidate path set, the generated candidate path set is pruned using the atomic expansion project and the sequence expansion project after adding the first differential privacy noise, and the pruned candidate path set is added with the second differential privacy noise; based on the candidate path set after adding the second differential privacy noise, a candidate path set whose noise support is greater than a support threshold is found as a frequent k-sequence pattern, including:
[0018] ① Based on the frequent (k-1)-sequence pattern, for any two treatment item sequence patterns, a candidate k-sequence C is generated by cross-connecting their vertical lists. K ;
[0019] ②Judge C K Is it empty? If it is empty, end the algorithm; otherwise, continue executing;
[0020] ③ Using the atomic expansion project and sequence expansion project after adding the first differential privacy noise to C K Prune the branches and calculate the C K Adding third differential privacy noise;
[0021] ④ Based on C after adding the third differential privacy noise K , find the candidate path set whose noise support is greater than the support threshold as the frequent k-sequence pattern; then k=k+1, go to step ② and continue.
[0022] Furthermore, pruning the generated candidate path set using the atomic expansion item and the sequence expansion item after adding the first differential privacy noise includes:
[0023] Let A i =P∪x,A j =P∪y, P is A i and A j The common prefix of i and A j For any two treatment item sequence patterns, let y be A i The expansion item of A i The last term of P, a is the last term of P, r=A i∪y; if y is in the atomic expansion item set of treatment project a, then the sequence pattern r is retained, otherwise the sequence pattern r is filtered out; if y is in the sequence expansion item set of treatment project a, then the sequence pattern r is retained, otherwise the sequence pattern r is filtered out.
[0024] Furthermore, the 1-sequence vertical list includes a sequence pattern corresponding to the list, the number of actual record nodes contained in the list, and each record; wherein each record structure is a sequence number and treatment time of the sequence pattern.
[0025] Furthermore, an atomic expansion project is a treatment project carried out together with a treatment project at the same time node, and a sequence expansion project is a treatment project carried out together with the treatment project at a subsequent time node.
[0026] Beneficial effects: The present invention designs a clinical pathway solution that satisfies differential privacy based on a vertical sequence pattern mining algorithm, which can complete the establishment of clinical pathways in sensitive data sets with fewer scans and meet differential privacy constraints; secondly, the present invention designs an optimization method based on co-occurrence pruning of treatment items, which reduces the noise required for privacy protection and improves the accuracy and generation efficiency of clinical pathways by filtering candidate patterns of obvious non-frequent sequence patterns in the data mining process and designing a strategy to reduce the number of candidate patterns. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 This is a schematic diagram of the existing principles for establishing clinical pathways based on data mining;
[0028] Figure 2 is a flow chart of a clinical pathway establishment and optimization method satisfying differential privacy constraints of the present invention;
[0029] Figure 3 This is a statistical diagram of the co-occurrence frequency of treatment items of the present invention;
[0030] Figure 4 It is a schematic diagram of a 1-sequence vertical list IDList corresponding to the SDB of the present invention;
[0031] Figure 5 It is a disturbance diagram of the IDList list of the present invention;
[0032] Figure 6 It is a schematic diagram of the query interface of the present invention. DETAILED DESCRIPTION
[0033] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be introduced below with reference to the accompanying drawings in the embodiments of the present invention.
[0034] In order to solve the problem of low accuracy of clinical pathways generated by existing methods, such as Figure 2 As shown, the clinical pathway establishment and optimization method that satisfies differential privacy constraints in this embodiment includes the following steps:
[0035] (1) Obtain hospitalization treatment items and time information of patients with target diseases and establish an anonymous treatment item sequence dataset corresponding to the target disease;
[0036] Establish a patient treatment project set for different diseases, and collect the treatment project and time data of the relevant patients, that is, what treatment project the patient received on which day. This embodiment takes hypertension treatment as an example, and the collected treatment project set is as follows:
[0037] For example, the treatment item set I = {i1: determination of serum total bilirubin, i2: injection of amikacin, i3: determination of serum total bile acid, i4: injection of deproteinized calf blood extract, i5: determination of inorganic phosphorus, i6: determination of serum uric acid, i7: blood routine test, ...};
[0038] The collected patient data is anonymized (quasi-identifiers such as name, ID number, and phone number are removed), retaining only the patient's treatment and time information. For each patient, a sequence of n time nodes is generated, where n represents the patient's hospitalization days. Each time node corresponds to a set of multiple treatments; N represents the maximum hospitalization day, and I represents the set of all treatments.
[0039] Based on the treatment item set, an anonymous patient treatment sequence dataset is constructed. Let SDB represent the generated anonymous treatment item sequence dataset, where each row represents the treatment item and time series for an anonymous patient. For example, S1 = <{i1,i2},{i3},{i6,i7},{i7},{i5}> represents the treatment item sequence for a patient hospitalized for 5 days. Each pair of curly braces {} in the sequence represents the treatment item set for a particular day, and 5 pairs of curly braces represent the treatment item sequence for 5 days.
[0040] Based on the above-mentioned hypertension disease treatment item set, an example of the SDB treatment sequence dataset is shown in Table 1. S1 represents the treatment item sequence of the first patient who was hospitalized for 5 days; S2 represents the treatment item sequence of the second patient who was hospitalized for 4 days; S3 represents the treatment item sequence of the third patient who was hospitalized for 4 days; and S4 represents the treatment item sequence of the fourth patient who was hospitalized for 2 days.
[0041] Table 1
[0042] Serial number Treatment program sequence <![CDATA[S1]]> <![CDATA[<{i1,i2},{i3},{i6,i7},{i7},{i5}>]]> <![CDATA[S2]]> <![CDATA[<{i1,i4},{i3},{i2},{i1,i2,i5,i6}>]]> <![CDATA[S3]]> <![CDATA[<{i1},{i2},{i6},{i5}>]]> <![CDATA[S4]]> <![CDATA[<{i2},{i6,i7}>]]>
[0043] (2) Counting the co-occurrence frequencies between different treatment items in the anonymous treatment item sequence dataset, obtaining the atomic expansion item and sequence expansion item of each treatment item in the anonymous treatment item sequence dataset according to the co-occurrence frequencies, and adding a first differential privacy noise to each atomic expansion item and each sequence expansion item;
[0044] Count the atomic expansion items, sequence expansion items and the actual frequency of co-occurrence of each treatment item in SDB. For an example of co-occurrence frequency statistics of treatment items, see Figure 3 As shown in the figure. An atomic expansion project refers to a treatment project that is implemented at the same time as a treatment project. A sequence expansion project refers to a treatment project that is implemented at a subsequent time point. For example, in Table 1, treatment project i1 can be atomically expanded to generate a 2-sequence <{i1, i2}>, and then sequence expansion to generate a 2-sequence <{i1}, {i3}>.
[0045] Since the above statistics are access to the original real sequence data, in order to avoid privacy leakage, it is necessary to add noise that satisfies the Laplace distribution to the statistical value to prevent privacy leakage. After calculation and analysis, the amount of Laplace noise added to each treatment item atom / sequence expansion item is: Laplace(b), Laplace(b) represents the addition of Laplace noise with a noise scale parameter of b, ω2 represents the maximum number of candidate 2-sequences contained in the longest SDB sequence, |C2| represents the actual number of candidate 2-sequences in the mining, ε represents the set privacy protection strength, α is the privacy budget allocation ratio, 0<α<1.
[0046] This step is essentially to count the number of times the sequences of two different treatment items co-occur. When generating candidate clinical pathways in the subsequent mining process, the candidate pathways will be pruned based on the co-occurrence frequency information, and candidate patterns that are obviously non-frequent patterns will be filtered out to reduce the size of the intermediate result set, thereby improving the accuracy of clinical pathway generation.
[0047] (3) Scan the SDB, establish the 1-sequence vertical list IDList corresponding to the SDB, and add differential privacy noise to the 1-sequence vertical list IDList; based on the 1-sequence vertical list after adding the second differential privacy noise, find the 1-sequence pattern whose noise support is greater than the support threshold and record it as the frequent 1-sequence pattern F1;
[0048] Scan the SDB and create a 1-sequence vertical list IDList. This list mainly records the serial number and time node information of each treatment item in the SDB. The formal structure is as follows: the sequence pattern corresponding to IDList, the number of actual record nodes contained in the list, and each record; among them, each record structure is the serial number SID and treatment time TID of the sequence pattern. The 1-sequence vertical list IDList corresponding to SDB is as follows Figure 4 shown.
[0049] The 1-sequence IDList is privacy-enhanced to obtain the frequent 1-sequence pattern F1 corresponding to the 1-sequence IDList. According to the implementation principle of the Laplace mechanism, one of the implementation mechanisms of differential privacy, noise that satisfies the Laplace distribution can be added to the IDList list for privacy-enhancing processing. The amount of noise added is m represents the maximum length of a frequent sequence; Δf represents the global sensitivity of calculating candidate 1-sequences, which is the maximum change in the number of candidate 1-sequences after adding or deleting any sequence; ε represents the set privacy protection strength; α is the privacy budget allocation ratio, 0<α<1.
[0050] Laplace perturbation includes positive noise perturbation and negative noise perturbation, which are specifically processed as follows:
[0051] Positive Noise Perturbation: Due to the structural characteristics of IDList, positive Laplace noise is added to the true count value of the header node. Simultaneously, dummy nodes are added to the IDList items based on the amount of noise (each dummy node item contains a dummy SID and a dummy TID). To ensure privacy protection without compromising mining accuracy, the addition of dummy nodes ensures that the dummy SIDs of candidate patterns that need to intersect in different pattern IDLists are different. This ensures that the number of nodes in the IDList meets differential privacy requirements while also ensuring the accuracy of the candidate k+1 patterns generated through cross-linking between different patterns.
[0052] Negative noise disturbance processing: According to the structural characteristics of IDList, it is necessary to add negative Laplace noise to the real count value of the header node, and delete the node records in IDList according to the amount of noise. However, the result of direct deletion will affect the accuracy of the longer sequence patterns generated subsequently. To this end, we design a deletion item list IDListoverflow arrays in this embodiment to record the records deleted in each pattern IDList. Its structure is the same as that of each IDList structure. When generating k+1 candidate patterns based on the IDList cross-connection of k-candidate patterns, the deletion list of the associated k pattern is used to improve the accuracy of k+1 pattern generation.
[0053] The disturbance example of the IDList list in this embodiment is as follows Figure 5 Based on the perturbed 1-sequence vertical list, find the 1-sequence pattern whose noise support is greater than the support threshold, which is the frequent 1-sequence pattern.
[0054] (4) Based on the frequent 1-sequence pattern, the candidate sequence pattern generation process is recursively called to generate a set of candidate paths. The generated candidate path set is pruned using the atomic expansion project and sequence expansion project after adding the first differential privacy noise, and the pruned candidate path set is added with the third differential privacy noise. Based on the candidate path set after adding the third differential privacy noise, the candidate path set with noise support greater than the support threshold is found as a frequent k-sequence pattern, k≥2; the frequent n-sequence pattern is provided to the untrusted clinical pathway researchers through the query interface, n≥1.
[0055] The detailed steps of the kth (k>=2) recursive call in this embodiment are as follows:
[0056] ① Based on frequent (k-1)-sequence F k-1 , for any two treatment item sequence patterns A i and A j , generate candidate k-sequence C by cross-connecting its IDList K ;
[0057] ②Judge C K Is it empty? If it is empty, end the algorithm; otherwise, continue executing;
[0058] ③Use the co-occurrence frequency of treatment items to analyze C k Pruning, computing and adding differential privacy noise;
[0059] ④ Based on C after adding differential privacy noise K , find the candidate path set whose noise support is greater than the support threshold as the frequent k-sequence pattern F K ; Then k=k+1, go to step ② and continue execution.
[0060] This embodiment uses the co-occurrence frequency of treatment items to prune the candidate pattern set. By filtering out some infrequent candidate patterns, the number of candidate patterns generated in the intermediate process is reduced, thereby reducing the amount of noise required for privacy protection and improving generation efficiency. The detailed steps are as follows:
[0061] ① Based on the frequency of item co-occurrence K Pruning,pruning rules are as follows, let A i =P∪x,A j =P∪y, P is A i and A jThe common prefix of A j The last term, y, can also be called A i The expansion item of A i The last term of P, a is the last term of P, r=A i ∪y. If y is in the atomic expansion item set of treatment project a, then the sequence pattern r is retained, otherwise the sequence pattern r is filtered out; if y is in the sequence expansion item set of treatment project a, then the sequence pattern r is retained, otherwise the sequence pattern r is filtered out.
[0062] ②Judge C K Is it empty? If it is empty, end the algorithm; otherwise, K The IDList of each candidate k-mode in the privatization process is performed: that is, Laplace noise is added to the number of record nodes in the IDList list of each candidate k-mode, and the amount of noise added is m represents the maximum length of the frequent sequence, Δf represents the global sensitivity of the k-sequence calculation task, where Δf = min(ω k , C k ).ω k Indicates the maximum number of candidate k-sequences contained in the longest SDB sequence, C k represents the set of candidate k-sequences actually generated during the mining process. ε represents the set privacy protection strength, and α is the privacy budget allocation ratio, where 0 < α < 1.
[0063] After obtaining the privacy-processed frequent sequence patterns through the above steps, untrusted clinical pathway researchers can query the privacy-processed frequent sequence patterns through the privacy query interface, such as Figure 6 As shown in the figure, untrusted clinical pathway researchers can evaluate and select private frequent sequence patterns and select patterns that meet the criteria as clinical pathways. Based on existing clinical pathway evaluation methods, such as the average length of stay and average drug-drug ratio of a disease-specific sequence pattern, researchers can select frequent sequence patterns with better performance as clinical pathway candidate sets and combine them with medical domain knowledge to create a clinical pathway knowledge base.
[0064] This embodiment provides a private sequence pattern query interface to untrusted clinical pathway researchers, allowing them to establish clinical pathways with the help of noise statistical information, avoiding privacy leaks caused by their access to raw data. This embodiment uses privacy processing based on differential privacy to prevent untrusted researchers from inferring sensitive patient data by analyzing changes in statistical values. This embodiment avoids multiple scans of raw data by designing a private vertical sequence pattern mining algorithm, thereby improving the efficiency of clinical pathway generation while reducing the risk of privacy leaks. This embodiment uses pruning processing based on the co-occurrence frequency of treatment items to improve the accuracy and efficiency of clinical pathway generation while ensuring the same degree of privacy protection.
[0065] The above are preferred embodiments of the present invention. However, once those skilled in the art know the basic inventive concept, they can make additional changes and modifications to these embodiments, and these changes and modifications all fall within the scope of protection of the present invention.
Claims
1. A clinical pathway establishment and optimization method that satisfies differential privacy constraints, characterized by: The following steps are involved: Obtain hospitalization treatment items and time information for patients with the target disease, and establish an anonymous treatment item sequence dataset corresponding to the target disease; Counting the co-occurrence frequencies between different treatment items in the anonymous treatment item sequence dataset, obtaining an atomic expansion item and a sequence expansion item for each treatment item in the anonymous treatment item sequence dataset according to the co-occurrence frequencies, and adding a first differential privacy noise to each atomic expansion item and each sequence expansion item; Scanning the anonymous treatment item sequence dataset, establishing a 1-sequence vertical list corresponding to the anonymous treatment item sequence dataset, and adding a second differential privacy noise to the 1-sequence vertical list; Based on the vertical list of 1-sequences after adding the second differential privacy noise, find the 1-sequence pattern whose noise support is greater than the support threshold and record it as a frequent 1-sequence pattern; Based on frequent 1-sequential patterns, the candidate sequence pattern generation process is recursively called to generate a set of candidate pathways. The generated candidate pathways are pruned using the atomic expansion project and sequence expansion project after adding the first differential privacy noise, and the pruned candidate pathways are added with the third differential privacy noise. Based on the candidate pathways after adding the third differential privacy noise, the candidate pathways whose noise support is greater than the support threshold are identified as frequent k-sequential patterns, k ≥ 2. The frequent n-sequential patterns are provided to untrusted clinical pathway researchers through a query interface, n ≥ 1. The first differential privacy noise is Laplace(b) Laplace(b) represents the Laplace noise with noise scale parameter b, ω2 represents the maximum number of candidate 2-sequences contained in the longest sequence of the anonymous treatment project sequence dataset, |C2| represents the number of true candidate 2-sequences in mining, ε represents the set privacy protection strength, α is the privacy budget allocation ratio, 0<α<1; Based on the frequent 1-sequence pattern, the candidate sequence pattern generation process is recursively called to generate a candidate path set, the generated candidate path set is pruned using the atomic expansion project and sequence expansion project after adding the first differential privacy noise, and the pruned candidate path set is added with the third differential privacy noise; Based on the candidate path set after adding the third differential privacy noise, find the candidate path set whose noise support is greater than the support threshold as a frequent k-sequence pattern, including: ① Based on the frequent (k-1)-sequence pattern, for any two treatment item sequence patterns, a candidate k-sequence C is generated by cross-connecting their vertical lists. K ; ②Judge C K Is it empty? If it is empty, end the algorithm; otherwise, continue executing; ③ Using the atomic expansion project and sequence expansion project after adding the first differential privacy noise to C K Prune the branches and calculate the C K Adding third differential privacy noise; ④ Based on C after adding the third differential privacy noise K , find the candidate path set whose noise support is greater than the support threshold as the frequent k-sequence pattern; then k=k+1, go to step ② and continue; The step of pruning the generated candidate path set by using the atomic expansion item and the sequence expansion item after adding the first differential privacy noise includes: Let A i =P∪x,A j =P∪y, P is A i and A j The common prefix of i and A j For any two treatment item sequence patterns, let y be A i The expansion item of A i The last term of P, a is the last term of P, r=A i ∪y; if y is in the atomic expansion item set of treatment project a, then the sequence pattern r is retained, otherwise the sequence pattern r is filtered out; if y is in the sequence expansion item set of treatment project a, then the sequence pattern r is retained, otherwise the sequence pattern r is filtered out; An atomic expansion project is a treatment project that is carried out together with a treatment project at the same time node, and a sequential expansion project is a treatment project that is carried out together with the treatment project at a subsequent time node.
2. The method for establishing and optimizing a clinical pathway that satisfies differential privacy constraints according to claim 1, wherein: The second differential privacy noise is m represents the maximum length of a frequent sequence; Δf represents the global sensitivity of calculating candidate 1-sequences, which is the maximum change in the number of candidate 1-sequences after adding or removing any sequence; ε represents the set privacy protection strength; α is the privacy budget allocation ratio, 0<α<1.
3. The method for establishing and optimizing a clinical pathway that satisfies differential privacy constraints according to claim 1, wherein: The third differential privacy noise is m represents the maximum length of a frequent sequence; Δf represents the global sensitivity of calculating the candidate k-sequence, which is the maximum change in the number of candidate k-sequences after adding or deleting any sequence; ε represents the set privacy protection strength; α is the privacy budget allocation ratio, 0<α<1.
4. The method for establishing and optimizing a clinical pathway that satisfies differential privacy constraints according to claim 2, wherein: The adding of a second differential privacy noise to the vertical list of 1-sequences includes: Positive noise perturbation processing: Based on the characteristics of the 1-sequence vertical list structure, positive Laplace noise is added to the real count value of the header node. At the same time, dummy nodes are added to the 1-sequence vertical list items according to the noise amount. Each dummy node item contains a dummy SID and a dummy TID. Negative noise disturbance processing: According to the structural characteristics of the 1-sequence vertical list, negative Laplace noise is added to the real count value of the header node. At the same time, the node records in the 1-sequence vertical list are deleted according to the amount of noise, and a deletion item list is designed to record the deleted records in each pattern IDList.
5. The method for establishing and optimizing a clinical pathway that satisfies differential privacy constraints according to claim 1, wherein: The 1-sequence vertical list includes the sequence pattern corresponding to the list, the number of real record nodes contained in the list, and each record; wherein each record structure is the sequence number and treatment time of the sequence pattern.