System and method for generating synthetic longitudinal dataset from original dataset
The method generates synthetic clinical trial data by embedding feature vectors and swapping event patterns to maintain privacy and fidelity, addressing privacy concerns and enhancing data utility in drug development.
Patent Information
- Application Number
- JP2024058952
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-04-12
- Filing Date
- 2024-04-01
- Publication Date
- 2024-11-13
- Estimated Expiration
- 2044-04-01
AI Technical Summary
Existing methods for generating synthetic clinical trial data from electronic data capture (EDC) datasets fail to adequately address privacy concerns and maintain data fidelity, particularly with small datasets, limiting their usefulness in drug development and research.
A method that generates high-fidelity synthetic longitudinal datasets by embedding feature vectors into a low-dimensional space, swapping similar event patterns, and adding noise to preserve privacy while maintaining statistical properties.
The method produces synthetic datasets that closely mimic original data characteristics, enabling secure and effective analysis for drug development and research without compromising patient privacy.
Smart Images

Figure 2024160204000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure generally relates to generating synthetic longitudinal (or time series) datasets from original datasets. [Background technology]
[0002] When it comes to health technology, innovation has recently been driven by the analysis of aggregated patient data. Currently, the most important category of such patient data is electronic health records (EHRs), and public and private aggregation of EHRs can enable a variety of uses, including prioritizing drug development programs, optimizing trial designs, and improving healthcare delivery processes. Another (but related) category of patient data that is becoming increasingly important for health technology is patient-level data collected using electronic data capture (EDC) technologies in clinical trials.
[0003] There are strong similarities between patient-level EHRs and EDCs, but EHRs are collected sporadically from patients visiting their health care providers, whereas EDCs are collected from patients who consent to participate in controlled experiments. Such experiments are designed to make high-value inferences (e.g., the effect of a drug versus a placebo on the survival rate of patients with a particular disease under a particular treatment regime). EDC data represent high-dimensional data that are collected repeatedly for each patient on a regular schedule under a carefully controlled and regulated process. EDC patient sample sizes are in the tens to hundreds, compared to the thousands for many EHR datasets. For applications such as estimating the power of clinical trial designs, discovering the effects of concomitant medications, and identifying patient factors that affect the success rate of clinical trials, EDCs are significantly more valuable as data resources than EHRs. Summary of the Invention [Problem to be solved by the invention]
[0004] Although the potential research value of clinical trial data has been well-recognized, to date, patient-level EDC remains largely unpublished. Of the hundreds of thousands of clinical trials conducted annually around the world, only a small fraction has been made widely available. As with other sensitive data, the main reason for this limited availability is that patient-level data from clinical trials are subject to regulation, privacy concerns, and require compliance with technical protection protocols. While anonymization or de-identification techniques can overcome some of these concerns, these techniques cannot fully address privacy and confidentiality concerns or patient consent issues, as traditional techniques for de-identifying and anonymizing EDC data cannot fully guarantee patient privacy for this type of data.
[0005] Traditional techniques such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Natural Language Processing (NLP) can provide privacy controls by using deep learning and machine learning to generate synthetic data from learned distributions. However, these techniques require large amounts of data provided by EHRs and do not work well with small datasets such as EDC data from clinical trials. Thus, privacy concerns remain for both subjects and clinical trial sponsors. [Means for solving the problem]
[0006] The disclosure herein provides methods, systems, and computer readable media for generating high fidelity longitudinal EDC data, including timed events, while retaining a high level of privacy at the subject level. The disclosed techniques balance data fidelity with data privacy by obfuscating the data in a manner that renders it unidentifiable to unauthorized users, while retaining the statistical properties of the data. This allows researchers to broadly access and analyze such data early in clinical trials, accelerating the development of new drugs and treatments.
[0007] The disclosed embodiments combine low-dimensional embedding of feature representations of source data with exchange of similar event patterns between records to identify nearest neighbors of records in a compressed feature space. The synthetically generated data is statistically similar to the source data and captures the underlying dependencies and event patterns of the source. The disclosed method does not store or retain records from the original data and does not reveal specific features from the original data. This enables practical use of the disclosed method given high privacy bar for patient-level EDC.
[0008] In one aspect, the disclosed embodiments provide a method, system, and computer-readable storage medium for generating a synthetic longitudinal dataset (X'). The method includes identifying subsequence patterns in N records (r) that define event sequences for N patients, resulting in m subsequence patterns, where m and N are integers. The method further includes determining feature vectors (Y'), each characterizing a corresponding one of the records (r), based at least in part on the m subsequence patterns. The method further includes embedding the feature vectors (Y') into a low-dimensional space (V) of dimension y to define a position of the records (r) in the low-dimensional space (V), where y is an integer and m>y>0. The method further includes identifying seed records (r's) from among the records (r). s) from the records (r) in a recursive manner. s Each time a seed record (r) is selected, the following steps are performed: Identifying a sub-sequence pattern in a subset of records (r) determined based at least in part on the position of the records (r) in the low-dimensional space (V). s ) with an instance of a corresponding subsequence pattern identified in a subset of records (r) based at least in part on the determined distance between the respective instances of the subsequence patterns to generate a modified seed record (r s A substitution step to form a new record (r '). s The selection, identification, and replacement steps are repeated until a modified seed record (r s The method includes a combining step of combining the data sets (X') to generate a synthetic data set (X').
[0009] Embodiments: One or more of the following features may be used alone or in combination.
[0010] The dataset (Y) may include event records for N patients, each event record including a patient identifier, a start date of the event, or a start date of the event and an end date of the event, and one or more attributes of the event. In such a case, the method may further include determining, for each of the N patients, an event sequence ordered by the start date of each event in the event records associated with each of the N patients to generate N records (r) defining the event sequences for the N patients.
[0011] The original dataset (X) may include multiple tables, each including a subset of the event records of the N patients, where the method may further concatenate the tables to form the dataset (Y). Each event record may include an event code identifying a corresponding event associated with the respective event record. The patient identifier of each event record may be uniquely associated with one of the N patients. For at least one of the event records, the event is one of a clinical procedure, intervention, or treatment, administration of a drug or medication, occurrence, report, onset, or termination of a symptom or condition, an adverse reaction to a drug, medication, or treatment, a laboratory test, measurement, evaluation, or diagnosis, and an output of a monitoring device, a patient report or complaint. For at least one of the event records, the start date may correspond to a date a procedure is performed on a patient identified by the patient identifier. For at least one of the event records, the start date may correspond to a first date that a patient identified by the patient identifier experiences a symptom, and the end date may correspond to a date that a patient identified by the patient identifier stops experiencing a symptom. For at least one of the event records, the attributes may include at least one of a categorical attribute and a numerical attribute.
[0012] The m subsequence patterns may include subsequence patterns having a frequency of occurrence greater than a predefined threshold. The m subsequence patterns may include the m most frequently occurring subsequence patterns, where m is a defined parameter. The feature vector (Y') may be of equal length. The feature vector (Y') may include the frequency of occurrence of the m subsequence patterns. The feature vector (Y') may include a binary presence / absence encoding of the m subsequence patterns.
[0013] The embedding step can be performed using one or more of t-distributed stochastic neighborhood embedding (t-SNE), uniform manifold approximation and projection (UMAP), and principal component analysis (PCA). A subset of records (r) is determined by finding the k nearest neighbors (rs ) can be supported. s The k nearest neighbors of a record (r) can be determined using the Hamming distance or edit distance. Let us consider one of the records (r) as the seed record (r s ) can be repeatedly selected randomly.
[0014] The replacement step may be performed only for those subsequence patterns for which the determined distance between respective instances of the subsequence patterns is less than a predetermined distance threshold. s ) with an instance of a corresponding subsequence pattern selected randomly from a group of subsequence patterns whose respective distances are less than a predetermined distance threshold. This replacement step may exclude subsequences whose total number of occurrences in the N event sequence is less than a predetermined threshold.
[0015] The method further comprises, in each iterative step, selecting a seed record (r s ) the start date of the first event in the seed record. The randomization may include adding noise sampled from a uniform distribution. The method may further include randomizing, at each iterative step, a gap between one or more pairs of events in the seed record (r s ), the randomization may include adding Gaussian noise.
[0016] The method may further include determining a conditional probability of each event in each of the N event sequences given that the preceding event occurred. The method may further include randomly replacing each event with its respective preceding event in each iterative step if the conditional probability is below a predetermined threshold. The method may further include outputting the synthetic dataset (X') to a storage system that provides encryption and access control. The synthetic dataset (X') may be output in at least one of CSV, JSON, or binary formats. [Brief description of the drawings]
[0017] [Figure 1] FIG. 1 is a block diagram of a system for generating and distributing synthetic longitudinal datasets according to disclosed embodiments. [Diagram 2] 1 is a block diagram of a synthetic data processing system according to disclosed embodiments; [Figure 3A] 1 illustrates an example in which a subsequence pattern of a seed record is replaced with a corresponding subsequence pattern of a record in the same cluster. [Figure 3B] 1 illustrates an example in which a subsequence pattern of a seed record is replaced with a corresponding subsequence pattern of a record in the same cluster. [Figure 4] It is an orthogonal representation of a feature vector (Y') embedded in a low-dimensional space (V). [Diagram 5] 1 is a flow chart of a method for generating a synthetic longitudinal dataset. [Figure 6A] Histograms of event sequence lengths for real and synthetic datasets from a clinical trial targeting non-small cell lung cancer (NSCLC). [Figure 6B] Histograms of event sequence length for real and synthetic datasets from a clinical trial targeting acute lymphoblastic leukemia (ALL). [Figure 6C] Histograms of event sequence lengths for real and synthetic datasets from a multiple myeloma clinical trial. [Figure 7A] This is a plot of the results of a cross-validation study of a clinical trial targeting non-small cell lung cancer (NSCLC). [Figure 7B] This is a plot of the results of a cross-validation study of a clinical trial targeting non-small cell lung cancer (NSCLC). [Figure 7C] This is a plot of the results of a cross-validation study of a clinical trial targeting non-small cell lung cancer (NSCLC). [Figure 7D] This is a plot of the results of a cross-validation study of a clinical trial targeting non-small cell lung cancer (NSCLC). [Figure 7E] This is a plot of the results of a cross-validation study of a clinical trial targeting non-small cell lung cancer (NSCLC). [Figure 7F] This is a plot of the results of a cross-validation study of a clinical trial targeting non-small cell lung cancer (NSCLC). [Figure 8A] 1 is a plot of cross-validation study results of a clinical trial targeting acute lymphoblastic leukemia (ALL). [Figure 8B] 1 is a plot of cross-validation study results of a clinical trial targeting acute lymphoblastic leukemia (ALL). [Figure 8C] 1 is a plot of cross-validation study results of a clinical trial targeting acute lymphoblastic leukemia (ALL). [Figure 8D] 1 is a plot of cross-validation study results of a clinical trial targeting acute lymphoblastic leukemia (ALL). [Figure 8E] 1 is a plot of cross-validation study results of a clinical trial targeting acute lymphoblastic leukemia (ALL). [Figure 8F] 1 is a plot of cross-validation study results of a clinical trial targeting acute lymphoblastic leukemia (ALL). [Figure 9A]1 is a plot of cross-validation study results of a clinical trial on multiple myeloma. [Figure 9B] 1 is a plot of cross-validation study results of a clinical trial on multiple myeloma. [Figure 9C] 1 is a plot of cross-validation study results of a clinical trial on multiple myeloma. [Figure 9D] 1 is a plot of cross-validation study results of a clinical trial on multiple myeloma. [Figure 9E] 1 is a plot of cross-validation study results of a clinical trial on multiple myeloma. [Figure 9F] 1 is a plot of cross-validation study results of a clinical trial on multiple myeloma. [Figure 10A] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of the disclosed method. [Figure 10B] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of the disclosed method. [Figure 10C] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of the disclosed method. [Figure 10D] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of the disclosed method. [Figure 10E] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of the disclosed method. [Figure 11A] 1 is a plot comparing real and synthetic datasets to illustrate the performance of n-gram models used in natural language processing (NLP). [Figure 11B] 1 is a plot comparing real and synthetic datasets to illustrate the performance of n-gram models used in natural language processing (NLP). [Figure 11C]1 is a plot comparing real and synthetic datasets to illustrate the performance of n-gram models used in natural language processing (NLP). [Figure 11D] 1 is a plot comparing real and synthetic datasets to illustrate the performance of n-gram models used in natural language processing (NLP). [Figure 11E] 1 is a plot comparing real and synthetic datasets to illustrate the performance of n-gram models used in natural language processing (NLP). [Figure 12A] 1 is a plot comparing real and synthetic datasets to illustrate the performance of a Long Short-Term Memory (LSTM) neural network. [Figure 12B] 1 is a plot comparing real and synthetic datasets to illustrate the performance of a Long Short-Term Memory (LSTM) neural network. [Figure 12C] 1 is a plot comparing real and synthetic datasets to illustrate the performance of a Long Short-Term Memory (LSTM) neural network. [Figure 12D] 1 is a plot comparing real and synthetic datasets to illustrate the performance of a Long Short-Term Memory (LSTM) neural network. [Figure 12E] 1 is a plot comparing real and synthetic datasets to illustrate the performance of a Long Short-Term Memory (LSTM) neural network. [Figure 13A] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of convolutional neural networks (CNNs). [Figure 13B]1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of convolutional neural networks (CNNs). [Figure 13C] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of convolutional neural networks (CNNs). [Figure 13D] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of convolutional neural networks (CNNs). [Figure 13E] 1 is a plot showing a comparison of real and synthetic datasets to illustrate the performance of convolutional neural networks (CNNs). [Figure 14A] 13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. [Figure 14B] 13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. [Figure 14C] 13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. [Figure 14D] 13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. [Figure 14E] 13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. [Figure 14F] 13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. [Figure 14G]13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. [Figure 14H] 13 is a plot comparing real and synthetic datasets to illustrate the performance of the technique in terms of event sequence length and event type frequency. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0018] In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of embodiments of the present invention. However, it will be understood by those skilled in the art that embodiments of the present invention may be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to obscure the present invention. Where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or similar elements. Furthermore, some of the blocks depicted in the figures may be combined into a single function.
[0019] The disclosed embodiments provide a system and method that addresses the above-mentioned privacy challenges by generating synthetic longitudinal datasets that preserve high fidelity. The method converts variable-length sequences into fixed-length feature vector representations using the frequencies of their event patterns, and then uses low-dimensional embedding of the feature vectors to identify neighborhoods of records in the compressed feature space, thereby identifying similar sequences, i.e., sets of records. The method then finds all subsequence patterns within the neighborhood regions of similar sequences, and swaps instances of patterns (i.e., subsequences) that are similar to each other across different sequences. The method generates high-fidelity synthetic longitudinal data while performing sufficient transformations on the original sequences to maintain privacy.
[0020] Maintaining data privacy for clinical trial data is a technical issue because it requires the development, implementation, and maintenance of complex methodologies and tools to ensure that sensitive patient information is protected and not exposed during data analysis and sharing. While the underlying motivation for ensuring data privacy may stem from ethical, legal, and regulatory concerns, the actual process of achieving data privacy involves addressing a number of technical challenges. Addressing data privacy concerns in clinical trial data involves several technical aspects, such as:
[0021] De-identification and anonymization: These processes involve removing personally identifiable information (PII) from datasets so that the remaining information cannot be used to identify individual patients.
[0022] Data Encryption: Secure storage and transmission of clinical trial data requires the use of strong encryption algorithms that protect the information from unauthorized access or interception.
[0023] Differential Privacy: This is a framework for ensuring privacy and quantifying the privacy risks associated with the disclosure of aggregate data derived from sensitive datasets. Differential privacy techniques introduce controlled modifications to data, making it harder to infer specific information about individuals while preserving the overall usefulness of the data for analysis.
[0024] The creation of synthetic datasets of clinical trial data provides a technical solution to the technical problems mentioned above, as it addresses some of the challenges inherent in working with real-world clinical trial data. Synthetic data creation techniques can create datasets that mimic the characteristics of the original data without exposing sensitive information, allowing researchers to work with the data while maintaining patient privacy. Access to clinical trial data can be limited by a variety of factors, including ownership, data silos, and strict access controls. Synthetic datasets help overcome such barriers by generating additional data with similar characteristics to the original data, allowing researchers to develop and validate models without direct access to the real data.
[0025] Moreover, the technical solution includes algorithms and / or rules with specific properties that generate highly obfuscated synthetic data while maintaining the statistical properties of the original dataset. As described below, the technical solution focuses on a specific improvement in the generation of synthetic datasets by using specific algorithms and / or rules. It is the implementation of the specific algorithms and / or rules, not merely the use of a computer, that improves existing technical processes by providing improved data obfuscation and thus improved data privacy. In this way, the technical solution provides a concrete approach to improve the relevant technology, as opposed to being merely directed at abstract results or effects that invoke generic processes or machines.
[0026] In summary, the creation of synthetic datasets for clinical trial data is a technical solution to a technical problem, addressing specific data privacy challenges that arise when dealing with real-world clinical trial data. The technical solution involves applying specific algorithms and / or rules to render information into a new dataset with specific properties (i.e., similar statistical properties as the original dataset) that can be used by researchers to perform analysis of the clinical trial data. The algorithms and / or rules described herein do not simply organize existing information into a new form, but generate entirely new datasets. The synthetic data generated in this way can be used to overcome the challenges mentioned above and enable more effective development, testing, and deployment of data-driven solutions in clinical research.
[0027] The following terms are used herein: "Event" is a term used in the broad sense to refer to an element or item in a series of elements or items. Examples include clinical procedures, interventions, or treatments; administration of a drug or medication; occurrence, report, onset, or termination of a symptom or condition; adverse reactions to a drug, medication, or treatment; laboratory tests, measurements, evaluations, or diagnoses; monitoring device outputs; and patient reports or complaints. "Event group" refers to an unordered or ordered collection of events that tend to occur together. "Sequence" refers to an ordered collection (usually of any length) of events or event groups. "Sub-sequence" refers to a collection of events or event groups that represents a portion of a sequence. "Sub-sequence pattern" refers to a sub-sequence that repeats with a particular frequency in a set or collection of sequences; the sub-sequence pattern may be of variable length and may contain gaps (i.e., "wildcards"). "Feature vector" refers to a numeric vector (usually fixed length) representation of an array. An "instance of a subsequence pattern" refers to the occurrence of a subsequence in a sequence that matches a subsequence pattern. A "Record" or "Event Record" refers to a set of events in a dataset that are associated with a particular patient. As used herein, the term "swap" or "swapping" may refer to a unidirectional operation, such as replacing a subsequence of a seed record with a corresponding similar subsequence of another record (e.g., based on a distance determination), or a bidirectional operation, such as reversing the order of two events in an event train based on their conditional probabilities. A "longitudinal dataset" refers to a collection or set of sequences with associated numerical, temporal, categorical, and / or continuous attributes (e.g., timestamps, parameter values, etc.).
[0028] FIG. 1 is a block diagram of a system 100 for generating and distributing synthetic longitudinal datasets. The system 100 receives clinical trial data, for example, via an Electronic Data Capture (EDC) system 110, which receives data from multiple data sources, such as clinical sites. The data from the Electronic Data Capture system 110 is a dataset, referred to as an original dataset 120, designated herein as X. As discussed in more detail below, the original dataset 120 is processed by a synthetic data processing system 140 to generate a synthetic dataset 150, designated herein as X′. The synthetic dataset 150 may be distributed to researchers for analysis, for example, via a secure data sharing platform, referred to as a synthetic data distribution system 160, which makes the data available via a network 170.
[0029] The original dataset 120 can contain many variable length sequences, each containing ordered or time-stamped items and item sets. An item can represent an event, which can have a start date / time and an optional end date / time. Each event can also have additional attributes, which can be numerical or categorical. The method disclosed herein is designed to work with both large and small (hundreds to thousands of subjects) datasets, but is particularly advantageous (compared to state-of-the-art algorithms) when working with small datasets, such as those obtained from clinical trials. Conventional approaches using neural networks, on the other hand, require very large training datasets.
[0030] The original dataset 120 consists of a number of tables (T1, T2...T n), each containing a subset of the event records for N patients. For example, clinical trial data may be stored as many separate tables (e.g., in the CSV format commonly used for spreadsheets) that contain one row per subject and specific information relevant to each subject, such as a table for subject-level information, according to a particular standard. Many supplementary tables may provide data on adverse events, medications, medical history, and sometimes contain multiple entries for a particular subject.
[0031] In another example, tables and records of the original dataset 120 may be combined to form a composite dataset 130 (referred to herein as Y), which may store data in various data structures such as rows and columns. The composite dataset Y has event records for N patients, with each row having a subject ID (i.e., patient identifier), an event code (identifying a corresponding event associated with the respective event record), an event start date, and optionally an event end date, as well as several other categorical and numerical values. The event start date of a particular record may correspond to a date on which a procedure is performed on a patient identified by the patient identifier of the particular record. In some cases, the start date of a particular record may correspond to a date on which a patient identified by the patient identifier experiences a particular continuous symptom, and the end date may correspond to a second date on which a patient identified by the patient identifier no longer experiences a particular symptom. The categorical and numerical values may include patient information and demographics (e.g., name, age, sex, country), vital statistics (e.g., height, weight, blood pressure, heart rate, etc.), as well as other types of data.
[0032] Here, a synthetic longitudinal dataset 150 is generated from data received from the electronic data collection system 110. The synthetic longitudinal dataset 150 is designed to mimic the statistical properties of the original dataset 120. This preserves the statistical properties of the original dataset 120, allowing researchers to draw accurate conclusions, while de-identifying and obfuscating the data to reduce the risk of identifying individual participants and violating their privacy. The statistical properties of the synthetic dataset 150 can be verified against the original dataset 120 to ensure that it is close to the original dataset. This can be done using various statistical tests, such as regression analysis.
[0033] The synthetic datasets can be distributed to researchers for analysis via a secure data sharing platform, such as, for example, a synthetic data distribution system 160, which may involve setting up private or public application program interfaces (APIs), creating data repositories, data portals, and / or using cloud-based data storage services with appropriate access control, authentication, and encryption mechanisms. Researchers can use the synthetic datasets 150 to conduct longitudinal analyses while minimizing the risk of identifying individual participants.
[0034] FIG. 2 is a block diagram of a synthetic data processing system 140. As described above, an original longitudinal data set (X) may be received that includes event records for a number N of patients, where N is an integer. The tables and records of the original data set (X) may be combined to form a composite data set (Y). Each record includes a start date (and optional start time) of the event, or both a start date and an end date of the event, and one or more attributes of the event. Based on the event information, the event sequencer 210 determines an event sequence for each of the N patients, ordered by the start date of each event in the event record associated with each patient, thereby generating a set of N records (r) that define the event sequences for the N patients. The length of the event sequence varies from patient to patient, since each patient undergoes a particular event sequence, and the length is determined by the number of events associated with each patient. Some patients may only experience, perform, undergo, etc., a few events, while other patients may be associated with a large number of events, depending on their clinical history.
[0035] The pattern mining unit 220 receives input data from the event sequencer 210 and performs an analysis to find subsequence patterns in all sequences, with or without gaps. Specifically, it identifies subsequence patterns in N records (r) that define event sequences for N patients, and generates m subsequence patterns, where m and N are integers. In some implementations, the m subsequence patterns may include all subsequence patterns with a frequency greater than a predefined threshold, or alternatively, the m patterns with the highest frequency, where m is a defined parameter. In this way, all frequent subsequence patterns in the entire dataset are identified by pattern mining (e.g., all subsequence patterns with a frequency greater than or equal to a predefined threshold). The patterns may contain gaps and wildcards. For example, a frequent subsequence pattern may be: <event1,event5,event10>
[0036] The feature vector generator 230 generates a set of feature vectors, designated herein as Y', based at least in part on the m subsequence patterns. The feature vector Y' represents each of the N records, designated herein as r, that define the event sequence as a vector that designates an instance of a unique subsequence pattern in each event sequence. In some implementations, the feature vectors Y' are of equal length. The feature vector Y' may be created using the frequencies of the subsequence patterns. For example, the feature vector Y' may be created using the frequency of occurrence of the m subsequence patterns or using a binary presence / absence encoding of the m subsequence patterns. Thus, each event sequence, i.e., record (r), has a corresponding feature vector representation, e.g., as a histogram of all patterns it contains, or simply as all items (e.g., events). In some implementations, one feature vector is generated for each record (r), and all feature vectors have the same length.
[0037] The embedding unit 240 takes the feature vectors Y' and embeds them into a low-dimensional manifold, i.e., a low-dimensional space with dimension y, designated herein as V, where m>y>0, where y is an integer, defines the location of the records (r) in the space (V). Preferably, y is small, e.g., y=2 or y=3, resulting in a low-dimensional feature space in which subsequent clustering, e.g., k-nearest neighbor clustering, works better. The embedding effectively helps determine which records are similar to each other.
[0038] In implementations, embedding can be performed using a variety of methods, including t-SNE, Uniform Manifold Approximation Projection (UMAP), and Principal Component Analysis (PCA). UMAP is similar to t-SNE, but assumes that the data is uniformly distributed on a locally connected Riemannian manifold and that the Riemannian metric is locally constant or nearly locally constant. PCA can be used to compute the principal components (i.e. dimensions in a low-dimensional space) of a set of records, and these can be used to perform basis changes on the data, for example, by using only the first few principal components and ignoring the rest.
[0039] Figure 4 shows an orthogonal representation of a feature vector (Y') embedded in a low-dimensional space (V) of dimension y = 2. Each point (i.e., feature vector position) corresponds to one of the records (r). The point (r s ) is surrounded by the cluster of points in space V that correspond to the nearest record (r).
[0040] 2, the modification engine 250 receives the records (r) and their embedding representations in the low-dimensional space (V) and uses them to generate modified records, i.e., synthetic data records, to generate the synthetic data set (X'). As will be described in more detail below, this involves selecting seed records (r) from among the records (r) and generating a synthetic data set (X'). s ) is repeatedly selected and the modified seed record (r s These iterations consist of running an algorithm that generates a seed record (r ') for every record (r). s ) is selected iteratively. The modified seed record (r s ') to generate a synthetic dataset (X'), which can then be used in place of dataset (Y), for example for research purposes.
[0041] The modification engine 250 uses a specific seed record r s , for example using the Hamming distance or edit distance, to find the seed record rs The clustering unit 260 defines a subset of records (r) related to the seed record r, such as the k nearest neighbors of the seed record r. In the implementation, in each iteration, one of the records r is clustered with the seed record r. s The value of k can be chosen heuristically based on a trade-off between data fidelity and the difficulty of decrypting the synthetic dataset, i.e., inferring records in the original dataset from the synthetic dataset (which essentially corresponds to data privacy), and based on the particular application. Larger k reduces fidelity but increases data privacy.
[0042] The modification engine 250 modifies the selected seed record r s , and its neighboring records r in the space V, i.e., based on the location of the embedded feature vector (Y') corresponding to record (r). Thus, the pattern mining unit 265 operates on a subset of the records (r) provided by the clustering unit 260. This pattern mining is in some sense a localized version of the operation of the pattern mining unit 220 described above, which operates on the entire dataset (Y).
[0043] The pattern mapping unit 270 identifies relationships between corresponding, e.g., similar, patterns. Similarity between instances of subsequence patterns identified by the pattern mining unit 265 may be determined based on a calculated distance (e.g., Hamming distance or edit distance) between the subsequences, where "distance" refers to a metric that assesses similarity, e.g., how many corresponding bit positions differ in equal length binary representations of the patterns. For example, a distance, e.g., Hamming distance, may be determined between the subsequence patterns identified by the pattern mining unit 265, the smaller the distance, the closer the seed record r sIt indicates that the subsequence pattern of V has a high similarity to a selected neighboring record r in the space V. In such a case, each subsequence pattern may be matched to a corresponding subsequence pattern in a cluster having a distance less than a threshold.
[0044] Next, the pattern exchange unit 275 converts the seed record r s Replace one or more instances of the subsequence pattern of with instances of similar patterns in the vicinity (i.e., within a cluster) to produce a modified seed record r s Generate a new seed record r s is determined and operation returns to the clustering unit 260. In an embodiment, the seed record r s The replacement of an instance of a subsequence pattern of V with a similar subsequence pattern of a neighboring record r in the space V is performed based at least in part on the determined distance between the respective instances of the subsequence patterns. In some implementations, the swapping is performed only for subsequences having a distance less than a predetermined distance threshold. In some cases, the replacement of an instance of a subsequence pattern of V with a similar subsequence pattern of a neighboring record r in the space V is performed based at least in part on the determined distance between the respective instances of the subsequence patterns. In some implementations, the swapping is performed only for subsequences having a distance less than a predetermined distance threshold. s The swapping may be performed by replacing each subsequence of with a subsequence randomly selected from a group of subsequences having a distance less than a predetermined distance threshold.
[0045] The correction seed records r generated by the correction engine 250 in an iteration s ' are combined and output as a synthetic dataset X'. The synthetic dataset X' may be in the form of a set or collection of event sequences of various lengths for N patients. In summary, these new synthetic records (r s ') is generated by going through every original sequence, i.e., record r, and replacing its subsequence patterns with subsequence patterns from each of its neighborhoods.
[0046] Modified seed record rs The generation of ' can provide for filtering or excluding from pattern swapping any subsequence patterns whose sum of occurrences across datasets Y is less than a determined threshold. The exclusion or filtering eliminates relatively rare events, e.g., rare adverse reactions or symptoms, because such events may facilitate deciphering the synthetic dataset, i.e., inferring records of the original dataset from the synthetic dataset.
[0047] Modified seed record r s ' is the seed record r s The modified seed record r may include randomizing the start date of the first event in the seed record. Such randomization may include adding noise sampled from a uniform distribution. s ' is the seed record r s This randomization of start dates and intervals between events makes the synthetic data set more difficult to decipher without substantially affecting the fidelity of the research data.
[0048] In other implementations, a determination of the conditional probability of each event (or group of events) given the occurrence of the preceding event (or group of events) in each of the N event sequences may be made. In such cases, the modified seed record r s Generating ' may involve randomly swapping each event with its respective preceding event if the conditional probability with respect to the preceding event is less than a predetermined threshold.
[0049] 3A and 3B show an example where an instance of a subsequence pattern in a seed record 310 is replaced with a corresponding subsequence pattern in a neighboring record 315 in the same cluster 320. A table 325 is shown listing examples of patterns found in the cluster 320 and their similar patterns determined using distance metrics such as Hamming distance and edit distance. An item set (e.g., an event group) in the seed record 310 or neighboring record 315 contains two or more items and is shown in parentheses. In this example, the seed record 310 contains a particular event sequence that contains both an event and an event group. Among the events and event groups are two instances of the subsequence pattern identified in the cluster 320, namely, "BC)(A" and "F)B" (see table 325). These subsequences are replaced with the corresponding similar subsequences to generate a modified seed record 330. In an embodiment, the subsequence patterns may have gaps and / or may be grouped into different event groups than the subsequence patterns they are replaced with, as in the depicted example.
[0050] FIG. 5 is a flow chart of a method 500 for generating a synthetic longitudinal data set (X'). The method includes identifying subsequence patterns resulting in m subsequence patterns in N records (r) defining event sequences for N patients, where m and N are integers (510). A feature vector (Y') each characterizing a corresponding one of the records (r) is determined (520) based at least in part on the m subsequence patterns. The feature vector (Y') is embedded in a low-dimensional space (V) of dimension y to define a position of the record (r) in the low-dimensional space (V). A seed record (r') from among the records (r) is identified as a subsequence pattern resulting in m subsequence patterns, where m and N are integers (520). s) are iteratively selected (540). At each iteration, subsequence patterns are identified in a subset of the records (r), the subset being determined (550) based at least in part on the position of the records (r) in the low-dimensional space (V). s ) is replaced with a corresponding instance of the subsequence pattern in a subset of records (r) based at least in part on the determined distance between the respective instances of the subsequence pattern, and a modified seed record (r s ') is generated (560). All records (r) are seed records (r s A determination is then made (570) as to whether the seed record (r s ), identification of subsequence patterns, and seed records (r s The replacement of instances of the subsequence in the set (r ') is repeated iteratively. Every record (r) is replaced by a seed record (r s ), the revised seed record (r s ) are combined to generate a synthetic data set (X') (580).
[0051] In addition to the above description and accompanying figures, the following methodological pseudocode 1-8 for generating longitudinal data from a clinical trial can be implemented at least in part based on the following methodological pseudocode 1-8.
[0052] 1. Original dataset: X consists of tables, each of which contains records of N patient events, start and end dates (optional), and other categorical and numerical attributes. 2. Join the table to dataset Y such that each row has a subject ID, an event code, an event start date, an optional event end date, and other categorical or numeric values. 3. Calculate the conditional probability of observing event code a given that event code b occurs in the patient's event timeline. 4. Arrange the events for each patient (N in total) in order of the start dates of the events to create an event sequence. 5. Use pattern mining to identify m frequent subsequence patterns across the event sequences to use as the feature vector Y'. 6. Embed Y' into the y-dimensional space V (e.g., t-SNE, UMAP, PCA, etc.) to generate embedded points (i.e. records). 7. Use the following properties: a. N>0 (size of synthetic data set) b. k>0 (number of neighbors, also called clusters) c. y (embedding dimension m>y>0) d. c>=1, C⊆R e. p>=0, p⊆R f. T(.) conversion mechanism 8. Generate a synthetic dataset X' from the longitudinal data of N patients While i is equal to or less than N, the following steps (1) to (5) are repeated. (1) Seed record s Generation of (2)r s Select the k nearest neighbors of (3) Create a new record by following (a) to (e) below. s Generate ' (a) The number of subjects in which event e exists is less than count c. s Drop event e in the event sequence (i) Seed record s Swapping event patterns between and its k nearest neighbors. Using pattern mining, s Identify patterns in and its k neighboring records, use distance to find instances of the patterns that are close to each other, and swap (c) Find all consecutive events in the order of (2) for which the conditional probability of event a occurring after event b is less than p, and randomly reverse the order of occurrence. (D) By adding noise sampled from the uniform distribution [-U,U], r sAnonymize the start date of the first event in the event sequence (e) Add noise to the lag between the start and end of the first event and all subsequent events so that the order of the sequence is preserved. The noise is added using Gaussian noise N(0,σ), but another mechanism could be used. (4) r to combine with other generated records to form X'. s ' to save. (5) i = i + 1;
[0053] To evaluate the fidelity of the generated longitudinal dataset, we conducted cross-validation tests that compared the properties of the synthetic dataset with those of the real dataset to evaluate the extent to which basic properties of the real dataset were preserved. As described in more detail below, many metrics can be used for cross-validation, such as event train length, mean event duration, and event frequency. Not depicted in the plots below are also the mean parameter values of continuous attributes for specific events, such as heart rate measurements. These means can be compared between real and synthetic data, as described below. For categorical attributes (e.g., ethnicity, race, and gender), modes and frequencies can be used as cross-validation measures.
[0054] Figures 6A-C show histograms of event sequence lengths in cross-validation of real and synthetic datasets from clinical trials targeting non-small cell lung cancer (NSCLC), acute lymphoblastic leukemia (ALL), and multiple myeloma, respectively. The histograms overlay the length of the records (r) defining the event sequences in the original dataset (or “real dataset”) and the synthetic dataset created according to the approach described herein. The histograms can be statistically compared, for example, using the Kolmogorov-Smirnov test. The Kolmogorov-Smirnov statistic quantifies the distance between the empirical distribution function of the sample (e.g., synthetic dataset) and the cumulative distribution function of the reference distribution (e.g., real dataset). The values of the Kolmogorov-Smirnov statistic in Figures 6A-C are 0.01, 0.01, and 0.02, respectively, and the value approaches zero as the difference between the distributions becomes smaller. The Kolmogorov-Smirnov test produced a measure of goodness of fit with values p=0.99, p=1.0, and p=0.97, respectively, with p=1.0 indicating perfect fit.
[0055] 7A to 7F are plots of the results of a cross-validation study of a clinical trial targeting NSCLC. FIG. 8A to 8F are plots of the results of a cross-validation study of a clinical trial related to ALL. FIG. 9A to 9F are plots of the results of a cross-validation study of a clinical trial related to multiple myeloma.
[0056] Figures 7A, 8A, and 9A show a comparison of the mean duration of events between real and synthetic datasets for events that have both a start date (or date / time) and an end date (or date / time). For example, the mean duration of an adverse event, such as nausea, can be calculated for subjects in the synthetic dataset and compared to that of the real dataset. In these plots, data from the original dataset is plotted on the x-axis and data from the synthetic dataset on the y-axis. Thus, the ideal distribution would fall along a straight line with a slope of 1.0 as shown in the figures. The fit of the data to the ideal profile can be assessed using simple linear regression, using the coefficient of determination (r 2 ) can be quantified using the r 2 The values of r are 0.56, 0.99, and 0.61, respectively. 2 A value of =1.00 indicates a perfect fit to the linear characteristic.
[0057] Figures 7B, 8B and 9B show a comparison of event frequencies. In this example, the frequency of all events is determined cumulatively, including, for example, nausea, headache, other symptoms, adverse events, blood pressure measurements, etc. Data from the real data set is plotted on the x-axis and data from the synthetic data set on the y-axis, in this example both axes have a logarithmic scale. An ideal distribution would fall along a line with a slope of 1.0 as shown in the figures. The r 2 The values of r are 1.0, 1.0, and 1.0, respectively. 2 A value of =1.00 indicates a perfect fit to the linear characteristic.
[0058] Figures 7C, 8C, and 9C show a "Bag of Word" comparison of event frequency feature vectors between real and synthetic datasets. In this example, the number of occurrences of each particular event is determined, such as nausea, headache, other symptoms, adverse events, and blood pressure measurements. Data from the real dataset is plotted on the x-axis and data from the synthetic dataset on the y-axis, with both axes having a logarithmic scale in this example. An ideal distribution would fall along a line with a slope of 1.0 as shown in the figures. The r in Figures 7C, 8C, and 9C 2 The values of r are 1.0, 1.0, and 1.0, respectively. 2 A value of =1.00 indicates a perfect fit to the linear characteristic.
[0059] Figures 7D, 8D, and 9D compare the 2-gram counts of the real and synthetic datasets. The term "2-gram" is used in a way similar to natural language processing to refer to a sequence of two consecutive events. In this example, the number of occurrences of each particular event "2-gram" is determined. The data from the real dataset is plotted on the x-axis and the data from the synthetic dataset on the y-axis. An ideal distribution would fall along a straight line with a slope of 1.0 as shown. The r 2 The values of r are 0.97, 0.98, and 0.97, respectively. 2 A value of =1.00 indicates a perfect fit to the linear characteristic.
[0060] Figures 7E, 8E, and 9E show a comparison of 4-gram counts in real and synthetic datasets. The term "4-gram" is used in a way similar to natural language processing to refer to a sequence of four consecutive events (i.e., four events in a particular order). In this example, the number of occurrences of each particular event, the "4-gram," is determined. The data from the real dataset is plotted on the x-axis and the data from the synthetic dataset on the y-axis. An ideal distribution would fall along a straight line with a slope of 1.0, as shown. The r 2 The values of r are 0.88, 0.74, and 0.72, respectively.2 A value of =1.00 indicates a perfect fit to the linear characteristic.
[0061] Figures 7F, 8F, and 9F show a comparison of 6-gram counts between the real and synthetic datasets. The term "6-gram" is used in a way similar to natural language processing to refer to a sequence of six consecutive events (i.e., six events in a particular order). In this example, the number of occurrences of each particular event, the "6-gram," is determined. The data from the real dataset is plotted on the x-axis and the data from the synthetic dataset on the y-axis. An ideal distribution would fall along a straight line with a slope of 1.0 as shown. The r 2 The values of r are 0.71, 0.89, and 0.56, respectively. 2 A value of =1.00 indicates a perfect fit to the linear characteristic.
[0062] 10A-10E, 11A-11E, 12A-12E, and 13A-13E are plots showing a performance comparison of the disclosed method (FIGS. 10A-10E) with other techniques, including non-neural network methods such as n-gram models used in natural language processing (NLP) (FIGS. 11A-11E), and neural network approaches such as long short-term memory (LSTM) (FIGS. 12A-12E) and convolutional neural networks (CNN) (FIGS. 13A-13E).
[0063] 10A-10E are plots showing a comparison of real (i.e., original) and synthetic datasets, including event frequency, 1-grams, 2-grams, 3-grams, and 4-grams, respectively, to illustrate the performance of the disclosed method. The term "n-gram" refers to a sequence of n consecutive events (or a single event in the case of 1-grams), in a usage similar to natural language processing.
[0064] 11A-11E are plots showing a comparison of real and synthetic datasets, including event frequency, 1-gram, 2-gram, 3-gram, and 4-gram, respectively, to illustrate the performance of n-gram models used in natural language processing (NLP).
[0065] 12A-12E are plots showing a comparison of real and synthetic datasets to illustrate the performance of a long short-term memory (LSTM) neural network, including event frequency, 1-gram, 2-gram, 3-gram, and 4-gram, respectively.
[0066] 13A-13E are plots showing a comparison of real and synthetic datasets, including event frequency, 1-grams, 2-grams, 3-grams, and 4-grams, respectively, to illustrate the performance of convolutional neural networks (CNNs).
[0067] 14A-14H are plots showing a comparison of real and synthetic datasets to illustrate the performance of the techniques described above with respect to event sequence length (FIGS. 14A, 14C, 14E, 14G) and event type frequency (FIGS. 14B, 14D, 14F, 14H). The plots compare the performance of the disclosed techniques (FIGS. 14A-14B) with other techniques, including non-neural network techniques (FIGS. 14E-14F), such as n-gram models used in natural language processing (NLP) (FIGS. 14C-14D), and neural network approaches such as long short-term memory (LSTM) (FIGS. 14G-14H) and convolutional neural networks (CNN) (FIGS. 14G-14H).
[0068] Aspects of the present invention may be embodied in the form of a system, a computer program product, or a method. Likewise, aspects of the present invention may be embodied as hardware, software, or a combination of both. Aspects of the present invention may also be embodied as a computer program product stored on one or more computer readable mediums in the form of computer readable program code embodied thereon.
[0069] The computer readable medium may be a computer readable storage medium, which may be, for example, an electronic, optical, magnetic, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.
[0070] The computer program code in the embodiments of the present invention may be written in any suitable programming language. The program code may be executed on a single computer or on multiple computers. The computer may include a processing unit in communication with a computer usable medium, the computer usable medium including a set of instructions, the processing unit designed to execute the set of instructions.
[0071] The above discussion is intended to illustrate the principles and various embodiments of the present invention. Numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. It is intended that the following claims be interpreted to embrace all such variations and modifications.
Claims
1. A method for generating a synthetic longitudinal data set (X′), comprising: an identification step of identifying an integer number m of subsequence patterns in N records (r) defining event sequences for an integer number N of patients; a determining step in which a determining unit determines a plurality of feature vectors (Y′) each characterizing a corresponding record among the N records (r) based at least in part on the m subsequence patterns; an embedding step in which an embedding unit embeds the feature vector (Y') into a low-dimensional space (V) of y dimension (y being an integer greater than 0 and less than m) to define positions of the N records (r) in the low-dimensional space (V); A selection unit selects one seed record (r) from the N records (r). s ) a selection step; an identification step in which an identification unit identifies a sub-sequence pattern in a subset of the N records (r), the sub-sequence pattern being determined based at least in part on a position of the N records (r) in the low-dimensional space (V); The replacement unit replaces the seed record (r s ) with an instance of a corresponding subsequence pattern identified in the subset of records (r) based at least in part on a distance between respective instances of the subsequence pattern, and generating a modified seed record (r s a substitution step to form The iterator determines whether all the records (r) are the seed records (r s an iterative step of repeatedly performing the selecting, identifying, and replacing steps until a first iteration of the first sub-sequence is selected as the first iteration of the first sub-sequence; The join is the seed record (r s a joining step for joining the first and second vertices together; The method includes:
2. each of the N patient event records included in the dataset (Y) includes a patient identifier, a start date of the event or a start date and an end date of the event, and one or more attributes of the event; 2. The method of claim 1, further comprising a determining step in which the determining unit determines, for each of the N patients, an event sequence ordered by a start date of each event in the event record associated with each of the N patients to generate the N records (r) defining event sequences for the N patients.
3. the original data set (X) includes a plurality of tables, each of which includes a subset of the event records for the N patients; The method of claim 2 , further comprising a forming step in which a forming unit concatenates the tables to form the data set (Y).
4. The method of claim 1 , wherein the m subsequence patterns include subsequence patterns that occur more frequently than a predetermined threshold.
5. The method of claim 1 , wherein the m subsequence patterns include the m most frequently occurring subsequence patterns, where m is a fixed parameter.
6. The method of claim 1 , wherein the plurality of feature vectors (Y′) are equal in length.
7. The method of claim 1 , wherein the plurality of feature vectors (Y′) comprises: (i) frequencies of occurrence of the m subsequence patterns; or (ii) a binary presence / absence encoding of the m subsequence patterns.
8. 2. The method of claim 1, wherein the embedding step is performed using one or more of t-SNE, Uniform Manifold Approximation Projection (UMAP), and Principal Component Analysis (PCA).
9. The subset of records (r) is the seed record (r) in the low-dimensional space (V). s 2. The method of claim 1, wherein the k nearest neighbors are represented by an integer number k.
10. The method of claim 1 , wherein the replacing step is performed only for those subsequence patterns where a distance between respective instances of the subsequence patterns is less than a predetermined threshold.
11. The replacement step is performed by replacing the seed record (r s 2. The method of claim 1, wherein the method is performed by replacing the one or more instances of the subsequence pattern in a sequence of a plurality of subsequence patterns with an instance of a corresponding subsequence pattern randomly selected from a group of subsequence patterns having respective distances less than a predetermined distance threshold.
12. The method of claim 1 , wherein the replacing step excludes subsequences whose occurrence in the sequence of N events is less than a predetermined threshold.
13. In the iterative step, the seed record (r s 2. The method of claim 1, further comprising the step of randomizing the start date of the first event in each event.
14. In the iterative step, the seed record (r s 2. The method of claim 1, further comprising randomizing a gap between one or more pairs of the plurality of events in the set of events.
15. determining a conditional probability of each event given that a preceding event has occurred in each of said N sequences of events; 2. The method of claim 1, wherein in the iterative step, each event is randomly swapped with its respective preceding event each time if the conditional probability is less than a defined threshold.
16. The method of claim 1 , further comprising the step of outputting the synthetic longitudinal data set (X′) to a storage system that provides encryption and access control.
17. 1. A system for generating a synthetic longitudinal data set (X′), comprising: a computer having a memory storing instructions executable by the one or more processors and one or more processors in communication with the memory; identifying an integer number m of subsequence patterns in N records (r) defining event sequences for an integer number N of patients; determining a plurality of feature vectors (Y′) based, at least in part, on the m subsequence patterns, each of the feature vectors characterizing a corresponding record among the N records (r); an embedding step of embedding the feature vector (Y') into a low-dimensional space (V) of y dimension (y is an integer greater than 0 and less than m) and defining the position of the record (r) in the low-dimensional space (V); From the record (r), a seed record (r s ) in an iterative manner; an identification step of identifying a sub-sequence pattern in a subset of the records (r) determined based at least in part on the position of the records (r) in the low-dimensional space (V); The seed record (r s ) with an instance of a corresponding subsequence pattern identified in the subset of records (r) based at least in part on a distance between respective instances of the subsequence pattern, and generating a modified seed record (r s a substitution step to form All the records (r) are the seed records (r s an iterative step of repeating the selecting, identifying, and replacing steps until a first iterative selection step is selected as The seed record (r s a joining step for joining the first and second vertices together; A system for causing the computer to execute the above.
18. each of the data sets (Y) containing event records for the N patients includes a patient identifier, a start date of the event or start and end dates of the event, and one or more attributes of the event; 20. The system of claim 17, wherein the memory further stores instructions executable for the one or more processors to determine, for each of the N patients, an event sequence ordered by a start date of each event in an event record associated with each of the N patients to generate the N records (r) defining event sequences for the N patients.
19. 18. The system of claim 17, wherein the plurality of feature vectors (Y') comprises: (i) frequencies of occurrence of the m subsequence patterns; or (ii) a binary presence / absence encoding of the m subsequence patterns.
20. The subset of records (r) is the seed record (r) in the low-dimensional space (V). s 20. The system of claim 17, wherein the k nearest seed records correspond to an integer number k of nearest seed records.
21. The memory comprises: determining a conditional probability of each event given that a preceding event has occurred in each of said N event sequences; a permutation step of randomly permuting each event with its respective preceding event each time when the conditional probability is less than a defined threshold in the iteration step; 20. The system of claim 17, further storing instructions executable by the one or more processors.
22. 1. A non-volatile computer-readable storage medium having stored thereon instructions that, when executed by one or more processors of a computer, cause the one or more processors to perform a method for generating a synthetic longitudinal data set (X′), the method comprising: identifying an integer number m of subsequence patterns in N records (r) defining event sequences for an integer number N of patients; determining a plurality of feature vectors (Y′) based, at least in part, on the m subsequence patterns, each of the feature vectors characterizing a corresponding record among the N records (r); an embedding step of embedding the feature vector (Y') into a low-dimensional space (V) of y dimension (y is an integer greater than 0 and less than m) and defining the position of the record (r) in the low-dimensional space (V); From the record (r), a seed record (r s ) in an iterative manner; an identification step of identifying a sub-sequence pattern in a subset of the records (r) determined based at least in part on the position of the records (r) in the low-dimensional space (V); The seed record (r s ) with an instance of a corresponding subsequence pattern identified in the subset of records (r) based at least in part on a distance between respective instances of the subsequence pattern, and generating a modified seed record (r s a substitution step to form All the records (r) are the seed records (r s an iterative step of repeating the selecting, identifying, and replacing steps until a first iterative selection step is selected as The seed record (r s a joining step for joining the first and second vertices together; A computer-readable storage medium having stored thereon instructions for executing the above.
23. each of the data sets (Y) containing event records for the N patients includes a patient identifier, a start date of the event or start and end dates of the event, and one or more attributes of the event; 23. The computer-readable storage medium of claim 22, further storing instructions executable by the one or more processors to determine, for each of the N patients, an event sequence ordered by a start date of each event in an event record associated with each of the N patients to generate the N records (r) defining event sequences for the N patients.
24. 23. The computer-readable storage medium of claim 22, wherein the feature vector (Y') comprises: (i) the frequency of occurrence of the m subsequence patterns; or (ii) a binary presence / absence encoding of the m subsequence patterns.
25. The subset of records (r) is the seed record (r) in the low-dimensional space (V). s 23. The computer-readable storage medium of claim 22, wherein the k nearest seed records correspond to an integer number k of nearest neighbor seed records.
26. determining a conditional probability of each event given that a preceding event has occurred in each of said N event sequences; a permutation step of randomly permuting each event with its respective preceding event each time when the conditional probability is less than a defined threshold in the iteration step; 23. The computer-readable storage medium of claim 22, further storing instructions executable by the one or more processors.
Citation Information
Patent Citations
Hospital matching for de-identified healthcare databases without explicit referencing information
JP2019514128A
System and method for generating synthetic dataset from original dataset
JP2023029263A
Generating synthetic patient health data
US20230010686A1
