Dissertation
Empowering health data insights: tools for anomaly detection and multivariate clustering
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2023
DOI: 10.25820/etd.006978
Abstract
Medical records have a long history; ancient stone tablets containing patient medical history have been unearthed. In modern times, patient data are more easily recorded in forms such as electronic medical records (EMR) and wearable device data. These efficient systems quickly create large datasets filled with latent features waiting to be discovered. To uncover these insights, clustering techniques can be used to identify patients with similar characteristics. However, these data have immutable properties that available techniques inadequately address. Specifically, wearable device data contain multiple responses that are highly correlated and EMRs tend to be recorded at inconsistent intervals. Furthermore, since the clustering results can have wide-ranging implications through guiding medical treatments, resource allocation and healthcare policy, ensuring data quality is of paramount importance. However in EMRs, data are typically measured and entered into the record by hand. This human element intensifies the risk of contaminating the data with anomalies. Thus, this dissertation introduces three tools to address the issues of anomaly detection and clustering.
In the second chapter, we provide a methodology for identifying and rectifying anomalies in EMR data. Rooted in a mean-shift model and utilizing studentized residuals, our approach is carefully crafted to adopt a conservative stance, specifically targeting anomalies that indicate mistakes in the data collection process. We incorporate strict control over the false discovery rate to bolster the dependability of identified anomalies. We substantiate our approach by presenting extensive simulation studies and a real-world application, affirming its capacity not only to pinpoint anomalies but also to facilitate precise data correction. Thus, it serves as an invaluable tool for analysts seeking to enhance data quality and reliability.
We then present a technique for clustering unaligned EMRs in the third chapter. We employ smoothing splines to model individual trends and introduce a model-based proximity, which is derived from the covariance matrices of fitted smoothing spline models. This approach ensures a sophisticated capture of subject-specific correlations, thereby enhancing the accuracy of clustering. By leveraging the k-medoids algorithm with this refined proximity, we achieve precise grouping of subjects who exhibit similar trends over time. Our methodology is validated through extensive simulation studies, highlighting its superior performance when compared to existing techniques. We further demonstrate the practical application and effectiveness of this approach through realworld case studies involving both EMR and wearable device data.
The fourth chapter tackles the challenge of addressing the highly correlated channels present in wearable device data. We introduce an innovative approach that employs a joint regression model to effectively characterize and integrate the correlations among the multiple channels over time. Utilizing B-splines, we approximate functional trends and calculate pairwise proximities using a model-based covariance matrix. This proximity, when combined with the k-medoids algorithm, streamlines the clustering of subjects into groups that exhibit consistent trends across various channels. Our method’s superiority in cases with substantial correlations, where conventional techniques struggle, is demonstrated through extensive simulation studies. The practical utility and enhanced accuracy of this approach are exemplified in an electroencephalography study.
Details
- Title: Subtitle
- Empowering health data insights: tools for anomaly detection and multivariate clustering
- Creators
- Daren Kuwaye
- Contributors
- Hyunkeun Ryan Cho (Advisor)Emine Bayman (Committee Member)Joseph E. Cavanaugh (Committee Member) - University of Iowa, Statistics and Actuarial ScienceJacob Oleson (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biostatistics
- Date degree season
- Autumn 2023
- DOI
- 10.25820/etd.006978
- Publisher
- University of Iowa
- Number of pages
- xii, 102 pages
- Copyright
- Copyright 2023 Daren Kuwaye
- Comment
This thesis has been optimized for improved web viewing. If you require the original version, contact the University Archives at the University of Iowa: https://www.lib.uiowa.edu/sc/contact/.
- Language
- English
- Date submitted
- 12/04/2023
- Description illustrations
- Illustrations, tables, graphs, charts
- Description bibliographic
- Includes bibliographical references (pages 68-71).
- Public Abstract (ETD)
Medical records have come a long way from ancient stone tablets to today’s electronic medical records (EMRs) and wearable device data. These modern systems generate vast datasets with hidden insights and clustering techniques can be used to uncover them. However, these data have challenges: wearable data consist of multiple channels that are highly correlated, and EMRs have inconsistent measurement schedules. Furthermore, EMR data can contain anomalies from human mistakes. Therefore ensuring data quality is crucial due to the far-reaching implications of clustering in medical treatment, resource allocation, and policy.
This dissertation addresses these challenges with three tools. First, a method to identify and correct anomalies in EMR data is introduced. This method is designed to be conservative, so that it focus on anomalies arising from human mistakes. Extensive simulation studies are used to validate this approach and a real-world application shows the value in this tool to improve data quality.
Then, a technique for clustering EMRs with an inconsistent measurement schedule is presented. A statistical model is used to capture the individual trends and estimate covariance over time, and a proximity which uses this information to improve clustering accuracy is presented. The k-medoids algorithm, in combination with this proximity, precisely groups subjects with similar trends. Extensive simulation studies demonstrate its superiority to existing methods and we substantiate its practical application in two real-world case studies.
Finally, the challenge of high correlations in wearable device data is addressed by modeling the correlated channels simultaneously. This allows the correlation to be captured so that it can better inform calculating the proximity between subjects. This approach, combined with the k-medoids algorithm, improves the clustering of subjects with consistent trends across various channels. Extensive simulation studies showcase its superiority in handling strong correlations. We also provide a practical application demonstrated in an electroencephalography study.
- Academic Unit
- Biostatistics
- Record Identifier
- 9984546849502771
Metrics
3 File views/ downloads
25 Record Views