Anomaly Detection Methodologies - Research Proposal Example

?Anomaly Detection Methodologies Introduction Medical records of patients contain thousands or millions of electronic entries that record informationon patient’s conditions, diagnosis and treatments. Tisdale (2010) says that while the storage and use of this data has been helpful in providing a detailed profile of individual patients, the benefits of using this information in identifying the most adverse cases are relatively unexplored and underutilized. Besides, current practices and procedures aimed at identifying such patients are slow, expensive and unsuitable for incorporating new analytical mechanisms. Buckeridge (2007) argues that Current algorithms used for achieving this risk stratification are dependent on the labelling of the patient data as positive or negative. This classification implies that determining trends and subsets that are rare in a given population requires an analysis of large data sets and the identification of positive aspects up to a threshold level. This process, as explained above, is not just slow or expensive, but puts additional burden on patients and hospital administrators, thereby affecting the validity and effectiveness of such practices. The proposed study aims to use appropriate anomaly detection methods that are known to be suitable for detecting interesting or unusual patterns in a given data set. Bohmer (2009) says that new frameworks allow anomaly detection to be applied towards determining anomalous patterns in subsets of attributes associated with a data set. In simpler words, anomaly detection methods identify unusual occurrences with the data that appear to deviate from the normal behaviour exhibited by a majority of the data set. Examples of such anomalies include an epidemic outbreak, traffic congestion in a certain section of roads or an attack on a network (Applegate, 2009). The proposed research aims to extend the standard approach to anomaly detection by devising techniques to identify partial patterns that exhibit anomalous behaviour with the remainder of the data set. Such techniques are believed to aid in the detection and assessment of unusual outcomes or decisions related to patient management in healthcare institutions. Anomaly Detection Several studies by researchers like Nurcan (2009) and Anderson (2007) have applied anomaly detection techniques to healthcare. In fact, anomaly detection has proved useful in areas under clinical behaviour and medical technology such as blood samples, vestibular information, mammograms and electroencephalographic signals (Brandt, 2007). However, the same principles have found little application in enhancing the quality of patient care or identifying existing deficiencies in the assistance extended to patients. The proposed study aims to improve and extend anomaly detection techniques to such relatively unexplored domains. While previous studies have relied primarily on detecting existing conditions such as diseases, the proposed research will apply similar methods to ascertain the level of risk that accompanies a potential outcome being analyzed. Thus, the measurement of this risk as a result of uncovering anomalies is likely to help in forecasting the vulnerability of patients to certain diseases or deficiencies. The study proposed to utilize several anomaly detection methods by applying them to existing clinical data on patients. In doing so, the number of outcomes and patients being analyzed will be much larger and wider than those adopted by previous studies. Some of the detection methods that will be included as part of the proposed study are listed below: Nearest Neighbour method As the name suggests, the nearest neighbour method helps detect patients (anomalies) from a given population based on information pertaining to their ‘n’ nearest neighbours. This method is based on the principle of vectors that are used to sum the distances between a point and it ‘n’ closes neighbours. As a result, dense and sparse regions are identified based on the total score which is lesser in the former case (Larson, 2007). In contrast, points that have the highest scores are candidates for being termed as anomalies. Classification Method This method is based on the concept of Support Vector Machines (SVM). Using this approach, a 1-class SVM is developed for the entire population which is then applied to a corresponding decision function to detect patients whose feature vectors continue to be aligned with the hyperplane (Zhang, 2004). Such patients are thereafter categorized as anomalies. Cluster Method Under the Cluster-based technique of anomaly detection, a count of the number of points that lie within a feature vector are calculated, which can then be used to calculate the proximity of a point to nearby points. Patients whose values are low are classified as anomalies (Herzlinger, 2004). Methodology The study proposed to adopt an elaborate methodology that is based on both primary and secondary sources of information. Primary sources include data obtained from existing patient records and through interaction with workers and patients at a few hospitals in the region (Tisdale, 2010). Information on surgical procedures (both general and vascular) performed on over 150,000 patients shall be used for the study and will be sourced from the 2008 NSQIP Participant File. Conditions such as post surgery scenarios and response will be analyzed to detect any anomalies. The data from the participant file contains valuable information on as many as 40 variables on patient demographics and morbidity. All data used for the study will be screened for any missing fields or inconsistencies using a set of predefined rules which will be elaborated in a detailed manner in the dissertation. The study shall also rely extensively on various secondary sources such as books, journals and articles from magazines. Based on Applegate (2009), secondary sources are important for this study as extensive information on different anomaly detection techniques, their benefits and constraints can be sourced from these resources and applied in the analysis. Further, experiences of previous researchers in applying these methods will also help improve the process of identifying and implementing the suitable detection methods in this study. Conclusion This research proposal introduces the interesting and useful concept of anomaly detection and explains its relevance and importance in detecting anomalies among patients, who are more vulnerable to adverse outcomes. Several anomaly detection techniques like the ones described in the preceding sections are very useful in identifying adverse conditions and behaviour within a large group of patients. Thus, the proposed study aims to expand the utility of anomaly detection towards improving the quality and effectiveness of patient care and help needy patients in the shortest time. References 1. Anderson (2007), Personal medical information: security, engineering, and ethics. New York: Springer. 2. Applegate (2009), Evidence-Based Imaging in Pediatrics: Optimizing Imaging in Pediatric Patient Care. New York: Springer. 3. Bohmer (2009), Designing care: aligning the nature and management of health care. Harvard Business Press. 4. Brandt (2007), Digital Government: E-Government Research, Case Studies, and Implementation. Torontoa: McGraw Hill. 5. Buckeridge (2007), Disease surveillance: a public health informatics approach. New York: Wiley. 6. Herzlinger (2004), Consumer-driven health care: implications for providers, payers, and policymakers. New York: John Wiley. 7. Larson (2007), Digital Government: E-Government Research, Case Studies, and Implementation. New York: Springer. 8. Nurcan (2009), Enterprise, Business-Process and Information Systems Modeling. Chicago: McGraw Hill. 9. Tisdale (2010), Drug-Induced Diseases: Prevention, Detection, and Management. Chicago: ASHP. 10. Zhang (2004), Toward a human-friendly assistive environment. London: IOS Press. Read More