Maintenance of Domain Knowledge for Nursing Care using Data in Hospital Information System

Transcription

1 Maintenance of Domain Knowledge for Nursing Care using Data in Hospital Information System Haruko Iwata, Shoji Hirano and Shusaku Tsumoto Department of Medical Informatics, School of Medicine, Faculty of Medicine Shimane University 89-1 Enya-cho Izumo Japan {tsumoto, haruko23, Abstract. Schedule management of hospitalization is important to maintain or improve the quality of medical care and application of a clinical pathway is one of the important solutions for the management. Although several kinds of deductive methods for construction for a clinical pathway have been proposed, the customization is one of the important problems. This research proposed an inductive approach to support the customization of existing clinical pathways by using data on nursing actions stored in a hospital information system. Since hospital data include temporal trends of clinical symptoms and medical services, we can discover not only knowledge about temporal evolution of disease, but also one about medical practice from hospital information system. This paper proposes temporal data mining process and applied the method to capture temporal knowledge about nursing practice. The results show that the reuse of stored data will give a powerful tool for management of nursing schedule and lead to improvement of hospital services. Keywords: temporal data mining; clustering; multidimensional scaling; hospital information system; visualization 1 Introduction Twenty years have passed since clinical data were stored electronically as a hospital information system (HIS)[1 3]. Stored data give all the histories of clinical activities in a hospital, including accounting information, laboratory data and electronic patient records. Due to the traceability of all the information, a hospital cannot function without the information system. All the clinical inputs are shared through the network service in which medical staff can retrieve their information from their terminals [4, 3]. Since all the clinical data are distributed stored and connected as a largescale network, HIS can be viewed as a cyberspace in a hospital: all the results of clinical actions are stored as histories. It is expected that similar techniques in data mining, web mining or network analysis can be applied to the data. Dealing with cyberspace in a hospital will give a new challenging problem in hospital management in which spatiotemporal data mining, social network analysis and DKKD'12 - International Workshop on Domain Knowledge in Knowledge Discovery The 20th International Symposium on Methodologies for Intelligent Systems - ISMIS 2012 Macau, December 4, 2012.

2 2 Haruko Iwata, Shoji Hirano and Shusaku Tsumoto other new data mining methods may play central roles[5, 2]. 1 This paper proposes a temporal data mining method to maintain a clinical pathway used for schedule management of clinincal care. Since the log data of clinical actions and plans are stored in hospital information system, these histories give temporal and procedural information about treatment for each patient. The method consists of the following four steps: first, histories of nursing orders are extracted from hospital information system. Second, orders are classified into several groups by using clustering and multidimensional scaling method. Third, by using the information on groups, feature selection is applied to the data and important features for classification are extracted. Finally, original temporal data are split into several groups and the first step will be repeated. The method was applied to a dataset whose patients had an operation of cataracta. The results show that the reuse of stored data will give a powerful tool for maintenance of clinical pathway, which can be viewed as data-oriented management of nursing schedule. The paper is organized as follows. Section 2 briefly explains clinical pathway, which is schedule management of clinincal process. Section 3 gives explanations on data preparation and mining process. Section 4 gives application of temporal data mining process to analysis of nursing orders, which will lead to revision of clinincal pathway. Finally, Section 5 concludes this paper. 2 Background: Clinical Pathway Since several clinincal actions should be repeated appropriately in the treatment of a disease, schedule management is very important for efficient clinincal process[8, 9]. Such a style of schedule management is called a clinincal pathway. Such each pathway is deductively constructed by doctors or nurses, according to their experiences. For example, Table 1 illustrates a clinincal pathway on cataracta in our university hospital. The whole process of admission will classified into three period: preoperation, operation and post-operation. The preoperation date is denoted by -1 day, and operation date is by 0 day. denotes Body Temporature and Pulse Rate, denotes Blood Pressure. 3 Data Preparation and Analysis 3.1 DWH Since data in hospital information systems are stored as histories of clinical actions, the raw data should be compiled to those accessible to data mining methods. Although this is usually called data-warehousing, medical datawarehousing is different from conventional ones in the following three points. First, since hospital information system consists of distributed and heterogeneous data sources. Second, temporal management is important for medical services, so summarization of data should include temporal information. Third, compilation with several levels of granularity is required. In this paper, we focus on the 1 Application of ordinary statistical methods are shown in [6, 7].

3 Maintenance of Domain Knowledge for Nursing Care 3 Table 1. An Example of Clinincal Pathway Preoperation Operation Postoperation -1day 0day 1day 2day 3day 4day 5day Nausea Nausea Nausea Nausea Nausea Nausea Vomitting Vomitting Vomitting Vomitting Vomitting Vomitting Coaching Coaching Coaching Coaching Coaching Preoperation Instruction Notations. : Body Temprature/Pulse Rate : Blood Pressure number of orders to capture temporal global characteristics of clinical activities, whose scheme is given as Fig 1. Here, data-warehousing has two stages: first, we compile the data from heterogeneous datasets with a given focus as the first DWH. Then, from this DWH, we split the primary DWH into two secondary DWHs: contents DWH and histories DWH. In this analysis, we focus on the latter DWH and we count the number of orders with a given temporal section, compiled into this DWH. Data mining process is applied to the generated data sets from this DWH. Records Prescription Labo Exam Radiology Hospital Information System Temporal Matching Order- History DWH First-DWH Summarization Contents History Second-DWH Data Mining Data Mining Fig. 1. Datawarehousing 3.2 Mining Process Except for the basic process, we will propose temporal data mining process, which consists of the following three steps, shown in Figure 2 We count tem-

4 4 Haruko Iwata, Shoji Hirano and Shusaku Tsumoto poral change of #orders per hour or per days in the second DWH. Then, since each order can be viewed as a temporal sequence, we compare these sequences by calculating similarities. Using similarities, clustering[10], multidimensional scalingmds, and other methods based on similarities are applied. In this paper, all the analysis is conducted by R Second DWH Temporal Datasets Calculation of Similarities (Eucledian Distance, DTW, Multiscale Matching,.) Feature Selection Clustering MDS Correspodence Analysis Decision Tree Induction Rule Induction Statistical Modeling Fig. 2. Mining Process 3.3 Clinincal Pathway Maintenance Process Figure 3 shows the process for maintenance of a clinincal pathway. The right cycle shows the repetition of procedure of temporal data mining process proposed in Fig reffig:process2. The procedure will be terminated when grouping becomes stable and a pathway based on the classification will be constructed. If a clinincal pathway is apriori given, it will be compared with the induced pathway. 4 Construction of Clinincal Pathway Here we apply the above temporal mining process to revision and construction of clinincal pathway Data Preparation We focused on histories of patients who were admitted to the hospital for operations of catracta. The number of patients is 121 in For this disease, the clinincal pathway mentioned in the above section is used to optimize the schedule of treatment. We counted the nursing orders during the stay of each patient and regard chronological change of each order as a temporal sequence. 2 The process follows the method proposed by Tsumoto [11]

5 Maintenance of Domain Knowledge for Nursing Care 5 Datasets Clustering/MDS Labeling Classification of Nursing Orders Yes Grouping Results: Stable? Comparison to Existing Pathway Yes A pathway Exists? No Decision Tree Induction Revision No Construction of Clinical Pathway Feature Selection Fig. 3. Maintenance Process 4.2 First Cycle: Grouping In the first cycle, clustering, MDS and correspondence analysis were applied as the second step in the temporal data mining process. Figure 4 to 6 show the results of clustering, MDS and correspondence analysis of nursing orders with respect to #orders. Clustering results gives two major groups: one includes the orders indispensable to this disease and the other includes those which are rather specific to the status of each patient, except for preoperation instruction. MDS gives further classification of the first group into three subgroups and the second one into two subgroups The former three subgroups consist of vital signs (, BT and PR), body care (Coaching and ), and watchlist (Eye Symptoms and Nausea). The latter two groups consist of preoperation instruction and other symptoms which may specific to the status to the patients. On the other hand, correspondence analysis did not give good information about correspondence between length of stay and nursing orders. This may be because the length of stay was too short (within a week) and all the nursing orders were well scheduled due to the application of clinical pathway. The results also give another interesting observation: comparison of the above results with the existing clinical pathway shows that the pathway lacks two orders, and Coach, although their temporal patterns are very similar to other orders indispensable to the treatment. So, they should be included into clinical pathway. In this way, these method can be used to evaluate the existing pathway for a disease and revise it. 4.3 First Cycle: Feature Selection With the labels obtained, decision tree induction[12] was applied to the dataset. Figure 7 shows the result where only selected attribute is d1, the first day of

6 6 Haruko Iwata, Shoji Hirano and Shusaku Tsumoto Height dist_iwata_opt3 hclust (*, "ward") Preoperation Instruction Shampoo BS(Ward) Nerve Others Bath SPO2 Skin/Nail (Ns)Vital Sign Vital Sign (Ns) DM management Fall Prevention Fall Prevention (Bed) (Ns) Guidance for Disease Eye Symptoms Ward s method Fig. 4. Clustering Results of Nursing Orders (Cataracta) iwata_opt_mds3[,2] Preoperation Instruction (Ns)Vital Sign Sign all Prevention BS(Ward) Guidance SPO2 Others Nerve (Ns) (Bed) Skin/Nail DM Bath management (Ns) for Disease Shampoo Eye Symptoms iwata_opt_mds3[,1] Fig. 5. MDS Results of Nursing Orders (Cataracta) postoperation period. In the induced tree, d1 < 23 to the right means that if the number of a given order is less than 23, the order belongs to the right class (which is specific to the status of patients.) In other words, Table 2 shows the list of information gain for the labels. These values show that for d1, d2 and d3 give complete classification for the labels, and preoperation, postoperation, d4 and d5 do not have enough information for complete classification. Thus, the data should be split into three groups: the first interval is

7 Maintenance of Domain Knowledge for Nursing Care Preoperation ( 1) ation Instruction BS(Ward) Operation (0) Vital Sign (Ns)Vital Sign Nerve SPO2 Fall Prevention Fall Prevention (Bed) 2 DM (Ns) management Others 1 3 Guidance (Ns) Eye Shampoo Symptoms for Disease Bath Skin/Nail Fig. 6. Results of Correspondence Analysis of Nursing Orders (Cataracta) d1>=23 c2 7/18 c1 7/0 c2 0/18 Fig. 7. Decision Tree (Cataracta) preoperation and operation, the second one is d1 to d3, and the final interval is d4 and d5. In the next cycle, the same procedures will applied to each subset of data. 4.4 Second Cycle: Grouping Figure 8 and 9 show the results for preoperation and operation (-1 to 0 days), which gives much clearer results. Here we can see three clusters, compared with

8 8 Haruko Iwata, Shoji Hirano and Shusaku Tsumoto Table 2. Information Gain for Each Attribute Node Information Gain preoperation < 8 to the right operation < 23 to the right d1 < 23 to the right d2 < 21 to the right d3 < 39 to the right d4 < 27.5 to the right 6.48 d5 < 18.5 to the right 6.48 Height dist_iwata_opt4a hclust (*, "ward") Bath Skin/Nail SPO2 Shampoo BS(Ward) Guidance for Disease Others Nerve (Ns)Vital Sign Vital Sign (Ns) DM management (Ns) Fall Prevention Fall Prevention (Bed) Eye Symptoms Preoperation Instruction Ward s method Fig. 8. Clustering Results of Nursing Orders (Cataracta: -1 to 0 days) the former results. Although Coaching and belongs to the same cluster in the first cycle, the former is close to Eye symptoms, Nausea/Vomiting and, the latter is close to, and Preoperation instruction. Since the former three observations are mainly for postoperation follow-up, Coaching may belong to the postoperation follow-up process. However, may belong to the whole process. Figure 9 supports this obeservation. Next, Figure 10 and 11 show the results for 1 to 3 days. Here the results are very similar to those obtained from the whole data. However, Coaching is similar to, and is close to and, and Eye symptoms. Finally, Figure 12 and 13 show the results for 4 and 5 days. Here the grouping is slightly different from the second one and Coaching and are not similar to,,, and Eye symptoms.

9 Maintenance of Domain Knowledge for Nursing Care 9 iwata_opt_mds4a[,2] Guidance (Ns) all Prevention DM Shampoo Skin/Nail BS(Ward) SPO2 Others Nerve Bath (Ns) management for Disease (Ns)Vital (Bed) Sign Vital Sign Eye Symptoms Preoperation Instruction iwata_opt_mds4a[,1] Fig. 9. MDS Results of Nursing Orders (Cataracta: -1 to 0 days) Height dist_iwata_opt4b2 hclust (*, "ward") Shampoo Others Nerve SPO2 BS(Ward) Bath Skin/Nail Preoperation Instruction (Ns) (Ns) DM management Fall Prevention (Bed) Guidance for Disease Fall Prevention (Ns)Vital Sign Vital Sign Eye Symptoms Ward s method Fig. 10. Clustering Results of Nursing Orders (Cataracta: 1 to 3 days) In either case, the labels were provided by the grouping methods for each period, and decision tree induction was applied. Each attirbute gave complete classification, so further data partition was not needed, and the repetition of grouping and feature selection was terminated.

10 10 Haruko Iwata, Shoji Hirano and Shusaku Tsumoto iwata_opt_mds4b2[,2] (Ns)Vital Sign Sign Guidance BS(Ward) SPO2 Others for Disease Shampoo Preoperation all Prevention all Prevention Skin/Nail DM Nerve Instruction (Bed) Bath management (Ns) (Ns) Eye Symptoms iwata_opt_mds4b2[,1] Fig. 11. MDS Results of Nursing Orders (Cataracta: 1 to 3 days) Height dist_iwata_opt4c hclust (*, "ward") Others Guidance for Disease Skin/Nail Vital Sign Bath Fall Prevention (Ns)Vital Sign Nerve Preoperation Instruction BS(Ward) SPO2 (Ns) Fall Prevention (Bed) Shampoo (Ns) DM management Eye Symptoms Ward s method Fig. 12. Clustering Results of Nursing Orders (Cataracta: 4 to 5 days) 4.5 Construction of Pathway From those results, coaching and wash, whose chronological characteristics are similar to the orders indispensable to the treatment of cataracta, were not included in the existing pathway, but should be added to the pathway. Furthermore, coaching and wash should be treated as postoperation follow-up and routine process, respectively. Thus, the pathway was revised as shown in Table 3.

11 Maintenance of Domain Knowledge for Nursing Care 11 Others Skin/Nail Guidance Vital Sign for Disease Preoperation BS(Ward) DM SPO2 management Instruction iwata_opt_mds4c[,2] (Ns)Vital all Prevention Nerve BathSign (Ns) all Prevention (Bed) (Ns) Shampoo Eye Symptoms iwata_opt_mds4c[,1] Fig. 13. MDS Results of Nursing Orders (Cataracta: 4 to 5 days) In summary, temporal data mining of nursing may be useful for to construct clinical pathway for a disease where a pathway is not introduced and to revise the existing pathway. 5 Conclusions In this paper, we propose a general framework on innovation of hospital services based on temporal data mining process. This process can be called similaritybased visualization approach in which similarity-based methods, such as clustering and multidimensional scaling (MDS) and correspondence analysis. We applied the process to datasets of #nursing orders for cases for operation of cataracta where clinincal pathway has been introduced. By using Clustering and MDS, we obtained two major groups in the nursing orders: ones were indispensable to the treatment, and the others were specific to the status of patients. Then, in the step for feature selection, the first day of postoperation could be be viewed as a threshold in the original datasets. Thus, periods before and after operation should be dealt as independent datasets. Repeating these steps, we could characterize the temporal aspects of nursing orders, and then found missing information in the existing pathway. This paper is a preliminary approach to data-mining hospital management towards a innovative process for hospital services. More detailed analysis will be reported in the near future. Acknowledgments This research is supported by Grant-in-Aid for Scientific Research (B) from Japan Society for the Promotion of Science(JSPS).

12 12 Haruko Iwata, Shoji Hirano and Shusaku Tsumoto Table 3. Revised Clinincal Pathway for Cataracta Preoperation Operation Postoperation -1day 0day 1day 2day 3day 4day 5day Nausea Nausea Nausea Nausea Nausea Nausea Vomitting Vomitting Vomitting Vomitting Vomitting Vomitting Eye Symp. Eye Symp. Eye Symp. Eye Symp. Eye Symp. Coaching Coaching Coaching Preoperation Instruction Notations. : Body Temprature/Pulse Rate. : Blood Pressure. Eye Symp: Eye Symptoms. References 1. Shortliffe, E., Cimino, J., eds.: Biomedical Informatics: Computer Applications in Health Care and Biomedicine. 3rd edn. Springer (2006) 2. Bichindaritz, I.: Memoire: A framework for semantic interoperability of case-based reasoning systems in biology and medicine. Artif Intell Med 36(2) (2006) Tsumoto, S., Hirano, S.: Risk mining in medicine: Application of data mining to medical risk management. Fundam. Inform. 98(1) (2010) Hanada, E., Tsumoto, S., Kobayashi, S.: A hubiquitous environmenth through wireless voice/data communication and a fully computerized hospital information system in a university hospital. In Takeda, H., ed.: E-Health. Volume 335 of IFIP Advances in Information and Communication Technology. Springer Boston (2010) Iakovidis, D., Smailis, C.: A semantic model for multimodal data mining in healthcare information systems. Stud Health Technol Inform 180 (2012) Tsumoto, Y., Tsumoto, S.: Exploratory univariate analysis on the characterization of a university hospital: A preliminary step to data-mining-based hospital management using an exploratory univariate analysis of a university hospital. The Review of Socionetwork Strategies 4(2) (2010) Tsumoto, Y., Tsumoto, S.: Correlation and regression analysis for characterization of university hospital (submitted). The Review of Socionetwork Strategies 5(2) (2011) Hyde, E., Murphy, B.: Computerized clinical pathways (care plans): piloting a strategy to enhance quality patient care. Clin Nurse Spec 26(4) (2012) Ward, M., Vartak, S., Schwichtenberg, T., Wakefield, D.: Nurses perceptions of how clinical information system implementation affects workflow and patient care. Comput Inform Nurs 29(9) (2011) Everitt, B.S., Landau, S., Leese, M., Stahl, D.: Cluster Analysis. 5th edn. Wiley (2011)

13 Maintenance of Domain Knowledge for Nursing Care Tsumoto, S., Hirano, S., Tsumoto, Y.: Clustering-based analysis in hospital information systems. In: Proceedings of GrC2011, IEEE Computer Society (2010) 12. Quinlan, J.: Induction of decision trees. Machine Learning 1 (1986)