Quality Assessment of Hemodialysis Services through Temporal Data Mining

Share Embed


Descrição do Produto

Quality assessment of hemodialysis services through temporal data mining 1

Riccardo Bellazzi1, Cristiana Larizza1, Paolo Magni , and Roberto Bellazzi2 1

Dip. Informatica e Sistemistica, Università di Pavia, via Ferrata 1, 27100, Pavia Italy {Riccardo.Bellazzi,Cristiana.Larizza,Paolo.Magni}@unipv.it 2 Unità Operativa di Nefrologia e Dialisi, S.O Vigevano, A.O. Pavia, Corso Milano 19, 27029, Vigevano Italy

Abstract. This paper describes a research project that deals with the definition of methods and tools for the assessment of the clinical performance of a hemodialysis service on the basis of time series data automatically collected during the monitoring of hemodialysis sessions. While simple statistical summaries are computed to assess basic outcomes, Intelligent Data Analysis and Temporal Data mining techniques are applied to gain insight and to discover knowledge on the causes of unsatisfactory clinical results. In particular, different techniques, comprising multi-scale filtering, Temporal Abstractions, association rules discovery and subgroup discovery are applied on the time series. The paper describes the application domain, the basic goals of the project and the methodological approach applied for time series data analysis. The current results of the project, obtained on the data coming from more than 2500 dialysis sessions of 33 patients monitored for seven months, are also shown.

1 Introduction Health care institutions routinely collect a large quantity of clinical information about patients status, physicians actions (therapies, surgeries) and health care processes (admissions, discharge, exams request). Despite the abundance of this kind of data, their practical use is still limited to reimbursement and accounting procedures and sometimes to epidemiological studies. The general claim of researchers is that the potentiality of generalization of those data, that we will refer to as process data, is very weak, since they are not collected in controlled clinical trials. However, the growing interest on knowledge management within health care institutions have highlighted the crucial role of process data for organizational learning [1,2]. One of the aspects of organizational learning is represented by the assessment of the quality of a hospital service, in particular in relationship to certain performance indicators [3]. Such performance indicators can be either related to the efficiency of the hospital department or to the efficacy of the treatment delivered. In this paper we are interested in the use of data mining tools for assessing the efficacy of treatment delivered by a Hospital Hemodialysis Department (HHD) on the basis of the process data routinely collected during hemodialysis sessions. HHD manage chronic patients that undergo

blood depuration (hemodialysis) through an extra-corporal circuit three times a week for four hours. The data accumulated over time for each patient contain the set of variables that are monitored during each dialysis session. In other words, the data collected are time series (inter-session data) of multidimensional time series (intrasession data). Those process data are typically neglected during clinical treatment, since they are synthesized by few clinical indicators observed at the beginning and at the end of each treatment session. Such clinical indicators are usually related to the well-being of patients, and do not contain detailed information about the quality of the treatment, in terms, for example, of blood depuration efficiency or nurse interventions during the dialysis itself. The goal of an auditing system for quality assessment is therefore to fully exploit the process data that may be automatically collected in order to: i) Assess the performance of the overall HHD; ii) Assess the performance achieved for each patient; iii) Highlight problems and understand their reasons. The steps i)-iii) need first to define a suitable set of automatically computable performance indicators and then to analyze the dialysis temporal patterns, by studying both interand intra-dialysis data. In particular, the design and implementation of this system requires the use of methodological tools to perform two different temporal data mining tasks [4]: a) the discovery of patient-specific relationships between the time patterns of monitoring variables and the dialysis performance indexes; b) the extraction of HDD-specific relationships between the time patterns of monitoring variables and the dialysis performance indexes. In this paper, we present both a new method for the discovery of patient-specific temporal pattern and a new system for quality assessment of dialysis sessions; the system is currently used in the clinical routine. In particular, the paper describes first the application domain and the basic goals of the project; then, it presents the methodological approach applied for time series data analysis and the results obtained.

2 End Stage Renal Failure and Hemodialysis End stage renal disease (ESRD) is a severe chronic condition that corresponds to the final stage of kidney failure. In ESRD, kidneys are not anymore able to clear blood from metabolites and to eliminate water from the body. Without medical intervention, ESRD leads to death. In 1999 the ESRD incidence in Italy was of 130 cases per million [5]. The elective treatment of ESRD is kidney transplant. Blood-filtering dialysis treatment is provided as a suitable alternative to transplant for people in waiting list or for people that cannot be transplanted at all, such as elderly people. Two main classes of dialysis treatments are nowadays available: hemodialysis (HD) and peritoneal dialysis. More than 80% of the ESRD patients are treated with HD. In the HD treatment the blood passes through an extra-corporal circuit where metabolites (e.g. urea) are eliminated, the acid-base equilibrium is re-established and the water in excess is removed. Typically, HD is performed by exchanging solutes through a semipermeable membrane (dialysis) and by removing water through a negative pressure gradient (ultrafiltration); a device called hemodialyzer regulates the overall procedure. In general, HD patients undergo a dialysis session for four hours three times a week.

The dialysis treatment has very high costs and it is extremely demanding from an organizational viewpoint [6]. Rather interestingly, a potential solution to improve the efficiency of dialysis services is represented by Information Technology, as reported in the literature [7-9]. In this paper we are interested in the exploitation of the recent advances in the implementation of hemodialyzers, that allow an automated monitoring of dialysis sessions [8]. In particular we have implemented an auditing system designed to summarize the dialysis sessions from a clinical quality viewpoint. This tool is aimed to automatically extract meaningful patterns from the data, that may be useful for assessing the dialysis sessions from an organizational learning perspective, i.e. for periodically evaluating the HDD performance, either for what concerns all patients or for what concerns each patient.

3 A System for quality assessment of hemodialysis centers 3.1 Measurements Our system for quality assessment of HD sessions is based on the automatic measurements of 13 variables, which reflect three main aspects of the HD process: 1Efficiency of the removal of protein catabolism products (urea, creatinine); such efficiency is indirectly evaluated by measuring the blood flow in the extracorporeal circuit (QB), the body weight loss (WL) and the dialysis time (T). 2- Efficiency of the extra-corporeal circuit of the dialyzer; this aspect is evaluated by measuring the pressure of the circuit before (arterial pressure, AP) and after (venous pressure VP) the dialyzer (i.e. the device where the exchange of water and solutes is performed) and the output pressure of the dialyzer (OP) 3- Body water reduction and hypotension episodes. The monitoring of body water and of patients arterial pressure is performed by measuring the water flow through the dialyzer (UF), systolic and diastolic pressures (SP, DP), the cardiac frequency (CF), the hemoglobin concentration (Hb) and the estimated plasma volume (PVol). The conducibility of the dialyzer solution (CD) is also monitored, to keep track of water exchanges due to osmosis. The body water reduction is monitored through the (already mentioned) weight loss, too. All those parameters are monitored during each session. Finally, for each dialysis session, the physician defines a set of prescriptions, that is the collection of hemodialyzer settings and HD targets that should be followed and reached at the end of the dialysis.

3.2 Data summaries for quality assessment. In order to assess the performance of each dialysis session, we compute a suitable summary of the intra-dialysis time series. In particular, each session is synthesized through the classical non parametric statistic indexes: the median and the 10th and 90th percentiles of each monitored variable. After having calculated the median values, we obtain a new multidimensional time series, in which each point is the vector of the median values of the 13 monitoring variables, computed on the data collected during a dialysis session. We will refer to this time series as the median time series.

Exploiting the median values, it is possible to assess the quality of a session by performing a comparison between a set of reference values and a set of dialysis outcome parameters. In more detail, we consider six parameters: i) the median levels of QB, VP, AP, that we will denote as QB*, VP*, AP*; ii) the time difference between the prescribed dialysis time and the effective one (∆T); iii) the difference between the prescribed weight loss and the weight loss measured at the end of the dialysis (∆L); iv) the difference between the weight reached at the end of the dialysis and the ideal weight prescribed by the physician (∆W). A general index of success is derived by judging as positive a treatment in which all (AND) the logic conditions of Table 1 are satisfied: Table 1. Outcome parameters and the corresponding logical conditions for their assessment Parameters

Condition

QB* VP* AP* ∆T ∆L ∆W

Not less than 2% of the prescription Less or equal to 350 ± 3 mmHg Greater or equal to –250 mg Less or equal than 2% of the prescription Less or equal than 7% of the prescription Less or equal than 5% of the prescription

If any of the conditions is not satisfied, the dialysis is considered to be failed. The failure is determined by one or more failure parameters, i.e. the outcome parameters that do not satisfy the condition. The parameters of Table 1 has been derived on the basis of the available background knowledge. In an audit session, that is typically performed monthly (but may be performed weekly or even daily), the physician can easily calculate the percentage of failures at the center or at the patient level. 3.3 Looking for temporal patterns and knowledge discovery A crucial aspect of our project is to be able to provide clinicians and nurses with a deeper insight in the HD temporal patterns, in order to discover the reasons of failures, derive associations between monitoring variables and understand if there are relationships between monitoring variables and failures that hold at the center (population) level. To this end, we have defined a temporal data mining strategy to analyze the data based on the time series of the median values of each dialysis session. Such strategy is based on the following steps: A) Extraction of basic Temporal Abstractions (State and Trends) from the median time series; the series are pre-processed for trend detection through a multi-scale filtering method; B) Search for associations between Temporal Abstractions and failures; these associations may be interpreted as association/classification rules, which may hold on a single patient; C) Discover subgroups of patients that show the same associations between monitoring variables and failure parameters. The remaining part of the paper will describe in detail the steps which have been implemented in our auditing system.

3.3.1 Representing the time series through Temporal Abstractions. Temporal Abstractions (TA) are techniques exploited to extract specific patterns from temporal data; such patterns should represent a meaningful summary of the data and should be conveniently used to derive features that characterize the dynamics of the system under observation [10,11]. The goal of the TA mechanisms is the identification of intervals corresponding to specific patterns shown by the data that should represent a condition that occurs and evolves during specific time periods. Our TA extraction technique is based on the analysis of time-stamped data and refers to two different categories of TAs: Basic and Complex TAs. Basic TAs represent simple patterns derived from numeric or symbolic uni-dimensional time series. In particular, we extract Trend abstractions, representing an increase, decrease or stationary trend of a numerical time series, and State abstractions, representing qualitative patterns of low, high, normal values of a numerical or symbolic time series. Complex TAs represent complex patterns of uni-dimensional or multi-dimensional time series which correspond to specific temporal relationships among Basic TAs. The relationships investigated with Complex TAs include the temporal operators defined in the Allen algebra [12]. In the hemodialysis problem, we use Basic TAs to summarize the dynamics of each variable during each session. Before running the TA mechanisms, the median time series is pre-processed in order to robustly detect trend TAs. 3.3.2 Pre-processing of the median time-series through multi-scale filtering methods. One of the major defects of applying trend detection algorithms directly to the raw data is the dependence of the result from the sampling frequency and from the measurement errors. Usually, trend detection is performed on a filtered series in order to reduce these problems; however, the choice of the filter can strongly condition the trend detection results. In particular, the filtering algorithms outputs depend on smoothing parameters, which reflect the prior knowledge available both on the process that generates the data and on the measurement noise. In our case, while it is possible to assume that the noise on the calculated median values is intrinsically low or absent, we do not have any information about the degree of regularity that can be expected in order to properly evaluate the trends from a clinical viewpoint. In other words, since the analysis of the median time series in HD is a new approach to dialysis quality control, we cannot rely on existing knowledge about the dynamics underlying the data generation process. For this reason, we resorted to a new robust strategy, based on a multi-scale smoothing filter. Multi-scale filtering can be performed through a variety of techniques, which have been proposed in the majority of cases in the image processing field. In our case, we resorted to the so-called discrete wavelets [13]. - The smoothing filter chosen is the discrete 1-D wavelet decomposition with Daubechies wavelets. - Thanks to the multi-scale nature of wavelet decomposition, five different smoothed series are reconstructed from the median time series using different wavelet transform coefficients. These series correspond to the first five wavelets scale values of the discrete wavelet transform. Each scale is related to a different smoothing level.

-

For each of the five times series, the trend is detected on the basis of a standard method [10]. In this way it is possible to assign to each time point of the filtered time series a TA within the set {decreasing, stationary, increasing}. - On the basis of a voting strategy, each time point of the median time series is assigned to one TA {decreasing, stationary, increasing}: the TA is confirmed if it is found at the majority of scales. The trend detection so obtained is robust since it is independent from the smoothing scale used for filtering the curve. 3.3.3 Search associations between TAs and outcomes. Once we have derived the state and trend TAs for each monitoring variable, we want to look for possible associations between the TAs and the failure parameters. We would like to obtain rules of the kind “IF the Trend of Venous pressure is increasing THEN dialysis fails due to insufficient weight loss”. To achieve this goal it is possible to search for the co-occurrences of TAs and failures and, then, to select the most frequent ones. The search algorithm described in this section has been inspired by the work of F. Hoppner [14] and by the well-known PRISM algorithm [15]. In order to describe the search algorithm we exploited, it is necessary to introduce some definitions and notations. Given the median time series Vj of a variable j, the state TAs for Vj can be represented as the collection of time intervals in which Vj is either high (H) or normal (N) or low (L), while the trend TAs for Vj can be represented as the collection of time intervals in which Vj is either increasing (I), stationary (S) or decreasing (D). Given two or more TAs, for example “Vj is N” and “Vi is I”, we can easily calculate their intersection as the intersection of their time intervals; the time span (TSji(N,I)) of such intersection corresponds to the number of dialysis sessions in which both TAs occur. The intersection can be calculated also for one or more TAs with any failure parameter. In this case, given the abstraction “Vj is S”, and the failure parameter O=Oi, we denote the TS of their intersection as TSjo(S, Oi). Finally, let us note that TSjjj(H,N,L) and TSjjj(I,S,D) are equal to zero. The search procedure aims to define rules of the form A Æ Oi where A is the body and Oi is the head of the rule: in our case A is any conjunction of TAs (e.g. “Vi is L” and “Vi is D” and “Vj is H”), and Oi is a failure parameter (i.e. failure due to ∆L)1. A is here interpreted as the intersection of TAs involved in the conjunction. It is therefore possible to calculate the time span of A (TSA) resorting to the definition given above. We define the support (SU) of a rule as the number of sessions in which the rule holds (i.e. SU= TSAOi) and we define the confidence (CF) of the rule as the conditional probability P(Oi|A), which may be calculated dividing SU by TSA, i.e. CF= TSAOi/TSA.

1

Although the search procedure looks for rules with the same head and may be thus interpreted as a supervised learning problem, its final goal is not to derive predictive rules, but to extract a description of the co-occurrences of abstractions and failures. For this reason, we use the term association rules to denote the outputs of the algorithm.

Resorting to the definitions given above, we run a search strategy which recursively constructs rules with the maximal body having SU>SUmin and CF>CFmin, being SUmin and CFmin suitable threshold values. The strategy works as follows: - Step 1. Select a single patient. Select a failure parameter Oi. Put all the TAs for all variables in the set A0. - Step 2. Compute the intersection of each member of A0 with Oi. Put the results with SU> SUmin and CF> CFmin in the set A1 and in the basic set B. Set the counter k to 1. - Step 3. Do: o Put in the set Ak+1 the intersection of each of the TAs in Ak with each of the TAs in B and with Oi, such that SU>SUmin and CF>CFmin. o Set k=k+1 while Ak is not empty. - Step 4. Put A=Ak-1. The rule A Æ Oi contains the maximum number of basic TAs, i.e. the rule with most complex body. Let us note that the algorithm allows to derive more than one rule for each Oi. The derived rules with their support can be shown to the users also in graphical form, thus highlighting the temporal location of the sessions in which each rule holds. Moreover, it is also possible to represent the association rules as Complex TAs, in which both the conjunction and the implications are interpreted as a contemporaneous occurrence of TAs and failures. 3.3.4 Search for predictive models at the population level. The algorithm described in the previous section is useful to derive a description of the single patient behaviour over time. In order to derive a model at the population level, it is possible to resort to a different strategy. Since in the approach described in the previous sections each variable in each dialysis has only one state and one trend abstraction which holds, we may easily derive a matrix M of features, where the columns represent the state and the trend value of each variable and each row represents a dialysis session. The matrix M is completed with the patient number, and the different values of the outcomes of each dialysis session. The matrix M can be used to investigate the relationships between the outcomes and the dynamic behavior of the monitored variables at the population level. However, such an investigation must take into account the fact that the rows of M are not independent from each other. In particular, two kinds of dependencies are present: the data may belong either to one patient or to different patients and the values of consecutive abstracted values may be related to each other since they belong to the same episodes. In order to avoid these problems, we randomly sampled the rows of M to obtain a submatrix Ms, in which the new data are not anymore correlated. On the basis of those data it is possible to apply a recently proposed algorithm for subgroup discovery [16], able to describe at the population level the subgroups which show peculiar behaviors.

4 Results The system described in the previous section is undergoing a clinical evaluation at the Limited Assistance Center Located in Mede, Italy, which is clinically managed by the Unit of Nephrology and dialysis of the Hospital of Vigevano, Italy. The current version of the software contains the basic auditing procedures that allow physicians in assessing the percentages of success and in visualizing both the time series of the median values and the time series of each variable in each session. Together with the association rules search, several graphical data navigation procedures have been implemented, based on simple plots, histograms and tables. The extraction of rules at the population level have been instead tested off-line using the Data Mining Server from the Rudjer Boskovic Institute, Croatia [17] The methods presented in the previous section have been applied to the data coming from 33 patients monitored for 3-8 months, for a total of 2527 dialysis sessions. Each of the monitored variables was sampled every [1–15] mins. Table 2 shows a synthesis of the dialysis center performance. For each outcome the number of failures and the percentage of the total number of dialysis is reported. Table 2. Outcomes assessment

# Failures % of total dialysis

QB*

∆T

∆L

VP*

AP*

∆W

Overall

620 23

321 12

169 6

152 6

1 0 (0.004)

60 2

992 39

Let us note that several times multiple failures occur. This explains why the overall number of failures is lower than the sum of the number of failures of each outcome. Search for associations between TAs and outcomes. The search strategy described in the previous section was implemented with a SUmin equal to the maximum value between 4 and the 40% of the sessions failed, while the minimum confidence was set to 0.5 for extracting the basic set B and to 0.9 for deriving the association rules. This step of the analysis allowed us to derive 18 association rules on the data of 7 patients, while for the other patients no rules have been found. Almost all rules are related to VP* (17 over 18). Two examples of the rules derived for two different patients are shown below: Patient 1: IF Trend of AP* is Decreasing, State of CF* is LOW and Trend of DP* is Increasing THEN VP* FAILS Support: 15 sessions, Confidence: 1, Total number of session failed: 36

Patient 9: IF State of SP* is HIGH and State of DP* is HIGH THEN VP* FAILS Support: 30 sessions, Confidence: 1, Total number of session failed: 56

The first rule describes a situation in which there is an increasing trend of both systemic pressure (DP*) and the hemodialyzer blood circuit pressure (AP*2) for patient 1; these are clinically relevant reasons to justify a value of VP* out of the normal range. The second rule describes the fact that the patient hypertension problems cause VP* failures for patient 9; this fact, although not proved by specific clinical studies, can be justified on the basis of available clinical knowledge. The overall set of extracted rules is currently under evaluation by physicians. We plan to carry on a formal evaluation of the results involving at least three nephrologists. Search for subgroups at the population level. Thanks to the subgroup discovery algorithm implemented in the Data Mining Server, it was possible to derive subgroups for several target attributes. The results for some causes of failure are reported below: Failure of ∆W (sensitivity 20%, specificity 100%): State of T* is LOW and State of Hb* is HIGH and State of SP* is LOW Failure of ∆T (sensitivity 35%, specificity 99.5%): State of WL* is LOW and State of T* is NOT NORMAL and State of OP* is NOT HIGH Failure of VP* (sensitivity 14%, specificity 100%): State of CD* is LOW and State of DP* is HIGH

Those rules turn out to be easily explainable on the basis of the available clinical knowledge. ∆L often fails due to hypothension problems (SP* is Low and Hb* is High); ∆T* is highly related to a low WL*, while VP failures are related to hypertension problems which cause an increase in the pressures of the hemodialyzer hematic circuit. The information extracted is clinically relevant, since it highlights what are the reasons of the problems that the dialysis center has to face with, and therefore, it may properly guide therapeutic decisions.

5 Discussion and future developments The project described in this paper applies a set of Artificial Intelligence techniques to address the needs of a clinical center in terms of data summarization and quality assessment. The auditing system is now in clinical use and it is planned to re-engineer the software for its widespread dissemination. It might be interesting to note that, if we consider the performance of the dialysis center from the beginning of the system use (17/06/2002), the percentage of failures decreased from 47.6% (first two months) to 32.5% (last two months). This result seems to show a potential impact of the use of the auditing system on the performance of the clinical center. Clearly, all the results described in this paper need to be assessed through an evaluation on a larger data set. From a methodological viewpoint several issues have to be still investigated. First, it must be formally evaluated the significance and usefulness of the association rules extracted. Moreover, it should be interesting to investigate the feasibility of extraction of rules in which the body of the rule is composed by a complex temporal relationship 2

Let us note that AP always assumes negative values, and a decreasing episode corresponds to an increasing episode of the absolute value.

between the TAs, instead of the conjunction of co-occurrent TAs. This may lead to rules such as, for example, “IF Trend of AP is decreasing BEFORE State of VP is High THEN VP fails” [14]. Finally, our aim will also be to identify patients at “risk of failures”, and to develop instruments to prevent unsuccessful outcomes. To this end, we are working on a probabilistic model for describing the temporal evolution of the patients in the TAs state space. Acknowledgements. This work is part of the project “Analysis, Information Visualization and Visual Query in Databases for Clinical Monitoring”, funded by the Italian Ministry of Education. We gratefully acknowledge Maria Grazia Albanesi, Daniele Pennisi, Andrea Pedotti, Antonio Panara and Davide Lazzari for their methodological and technical contributions. We also thank the team of the Unit of Nephrology and dialysis of the Mede and Vigevano hospitals.

References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17.

Stefanelli, M.: The socio-organizational age of artificial intelligence in medicine. Artif Intell Med. 23 (2001) 25-47. Abidi, S.S.: Knowledge management in healthcare: towards 'knowledge-driven' decision-support services. Int J Med Inf. 63 (2001) 5-18. Zoccali, C.: Medical knowledge, quality of life and accreditation of quality in health care. The perspective of the clinical nephrologist. Int J Artif Organs. 11 (1998) 71720. Roddick, J.F., Spiliopoulou, M.: A survey of temporal knowledge discovery paradigms and methods. IEEE T. Knowl. Data En., 14 (2002) 750-766. Registro Italiano di Dialisi e Trapianto, http://www.sin-italia.org McFarlane, P.A., Mendelssohn, D.C.: A call to arms: economic barriers to optimal dialysis care. Perit Dial Int 20 (2000) 7-12. Moncrief, J.W.:Telemedicine in the care of the end-stage renal disease patients. Adv Ren Replace Ther 5 (1998) 286-291. Ronco, C., Brendolan. A., Bellomo. R.:Online monitoring in continuous renal replacement therapies. Kidney Int 56 (1999) S8-14. Bellazzi, R., Magni, P., Bellazzi, R.: Improving dialysis services through information technology: from telemedicine to data mining. Medinfo. 10(Pt 1) (2001) 795-9. Shahar, Y.: A Framework for Knowledge-Based Temporal Abstraction, Art. Int. 90 (1997) 79-133. Bellazzi, R., Larizza, C., Riva, A.: Temporal Abstractions for Interpreting Diabetic patients monitoring data, Intelligent Data Analysis, 2 (1998) 97-122. Allen, J. F.: Towards a general theory of action and time. Artificial Intelligence 23 (1984) 123-154. Cohen, A., Daubechies, I., Jawert, B., Vial, P.: Bioorthogonal basis of compactly supported wavelets. Comm. Pure Aplli. Math. (1992) 45 485-560. Höppner, F.: Discovery of Temporal Patterns - Learning Rules about the Qualitative Behaviour of Time Series. Proc. of the 5th PPKDD, LNAI 2168 (2001) 192-203. Witten, I., Frank, E.: Data Mining, Academic Press, 2000. Gamberger, D., Lavrac, N.: Expert-guided subgroup discovery: Methodology and Application. J Artif Intell Res. 17 (2002) 501--527. Gamberger, D. & Šmuc, T. (2001). Data Mining Server [http://dms.irb.hr/]. Zagreb, Croatia: Rudjer Boskovic Institute, Laboratory for Information Systems.

Lihat lebih banyak...

Comentários

Copyright © 2017 DADOSPDF Inc.