This paper presents a process model to guide the data mining projects in the health care sector. The process model (PMH) is a specialization of CRISP-DM methodology and presents di erent issues associated to the data analysis and management. This proposal was validated in order to address real problems related to health care in Colombia. The results show that it is possible to establish new hypothesis about the clinical data, and revalidate these a rmations using the proposed process model.
In the health care sector, there are many opportunities to apply data mining. Some of them are related to the improvement of the quality control in health care. In particular, analysis to detect and diagnose diseases, predict the responses of the organism to speci c treatments and to identify epidemiological pro les, are relevant themes for the health care community.
There are many methodologies to tackle data mining
opportunities such as CRISP-DM[
This paper presents a process model to guide the data mining process in the health care sector. This process model allows to reduce the costs and resources used in data mining projects. The process model was evaluated by analyzing 49,000,000 individual register of health care (RIPS) obtained from di erent sources, such as HMOs and the Minister of Social Protection in Colombia from 2003 to 2006. The evaluation goal was to compare treatments among Health Maintenance Organizations (HMOs), as well as verifying the compliance to standards of evidence gained from the scienti c method (EBM- Evidence-based medicine); which will support the quality process of health services. The analysis was focused on hypertension diagnosis and allows to evidence similarities between the national guidelines and the health service. Speci cally it was possible to identify the use of captopril an Angiotensin-Converter enzyme inhibitor medicament in the hypertension treatment. This medicament is cheaper according to other medicines of this type. One hypothesis is that it is pre-scripted for economical reasons. However, patients with this kind of treatment, returns to the healthcare institution with complications in the hypertension disease. This kind of complications increase the illness costs. Finally, validations about this process model enable new opportunities to establish public health policies in Colombia.
This paper is organized as follows. Section 2 describes problems related to the data mining process in healthcare. Section 3 presents Health care data management characteristics and the PMH the Process Model proposed for data mining in Healthcare sector. Section 4 exposes the validation method of the proposed process. Finally, Section 5 concludes the paper and presents other research issues.
There are many studies that evidence the relevance of data mining techniques in the health care sector. These studies are associated to the treatment of patients and generally, to the identi cation of best practices in the treatments of speci c diseases.
Some works such as [
Abidi and Stolba in [
Although some of the works such as [
This section presents a brief description about health care context in subsection 3.1, with the purpose to highlight the opportunity to provide new guides associated to data mining applications to improve di erent kind of projects in health care. Subsection 3.2 describes a new process model to guide these projects.
The health care sector evidences challenges w.r.t. data management because of data characteristics such as volume, quality, availability, accessibility, and the relevance of the decisions involved during the process analysis.
Moreover, there are clinical guidelines that describe a set of steps to treat a speci c disease. From these guides, it is possible to determine the service e ciency, the time involved in the treatment, and typical practices such as treatments and medications. This kind of information provides important elements to the decision's maker.
According to the volume of data analyzed, it is important to highlight that all clinical cases are relevant during process analysis. This is contrary to other sectors. In other domains a rule is meaningful when the support of the data is relatively high, whereas in the health care domain, analysis involving mortalities will be accounted for although the number of cases will not be signi cant, statistically speaking. The ideas mentioned before motivates the development of the process model for healthcare (PMH), which is described in the next sections.
The knowledge in data mining projects frequently remains in few people like consultants and experts in speci c domains. For this reason, the experiences and processes cannot be reusable and applied to similar projects in health care. Furthermore, it is necessary to know di erent guidelines, references and standards related to quality of service, with the purpose of understand the main characteristics of the health care domain.
As a result, this paper presents PMH (Process Model for the health care context). This process is an specialization of the CRISP-DM methodology proposed in Colombian health care context, based on the veri cation carried out in the pharmacological and non-pharmacological treatments in Colombia's health care institutions (IPS), through the application of data mining techniques on RIPS les. This process allows to reduce time and resources with respect to develop mining projects in this domain without a speci c knowledge. This guide proposes seven steps in an iterative way tacking into account health care context. The following paragraphs provide a description about the di erent steps involved in this process model. The numbers used in this description correspond to the number used in the gure 1. I. Scope De nition of the exercise. This step allows to de ne the business problem to be analysed. Several fundamental aspects must be clari ed in this step: what, how, and why the assessment is done, as well as de ning the criteria for the success of the exercise. Some of the typical questions proposed by experts are focused on problems to validate the e ectiveness of pharmacological treatments in the emergency room according to IPS's best practices, others in the control and monitoring of chronic diseases according to the standards. In this step, these questions will be contextualized according to the service(hospitalizations, urgencies, procedures or medical appointments), and the diagnosis to be monitoring.
II. Selection of the reference guide. This step II consists on the selection of the reference guide(s) to evaluate the question proposed in the rst step; for example the medical treatment used for a diagnosis. Currently, it is possible to use the expert advise to validate the quality of service in health care. On the other hand, there is speci c manuals such as standards, protocols, and clinical guidelines proposed by governments and organizations that can be used as a reference guide.
The standards are evidence based on references used to
evaluate the quality and performance of services, while protocols
are documents that describe the rules of action depending
on a speci c circumstance [
About clinical guidelines, the World Health Organization
(WHO) presents di erent guides related to the diagnosis,
treatment and prevention of speci c diseases such as obesity,
malnutrition or diabetes. Furthermore, di erent countries
have established national standards to treat a disease. For
example, Colombia has the 412 resolution which suggests
the set of activities and procedures that should be used in
public health diseases [
Although clinical guidelines vary in content, they have essentially the following structure:
Clinical guideline structure 0. Authors 1. Introduction 2. Disease detection 3. Diagnosis 4. Classi cation and Tracking 5. Disease evaluation 6. Non-pharmacological treatment 7. Pharmacological treatment 8. Disease complications 9. Disease special situations 10. Hospital treatment 11. Emergency treatment 12. Clinical guidelines future review recommendations 13. Bibliography In the structure above, the interest lies (mainly) in points 4 to 11. The quality control proposed is based on the comparison between treatments with a speci c admission diagnosis and an established clinical guide diagnosis.
The suggestion is to choose the reference guide that has been established in the national policies or regulations, because international clinical guidelines may not have validity in a speci c country, or may not be applied in certain IPS because of socio-economic or epidemiological factors. III. Identi cation of information sources. This step consists of the identi cation of useful information sources according to the scope of the project and the selected reference guides. This identi cation depends on the data quality and availability. These issues will be tacking into account to decide the use of these sources during the analysis process. In the healthcare sector, there are di erent sources that can be obtained and used for the development of data mining projects.
Generally, countries have an individual healthcare register corresponding to every hospitalization, urgency or procedure associated to a patient. For example, the United States has the Electronic Health Record (EHR) which includes demographics, medical history, medication and allergies, immunization status and observations (among others). On the other hand, Colombia has the Individuals Registers of Health Care (RIPS in Spanish), that provides information related to the delivery of health services and demographic variables.
There are other kind of information sources related to national behavior such as national survey or naming standards. Some national surveys contains demographic and health information that can be used to support the data mining process. Furthermore, the WHO de nes the CIE10 standard. This standard de nes the classi cation and organization of diseases based on a unique code that represents the category and the speci c a ection.
Some naming standards are associated to speci c health interventions available for each country. For example, the Australian Classi cation of Health Interventions (ACHI) contains all the procedures that are realized by HMOs in the country. The Unique Procedures Classi cation in healthcare (CUPS in Spanish) is the Colombian classi cation system for this information.
IV. Selection and preparation of healthcare information. It is necessary to have the support of health experts to select the information that is highly relevant to solve the proposed problem in step I. Each disease presents di erent characteristics. For example, there are diseases like prostate cancer or pre-eclampsia in which sex is not a determining factor. The rst a ects men, and the latter applies only for pregnant women. In chronic diseases like hypertension, time is a signi cant variable. Its treatment is based on monitoring the patient periodically with formulated procedures and medicines according to a speci c order and to patient's evolution over time. On contrary, appendicitis treatment is considered relatively short. In this case the time variable is not relevant. These considerations must be taken into account in selection step.
The selected information follow a data cleaning process. Generally, health information has problems related to data management such as replication of records. Moreover, medicines data management proposed a new challenge to data mining experts. Usually, this data does not have a standard to be lled. For example, the medicine "amoxycillin" can be lled as "amox" or "amoxycilin". In these cases, the similarity word analysis can be used to solve the issue.
The corresponding discretization of continuous data and the standardization of information must be made, necessary procedures for the execution of mining algorithms, which should be in line with own business rules of the selected diagnosis. For example, in the case of Alzheimer's disease, age categories should be created after 40 years, being consistent with the characteristics of vulnerable populations, and the evolution of the disease over time. Di erent from Appendicitis disease, where age ranges should be used much broader, since it is a disease that can occur at any age. The complexity of both the discretization and the standardization of data may depend largely on the amount of selected information sources and the absence of the use of ontologies for the uni cation and standardization of medical terms and concepts.
An statistical analysis of this step is necessary for physicians because it's important to know the data percentage that must be cleaned and the problems that arise the datasets. V. Information adjustment and preliminary analysis. In this step is important to analyse the resulting dataset of the previous phase. This analysis concerns to identify the main characteristics of the data and the discovering of new variables that are relevant for an speci c situation. In healthcare, variables such as hospitalization window, total cost of treatment and patient satisfaction are relevant in many situations. For example, to solve questions like, what is the most expensive treatment?, which is the one with the lowest satisfaction?, which one represents a lower rate of days of staying? these variables are relevant in this context. The preliminary analysis is based on descriptive statistics. The objective is to review the statistical data distribution in order to avoid biased results. In this step can be identi ed how many men or women are involved in the dataset, how is the age distribution or if the treatment is based signi cantly on drugs rather than procedures. The health experts evaluate the results and if necessary, this step is review again. VI. Selection and implementation of data mining algorithms. This step consists in the identi cation of data mining algorithms to achieve the objectives proposed. In healthcare, there are probable classi cations of the mining problems. The rst, is related to the analysis of treatments, which is feasible to predict the organism response against speci c procedures. Next, the monitoring and evolution of patients in infectious and chronic disease. The latter, is based on verify the proper provision of health services. For example, for the last classi cation there situations in which is appropriate to know the percentage of compliance of treatments with respect to a clinical guidelines or which are the implications of meeting/failing the clinical guidelines in terms of costs. The data mining algorithms that are proposed are association rules, sequential patterns and clustering techniques to solve speci c problems in healthcare sector.
Association Rules: this technique identi es the cause- effect relations between variables. It is possible to characterize an speci c group using clustering techniques, and determine the behaviour of an speci c cluster using association rules. In healthcare, it's interesting to discover relations among events. For example, if a patient's disease evolves to a chronic phase, the probability to have a decease based on speci c characteristics can be found. Moreover, In the context of quality control, the association rules allow us to discover rules that may or may not correspond to established treatments in clinical guidelines It is recommendable to use this technique in the diagnosis of acute illness. This kind of illness usually are applied during a patient's admission to the IPS and does not require periodic monitoring to ensure new procedures or drugs depending on the patient progress. Furthermore, association rules could be used in problems that analyze the treatment's behaviour.
The possible way to do in quality control in the health care it is as follows: Two data sets are created, one for drug treatment information and one for non-drug information for each patient with the same admission diagnosis, and with the same method of admission. It is important that the information of the admission method is not mixed, since the in each treatments can be di erent. Thus, given an X diagnosis in hospitalization, we have the following data set of medications: For this example, an association rule would be: "In the pharmacological treatment of a patient with an X diagnostic, if medicines M30 and M90 are provided, then M70 medicine will also be provided." This technique required some parameters to be customized: The support and the con dence. The rst one is the proportion of transactions in the data set which contain the itemset: supp(A) = occurrence(A)
size(dataset).
The second one is de ned as the conditional probability ( P (BjA)): the occurrence of A, given the occurrence of B: conf (A ) B) = supp(A S B) supp(A).
Sequential Patterns: This technique detects cause-e ect relations considering the time periods in which transactions occurred. In health care context, there are many situations that involves periodic controls and patient's monitoring. As mentioned before, standards, clinical guidelines and protocols specify sequences of treatments that should be applied in a explicit situation. For this reason, it is feasible to compare the patient's procedures versus reference guides using sequential patterns. Chronic diseases like hypertension or diabetes requires that patients return to the IPS several times with the same diagnosis. This kind of problems involves many variables in each time period. The recommendation is to use sequential patterns to analyze the evolution of patient's health based on the treatments applied. Based on the concept that, at any given time, a clinical event is the formulation of one or more drugs or procedures we propose the creation of two data sets, drugs and procedures datasets. In this case the records should be grouped by patient, ordering clinical events in ascending order by date and time. All clinical events from a patient arranged in order, can be seen together as a sequence, where each event corresponds to a set of drugs or procedures. Figure 4 shows the sequences found in previous patients. Clustering: to found groups of elements with similar characteristics. In healthcare, is very common to analyze populations based on speci c characteristics. Nevertheless, it's possible to use this technique in the validation of right treatments to right people. The reference guides de ne speci c treatments for people with Speci c demographic characteristics. In this case, the use of clustering determines subsets based on procedures and demographic information. In addition, there are situations in which data quality is poor. For this reason, its impossible to analyze particular issues in the dataset because of the con dence of data. The suggestion is to use clustering techniques to generalize the main characteristics of an speci c group. To perform these analysis, a data set must be created which includes patient information such as gender, age, marital status and race, (among others), as well as drugs and/or procedures provided, grouped into treatments found in the previous section, as shown in gure 5: As in previous algorithms, it is important to analyze the number of people supporting each cluster, before making any conclusions. It is also essential to understand, that a cluster represents a very small percentage of the population does not necessarily implies that should be discarded. It all depends on the clinical context that is being evaluated and the criteria of the medical expert.
VII. Result Validation and Impact Analysis. This nal steps concerns to tunning up the data mining model based on the health expert feedback and the results achieved. A detailed study of the results is made by a board of medical experts in the area supported by a technical group of people to determine whether or not to repeat some of the above steps, possibly with a change in strategy or range, or by a re nement of the data used, to get speci c conclusions.
This proposal is validated based on the PMH process and the product obtained by applying this steps in the solution of a real problem in Colombia's health care sector. Our proposal is a specialization of the CRISP-DM methodology. The CRISP-DM process has been validated in several domains and modi ed during more than a decade, based on the application of the process in many data mining projects. In that sense, we re ned the CRISP phases in order to improve the knowledge and special issues of the health care sector, reducing the time and resources that have to be used for understand this particular domain.
The PMH process incorporates the diagnosis and reference guide selection in the business understanding phase of CRISPDM, and presents the selection criteria for these steps. Furthermore, in the other phases is possible to understand the principal problems associated with data quality in health care, and to determine which algorithms should be used to solve several problems in this sector.
The next sections are focused on the product validation. It applies the PMH process in the quality control of Colombia's health care sector and validate the results based on the experts criteria. The development of this exercise consists in two iterations of the process. The speci cs steps of PMH are described below.
Hypertension is a chronic disease that a ects 20% of people in the world. This is considered the rst cause of morbility and the most representative disease related to cardiovascular a ections. For this reason, the objective of this exercise was to evaluate the pharmacological and non-pharmacological treatment for this disease in Colombia. plemented using the IBM Intelligent Miner 8.1 data mining tool.
The gure 6 shows the results of the mining model.
Step I consists on the description of the problem according to
the context, related to the diagnosis and type of disease that
is going to be tackled. In this case, it is relevant to analyze
the characteristics of hypertension. This is a chronic disease,
generally asymptomatic and it requires continuous medical
assistance. The typical complications of hypertension are
related to cardiac failures. This complications implies
hospitalizations, urgencies and complex procedures.
Based on the health experts support, there are di erent
reference guides related to hypertension, but its treatment may
di er from one country to another. For this reason, the
selection of the reference guide was focused on the "Clinical
guideline for hypertension disease" proposed by the
Association of Faculties of Medicine of Colombia [
As mentioned in section III, Colombia has the Individual Registers of Health Care. The objective of this data is related to bill the delivery of health services made by the Colombian's IPS. The information was collected from HMOs and the Minister of Social Protection from years 2003 to 2006 (49,000,000 of individual register of health care). Furthermore,the RIPS data are lled based on the CIE10 and CUPS standards. For this reason, these information is also taken into account.
According to the expert opinion, the relevant information for this analysis is described in table 2:
Variables Principal diagnosis IPS Identi cation Date Sex Department Type of medical service Medicine Procedures Length of stay decease diagnosis related diagnosis The statistical analysis results for the dataset shows that a 77% of the registers correspond to procedures, 17% are medical appointments and a 6% of the data to medicines prescriptions. Moreover, a 67% correspond to men and a 33% to women. On the other hand, a 16% of the patients return to the IPS for health controls.
In the pharmacological treatment of hypertension, we want to analyze the most relevant medication sequences. In this case, the time variable is highly relevant because we want to trace the prescription of medicines for a speci c disease. For this reason, we used sequential patterns, according to the ideas presented in subsection 3.2. Our model was imThe main result is the use of captopril. More than 40% of the patients were medicated with this medicine. Based on health experts opinions, the captopril is a medicine used in mono-therapy treatments and is prescripted for economical reasons. Other important conclusion is related to data quality of medicines; a speci c naming standarisation was used to resolve the problem. In the same way, several manual process was realised to mitigate replication of records problem.
Based on the results of the rst iteration, a second iteration of the steps suggested in PMH was performed. Therefore, the health experts proposed to analyze if the health system may incur in higher costs because of the prescription of captopril to the patients. Using the RIPS data, we want to determine the complications related to these patients. According to the PMH, it is necessary to prepare the data that will be used by the model. For this reason, the data presents demographic information and relevant aspects about the evolution of a patient's disease.
In this case, new variables have to be include based on the health expert recommendation. This variables are associated to the evolution to a chronic phase in the clinical history of patients. For this reason, were introduced the date of the patient's complication, the associated diagnosis, the number of hospitalizations or procedures before and after the complication of hypertension. The next step of the methodology, proposes a preliminary analysis of the information. In this case, the information consists on all the records of the patients that have su ered this disease.
To describe the complication of a patient, the sequential clustering technique was used. As described in section 3,with this technique it is possible to nd clusters of patients with similar sequences and characteristics.
The results of the second iteration shows that patients with similar sequences associated to the use of captopril have complications such as chronic cardiac failures, hypertensive crisis or heart attacks.
In general, it is important to analyze that the use of captopril is pre-scripted for economical reasons. This strategy is useful in a short term period, but in a long term, we can observe that patients with this kind of treatment, returns to the healthcare institution with complications in the hypertension disease. This kind of complications increase the illness costs.
This paper proposes a process model to guide the data mining process in the health care sector. It suggests a set of iterative and facultative steps to improve the results of the mining process. This process model was evaluated using the analysis of quality of service for the treatment of hypertension in Colombia. The results shows that it is possible to establish new hypothesis about the datasets, and revalidate this a rmations using the proposed process model. At the same way, these results evidence some facilities provided to the data mining expert to guide their process, specially associated to the knowledge about healthcare context such as data sources, reference guides and data mining techniques. An exhaustive validation of the process model is considered as future work, in terms of a formal comparison between the use of CRISP-DM and PMH. At the same time, new kind of question from the expert point of view will be interesting to resolve using this process model. In particular, the identi cation of an epidemiological pro le for Colombian population.
The authors appreciate the support of Jose Abasolo, professor at Los Andes University, which provides the initial ideas and suggestions for the development of this article.