Applying sequential pattern mining to investigate cerebrovascular health outpatients’ re-visit patterns

Sequential pattern mining (SPM)

SPM is a method to discover all sequential patterns with user-specified minimum support. The support value of a pattern is the number of data sequences that contain the pattern (Agrawal & Srikant, 1995). The SPM method is commonly used to identify frequent patterns or behaviors that appear in a dataset. In the medical area, some methods have been proposed. For instance, the Time-Annotated Sequence algorithm was used to assess the effectiveness of extracorporeal photopheresis (ECP) from a liver transplant dataset (Berlingerio et al., 2007), and Ali et al. (2008) mined routine behaviors in a pervasive healthcare environment using the Routine Tree approach.

Association rule mining (ARM)

ARM, a task that discovers hidden correlations among items in a large database (Wang & Shao, 2004), has received considerable attention, particularly since the publication of the AIS and Apriori algorithm (Agrawal, Imielinski & Swami, 1993; Agrawal & Srikant, 1994). In the medical area, it has been applied to identify relationships among specific drugs or specific diseases. For instance, Soni et al. (2011) used association mining to predict or diagnose a heart attack by seeing the correlation between some attributes associated with the heart attack. Ji et al. (2011) discussed implementing association mining for the early detection of unknown adverse drug reactions (ADRs). In Taiwan, the National Health Insurance Research Database managed by the Taiwan National Health Insurance has implemented association mining in several studies, for instance analyzing the co-prescribing pattern of antacid drugs (Chen, Liu & Hwang, 2002), exploring the comorbidity of attention deficit/hyperactivity disorder (ADHD) (Tai & Chiu, 2009), and investigating pairs of drugs associated with Stevens-Johnson syndrome (SJS) re-hospitalization (Ou-Yang, Agustianty & Wang, 2013).

In general, several studies have shown the importance of predicting demand in medical practice due to its usefulness for strategic planning. For instance, predicting demand in an emergency department (McCarthy et al., 2008) and healthcare services (Al Nuaimi, 2014), and predicting patients’ visits at an urgent care clinic (Batal et al., 2001) were discussed. However, to the best of our knowledge, only a few studies have investigated characteristics of re-visit patterns of patients, especially using a cerebrovascular health examination dataset. Therefore, this study combined some methods used by previous research to delve into the issue. The proposed methodology consists of three main parts: data discretization, SPM, and association mining.

The data discretization portion of our proposed method employs equal-frequency discretization (EFD), adapted from Wong & Chiu (1987) and Garcia et al. (2013). The main part of our methodology is a priori-based sequential pattern mining first proposed by Agrawal & Srikant (1995). This method is used to detect useful information of frequent (totally or partially ordered) subsequences, and is commonly used in marketing to identifying customer behaviors. This study adapted this method for identifying outpatients’ re-visit behaviors towards regular medical check-up activity, in terms of a risk of stroke. The last portion of our proposed method adopted a priori-based association mining, which was used to investigate pairs of drug associated with SJS re-hospitalization (Ou-Yang, Agustianty & Wang, 2013). In our study, this method was used to investigate pairs of health factor variables associated with outpatients’ re-visit behaviors. Some adjustment needs to be considered to combine the association mining method with pattern mining in the previous step.

Data collection and data preprocessing

A seven-year (2004–2011) collection of cerebrovascular health examination data records from a local medical center in Taiwan was used in this study. It consisted of 13,692 medical check-up records from 12,062 outpatients. A detailed description of the dataset is presented in Table 1.

In the beginning of the process, data preprocessing was applied to remove noise from the data, extract and merge data from different sources, and convert the data into a proper format (Hu, 2003). The procedures to conduct data preprocessing are described as follows.

Data grouping of ABD variables

Mining sequential patterns

After the final grouping result was obtained, the next step was extracting frequency patterns by employing SPM (Mooney & Roddick, 2013). This research tried to develop a new type of SPM method called a two-phase SPM. In the first phase of the method, frequency patterns were separately extracted by obtaining candidate sequence item sets of each characteristic group for each of the c- visits. The traditional priori-based SPM (Srikant & Agrawal, 1996) was employed in this phase by converting the dataset into candidate sequence item sets, in which procedures for each characteristic group are described in pseudocode in Table S5 and illustrated in Fig. 2.

Mining association patterns

Regarding the complete frequent sequence patterns of each of the c- visit groups, the association mining method can be employed to obtain combinations of frequent sequence patterns which have high association values. The association mining method can be conducted using the following procedures (Ou-Yang, Agustianty & Wang, 2013).

Results of final data grouping of ABD variables

According to Table S3, there were some groups in which cardinality did not satisfy the cardinality threshold. Hence, merging such a group with another group, which was closest in distance, was necessary. Using SPSS software (SPSS, Chicago, IL, USA), the centroid coordinate of candidate merging groups was obtained. The distance between candidate merging groups was computed by means of the Euclidean distance. The merging process was iteratively performed until all groups’ cardinality satisfied the threshold. The group merging processes of the two-visit and three-visit groups from dataset are respectively illustrated in Figs. S1 and S2.

The final characteristic of the ABD-based grouping result and the percentage value of the explained characteristic are presented in Table 3. Group characteristics, which are presented in columns three to five of Table 3, are defined by identifying characteristics of a group’s members that mostly appear in the corresponding merged group. The explained character denotes the total number of members which contained the corresponding group characteristic.

Furthermore, an evaluation of the significant difference between groups is required. The purpose is to ensure that ABD-based grouping results actually split the data in accordance with the similarity. Box’s M test in the discriminant analysis was employed to estimate the within-group covariance, with a critical value of 50% (Hair et al., 2006), and was performed using SPSS software. Table 4 presents results of the discriminant analysis, in which significant values of group characteristics from the two-visit and three-visit data records were <0.05. This implies that the grouping result was good. However, an error result occurred for the four-visit data records because they contained a large number of variables with a small number of data. In the discriminant analysis, the more variables used, the more data are required. Henceforth, the two-visit and three-visit group datasets were used to generate frequent sequence patterns and association mining.

SPM analysis and results

SPM was conducted on a cerebrovascular health medical records dataset to generate hidden knowledge, which could possibly represent outpatients’ re-visit behavior patterns toward regular medical check-ups. In general, the generated results of frequent sequence patterns can be analyzed by identifying characteristics of an outpatient’s return based on different personal ABD health histories.

For instance, in the frequency sequence pattern results of two-visit data records presented in Table 5, groups II-23, II-39, and II-40 were considered to be middle-aged people (aged over 47 years), with no history of abnormal blood pressure or diabetes. People in group II-40 (aged over 59 years) were more likely to come in for a second visit because of an abnormal blood pressure or abnormal cholesterol which occurred at their first clinical visit. The former frequent sequence pattern contributed the highest support value of 63%, while the support value for the latter was 61%. Another high support frequent pattern showed that 62% of middle-aged (aged 53–58 years) people in group II-23 came for a second visit after having been diagnosed with abnormal cholesterol.

On the other hand, results also showed that outpatients with the characteristics of group II-44, i.e., being aged over 59 years and with a history of diabetes, were more likely to re-visit the clinic for a second time. The frequency sequence pattern contained abnormal blood pressure, abnormal blood sugar, as well as an abnormal neck vasculature ultrasound on the first visit, and had a 59% support value. Results of group II-44 imply that the pattern occurred often, and the corresponding outpatients could probably be diagnosed with a stroke in accordance with an abnormal result from the radiology test, especially through the neck vasculature ultrasound. Based on this condition, people of this age and in a poor health condition are at high risk of stroke according to the domain knowledge. Moreover, physicians could probably prepare appropriate medications for outpatients or suggest special treatments such as an MRI brain test to ensure the diagnosis, if necessary.

In Table S6, groups III-53, III-60, and III-68 showed no frequent sequence pattern. This implies that with no history of hypertension or diabetes, and no abnormal conditions in their health factor variables even on their second visit, it is basically their habit to seek regular medical check-ups. This indicates that those people can maintain their health from a young age and into old age.

Regarding the SPM results of the three-visit group presented in Table S6, we highlighted some patterns which had high support values. Groups III-66 and III-72 showed interesting frequent sequence patterns which were quite similar to one another. The former group contained changes in characteristic in terms of age, from Age1 to Age2 and Age2 to Age3, with no history of hypertension or diabetes. In group III-72, people under age 26–52 years with a history of hypertension were more likely to re-visit the clinic for a medical check-up and to find the same result on each visit. For instance, if they found abnormal blood pressure or abnormal BMI on the first and second visits, they would be more likely to visit the clinic a third time. The support value of these patterns was up to 68%. Learning by this condition, the physician could probably begin suggesting that outpatients go for a consultation with a nutritionist on the third visit, as a marketing purpose of medical services to take care of the abnormal BMI condition. In addition, they could also suggest other medications to outpatients, whose blood pressure remained in an abnormal condition through their second visit.

Results and analysis of association mining

Results of association mining that consisted of a single variable can be seen in columns 2 and 4 of Table 6 and Table S7 for the two-visit and three-visit groups, respectively. In general, abnormal blood pressure, abnormal BMI, and an abnormal neck vasculature ultrasound test were single variables that were found to be very indicative of the re-visit behavior in the two-visit and three-visit datasets, due to their large support counts. However, considering a single variable directly as outpatients’ re-visit behavior characteristics might be less useful, because the knowledge is too general. Hence, in this section, an analysis of the results focuses on some pairs of health factor variables that can be measured to determine pairs of variables which have a high association leading to re-visit behavior patterns. Each pair of health factor variables can be identified in accordance with the general behaviors of each of the c-visits and its degree of association (Jaccard and Cosine measurements). As presented in Table 6 and Table S7, every pair represents an association between variables that occurred in frequent item sets of the c-visit group.

In general, behavior {f₁}∧{f₁} of the two-visit group, a variable abnormal result of blood pressure paired with an abnormal BMI, at the first visit, showed a quite high association degree with the outpatients’ re-visit behavior. The variable pair was considered to coincide with the Cosine value of 67.6% and covered the re-visit behavior pattern with 47.5% for the Jaccard value. On the other hand, variable blood pressure also had quite a high association when it was paired with the radiology diagnosis. The value of the cosine for BPA₁∧NVA₁ was 61.5%, which was two-times higher than BPA₁∧MBA₁. This means that patients with the pairing of abnormal blood pressure and an abnormal neck vasculature ultrasound result were more likely to represent the behavior than MRI of the brain. The reason might be that ultrasound is available in many places throughout the country and is less expensive than MRI; hence it is still reasonable to perform a radiology diagnosis on the first visit. Another high association in the two-visit dataset was found in the variable pair of abnormal blood sugar (indicating diabetes) and abnormal triglycerides (indicating hypertension) on the first visit. The degree of association for this pair reached a Cosine value of 53.4%, which implies that those two variables coincided but covered the re-visit behavior pattern by only 33.9% according to the Jaccard value. However, a high degree of association was obtained in the case of the three-visit dataset, in which abnormal blood sugar occurred along with abnormal BMI, both of which are related to diabetes.

A similar result was also shown in the variable pair III-1-2, from general behavior {f11}∧{f11} of the three-visit group presented in Table S7, with respective Cosine and Jaccard values of 43.3% and 18.8%. Nevertheless, in the three-visit group, outpatients with abnormal blood pressure who had an abnormal MRI of the brain result on their second visit had a higher degree of association than those outpatients who continuously had abnormal neck vascular ultrasound results in every visit. In this case, the physician could recommend that outpatients have an MRI to follow-up on the abnormal findings and get a more-detailed view. Once the results show an abnormality, outpatients tend to begin scheduling regular medical check-ups to monitor the progress of their health conditions.

Our findings show that variable pairs with the highest degree of association were found in general behavior {f11}∧{f01} followed by {f10}∧{f01}. Those variable pairs had Cosine and Jaccard values of 100%. The former general behavior implies that the presence of an abnormal variable f still continued in the subsequent visit, while the latter shows the change in variable f from normal at the first visit to abnormal in the second visit. Those were associated with outpatients’ re-visit behavior patterns. In particular, a patient who had an abnormal blood pressure tended to begin monitoring their health condition only if it was paired with obesity and/or abnormal blood sugar. The measurement values show that each variable pair in general behavior {f11}∧{f01} coincided with and covered 100% of the re-visit behavior pattern.

Moreover, variable pair III-3-1 emphasizes that any patient who receives an MRI of the brain will tend to return for another medical check-up. This interpretation was also supported by variable pair III-2-2 to variable pair III-2-5 in general behavior {f11}∧{f01}, which contained either MBA or NVA variables. This implies that most outpatients will begin monitoring their health after receiving a radiology diagnosis such as an MRI of the brain or a neck vasculature ultrasound test. Even though the degrees of association of these variable pairs were not as high as general behavior {f11}∧{f01}, which was only around 50%∼70% of the Cosine and Jaccard measures, it still can be implied that a radiology diagnosis plays an important role in the association with outpatients’ re-visit behavior patterns. Whether the MBA or NVA continuously occurs in every visit in the sequence or only in the first or second visit, these features can encourage outpatients to return to check the progress of their health conditions or to ensure that the result is actually a matter of concern.

Result validation

Evaluation of the obtained results in clinical practice is often based on interpretations by domain experts. In this work, we created a questionnaire with 10 questions to determine the degree of agreement among experts toward our findings (Appendix S1). Those experts were from six different divisions such as general medicine, general surgery, neurology, plastic surgery, rehabilitation, and pediatrics. An index of inter-rater agreement in statistical methods, the so-called content validity index (CVI) (Horgas et al., 2008; Jakobsson & Westergren, 2005; Larsson et al., 2015) and the modified Kappa value (Larsson et al., 2015) were used to measure agreement for ordinal data. The experts were asked to what extent they agreed with each item of our findings on a 5-point Likert scale, which ranged from 1 (strongly disagree) to 5 (strongly agree). For each item, the item CVI (I-CVI) was computed as the number of experts giving a rating of either 4 (agree) or 5 (strongly agree), divided by the total number of experts. The evaluation criteria of the agreement, proposed by Landis & Koch, 1977, were used to interpret the strength of agreement beyond chance based on a modified Kappa value (k*). A value of k* <0 is poor (P); k* = 0∼0.20 is slight (Sl); k* = 0.21∼0.40 is considered fair (F); k* = 0.41∼0.60 is moderate (M); k* = 0.61∼0.80 is substantial (S); and k* = 0.81∼1.00 is almost perfect (A).

Table 7 presents the overall degree of agreement of 10 items by 16 experts. According to the results, the experts tended to agree with items that focused on re-visit behavior patterns of middle-age and old people, who underwent a regular medical check-up, due to their experience with hypertension and/or diabetes as well as abnormalities found on the radiology test. Those items obtained the strength of agreement of “substantial” to “almost perfect” in the range of 60%–100%. This situation seems to be common knowledge among physicians. On the other hand, evaluation criteria for items Q1, Q3, and Q10 were slight (Sl), in which the experts found this behavior to be uncommon in their knowledge. However, according to our dataset, the aforementioned items were a fact of outpatients’ re-visit behavior patterns that might still be uncommon knowledge for experts, yet could be expected to provide useful insights for medical services to provide appropriate treatment and increase the service quality. Additionally, some of the items from Q6 to Q9 represented the association of variable pairs, while some others represented the radiology test and its association with blood pressure or blood sugar. The strength of agreement among experts toward these items of our findings was considered to be fair (F) or moderate (M) in the range of 20%–60%.