For patients to choose a preferred hospital, information transmission using a hospital website is important because the external environment is constantly changing, owing to improvements to the internet and new competitive relationships as a result of the declining trend of patients [1,2]. It is assumed that there are 2 major reasons why information on websites has become increasingly important for hospital management. First, patients can choose a hospital based on their personal preferences. The “paternalism model,” where physicians make decisions, was previously the mainstream; however, patients have shifted to an “informed decision model,” where they make decisions themselves [3]. A previous study conducted by the Japanese Ministry of Health, Labor, and Welfare revealed that the percentage of those who responded that “I have obtained information” for the item “Is information usually obtained when you visit the hospital?” increased by 30% compared to 2011 [4]. Second, the popularization of smartphones among all age groups has led to significant lifestyle changes [5,6]. Patients can select a hospital based on the information available on a website. According to the National Federation of Health Insurance Societies [7], the information from a hospital website has the largest impact on hospital selection. The internet serves as an important source of medical information [8]. A study showed that “hospital websites” and “information websites about hospitals”’ account for a high proportion of the available credible information.

In previous research targeting hospitals in Japan, the introduction of marketing departments in hospitals has recently been proposed [9]. However, medical advertising guidelines make it difficult for hospitals to advertise as a company [10]. In today’s society, with advances in the internet, it is expected that the proportion of individuals who choose to seek information from such sources when deciding on a hospital for an examination or treatment will continue to increase, despite the strict regulations. Therefore, hospitals need their own web marketing methods.

Through recent marketing, CRM, one-to-one marketing, and personalization aimed at increasing patient satisfaction and loyalty by reaching individual consumers through the spread of the internet and intensified competition are attracting attention [11,12]. It is therefore desirable to transmit appropriate information to the websites according to the needs of each visitor. For this reason, to improve information provisioning on websites, it is common to analyze the current state of websites using quantitative data such as access logs. As advertising regulations were adopted on websites in Japan, the study focused on whether the information posted on websites followed the regulations [13,14]. In the United States, there are weaker advertising regulations on websites than Japan; a study evaluated the accessibility of websites to provide high-fidelity information from the viewpoint of public health [15]. Few have focused on the browsing behaviors of visitors on hospital websites. Suzuki et al [16] attempted to develop a method for creating browsing behavior models and assessing the intentions of patients on hospital websites. However, no models have been proposed that allow medical staff to intuitively understand the browsing behavior of patients and propose improvements to meet the needs of visitors seeking radiology-related information. Therefore, we focused on a Bayesian network that can easily understand variable relationships.

Owing to the market trends, we decided to analyze not all visitors but specific visitors. Our research target, Hokkaido University Hospital, is one of the largest hospitals in Japan with 944 beds. Furthermore, Hokkaido University Hospital has one of the 17 proton therapy facilities in the country and therefore provides advanced medical care in the field of radiotherapy [17]. It is possible that there was a certain need at this hospital to provide information in response to citizen requests for the transmission of radiology-related information. In addition, after the Great East Japan Earthquake (March 11, 2011), several individuals have become interested in radiology-related information. Some studies have reported that few people have sufficient knowledge of radiology, and many people are concerned about exposure to radiation [18,19]. Owing to these situations, several patients have sought information on the safety and effects of radiotherapy and radiological diagnosis. Furthermore, there are various types of radiological diagnoses, such as radiography, computed tomography, and magnetic resonance imaging, which make it difficult for patients to understand each type of examination. This study conducted a behavioral analysis focusing on visitors seeking radiology-related information.

The purpose of this study is to propose a hospital website search behavior model in which medical staff can easily understand the browsing scenario by analyzing the browsing behavior of visitors who are looking for radiology-related information based on the results of a Bayesian network from the perspective of one-to-one marketing.

A flowchart of our research is shown in Figure 1. The page that we believe the target wants to browse was defined as the “request page (RP).” In addition, a target reaching the requested page was defined as a “conversion,” and the ratio of conversion as the “conversion rate (CVR),” as shown in .We attempted to enhance the CVR using a Bayesian network with an access log. We used the website data access log of the Hokkaido University Hospital website (Figures 2 and 3), which has 208 pages. Our research focused on the patients seeking radiology-related information at Hokkaido University Hospital as the target visitors.

First, we collected the website access log of Hokkaido University Hospital from September 1, 2016, to August 31, 2017. Knowing the purpose and needs of the visitors is useful for improving a website; however, this information usually cannot be observed, and it is necessary to understand this information through an analysis of the data access log. This research analyzed the access log using Google Analytics. Google Analytics is a tool that allows users to visualize the access status without a server that even medical clerks can easily handle. Google Analytics has been widely used in public institutions, including hospitals [16,20]. We used “page view,” which shows the number of pages viewed by visitors, “session,” which shows the number of times any visitor accessed the website, and “page depth,” which shows the number of pages viewed in a single session.

There were 2,521,279 sessions during the study period, as presented in Table 3. Access to the department pages was concentrated next to the TOP page, as listed in Table 4. After data extraction, a behavior analysis was conducted for a total of 90 sessions, including 35 sessions for the radiotherapy interest group, 35 sessions for the nuclear medicine examination interest group, and 20 sessions for the radiological diagnosis interest group. The CVR for the 3 groups is presented in Table 5. In our study, 74% (67/90) of the targets could reach the RP.

The model described in Figure 5 was structured using 35 sessions from the learning data. The browsing of RP2, the “Radiotherapy Department page,” was affected by the browsing of the Proton Therapy News page, which affected the browsing of the radiology-related department and Medical Expenses pages. The results of the sensitivity analysis are illustrated in Figure 6. The mutual information volume of the Proton Therapy News page was the largest, indicating that browsing this page had the greatest effect on the browsing of RP2. Based on the results of the sensitivity analysis, we calculated the change in the browsing probability of RP2 owing to browsing or nonbrowsing of the pages based on a probability inference, as shown in Figure 7. When the target visitor browses the Radiological Diagnosis Department page, the probability of RP2 increased by 17% from the prior probability.

The model, as described in Figure 8, was structured using 35 sessions from the learning data. Browsing of the RP2 “Nuclear Medicine Department page” was affected by the browsing of RP4 “PRT-CT and Nuclear Medicine Examination” and was related to the browsing of the reservation and the medical staff pages. The results of the sensitivity analysis are shown in Figure 9. The interaction between the browsing of RP2 and that of RP4 was also demonstrated. We then calculated the change in the browsing probability of RP2 and RP3 owing to the browsing or nonbrowsing of the pages, as depicted in Figures 10 and 11. The browsing probability of RP2 increased when browsing the Medical Staff page. When the target visitors browsed the FAQ page, the browsing probability of RP3 increased by 15% from the prior probability.

The model described in Figure 12 was structured using 20 sessions from the learning data. The browsing of the RP2 “Radiological Diagnosis Department page” affected the browsing of the Radiotherapy Department and related department pages, such as Orthopedics. The results of the sensitivity analysis are shown in Figure 13. Browsing of the Radiotherapy Department page had the greatest impact on RP2 viewing. We then calculated the change in the browsing probability of RP2 owing to the browsing or nonbrowsing of the Radiotherapy Department page, as shown in Figure 14. When the target visitors browsed the Radiotherapy Department page, the probability of RP2 increased by 35% from the prior probability.

One of the limitations of this study is the low accuracy of the models applied. The number of samples (20-35 sessions) used to build the models was small compared with the actual usership for Hokkaido University Hospital’s website. Owing to the recent shift to Secure Sockets Layer for use in search engines, the sample size has remained limited. Future studies should substitute search keywords. Furthermore, some data may have been missing because there were some pages set with the no web tracking code in Google Analytics. By increasing the sample size, factors such as browsing time can be incorporated into the model, improving the prediction accuracy for the probability distribution of each page. However, the effects of the visitor’s interest on each page regarding factors such as the browsing time and browsing devices for hospital websites are still unknown. Therefore, it is necessary to clarify the effects of the browsing factors before modeling.

As another limitation, we were unable to verify the accuracy of the proposed model compared to other approaches. To evaluate our model, we need to accurately set the web tracking code and extract data without using retrieval keywords because we have obtained few retrieval data in recent years.

Third, the RP we set during the page-evaluation phase is uncertain. Thus, setting an RP objectively to obtain information by means of questionnaires to accurately understand visitor needs will be an issue in the future. Similarly, it will be possible to evaluate pages suitable for each visitor by clarifying the above relationship between visitor interests for each page and access indicators such as the browsing time. In recent years, Japan has been rapidly aging, and older adults are using hospital websites. However, previous studies have reported that age-related deterioration in visual acuity and changes in color vision affect the use of older adults to use such websites [30,31]. Therefore, we hope that the factors affecting CVR differ between older adults and young adults. In the future, we would like to consider improving the CVR by modeling the categorization of young and older adults.

This study structured 3 browsing models, each based on specific patient needs: radiotherapy, nuclear medicine examination, and radiological diagnosis. A total of 74% (67/90) of the target visitors could reach their requested page, but only 2% (2/90) could reach the inspection information page owing to the site structure. Furthermore, for the radiological diagnosis interest group, browsing of the “Radiological Diagnosis Department page” affected the browsing of the related department pages, such as Orthopedics.

Plans were considered for improving the browsing probability of the inspection information page such as summarizing radiology-related information on the website rather than posting according to the hospital’s organizational structure. Thus, our method has the potential to increase the probability of delivering the desired information to the users. In the future, we plan to improve the accuracy of the models by using a large number of samples. We believe that an increase in the number of subjects will reveal the characteristics of the browsing scenario in more detail.

We built models with the objective of improving the provisioning of radiology-related information; however, because the structure of a hospital website is not unique, we hope to apply the proposed method to other hospitals. It is expected that the provisioning of information on hospital websites will be improved using this approach.