General practitioners’ attitudes towards decision-supported prescribing: An analysis of the Dutch primary care sector

Clinical decision support systems

Clinical decision support systems in the primary care sector have the potential to improve the decision-making of GPs and pharmacists.^1–4 In their extensive literature study, Kuperman et al. ¹ conclude that in order to realize the benefits of these systems, context-specific implementation complexities have to be addressed; in their words, “there is not yet a ‘one-size-fits-all’ approach.” Other studies that have been undertaken specifically in the Dutch primary care sector show mixed results regarding the influence implemented decision support systems have. While Martens et al. ²⁰ found no favorable effects for computerized messages reminding GPs to prescribe certain drugs, messages not to prescribe them “sometimes positively influence[d] the prescribing behaviour [sic] of GPs.” A study by Kuilboer et al. ²¹ showed that the system they tested “changed the manners in which the physicians monitored their patients and, to a lesser extent, their treatment behavior.”

Due to these developments, CPOE software in the primary care sector has increased in diversity in recent years. From systems that were mostly organizational in nature, they have been enhanced to facilitate consultation of electronic medical records and clinical decision support. ²²

Software market diversity

Consequently, the Dutch software market of GPs’ CPOE systems is diverse. Up to 10 different systems are currently in use, all providing their own distinct features and tools in addition to the common patient record management facilities. ²² Some of these systems include their own digital prescription aids and decision support tools, while others do not. As a result, third-party developers have produced additional software that, integrated with their existing systems, assists GPs in prescribing drugs. The major plug-ins on the market are Prescriptor, which can communicate with most existing GPs’ CPOE systems, and NHGDoc, which currently has limited integration capabilities.^23,24

Technology adoption

The adoption of technology by potential users has been extensively studied and has led to the development of various predictive models, including the technology adoption model (TAM), TAM2, and unified theory of technology acceptance model (UTAUT).^25–27 These models have not, however, performed equally well at predicting acceptance behavior for all kinds of user groups. Subsequently, the potential influence of various different aspects has been explored, and the models extended accordingly. Physicians, in general, and GPs, in particular, have been studied extensively, and elements as diverse as financial stimuli or psychological ownership have been proposed.^28,29 The deviations from elements common in most models that were found in multiple studies regarding this particular group are elaborated upon in the following.

When applied to GPs, some theoretical elements common in technology acceptance models do not seem to be supported. Notwithstanding new research by Ketikidis et al. ³⁰ suggesting otherwise, many studies show an application’s ease-of-use is of minor or no importance to their acceptance by GPs, even though in literature, this is usually considered a core element.^31,32

In contrast, time reduction is mentioned by various authors as a highly relevant and influential factor regarding GPs’ technology adoption, while this is not an element commonly found in any of the models.^28,32,33 In their study, Chismar and Wiley-Patton ³¹ tested the TAM2 model on a group of pediatricians and found that, of the inherent constructs, only output quality and job relevance were determinant factors that influenced the respondents’ adoption.

In their study on the low acceptance rates of prescription software, Boonstra et al. ²⁸ tested the original TAM and concluded that its factors were of hardly noticeable influence, and that adoption was mainly affected by systems’ embedding in consultation processes, financial stimuli, policy decisions, and cultural boundaries instead.

Hypotheses

The hypotheses that were tested in this study are listed below. As will be explained in more detail in the “Research Design” section, the hypotheses were tested for correlations, and thus, they do not contain causality implications. Therefore, the survey findings are discussed in the “Results” section, whereas the meaning of the data analysis will be elaborated on separately in the “Discussion” section.

Based on the consensus in the literature concerning the benefits of clinical information systems, the following hypotheses were assumed and tested:

Following the model in Figure 1 on the aspects influential on behavioral intention, these hypotheses were formulated and tested:

Finally, also according to the model in Figure 1, the influence of age on the various aspects was investigated through the following hypotheses:

Method

A survey was deemed an appropriate method to examine potential users’ attitudes toward a system assisting them with polypharmacy prescribing. The fact that many potential users could be surveyed was a decisive factor.

First, the questionnaire (Appendix I) investigated demographic information, such as respondents’ ages and potential partnerships (Appendix 1). The next section was dedicated to surveying GPs on their current use of information systems and plug-ins and included questions about formularies and common problems. The last section started with a short explanation of the features the POMP would include, and investigated their opinions toward the newly proposed system. In this last section, only a conceptual idea of the POMP was provided, without specific details regarding its implementation or user interface.

To ensure an optimal response rate, the number of questions was kept to a minimum. Consequently, only a limited number of questions on technology adoption could be included. Consistent with the theoretical framework, questions regarding perceived ease-of-use were disregarded. The oft-mentioned time constraint was surveyed, as well as the perceived influence of financial stimuli. Specific questions on usefulness and output quality were included, but job relevance was not directly surveyed; the authors considered it unfeasible to ask about both relevance and usefulness, as the subtle difference could be lost in a questionnaire. Instead, the job relevance construct was divided over two questions: one asking about the number of polypharmacy patients each GP treated (phrased as “patients using five or more drugs”) and another on the problems they encountered with these patients.

The questionnaire limitations restricted the explicit inclusion of less often investigated aspects, such as cultural biases mentioned by Boonstra et al. ²⁸ or psychological ownership proposed by Paré et al. ²⁹ A comment section, however, was provided where respondents could fill out any additional remarks. The survey in its final form contained 15 questions.

The number of respondents necessary to generalize results was calculated with the “proportion sample” and “finite population correction” formulas provided by Israel, ³⁵ which showed that 95 respondents were required to fill out the questionnaire. After taking into account an expected response rate of 20 percent, the survey was sent by postal mail to 500 practicing GPs, who were selected through a process of simple random sampling (i.e. each individual was chosen randomly and entirely by chance). The addresses were geographically spread across the Netherlands; addresses included more practices in the country’s more densely populated western provinces than elsewhere, but this can be explained by the fact that more GPs work there. A prepaid return envelope was included, giving respondents the opportunity to fill it out on paper or online. Ethical concerns described by Swanwick ³⁶ were taken into consideration in the conduction of the survey. The responses were gathered anonymously, and the respondents were assured that their results would be treated confidentially. No coercion of any kind was used to persuade potential respondents to participate nor were any incentives provided in return for their answers. Responses were gathered during a period of 3 months, from mid-February to mid-May 2011. After this period, no additional responses were received. During or after this period, no attempts to increase the response rate were made.

Validity

The internal validity of most of the survey’s questions on technology adoption followed naturally from their close relations to the theoretical constructs. In cases where the constructs were divided over multiple questions (such as those on job relevance and financial stimuli), results were not combined into single values to avoid distortion of the concepts.

As the respondents were selected through a process of simple random sampling, this study’s findings are generalizable to the whole population of 8921 practicing GPs in the Dutch primary care sector, insofar as a confidence interval of 7.15 and a confidence level of 95 percent are maintained. As it turned out, multiple questions were answered unevenly. As the frequency of answers in contingency tables was often lower than five, Fisher’s exact test (FET) was favored over Pearson’s chi square test to test the hypotheses. ³⁷ The setup of the study required the authors to test for correlations only. As correlation by itself does not imply causality, the influences in Figure 1 were not directly tested. The implication this restriction has for the results interpretation will be further elaborated upon in the “Discussion” section.

When testing a large number of hypotheses, the potential problem of multiple comparison analysis arises. This problem states that the probability of a tested result being significant by chance increases when testing multiple hypotheses. Different approaches to cope with this problem exist, but they are disputed; in their pursuit to minimize type 1 errors, they might erroneously introduce type 2 errors, that is, falsely accepting a false hypothesis.^38,39 The techniques are most widely employed when two or more different samples are used. In this single-sample study, the Bonferroni–Holm correction technique was employed to check the results’ validity whenever subgroups’ correlations were tested. In order to avoid type 2 errors, however, these corrected p-values or significance levels are not incorporated into the results reported below; applying the Bonferroni–Holm correction ensured, however, that the hypotheses from which the main conclusions were drawn remained significant after correction.

Research group

The survey was sent by mail to 500 practicing GPs, accompanied by a cover letter. Potential respondents were invited to either fill out the survey online or return their copy by mail. The vast majority (98%) chose for the latter option.

A total of 184 GPs returned filled-out questionnaires, which made a response rate of 37 percent. Of these respondents, only 15 percent were younger than 40 years; 27 percent were in their 40s, while 47 percent were in their 50s. Ten percent were 60 years or older, with only one respondent being above 65 years of age. Twenty-one percent of the respondents worked alone, 37 percent shared their practice with one colleague, but the majority of 42 percent worked with several others. All but one of the surveyed GPs reported having a professional information system; one was unsure.

Software market diversity

The earlier assumption regarding the diversity of the CPOE systems market was proven to be correct by the survey results. Two of the companies dominated the market, as they were used by 58 percent of the respondents (see Figure 2). Seven smaller manufacturers each had market shares between 1 and 12 percent. Twenty-nine percent of the respondents indicated owning Prescriptor, while 25 percent were unsure of that; the remainder indicated not owning the product. Fewer respondents (13%) owned the comparable decision support system NHGDoc, with an additional 26 percent being unsure.

OPEN IN VIEWER

Figure 2.

Model depicting the relative ownership of CPOE software by Dutch GPs.

Of the respondents indicating having one of these systems, the actual use differed. Of the owners of Prescriptor, the majority (55%) indicated using it never or rarely (less than once a week). Thirty-nine percent used it once or multiple times per consultation, with only 6 percent using it once per week. NHGDoc numbers were comparable: 54 percent used the software package rarely or never, 38 percent used it intensively, and 8 percent just once per week.

Decision support systems and medical formularies

Statistical analysis showed that owners using one decision support product (either Prescriptor or NHGDoc) are more likely to use the other one as well than GPs who do not (p = .002, FET); this led to acceptance of H1. The medical formularies, often integrated with GPs’ CPOE systems and facilitated by Prescriptor and NHGDoc, were widely used. Seventy-eight percent specified using the Pharmaceutical Compass, 38 percent used the guidelines by the Nederlands Huisartsen Genootschap (NHG; Dutch College of General Practitioners) and 13 and 18 percent also adhered to regional rules as established in Groningen and Nijmegen, respectively. There appeared to be some evidence that ownership and use of plug-ins was related to GPs’ use of medical formularies. Owning Prescriptor correlated with using the NHG-guidelines (p = .020, FET) and the regional guidelines from Groningen (p = .029, FET). The use of Prescriptor was correlated with the use of NHG-guidelines only (p = .006, FET). Additionally, the number of formularies used correlated with Prescriptor ownership (p = .015, FET).

Relations including ownership or use of NHGDoc did not occur in the same number; use of this plug-in correlated with use of the NHG-guidelines (p = .009, FET), but none of the other relations. These results validated H4 in the above-mentioned cases. H2 and H3 were accepted for Prescriptor but not for NHGDoc, while H5 was rejected altogether.

Problem recognition

The problems commonly associated with polypharmacy were widely recognized. Ninety-five percent of respondents recognized a higher risk of adverse effects with polypharmacy patients. Patients of 83 percent of the surveyed GPs showed signs of decreased adherence to their drug prescriptions. Forty-nine and 40 percent of the respondents recognized increased risks of under- and overmedication, respectively. Overall, after correcting for double counting due to the possible inclusion of GPs sharing the same clientele, GPs each treated 159 polypharmacy patients on average.

The sixth and seventh hypotheses were tested by checking whether the use of any of the formularies was related to the problems encountered. Especially relations including the NHG-formulary showed significant results; underprescription (p = .002, FET) and decreased adherence (p = .044, FET) were more often reported by GPs using the NHG-guidelines. Overall, GPs using the NHG-formulary reported more problems than those who did not (p = .016, FET). None of the other formularies showed any relations, except for the Nijmegen-guidelines, which showed that use of those clinical rules correlated with the recognized occurrences of underprescription (p = .036, FET). This partly validated H6 for the cases involving the NHG- and the Nijmegen-guidelines. H7 was valid only for the NHG-guidelines.

With the testing of these hypotheses, the current situation of the use and impact of decision support systems in the Dutch primary care sector was investigated. The results are visually represented in Figure 3.

OPEN IN VIEWER

Figure 3.

Model depicting the probability values of the statistically significant relationships between decision support systems, medical formularies, and recognized problems.

Perceived usefulness

When informed about the possible new system aiding them with treating polypharmacy, the majority (57%) of GPs responded positively; only 5 percent responded outward negatively. Subsequently, the constructs presented in the model on technology acceptance in Figure 1 were tested. The two questions on job relevance showed different results; the number of problems encountered with polypharmacy patients did not appear to have any relation with the POMP’s perceived usefulness (p = .759, FET); neither did any of the specific problems themselves appear to correlate with GPs’ attitudes (adverse effects: p = .186, FET; undertreatment: p = .676, FET; overtreatment: p = .394, FET; and adherence: p = 1.000, FET). The total number of patients using five or more different drugs that each GP reported did not significantly correlate either (p = .079, FET). The aspect of output quality, however, appeared to correlate strongly (p = .000, FET) with the POMP’s usefulness as perceived by the GPs. A similar result was found for the relation between time investment (p = .000, FET) and the system’s usefulness. The two questions used to measure the financial phenomenon, on payability and reimbursement, showed significant correlations with p = .000 (FET) and p = .015 (FET), respectively. This led to the acceptance of H11, H12, H13, and H14. H8, H9, and H10 were rejected.

Age

Finally, the hypotheses regarding the impact of respondents’ ages on the constructs were tested. As it appeared, age correlated significantly with the constructs of time (p = .016, FET) and payability (p = .022, FET). The aspects of output quality (p = .066, FET) and reimbursement (p = .098) were not related. Age appeared to significantly correlate neither with the number of encountered problems (p = .725, FET) nor with the number of patients (p = .963, FET); no evidence was found to assume any relations between age and the recognition of specific problems (adverse effects: p = .960, FET; undertreatment: p = .827, FET; overtreatment: p = .191, FET; and adherence: p = .203, FET). This validated H19 and H21 but led to the rejection of H15; H16; and H17, H18, and H20. All probabilities found with FET that are part of the model in Figure 1 are depicted in Figure 4.

OPEN IN VIEWER

Figure 4.

Model depicting the probability values of all tested relationships between the factors relating to use behavior, according to Figure 1.

Limitations

Even though utmost care was taken in the conduct of this study in order to ensure its reliability and validity, some considerations should be taken into account when applying its findings. The survey employed in the study contained some questions that did not directly represent theoretical constructs, most notably the one on job relevance. This choice was made because of the possible ambiguity respondents may perceive between that concept and usefulness, but ultimately, it hinders generalizability. The large differences between the rough estimates given by the respondents may explain the lack of significance in the relations measured. Considering that this study’s outcomes on job relevance do not concur with scientific consensus, the authors suggest that these specific findings should be generalized outside of this study with caution.

Additionally, given that respondents voluntarily participated in the study, they may be more interested in, and thus hold more positive attitudes toward clinical decision support systems than nonparticipants.

Further research

This study has investigated factors of influence on GPs’ attitudes toward adopting decision support systems. Factors commonly not included in TAMs, such as time or financial investments, have been proven to be influential in shaping GPs’ attitudes toward such systems.

Further research should focus on investigating the wider generalizability of these claims. Additionally, the salient beliefs held by health-care professionals should be explored, as proposed by Holden and Karsh, ⁴² in order to find out whether indeed some fundamental concepts of technology adoption, most notably perceived ease-of-use, do not influence this group’s attitudes, while other aspects generally not included in such models do.