New digital technologies offer the opportunity to provide more connected health services to individuals [1]. Health information systems (HIS) building on these digital technologies provide different health services to individuals. They can help to identify diseases and support individual treatment faster [2]. HIS exist in various forms with different levels of automation, as the share of HIS using artificial intelligence (AI) is rising. AI-based HIS are self-learning and improve their own algorithms on the basis of an ongoing evaluation of the data entered [3]. As such, and until now, AI-based HIS have mostly been used to support physicians in their decisions concerning medical treatments toward patients, for example, in radiology [4]. However, AI-based HIS can also be used by patients to monitor their long-term treatments, for example, concerning cancer detection and treatment [5]. Nonetheless, the use of AI-based HIS nowadays is still substantially lower than non–AI-based HIS [1]. Thus, while predicting the use of such future systems is important [6], investigating the use of AI-based HIS is of particular importance to maintaining successful HIS adoption in the future [7]. By doing so, AI-based HIS shall be able to adapt to new situations, which is especially useful and more efficient in health care applications [8].

In health care, there are 2 different perspectives that have to be balanced to explain intention and behavior and thus are relevant for investigating health-related perspectives. The term for these 2 perspectives is “duality approach” [9,10]. First, the disease is considered, and how individuals root their behavior in this regard is referred to as “adaptive appraisals.” Second, individuals evaluate the behavior toward the intervention in question to counter the disease, which is called “nonadaptive appraisals” [11]. Current research shows that the duality approach of balancing adaptive and nonadaptive appraisals is sufficient and important for AI-based HIS [3]. Investigating such appraisals in health care is of particular importance, as individuals often come into physical contact with information systems [12].

To investigate these 2 perspectives, we combined theories of different areas to have an interdisciplinary view. Namely, they are the Health Belief Model (HBM), the Protection Motivation Theory (PMT), the Extended Parallel Process Model (EPPM) from a health care perspective, the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) from a technological perspective, and the reasoned action approach (RAA) from a psychological perspective. By combining these theories, it becomes evident that the theories share variables for their explanation process, although they are arranged in different ways. Such combinations have already been used to develop integrated frameworks to predict the intention to use AI-based HIS, for example, by Gesk et al [13,14] and Chu and Liu [15]. Nonetheless, all the approaches only consider short- or long-term treatments while failing to compare and draw conclusions for both. This is important as one-point-in-time decisions differ from ongoing decisions that accompany individuals every day [16,17].

Thus, our research question to fill this gap is which factors are relevant for the intention to use AI-based HIS for (1) short-term treatments and (2) long-term treatments?

To answer this question, we design an integrated framework using the duality approach of adaptive and nonadaptive appraisals. We do so as we apply our framework to the diseases such as cataracts, arthrosis, and diabetes. All diseases are non–life-threatening but limit the quality of life [18-20]. They, however, differ in terms of standard procedures. Arthrosis operations typically follow less-standardized protocols as fewer people are affected [13,19,21], cataract surgeries are conducted with standard procedures for a large number of individuals [13,22,23], and diabetes follows a continuous treatment approach for which the options are also quite standardized. Hence, cataracts and arthrosis are considered short term; for example, surgeries that are supported by AI can cure these diseases for at least a certain amount of time [19,22]. In contrast, diabetes is considered long term; for example, treatment plans that are supported by AI can at least reduce complaints that are caused by diabetes [24].

To investigate the intention to use AI-based HIS for short- and long-term treatments, we surveyed 693 individuals in total, who were assigned to the 3 diseases and 3 related AI-based HIS scenarios. First, physicians perform the medical treatment on their own. Second, physicians are supported by AI. Third, AI performs medical treatment on its own. Our results showed that with the increasing support of AI, adaptive appraisals become more important than nonadaptive appraisals for both short- and long-term treatments. Thus, individuals do not consider the disease in balancing their appraisals anymore; instead, they evaluate the treatment based on information concerning the AI-based HIS. Hence, this study provides important implications for business and research as we give insights into how information for AI-based HIS should be distributed.

This study is organized as follows: first, we define HIS. Second, we provide an overview of relevant theories in health care, information systems, and psychology and combine the relevant elements to design our integrated theoretical model. Third, we present materials and methods, in which we apply our integrated model to short- and long-term treatments with different scenarios. We then report the results of our study, which are then discussed concerning short- and long-term treatments and our integrated theoretical model. This study concludes with theoretical and practical implications and limitations, as well as directions for future work.

For our study, we used 3 different scenarios: (S1) arthrosis, (S2) cataracts, and (S3) diabetes type 1. Each scenario investigates different treatments that are performed (1) by physicians, (2) by physicians with AI support, and (3) by AI only. We used quantitative surveys for our study. First, we introduced the respective scenario using explanations shown in Multimedia Appendix 1. To ensure that our questions were answered attentively and correctly, we used attention checks according to Goodman et al [67]. Specifically, participants had to pass initial test questions to continue. We aligned the questionnaire according to the variables in the integrated model. The reflective variables are intention to use (3 items adapted from Fishbein and Ajzen [41]), attitude (HIS and disease; 5 items each adapted from Fishbein and Ajzen [41]), perceived norms (4 items adapted from Fishbein and Ajzen [41]), perceived behavioral control (4 items adapted from Fishbein and Ajzen [41]), perceived response efficacy (3 items adapted from Taheri-Kharameh et al [68]) as well as fear (HIS and disease; 3 items each adapted from Izard et al [69]), and nonadaptive rewards (3 items adapted from Vance et al [61]). The formative variables are perceived vulnerability (4 items adapted from Ku and Hsieh [17]) and perceived severity (7 items adapted from Ku and Hsieh [17]). All items were measured on a 7-point Likert scale. In addition, experience with AI-based services, general computer skills, and technical knowledge about AI are self-developed single-item control variables, which were measured with a 5-point Likert scale next to basic demographics. The complete questionnaire can be found in Multimedia Appendix 1.

For gathering empirical data, we used the crowdworking platform Clickworker, which is similar to Amazon MTurk. On the one side, we asked 496 probands hypothetically having (S1) arthrosis and (S2) cataracts. We used this approach as we aimed to investigate medical decisions that are short term, that is, individuals had to decide whether they intend to use AI or physicians for treatment. Furthermore, they should be able to comply with nonadaptive appraisals, that is, to decide to have no treatment, which is possible for arthrosis and cataracts [19,22]. For S1 and S2, a total of 40.7% (203/498) of the participants were female, 59.1% (294/496) were male, and 0.2% (1/496) did not specify their sex. On the other side, 197 probands actually had type 1 diabetes (S3). Hence, we took a different approach to long-term treatments, as we had to ensure that individuals actually have the disease to be able to judge differences concerning medical treatments performed by AI or physicians. Hence, we investigated participants who received recommendations on how to handle their type 1 diabetes. For S3, a total of 40.9% (81/197) of the participants were female, 58.1% (114/197) of the participants were male, and 1% (2/197) of the participants did not specify their sex. We gathered a total of 693 participants, who were distributed among the scenarios (a) physician only, (b) physician with AI support, and (c) AI only as follows: S1a: n=101, 14.6%; S1b: n=74, 10.7%; S1c: n=83, 12%; S2a: n=78, 11.3%; S2b: n=83, 12%; S2c: n=77, 11.1%; S3a: n=67, 9.7%; S3b: n=60, 8.7%; and S3c: n=70, 10.1%. The mean average ages were 37.08 (SD 12.07) years for S1 and S2 and 36.04 (SD 11.52) years for S3. Our participants were aged between 18 and 70 years. We further measured some descriptives that included general computer skills (S1 and S2: mean 4.25, SD 0.744 and S3: mean 4.29, SD 0.744); experience with AI-based services (S1 and S2: mean 3.06, SD 0.917 and S3: mean 3.24, SD 0.993); and technical knowledge about AI (S1 and S2: mean 3.34, SD 0.850 and S3: mean 3.47, SD 0.867).

A partial least squares (PLS) approach to structural equation modeling (SEM) was used to test individuals’ intention to use AI-based HIS short- and long-term medical treatments. We used the Software SmartPLS 4.1.0.4. Variance-based SEM is more suitable than covariance-based SEM in cases when target constructs are to be predicted and explained with the constructs and the main drivers in structural models [70]. For our investigation, we used bootstrapping and permutation procedures with 5000 subsamples using recommendations of Hair et al [70,71] for PLS-SEM for short- and long-term treatments alike. Thus, we had to evaluate our measurement and structural models to ensure that we could compare our scenarios appropriately and report on the predictive power of our PLS-SEM. First, we investigated our reflective measurement models by considering the respective loadings, internal consistency reliability, convergent validity, and discriminant validity. The reflective variables in our investigation are “abilities (HIS),” “intention to use,” “attitude (HIS),” “attitude (disease),” “perceived norms,” “perceived behavior control,” “perceived response efficacy,” “perceived costs/barriers (HIS),” “perceived threat vulnerability,” “fear (HIS),” and “fear (disease).” We determined that all reflective variables and items surpassed the respective thresholds. Second, we examined our formative measurement models considering convergent validity, collinearity, and loadings and weights of our formative items. Our sole formative variable is “perceived threat severity.” The formative variable and items surpassed the respective thresholds. Third, we investigated our structural models by reporting on collinearity and R² values to display the explanation power of our study, as well as the standardized root mean square residual [72]. We determined values of 0.087 (S1 and S2) and 0.089 (S3) for our saturated models.

Furthermore, we performed PLSpredict and cross-validated predictive ability test examinations to report on the predictive power of our models and used Measurement Invariance Assessment procedures to report on significant differences between groups. To report on those significances, Hair et al [70,71] stated that f² values have to be considered that were adapted from Cohen [73]. According to Cohen [73], f² values of 0.02, 0.15, and 0.35 represent small, medium, and large effect sizes, respectively. Furthermore, Kock and Hadaya [74] proposed that f² values depend on sample size and therefore need to be adjusted based on the sample used. Accordingly, we determined the following f² values for our scenarios by using G*Power (version 3.1.9.7 [75]): S1a: 0.0621478, S1b: 0.0853821, S1c: 0.0759154, S2a: 0.0808975, S2b: 0.0759154, S2c: 0.0819737, S3a: 0.0945622, S3b: 0.105969, and S3c: 0.0903957. By investigating PLSpredict, cross-validated predictive ability test, Measurement Invariance Assessment, and f² values, we determined that our integrated theoretical model was sufficient to report on group differences and to predict further investigations toward the intention to use.

All procedures performed in our survey involving human participants were in accordance with the ethical standards of the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. As stated, the data were collected with a platform (Clickworker) on which participants contribute anonymously and receive a money award for their time spent, which was about €3 (US $3) per survey. Furthermore, we ensured that our participants provided informed consent, permitting the use of the survey data for research purposes. We ensured that our data analysis and presentation is exclusively anonymized and therefore no conclusions can be drawn about certain individuals. As a consequence, the Ethics Committee of the School of Business and Economics at the University of Marburg evaluated the research design and process and granted an exempt status.

In our investigation, we designed an integrated model to determine intentions to use AI-based HIS for short- and long-term treatments. Our results revealed that intention to use is well predicted by our model. Furthermore, to answer our research question, we estimated several variables that predict the intention to use for short- and long-term treatments. Thus, our model is sufficient for this and further investigations addressing the intention to use AI-based HIS.

By comparing short- and long-term treatments, our results showed that there were several differences in estimating the significance of our hypotheses. First, perceived threats of fear (disease) are important to consider. Thus, we support the propositions of Gao et al [18], who investigated wearable devices in health care and stated that perceived threats and perceived effectiveness are mandatory for the intention to use HIS. While both of our variables, threats and fear (disease), are significant variables while physicians participate in the treatment, threats are less important when the treatment is performed by AI only. In our case, treating arthrosis, cataracts, and diabetes by physicians only (or supported by AI-based HIS) are treatments that are known to individuals. We conclude that treating such diseases by humans is the most common appearance in individuals’ perceptions of surgeries [34]. Furthermore, we conclude that nonadaptive appraisals are more important when the treatment is known. We do so as we support the propositions of Longoni et al [1], who stated that individuals are more likely to rely on recommendations from physicians than from AI. According to them, physicians perform better than AI in evaluating an individual’s medical history. In contrast, adaptive appraisals are more important when the treatment is unknown to individuals, for example, as it is supported by AI only. This also matches with further differences in the significance we determined with our integrated model.

Second, perceived efficacy toward attitude (HIS) is important for all scenarios. Hence, we support the proposition of Venkatesh et al [49], who stated that perceived efficacy is one of the major predictors of determining the intention to use new information systems. Furthermore, Palmisciano et al [30] supported this proposition as they investigated the attitudes of patients and their relatives concerning the use of AI in neurosurgery. They stated that the efficacy individuals perceive toward the efficiency of AI-based HIS is dependent on their attitude toward the HIS.

Third, our results revealed that attitude (HIS) is important to determine the intention to use AI-based HIS. Thus, we support the propositions of Fishbein and Ajzen [41], who stated that attitude is one of the major predictors in determining the intention to use toward a certain behavior. In addition, Esmaeilzadeh [7] supports the propositions of Fishbein and Ajzen [41] and Palmisciano et al [30], as she estimated that attitude and perceived effectiveness are important for the intention to use. Furthermore, physicians are relevant with regard to the intention to use AI-based HIS; for example, training leads to more positive attitudes toward HIS and is beneficial for the intention to use. Furthermore, our proposition is supported by Yakar et al [35], who determined that attitude toward AI is important for intention to use in a Dutch society. In contrast to our study, their participants actually stayed in a Dutch hospital and were surveyed concerning treatments in radiology, robotic surgery, and dermatology. Hence, and for our propositions concerning arthrosis and cataracts, attitudes might change if individuals actually fall ill with the relevant diseases, which might also lead to changes in their behavioral intention. Fourth, we determined that perceived norms are important in our integrated model, as we estimated the significance of attitude (HIS) as well as perceived efficacy. Hence, we support the propositions of Broadbent et al [32] and Dziergwa et al [12]. While Broadbent et al [32] determined that the acceptance of social care robots by older adults was low in 2009, Dziergwa et al [12] estimated the opposite in 2017, reflecting high acceptance by older adults in 2017. They stated that society plays an important role in the perception of new information systems and that this perception may vary over time due to experiences. As such, the acceptance of social care robots changed in society over time, reflecting that perceived norms as opinions of others are important in the determination of intention to use and might lead to a change in perception. Fifth, our results showed that once AI is involved in our treatments, fear toward HIS is important, while fear toward the disease is important for all scenarios. Hence, our results support the propositions of Yun et al [36] and Fritsch et al [40] that emotions are important when determining the intention to use medical treatments by both physicians and AI. In accordance with Yun et al [36] and Fritsch et al [40], fear (HIS) is important for all scenarios with AI involvement. Hence, we conclude that our participants are concerned about whether AI is able to perform proper treatments. Sixth, we determined that perceived efficacy toward intention to use differs in significance between short- and long-term treatments. While significance was determined for treatments by physicians and physicians with AI support concerning short-term treatments, it is not significant for long-term treatments. In contrast, perceived efficacy toward intention to use is significant for long-term treatments performed by AI only, while it is not significant for short-term treatments. Hence, our results support the propositions of Palmisciano et al [30] and Esmaeilzadeh [7], who stated that patients want physicians to monitor autonomous AI treatments as they doubt the efficacy of AI-based HIS.

Our study leads to several theoretical implications. First, we propose an integrated theoretical model that is helpful for the explanation and prediction of the intention to use AI-based HIS. With positive results of various tests concerning measurement and structural models as well as model predictions, we recommend that our model be used in future studies that address the intention to use AI-based HIS. Compared to models from information systems, the prediction gets better due to the duality approach. In comparison to the health care models, we add the specifics of information systems, which enhance the duality approach beyond general treatments. Second, our integrated model benefits the theories we addressed from health care, information systems, and psychology. In particular, we show how adaptive and nonadaptive appraisals are balanced when individuals determine the intention to use information system–based medical treatments that are unknown and not available yet. Third, our empirical results contribute to research that addresses short- and long-term medical treatments. We provide certain scenarios that can be used in other studies to estimate the intention to use AI-based HIS. Fourth, we benefit from studies that address different technologies, that is, concerning different types of AI involvement in medical treatments. In particular, we show how the intention to use medical treatments differs when AI is not or only partly existent or performs the treatment on its own.

Our results have several practical implications. First, health care professionals and physicians can use our results in their daily work. As we determined that perceived efficacy and attitude (HIS) are mandatory for the intention to use AI-based HIS, we propose that health care professionals and physicians consider those variables when advertising novel HIS-based treatments. In particular, they should address information that influences attitude (HIS) positively and raise perceived efficacy in that they clarify risks and show the benefits of AI-based treatments. Second, information system professionals benefit from our study, as we investigated patients as important stakeholders in AI-based HIS. Specifically, we estimated that many patients intend to use AI-based HIS and thus support the propositions of Tran et al [31] that the demand for AI-based HIS is high. Thus, we propose that information system professionals consider fear (HIS) in their development, as they should—combined with perceived efficacy—ensure that AI-based HIS are error-prone. Then, error-prone HIS should lower fear (HIS) and improve perceived efficacy. Third, our results benefit health care management. As we ascertained that the overall perception of AI-based HIS in society changes over time and thus acceptance is rising, we propose that management performs AI training. In particular, health care management should organize training for health care professionals and physicians to improve treatments that are supported or performed by AI-based HIS. Ultimately, such training should lower error rates and hence lead to more positive attitudes toward such HIS and higher perceived efficacies.

Our study has some limitations. First, we investigated scenario descriptions for our study and our participants for short-term treatments, and our participants had to imagine that they fell ill with the disease. Thus, in contrast to our long-term treatments, our participants did not have arthrosis and cataracts, and their behavioral intention might change when they actually fall ill. Nonetheless, other studies show that attitude is important for individuals who actually stay in a hospital [35]. Second, the participants in our empirical study had an average experience with AI and new technologies in general, while we acquired individuals from across German society. Although these variables had no significant impact, groups with more or less experience could lead to different results than ours. Other studies from Germany determined similar results from a patient point of view [40], while AI is perceived better by German physicians, who had better abilities toward AI use [39]. Furthermore, recruiting participants from other cultural backgrounds could lead to different results. Third, our investigation is subject to the famous intention-behavior gap [41]. Although intention is the most important to predict human behavior, individuals might act differently in their actual behavior, as there might be other variables that influence their behavior. Hence, future studies should address various factors that influence intention to use AI. They could use duality approaches when patients are actually skilled in AI use due to various utilizations for medical purposes instead of using hypothetical scenarios like in this study. Furthermore, combined studies would be interesting to consider existing fear toward AI-based HIS. Those studies could then investigate how AI-based HIS perform while using augmented reality to reduce pain as proposed by Ahmadpour et al [37] and Spiegel [38] simultaneously.

We investigated short- and long-term medical treatments with AI-based HIS using an integrated theoretical model to predict human behavior. To do so, we draw on theories from health care, information systems, and psychology. On the basis of the high explanation powers and significant variables that we determined, we conclude that our integrated theoretical model is sufficient for such investigations. Furthermore, we answer our research question by determining several factors that significantly influence the intention to use AI-based HIS for both adaptive and nonadaptive appraisals. Hence, individuals indeed balance between adaptive and nonadaptive appraisals when determining the intention to use new health care inventions. While nonadaptive appraisals are important when the treatment is known, adaptive appraisals are mandatory when the treatment is unknown. As key takeaways, perceived efficacy and attitude (HIS) are important factors to determine the intention to use AI-based HIS. Hence, we conclude that future AI-based HIS developments and research should specifically address perceived efficacy as well as attitude (HIS) of new AI-based HIS. Furthermore, our integrated model should be used to investigate the intention to use AI-based HIS in the future. Ultimately, we derive theoretical and practical implications, show future research directions, and discuss the limitations of our study that we have presented in the previous sections.