Patients' Attitudes Towards the Use of AI-Based Decision Aids for Breast Cancer Treatment Decision-Making: A Qualitative Study

doi:10.21203/rs.3.rs-3166678/v1

Objectives:

While AI has the potential to transform cancer care, there has been limited progress in incorporating AI tools into clinical practice. As healthcare providers work towards enhancing patient satisfaction and care quality, understanding patients' attitudes towards AI is crucial to facilitate the adoption of these tools in clinical settings. Despite this, few studies have explored patients' views on AI-based decision aids. The aim of this research is to explore the perceptions of cancer patients towards the use of AI-powered decision aids in medical decision-making.

Methods:
To explore the patient perspective on AI-based decision aids, the study conducted 12 semi-structured interviews with former breast cancer patients recruited through the Dutch Breast Cancer Association (BVN). The interviews covered a range of topics such as treatment recommendations, side effect prediction, survival, and recurrence. After transcription, the interviews were analyzed using thematic analysis to identify recurring themes and relevant quotes associated with each theme. The study analyzed the patients' responses in three primary domains: their familiarity with AI, the use of AI in various scenarios related to outcomes, and a comparison of AI and MD.

Results:

Patients' familiarity with AI was found to vary depending on their demographics, with younger and highly educated patients demonstrating a better understanding of AI. Generally, patients had a positive attitude towards AI when used for less critical scenarios such as side effects and treatment recommendations. However, when it came to more severe cases like the prediction of survival and recurrence after treatment, patients were hesitant to trust AI. The participants identified trust as a crucial factor affecting their willingness to use AI, with most of them being positive towards using AI only if they had the chance to consult with an MD. Despite the recognition of the human nature of MDs and their potential to make errors, patients still trusted them more than AI. Participants’ reluctance to accept AI was also partly attributed to the belief that AI cannot consider individuals' unique circumstances, making it more suitable for the average population. Moreover, lack of health literacy and digital skills, as well as ambiguity about accountability in case of errors, were identified as barriers to the adoption of AI in healthcare.

Conclusion:
This qualitative study sheds light on the perceptions of former breast cancer patients in the Netherlands regarding the use of AI in medical decision-making. The findings suggest that patients are generally open to the idea of utilizing AI-based programs to aid in decision-making, but have reservations about using them in high-stakes situations like survival and recurrence predictions. To address these concerns, the study highlights the significance of increasing awareness and understanding of AI's potential in personalized medicine, and creating educational resources for various health areas. Collaboration between healthcare providers, systems, and AI developers is essential, as well as well-defined protocols for accountability and liability in cases of patient harm. Future research should aim to diversify the patient population and provide an accurate representation of the AI program's capabilities to prevent misinterpretation.

Health sciences/Health care/Quality of life

Health sciences/Health care/Patient education

Health sciences/Medical research/Translational research

Artificial intelligence (AI) is defined as a computer program that mimics human intelligence (1). Essentially, AI algorithms are designed to recognize patterns in large amounts of data, learn from that data, and make predictions or decisions based on what they have learned (2). In recent years, AI products have gained the attention of the healthcare sector (3) due to their potential for improving care, reducing human error, and enhancing patient satisfaction (4). One of the key advantages of AI in healthcare is its ability to analyze large amounts of data quickly and accurately (5). Accurate and timely information in oncology can have a profound impact on patient outcomes if it is combined with appropriate clinical decision-making and personalized treatment. (6–8).

Breast cancer is the most common cancer in women worldwide with around two million new cases per year (9). The Netherlands has a breast cancer rate of one in eight women (10). Thanks to public awareness of the disease and medical advances in different stages from screening to follow-up, the mortality rate has decreased (11). However, treatment decision-making can be a challenging task, as there are often trade-offs between the potential benefits and harms of each option. This complexity can make it difficult for patients to choose the optimal treatment for their individual needs (12). To promote patient participation and guide patients toward optimal treatment, patient decision aids (PDA) were developed (13, 14). However, the first generation of decision aids only supports patients with general information and risks, without considering an individual's profile. Therefore, having access to a program that provides personalized information has been receiving an increasing amount of interest, allowing patients to make informed decisions (15, 16). By applying advanced machine learning (ML) techniques to patients' data, an individual's risk of a particular outcome can be estimated (17). In some areas such as the prediction of 3-month mortality in patients with metastatic cancer, AI outperformed medical doctors (MD) (18). Moreover, incorporating patient-reported outcome measures (PROMs) into AI models can provide even more personalized information that considers a patient's unique circumstances and experiences. This can help guide patients toward the best treatment options for them and improve their outcomes (19).

Despite promising results from clinical research, the implementation and adoption of AI models have been slow. A number of factors influence the adoption of AI applications, including patients' acceptance. Patients' perspectives were explored in previous studies and revealed that, in general, patients are positive about using AI but there are some concerns such as the risk of miscommunication and privacy (20). However existing studies on patients' perceptions toward AI are limited to its applications in screening and treatment, therefore how patients perceive AI in different stages of the care path such as for their treatment-decision making has not been studied yet.

In this study, we aimed to investigate the factors that influence patients' perceptions of AI in the context of breast cancer treatment decision-making. We hypothesized that patients' demographics, the scenario being studied, and the way patients compare AI to MDs will affect their perceptions of AI. By exploring these factors, we gain a deeper understanding of patients' attitudes towards AI and provide insights into how AI can be effectively integrated into the breast cancer treatment decision-making process.

Ethics approval and consent to participate

The study adhered to the principles of the Helsinki Declaration and Good Clinical Research Practice (21). Prior to commencing the interviews, all participants provided their signature on a written informed consent form. Approval was granted by the Internal review board of the Department of Radiation Oncology (MAASTRO) and the Medical Ethics Review Committee of Maastricht University Medical Center (MUMC) (Request ID: METC 2022-3124). The inclusion criteria for study participation were defined as being diagnosed with breast cancer and having adequate Dutch language skills. The sole exclusion criterion was patients who were still receiving medical treatment for breast cancer, with the exception of hormone therapy.

Between March and April 2022, former breast cancer patients were recruited with voluntary response sampling via email through the Dutch Breast Cancer Association (Borstkanker Vereniging Nederland; BVN). Before the interview session, patients were asked to answer questions about their demographics, type of cancer, and year of diagnosis (Supplementary file 1).

Patient interviews

Before developing the interview guide, the important themes were identified based on the study's research objectives and literature review. As the primary goal of the study was to investigate the patient perspective on AI-based decision aids, the identified themes were specifically chosen to gather information on the patients' familiarity with AI, their views on using AI in different scenarios related to breast cancer treatment, and their perspectives on using AI independently or in conjunction with their treating MD. These themes were considered crucial to gaining a comprehensive understanding of how patients perceive and approach AI-based decision-making in breast cancer treatment.

We adapted Nelson et al's (22) semi structured interview guide for our study, and through four mock interviews, the interview guide was tested and improved (Supplementary file 2). All the interviews were conducted online using Microsoft Teams and were recorded for subsequent analysis. The interviews lasted approximately one hour and were conducted in Dutch by one researcher. Interviews occurred between March and April2022. The interviews were transcribed verbatim (SONIX (23) and reviewed for accuracy by an independent researcher (EE). A thematic analysis was then conducted by two of the authors (EE and HH) and disagreements were discussed with the third researcher (CO).

Data analysis

Using thematic analysis, important subthemes to participants were identified. The grounded theory approach (24) was adopted to develop the codes. Following the three-step process of Strauss and Corbin we completed open coding, axial coding and selective coding. For qualitative data analysis, we used Microsoft Word and Microsoft Excel and adhered to the COREQ criteria (25).

To identify similarities and differences in participants' opinions, codes were compared with transcripts. Data were grouped into themes when they appeared to represent significant conceptual connections. Concepts related to a single theme were grouped into subthemes when necessary.

Availability of Data and Materials

For this qualitative study, the data consists of the complete transcripts from 12 interviews with participants. In accordance with patient privacy regulations, we are unable to publicly share the interview transcriptions. However, to align with the principles of open science, we have included the coding tree in the supplementary materials 3. Additionally, researchers can send requests for confidential data access to the corresponding author.

In total, thirteen patients responded to the invitation for participation, of which one patient was excluded as she was still undergoing treatment. Twelve patients were interviewed with ages ranging between 51 and 70 years (mean 61.6 ± 6.7). The duration of time since participants were primarily diagnosed with breast cancer was on average 6.5 ± 3.3 years, with a median duration of 6 years. Our findings showed that TNM varied in participants from 0 to 4. In addition, the majority of interviewed patients (n=11) underwent surgery, and some received additional therapies such as hormone therapy. Ten participants had a high education level (college or university), and all were employed in different sectors varying from justice to healthcare (see Table 1 for details). As a result of the participants' elaborations on AI's use in multiple scenarios, three major themes emerged: (i) familiarity with AI (ii) preferences for utilizing AI (iii) comparison of AI and MD. Here, we describe the characteristics of the participants, followed by a detailed discussion of each theme.

Table 1. Summary of participants’ characteristics (n = 12).

Characteristics	Descriptives, n (%)	mean ± SD
Age (years)		61.6 ± 6.7
Time since primary diagnosis (years)		6.5 ± 3.3
Tumor (TNM) stage
0	3 (25)
I	2 (17)
II	3 (25)
III	2 (17)
IV	1 (8)
Unknown	1 (8)
Treatment received ^a
Chemotherapy	7 (58)
Hormone therapy	10 (83)
Immunotherapy	2 (17)
Radiotherapy	9 (75)
Surgery	11 (92)
Educational level
Medium vocational training (MBO)	2 (17)
Higher vocational education (HBO)	6 (50)
Academic education (University)	1 (8)
Ph.D.	3 (25)
Employment sector ^a
Commerce and service	3 (25)
Education, culture and science	2 (17)
Environment and agriculture	1 (8)
Healthcare and wellbeing	7 (58)
Justice, security and public administration	3 (25)
Media and communication	1 (8)

^aThe total number exceeds the participant count, as categories are not exclusive.

Domain 1: Familiarity with AI

The first theme, familiarity with AI, refers to the participants' knowledge and experience with AI technology, which was explored through their knowledge and prior experience with AI systems in different contexts. Although most participants had some degree of awareness regarding AI, a few participants had never heard of AI or were unsure of what it actually is. Of all participant characteristics, we identified two demographic variables (educational level and age) and former exposure to AI as the two main determining factors.

Demographics

Participants who were either highly educated (postdoc, n=3) or young (n=5) provided more detailed descriptions of AI. They characterized it as a prediction model that relies on analyzing data from past patients, or as a tool that can identify connections in data that are not easily discernible by humans.

“I think that based on a lot of data that you have from surveys or from people, from patients in this case, you might be able to make certain predictions with it in the future.’’[A12]

“to me it's the closest thing to how algorithms work. By computer, any kind of computer records human behavior keeps track of what patterns. and the computer is programmed to develop new patterns from that. The first part just is very good programming. And the second part I would call artificial intelligence. It no longer is inspired by people, but it is the computer which programs patterns. “

Former exposure to AI

Participants’ familiarity with AI in healthcare was influenced by their jobs. For instance, A 70-year-old participant with a background in nursing defined AI as a computer algorithm and provided one use case (in radiotherapy).

“I do read from time to time the radiology, for example, they are already using artificial intelligence to assess photos [radiology report/medical images].” [A1]

Half of the participants (n=6) viewed AI as statistical information that is generalized to fit average populations. On the other hand, only a few patients (n=3) could provide examples of personalized outcomes with AI in areas such as social media, suggesting that personalized AI applications are not yet widely understood or experienced.

“What I found annoying is that everything is based on statistics and averages. And no human being is average. So I was also like yeah. What good is all that information anyway? If you're just your own person. And, how reliable is it …” [A12]

“Yeah. That's exactly the downside of artificial intelligence. That it does give advice, but it does not address your personal needs at that moment and your situation. They are all averages.”

Domain 2: Preferences for utilizing AI

The second theme, preferences for utilizing AI, relates to the participants' attitudes and opinions towards the use of AI in various scenarios. In exploring these desirable scenarios, participants expressed greater enthusiasm for using AI applications in scenarios with low impact like predicting side effects (n=10) over situations that could severely impact their lives, such as predicting survival after treatment (n=5) or recurrence (n=4) (see Figure 1).

Low risk Scenarios

In our study, low-risk scenarios consisted of treatment recommendations and predicting side effects after treatment. The majority of the participants (N = 10) were willing to use AI for predicting the risk of side effects, with some citing their prior experience of encountering unexpected complications during treatment as the reason for their willingness to use AI.

“I 'd love to. I do hope it contains a lot of information, because if you mention side effects. Unfortunately I've had so many side effects from everything that every doctor says: you've had bad luck every time! …. I hope artificial intelligence gives a complete answer.” [A11]

“Yeah, what I mentioned of I have so many side effects, and still suffer from those radiation treatments. And that would be really nice, Then you hope that the app has an answer to that, that it's set right. Yes, then you don't have to worry for so long.“ [A3]

Out of 12 patients, seven expressed interest in using AI software for treatment recommendations, while four were uncertain and one participant showed no interest. Upon further analysis, it was found that participants were more inclined to use AI as an additional source of information rather than relying solely on it for decision-making purposes.

“On the one hand, I think it's very positive. Because then you get neutral advice at least. On the other hand, I would still think that I would go for it, but I would also want to talk to a specialist.” [A7]

“What does it add to the oncologist? Look, if he's as good as the oncologist I'd rather have a real human in front of me because it's a pretty intense disease with intense risks. And then when I think back to what you mentioned as examples, just of applications of artificial intelligence, the chat box and ads and stuff. That can be very useful, but you can also end up in a kind of bubble or keep getting the wrong answers because a machine doesn't really understand what you want to ask. And in that case I'd rather have a real person in front of me.” [A12]

High risk scenarios

Out of the 12 interviewees surveyed, around 42% expressed an interest in utilizing AI programs for survival prediction, whereas approximately 33% showed interest in using AI for recurrence prediction. Specifically, those who expressed interest in using AI argued that healthcare providers also utilize the same program and it is vital that patients know their odds. However, they also expressed a desire to have a conversation with their doctor.

“What with or without treatment. What if I just this treatment? What if I...? What are the chances of survival? Yes, I find that quite crucial to be able to make a decision.” [A7]

“I would be very excited… Because that's the question that's always on your mind and they can't say anything about that. But just imagine there was a program that could calculate it properly. I would like that.” [A7]

“I myself would find it very pleasant to have thought about the various possibilities in advance. And I don't see many dangers in that.” [A11]

Patients were negative or doubtful for a variety of reasons. Since the logic behind the predictions is not clear, participants were skeptical about the reliability of the AI program (n=2) or believed it is hard to predict survival or recurrence odds (n=6). Three out of twelve were reluctant to use AI programs since it requires digital skills or medical knowledge. In both scenarios, there were three patients who stated that they do not want to know the odds of their treatment outcomes, regardless of whether AI is involved or not.

“it depends on how such a program functions and on what basis. Yes, and if it's based on those statistics, then I don't trust it. Whether it's favorable or unfavorable doesn't matter that much.” [A12]

“ I don't need to know, I don't value it…. and it has to do with the unpredictable nature of cancer. That has nothing to do with the program.” [A9]

Figure 2 displays the variations in participant attitudes regarding the utilization of AI in medical decision-making across different scenarios. Among the total participants (n=12), only five individuals maintained a consistent attitude in both scenarios, while the remaining seven respondents exhibited shifts in their opinions. Notably, these changes in attitudes were observed even though participants employed similar reasoning. The motivations behind these shifts remain unclear, but it is worth noting that personal experiences may have played a role. For instance, one patient who had previously experienced a recurrence of breast cancer altered their initial negative stance on survival prediction to a positive outlook for disease recurrence. The findings depicted in this figure highlight the dynamic nature of patient attitudes towards AI-based decision-making and suggest that individual experiences may influence their perspectives.

“ [survival rate] But when I first got breast cancer, a year and a half ago, the prognosis was 95 percent will get better from this, that all looked very good. And three months after the oncologist had said that, there was suddenly a metastasis. And I have not even asked about the prognosis… So I don't need to know, I guess. You know, I don't value it anymore because the first time was disappointing.” [A9]

“[for recurrence] that's information you really want to know… I think that would give you peace of mind if you could…..’’ [A9]

Domain 3: Comparison of AI and MD

The third theme focuses specifically on participants' preferences for consulting AI, an MD, or a combination of both, and how these preferences are influenced by the (i) perceived performance, (ii) reliability, (iii) accountability, (iv) disagreement between the AI and MD, (v) interaction, and (vi) collaboration of both.

Perceived performance of the AI model

Perceived performance relates to how effective and successful participants believe the AI and MD to be and how this perception can influence their preferences. We found that patient attitudes toward the performance of AI and MDs differed based on their educational level.

Highly educated participants were more likely than those with lower educational levels to have a nuanced understanding of the differences between the two systems. They were aware that AI utilizes data and statistical analyses to generate predictions, while MDs rely on established protocols and previous experience to make treatment recommendations. However, this heightened level of education did not necessarily translate into greater acceptance of AI within this group.

The program has a lot more information and a lot more data. And the doctor, as far as I understand it, works mainly according to the protocol. So of course he looks at the specific situation, but he just has a protocol next to it, which steps to take. At least I heard the word protocol quite often during that period. And that means that if the protocol changes, the treatment will also change. [A10]

Reliability of the AI model

Reliability concerns the consistency and dependability of both AI and MD, particularly with respect to mistakes, and how this can influence preferences.The majority of participants (n=10) believed that MDs are more likely to make mistakes compared to AI programs. The participants frequently attributed human factors, such as fatigue, as causes of MD mistakes. While AI is often perceived as more accurate and reliable, all participants acknowledged that it is not infallible, due to reasons such as incorrect input data, system development based on erroneous information, and technical errors in the AI algorithm.

“A doctor's a human being. Nothing human is alien to us. And I think doctors are more likely to make mistakes … because of personal circumstances or assumptions or experiences that a doctor has that makes them think” [A7]

“Well, I would say that basically if the computer program works well that it shouldn't make mistakes. But again, it's only as good as the people entering the information. And if there's a malicious person who suddenly enters all kinds of data into that algorithm that's not correct, then of course you do have a very wrong program.’’[A1]

Accountability

Accountability refers to how responsible participants perceive the AI and MD to be for their outcomes. Participants generally believed that MDs are more accountable for their decisions and actions than AI, but accountability for AI is complex and multifaceted. Concerns were raised about the lack of transparency in AI decision-making processes and the difficulty in holding AI systems accountable for errors or biases.

“Can't really pinpoint anyone I think. Was just your trust in that program at that time. And who is responsible is so difficult. Yes, it's not the person who introduced it, it's not the program itself. Wherever work is done, mistakes are made. I find it difficult to really point the finger at someone. It is very, very difficult.“ [A4]

Disagreement between the AI and MD

While both AI and MD can provide advice or recommendations based on their analysis of medical data and information, there may be instances where the recommendations provided by AI and MD differ. Participants expressed that when faced with conflicting advice from the AI program and their MD, decisional conflict arises. They stated that in such instances, they may choose to seek a second opinion or engage in a discussion with their MD to better comprehend the reasoning behind the conflicting opinions

Yeah, but why did you develop that app [AI program]? Yeah. We're not gonna do that. Because then I wouldn't trust the app or the doctor. That sounds strange but then I might still want a second opinion. No, then I do not know in whom I no longer have confidence. That wouldn't be possible. No. Unless the doctor comes up with such good arguments. {A3}

Interaction between the MD and AI

The interaction between a doctor and patient involves communication and information exchange related to the patient's medical condition, treatment options, and healthcare decisions. Most patients (n=7) would prefer to have MDs involved in their medical decision-making process, even when an AI program is available and may outperform the MD. These patients expressed that they placed a high value on human connection, empathy, and the ability to ask questions and receive personalized explanations from an MD.

With all the human failings involved. Why? Because at least in my experience I can use my own brain and ask 'yes but' questions. And if he says, on the basis of that conversation, "Yes, but that would be my advice. Then it has been an interaction. And with the computer you can't do interaction. At the most you can go back to the beginning of the program and fill in some other data. {A11}

Collaboration

Collaboration between AI and MD in healthcare involves utilizing AI systems to assist MDs with various aspects of medical practice, such as treatment planning. This approach takes advantage of the unique strengths of both AI and MD, combining the precision and speed of AI algorithms with the human expertise and judgment of MDs to improve patient care. All the participants agreed that the idea of MD using AI is the optimal approach for patients, as both can complement each other's strengths and weaknesses. In this scenario, patients believe that even if a mistake was made, they had given their best effort, and no one is to blame.

“I think that's going to be the future when I see it like this…. That would put my mind at ease. Then I think, well, all the information that's been put into it brings me this. And the doctor has looked at it and he says well, I can't think of any other ways, this is the right way for you… it's the best of both worlds.” [A8]

This qualitative study was conducted with twelve former breast cancer patients in the Netherlands to explore patients’ perspectives towards AI to assist in decision making. Patients’ responses were analyzed in three main domains: 1) familiarity with AI, 2) Use of AI in scenarios related to outcomes, and 3) comparison of AI and MD. Overall, our participants viewed AI programs as beneficial under certain conditions. In our study, participants’ familiarity with AI varied based on their demographics, i.e. younger, highly educated patients defined AI more precisely. This finding aligns with prior research indicating that young individuals with higher levels of education are more inclined to adopt AI-assistance in healthcare (26).

Participants’ willingness to use AI varied depending on the examined scenario and the experience they had during their care path. In fact, in less impactful situations such as the prediction of side effects, patients were more willing to use AI compared to situations that were a matter of life and death like the prediction of survival after treatment. Among the four scenarios, optimal treatment recommendations by AI gained the most support. On the other hand, predictions of recurrence were the least supported in light of the skepticism of the accuracy of prediction and the unfavorable nature of negative results. These findings align with Sung et al.'s (27) survey on AI utilization in medical decision-making, highlighting greater acceptance in low-risk scenarios. Similarly, Yakar et al. (28) noted contrasting attitudes towards AI in radiology and dermatology versus robotic surgery, possibly influenced by the perception of less invasiveness in image-based specialties.

How patients perceive AI also depends on the role AI plays in the decision-making process. To this end, a comparison was made between the reaction of patients in different scenarios: AI used alone, MD alone, and MD with AI. Our findings showed that trust is a key factor in all scenarios that affect patients' desire to use AI. Generally, patients' trust in their physician is a crucial component of patient-centered care and has a significant impact on their quality of life and outcome. In previous research, breast cancer patients expressed high trust (83%) in their doctors for treatment planning, and both patients and clinicians regarded trust as essential to working together (29). In our study, most participants were positive toward using AI only if they have the opportunity to consult with MD. In the same way, Brennan et al,. (30) concluded that while AI offers high accuracy in healthcare, human relationships and intuition cannot be replaced. Our findings align with the doctor-in-the-loop AI framework that combines AI algorithms and medical expertise to enable more accurate and personalized diagnoses and treatments. The AI system provides recommendations to the doctor (and possibly the patient), who can modify them based on their medical knowledge (31).

Assessing patients' perceptions and reactions to doctors' mistakes compared to AI errors reveals that patients are aware that the human nature of MDs makes them more likely to make errors than AI. Indeed, medical errors, stemming from inadequate training, busy schedules, and tiredness, are a known contributing factor to the third leading cause of death (32). Nevertheless, patients continue to exhibit lower levels of trust in AI (33). It was reported by Yokoi, et al., (33) that patients prefer human doctors and are less likely to trust AI's advice even when it outperforms MDs. Whether patients' willingness to use AI over MDs is affected by information about the accuracy of AI programs was examined in our study. Our findings aligned with Yokoi, et al., where the most patients stated MD’s approval is necessary even if the AI outperforms human doctors. Reluctance to accept AI can partly be attributed to the belief that AI cannot consider people individually (34). Likewise, one of our key findings was that many participants thought that AI's advice is not specific to individual patients, so results are suitable for the average population rather than those of single individuals. As this is not correct, it is necessary to increase awareness and knowledge regarding AI's applications in personalized medicine. Despite this, participants' support for the AI program was highest when it was used in combination with expert MD advice. This is consistent with the findings of Soellner et al., about treatment recommendations for skin cancer where patients were more receptive to MD with AI compared to using it alone (35).

Patients worry about the use of AI not only due to its limitations, but also because of other issues, such as inadequate health literacy and digital skills. Additionally, there is a lack of clarity about who is accountable in case of errors, which further adds to their concerns. The mean age of our participants is 61.6 years which is close to the average age of breast cancer patients (age=63) (36), and their ability to use these aids effectively may vary depending on individual factors such as cognitive abilities, health literacy, and familiarity with technology. It is important to consider the specific needs and capabilities of older individuals when designing and implementing patient decision aids, ensuring they are accessible, user-friendly, and tailored to their unique requirements. A long-term plan such as developing educational materials and making terminology lists for each health area is recommended to overcome this (37). Moreover, our study participants emphasized the importance of collaborative involvement among patients, doctors, healthcare systems, and AI developers in the development of AI programs. This collaborative process brings attention to the unresolved issue of accountability in cases of AI-related errors, as previously discussed (38). This finding underscores the need for further exploration and establishment of clear frameworks to address accountability in the context of AI development and implementation in healthcare settings. Currently, Despite the fact that the European General Data Protection Regulation and Medical Device Regulation (39) cover AI systems, there still remains a lack of clear procedures regarding liability and accountability for injured patients (40). To be widely adopted, AI programs need the legal system to strike a balance between accountability, and the safety of patients.

This study has some limitations. While the recruiting procedure was performed on a random subset of former breast cancer patients through BVN, mostly highly educated women volunteered to participate. This lack of diversity in education level may influence the generalizability of our results to breast cancer patients. Future studies are needed to recruit patients with different demographics. Secondly, the interview was conducted by providing hypothetical scenarios, however, how patients conceptualize the AI program might be different than the intended use case. For future studies, it is suggested to show an AI program that represents the functionality of the AI program, avoiding misinterpretation.

To the best of our knowledge this is the first qualitative study that explored breast cancer patients’ views toward AI in treatment decision-making. Our results indicated that former breast cancer patients' willingness of using AI programs depends on the degree of familiarity with AI, the scenario under study, and the role that AI will play in their decision-making process. While AI programs are interesting for treatment recommendations and side effects, patients are hesitant to use them for vital decisions such as survival and recurrence. Additionally, patients are more inclined to accept AI systems when they are employed by their doctors rather than used independently, even when AI is superior to MD.

Čerka P, Grigienė J, Sirbikytė G. Liability for damages caused by artificial intelligence. Comput Law Secur Rev. 2015;31(3):376–89.
Ghazal TM, Hasan MK, Alshurideh MT, Alzoubi HM, Ahmad M, Akbar SS, et al. IoT for Smart Cities: Machine Learning Approaches in Smart Healthcare—A Review. Future Internet. 2021;13(8):218.
Buch VH, Ahmed I, Maruthappu M. Artificial intelligence in medicine: current trends and future possibilities. Br J Gen Pract J R Coll Gen Pract. 2018;68(668):143–4.
Haleem A, Vaishya R, Javaid M, Khan IH. Artificial Intelligence (AI) applications in orthopaedics: An innovative technology to embrace. J Clin Orthop Trauma. 2020;11(Suppl 1):S80–1.
Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56.
Aapro M, Arends J, Bozzetti F, Fearon K, Grunberg SM, Herrstedt J, et al. Early recognition of malnutrition and cachexia in the cancer patient: a position paper of a European School of Oncology Task Force. Ann Oncol. 2014;25(8):1492–9.
Every month delayed in cancer treatment can raise risk of death by around 10% | BMJ [Internet]. [cited 2023 Mar 21]. Available from: https://www.bmj.com/company/newsroom/every-month-delayed-in-cancer-treatment-can-raise-risk-of-death-by-around-10/
Sadik H, Pritchard D, Keeling DM, Policht F, Riccelli P, Stone G, et al. Impact of Clinical Practice Gaps on the Implementation of Personalized Medicine in Advanced Non–Small-Cell Lung Cancer. JCO Precis Oncol. 2022;(6):e2200246.
Arnold M, Morgan E, Rumgay H, Mafra A, Singh D, Laversanne M, et al. Current and future burden of breast cancer: Global statistics for 2020 and 2040. The Breast. 2022;66:15–23.
Breast cancer in the Netherlands | RIVM [Internet]. [cited 2023 Mar 21]. Available from: https://www.rivm.nl/en/breast-cancer-screening-programme/breast-cancer-in-netherlands
Narod SA, Iqbal J, Miller AB. Why have breast cancer mortality rates declined? J Cancer Policy. 2015;5:8–17.
Krist AH, Tong ST, Aycock RA, Longo DR. Engaging patients in decision-making and behavior change to promote prevention. Stud Health Technol Inform. 2017;240:284–302.
Ankolekar A, Dekker A, Fijten R, Berlanga A. The Benefits and Challenges of Using Patient Decision Aids to Support Shared Decision Making in Health Care. JCO Clin Cancer Inform. 2018;(2):1–10.
Grüne B, Kriegmair MC, Lenhart M, Michel MS, Huber J, Köther AK, et al. Decision Aids for Shared Decision-making in Uro-oncology: A Systematic Review. Eur Urol Focus. 2022;8(3):851–69.
Vromans RD, Tillier CN, Pauws SC, van der Poel HG, van de Poll-Franse LV, Krahmer EJ. Communication, perception, and use of personalized side-effect risks in prostate cancer treatment-decision making: An observational and interview study. Patient Educ Couns. 2022;105(8):2731–9.
Vromans RD, Eenbergen MC van, Geleijnse G, Pauws S, Poll-Franse LV van de, Krahmer EJ. Exploring Cancer Survivor Needs and Preferences for Communicating Personalized Cancer Statistics From Registry Data: Qualitative Multimethod Study. JMIR Cancer. 2021;7(4):e25659.
Hasannejadasl H, Roumen C, van der Poel H, Vanneste B, van Roermund J, Aben K, et al. Development and external validation of multivariate prediction models for erectile dysfunction in men with localized prostate cancer. PLOS ONE. 2023;18(3):e0276815.
Zachariah FJ, Rossi LA, Roberts LM, Bosserman LD. Prospective Comparison of Medical Oncologists and a Machine Learning Model to Predict 3-Month Mortality in Patients With Metastatic Solid Tumors. JAMA Netw Open. 2022;5(5):e2214514.
Verma D, Jansen D, Bach K, Poel M, Mork PJ, d’Hollosy WON. Exploratory application of machine learning methods on patient reported data in the development of supervised models for predicting outcomes. BMC Med Inform Decis Mak. 2022;22(1):227.
Young AT, Amara D, Bhattacharya A, Wei ML. Patient and general public attitudes towards clinical artificial intelligence: a mixed methods systematic review. Lancet Digit Health. 2021;3(9):e599–611.
World Medical Association. World Medical Association Declaration of Helsinki: Ethical Principles for Medical Research Involving Human Subjects. JAMA. 2013;310(20):2191–4.
Nelson CA, Pérez-Chada LM, Creadore A, Li SJ, Lo K, Manjaly P, et al. Patient Perspectives on the Use of Artificial Intelligence for Skin Cancer Screening. JAMA Dermatol. 2020;156(5):1–12.
Automatically convert audio and video to text: Fast, Accurate, & Affordable | Sonix [Internet]. [cited 2023 Mar 22]. Available from: https://sonix.ai/
Grounded theory in practice. Thousand Oaks, CA, US: Sage Publications, Inc; 1997. viii, 280 p. (Strauss AL, Corbin JM, editors. Grounded theory in practice).
Consolidated criteria for reporting qualitative research (COREQ): a 32-item checklist for interviews and focus groups | International Journal for Quality in Health Care | Oxford Academic [Internet]. [cited 2023 Mar 22]. Available from: https://academic.oup.com/intqhc/article/19/6/349/1791966
York T, Jenney H, Jones G. Clinician and computer: a study on patient perceptions of artificial intelligence in skeletal radiography. BMJ Health Care Inform. 2020;27(3):e100233.
Sung J, Portales-Casamar E, Görges M. Perceptions of expert and lay users on trust in the use of artificial intelligence for medical decision-making and risk prediction. 2020.
Yakar D, Ongena YP, Kwee TC, Haan M. Do People Favor Artificial Intelligence Over Physicians? A Survey Among the General Population and Their View on Artificial Intelligence in Medicine. Value Health. 2022;25(3):374–81.
Lansdown M, Martin L, Fallowfield L. Patient–physician interactions during early breast-cancer treatment: results from an international online survey. Curr Med Res Opin. 2008;24(7):1891–904.
Brennan HL, Kirby SD. Barriers of artificial intelligence implementation in the diagnosis of obstructive sleep apnea. J Otolaryngol - Head Neck Surg. 2022;51(1):16.
Holzinger A, Plass M, Holzinger K, Crişan GC, Pintea CM, Palade V. Towards interactive Machine Learning (iML): Applying Ant Colony Algorithms to Solve the Traveling Salesman Problem with the Human-in-the-Loop Approach. In: Buccafurri F, Holzinger A, Kieseberg P, Tjoa AM, Weippl E, editors. Availability, Reliability, and Security in Information Systems. Cham: Springer International Publishing; 2016. p. 81–95. (Lecture Notes in Computer Science).
Makary MA, Daniel M. Medical error—the third leading cause of death in the US. BMJ. 2016;353:i2139.
Yokoi R, Eguchi Y, Fujita T, Nakayachi K. Artificial Intelligence Is Trusted Less than a Doctor in Medical Treatment Decisions: Influence of Perceived Care and Value Similarity. Int J Human–Computer Interact. 2021;37(10):981–90.
Longoni C, Bonezzi A, Morewedge CK. Resistance to Medical Artificial Intelligence. J Consum Res. 2019;46(4):629–50.
Soellner M, Koenigstorfer J. Compliance with medical recommendations depending on the use of artificial intelligence as a diagnostic method. BMC Med Inform Decis Mak. 2021;21(1):236.
Muss HB. Coming of Age: Breast Cancer in Seniors. The Oncologist. 2011;16(S1):79–87.
Hasannejadasl H, Roumen C, Smit Y, Dekker A, Fijten R. Health Literacy and eHealth: Challenges and Strategies. JCO Clin Cancer Inform. 2022;(6):e2200005.
Ongena YP, Yakar D, Haan M, Kwee TC. Artificial Intelligence in Screening Mammography: A Population Survey of Women’s Preferences. J Am Coll Radiol JACR. 2021;18(1 Pt A):79–86.
Medical devices [Internet]. [cited 2023 Jul 10]. Available from: https://single-market-economy.ec.europa.eu/single-market/european-standards/harmonised-standards/medical-devices_en
Maliha G, Gerke S, Cohen IG, Parikh RB. Artificial Intelligence and Liability in Medicine: Balancing Safety and Innovation. Milbank Q. 2021;99(3):629–47.

No competing interests reported.

Patients' Attitudes Towards the Use of AI-Based Decision Aids for Breast Cancer Treatment Decision-Making: A Qualitative Study

Status:

Version 1

Abstract

Figures

Introduction

Methods

Results