This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Informatics, is properly cited. The complete bibliographic information, a link to the original publication on https://medinform.jmir.org/, as well as this copyright and license information must be included.
Artificial intelligence (AI) has shown promising results in various fields of medicine. It has the potential to facilitate shared decision making (SDM). However, there is no comprehensive mapping of how AI may be used for SDM.
We aimed to identify and evaluate published studies that have tested or implemented AI to facilitate SDM.
We performed a scoping review informed by the methodological framework proposed by Levac et al, modifications to the original Arksey and O'Malley framework of a scoping review, and the Joanna Briggs Institute scoping review framework. We reported our results based on the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) reporting guideline. At the identification stage, an information specialist performed a comprehensive search of 6 electronic databases from their inception to May 2021. The inclusion criteria were: all populations; all AI interventions that were used to facilitate SDM, and if the AI intervention was not used for the decision-making point in SDM, it was excluded; any outcome related to patients, health care providers, or health care systems; studies in any health care setting, only studies published in the English language, and all study types. Overall, 2 reviewers independently performed the study selection process and extracted data. Any disagreements were resolved by a third reviewer. A descriptive analysis was performed.
The search process yielded 1445 records. After removing duplicates, 894 documents were screened, and 6 peer-reviewed publications met our inclusion criteria. Overall, 2 of them were conducted in North America, 2 in Europe, 1 in Australia, and 1 in Asia. Most articles were published after 2017. Overall, 3 articles focused on primary care, and 3 articles focused on secondary care. All studies used machine learning methods. Moreover, 3 articles included health care providers in the validation stage of the AI intervention, and 1 article included both health care providers and patients in clinical validation, but none of the articles included health care providers or patients in the design and development of the AI intervention. All used AI to support SDM by providing clinical recommendations or predictions.
Evidence of the use of AI in SDM is in its infancy. We found AI supporting SDM in similar ways across the included articles. We observed a lack of emphasis on patients’ values and preferences, as well as poor reporting of AI interventions, resulting in a lack of clarity about different aspects. Little effort was made to address the topics of explainability of AI interventions and to include end-users in the design and development of the interventions. Further efforts are required to strengthen and standardize the use of AI in different steps of SDM and to evaluate its impact on various decisions, populations, and settings.
Shared decision making (SDM) is the process in which patients and health care providers collaborate to make decisions based on the latest medical evidence and patients’ preferences and values [
Elwyn et al [
AI, defined as “computational intelligence” or the “science and engineering of making intelligent machines” [
In the last 2 decades, AI has been applied in various fields, such as telecommunications [
AI has the potential to foster SDM by informing decision-making and allowing health care providers to focus their energy on spending more time with the patient [
The objective of the scoping review is to examine evidence on the use of AI in SDM, namely, to explore what has already been done and what future roles may exist for the use of AI in SDM.
Our specific research questions are as follows: (1) What is the available knowledge on the use of AI interventions for SDM? (2) How is AI being used for the decision-making point of SDM?
The scoping review methodological framework proposed by Levac et al [
We defined the eligibility criteria for our search using the
Any population that provided health care (eg, general practitioners, nurses, social workers, pharmacists, and public health practitioners) and any individual who received care (eg, patients and their families and caregivers) were included.
Any AI intervention implemented or tested during an SDM process in a clinical context was included in the study. AI was defined according to the definition provided by McCarthy [
No limitation.
Any outcome related to patients, health care providers, or health care systems were included in this study.
Studies in any health care setting (eg, primary care and secondary care); all studies using qualitative, quantitative, and mixed methods designs; and only studies published in the English language were included. Reviews, opinion pieces, editorials, comments, news articles, letters, and conference abstracts were excluded.
A comprehensive literature search was designed and conducted by an experienced information specialist in consultation with the research team. The seed articles were identified by experts on the team, and the final search strategy was reviewed by the lead author. The process of the literature search was iterative. The following six electronic databases were searched from their inception to May 2021: MEDLINE (Ovid), EMBASE (Ovid), Web of Science Core Collection, CINAHL, Cochrane Library (CENTRAL), and IEEE Xplore Digital Library. The reference lists of the included studies were searched manually. Retrieved records were managed with EndNote X9.2 (Clarivate) and imported into the DistillerSR review software (Evidence Partners) to facilitate the selection process. The final search strategies and key terms for each database are available upon request.
We removed duplicates and then applied the inclusion criteria for level 1 (title and abstract) and level 2 (full text) screening using a standardized inclusion criteria grid. A pilot test of 55 studies (12% of the total 458 citations) for level 1 screening was conducted. Once familiar with the literature of interest, we modified the a priori eligibility criteria to adjust our study selection where necessary. Subsequently, 2 reviewers (PG, MC, and YH) independently screened the titles and abstracts. The reasons for exclusion were recorded for full-text selection. Any disagreements regarding study inclusion were resolved by a third reviewer (SAR).
A data extraction form was drafted and finalized with feedback from the team members. Elements for data extraction included study characteristics (eg, year published, country of the corresponding author, and study setting), characteristics of the AI intervention (eg, purpose of the intervention, methods/techniques used, data sources, and performance), involvement of end users in the development of the intervention (eg, health care providers and patients), aspects of the AI intervention (eg, explainability of AI and reproducibility of intervention), whether AI was implemented or tested, how the AI intervention was used for decision-making in SDM, and outcomes (eg, related to patients, health care providers, and health care systems). A total of 2 reviewers (YH, PG, and MC) independently extracted relevant data from each included study. All data were verified by a third reviewer (SAR).
In alignment with the proposed framework for methodological guidance in scoping reviews, we did not conduct a quality appraisal. Critical appraisal in scoping reviews is not considered mandatory [
We summarized our findings using descriptive statistics and performed a narrative synthesis describing the characteristics of the AI intervention, whether end users were involved in the development and/or its validation, how the AI intervention supported the decision point of SDM, and what the outcomes were if it was implemented in a clinical setting. We informed our synthesis through the work and toolkits published by Popay et al [
The results were provided to the team members for their feedback. Study updates were also provided to the researchers and health care providers during 2 workshops led by the first author (SAR) at 2 international scientific conferences, that is, the 10th International Shared Decision Making Conference and the annual meeting of the North American Primary Care Research Group.
The search process resulted in 1445 records from the selected electronic databases, 551 of which were excluded as duplicates. Of the remaining 894 studies, we excluded 677 at level 1 screening because they did not meet the inclusion criteria and the remaining 217 underwent full-text review. Citations were manually searched (n=227), of which 3 studies were sought for retrieval and was assessed for eligibility. No eligible studies were found in the reference search. Ultimately, 6 articles met our inclusion criteria (
PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram. Adapted from Page et al [
The number of studies published annually has increased since 2017, with the majority conducted in North America and Europe. The distribution and publication dates of the included studies are shown in
Years of publication and countries where studies are outlined in the included papers.
In
Characteristics of artificial intelligence (AI) interventions.
Study | AI method | Data set and its characteristics | Performance |
Frize et al [ |
Machine learning, artificial neural networks, and case-based reasoning |
Not provided |
Not provided |
Wang et al [ |
Machine learning, multilabel classification methods, k-nearest neighbors, and random k-label sets |
Electronic health records 2542 patients 65.6% male, 34.4% female Mean age 66.46 (SD 13.81) years 70% of this was used for training, and 30% was used for testing |
Performance accuracy of 0.76 |
Twiggs et al [ |
Machine learning, Bayesian belief network, and Bayes network |
Data from the National Institutes of Health Osteoarthritis Initiative 330 patients, between the ages of 45 and 79 years, have undergone total knee arthroplasty |
Not provided |
Jayakumar et al [ |
Machine learning (type not specified) |
Not provided |
Not provided |
Kökciyan et ala [ |
Metalevel argumentation frameworks |
Not provided |
Not provided |
aThis refers to both articles describing the system developed by Kökciyan et al [
Of the included articles, all used machine learning as the type of AI. Only 2 articles presented information on the data set used to develop the AI intervention [
Most of the included articles (n=4) did not report on the data set used to develop the AI intervention; among those that did (n=2), only 1 reported on the sex distribution of the patient data [
Explainable AI is a broad and new domain and is being studied in AI. In general, we can consider explainability throughout AI development: (1)
In health care, explainability and interpretability are required for patients and health care providers to understand why AI interventions produce a certain prediction or suggestion and to trust this output [
Frize et al [
Wang et al [
Twiggs et al [
Jayakumar et al [
Kökciyan et al [
In terms of end user (ie, patients and health care providers) involvement in the design, development, and/or validation of AI systems, we found that 3 of the articles [
One of the articles [
In total, 4 of the included articles tested their interventions for usability and acceptability [
All the included articles provided some level of detail related to the population of the data sets that they used to train or test their algorithm. Only 1 article provided a thorough presentation of the population by reporting the sociodemographic characteristics of the participants involved [
Frize et al [
Kökciyan et al [
Wang et al [
Jayakumar et al [
Twiggs et al [
Of the included articles, 3 designed AI interventions for primary care [
Summary of artificial intelligence interventions and how they are being used for decision-making in the included studies.
Study | Setting | Decision-making problem | AIa for decision-making |
Wang et al [ |
Primary care | Knowledge and choices about antihyperglycemic medications | The tool provides patients and health care providers with tailored knowledge and choices about antihyperglycemic medications through the integration of electronic health record data. Patients and physicians can review patients’ conditions more comprehensively and tailor consultations to the patient’s current condition. |
Frize et al [ |
Secondary care | Neonatal intensive care decisions | The tool allows health care providers to predict outcomes in neonatal intensive care and counsel families on the pros and cons of deciding to initiate or withdraw treatment. The tool also promotes parental involvement in the decision-making process. |
Twiggs et al [ |
Secondary care | The decision about total knee arthroplasty | The AI intervention presents end users (patients and surgeons) with interpretable information relating to the risk of no improvement after total knee arthroplasty. This helps them decide whether to proceed with total knee arthroplasty. |
Jayakumar et al [ |
Secondary care | The decision about total knee replacement | AI system provides patients with a personalized outcome report, which is then discussed with the surgeon during decision-making discussions. |
Kökciyan et al [ |
Primary care | The decision about treatment plans and options for stroke survivors | This tool supports the decision-making point by providing an up-to-date view of the patients’ situation based on personalized metrics and provides explanations for its recommendations. |
aAI: artificial intelligence.
bThis refers to both articles describing the system developed by Kökciyan et al [
The AI intervention by Wang et al [
The AI intervention by Frize et al [
The tool presented by Twiggs et al [
The intervention by Jayakumar et al [
The CONSULT system by Kökciyan et al [
We conducted a scoping review as a first step toward a comprehensive overview of the literature on the use of AI in SDM. This overview provides a basis for future systematic review. The results of our study lead us to make the following observations.
The included articles presented AI interventions used for decision-making during SDM in similar ways. Within the included articles, AI interventions were specifically applied to predict outcomes of clinical significance and for clinical recommendations. The decision-making step can benefit from AI interventions because AI can present a comprehensive and personalized list of treatment options, as well as risks and benefits, thus increasing the amount of knowledge related to the condition, treatment, side effects, risks, and outcomes. AI models are capable of learning and processing all information related to a patient’s care and can generate evidence-based recommendations to support SDM [
The decision-making step is a core step of SDM, in which patient–health care provider interaction is essential and should remain independent of and unrestrained by AI intervention. Patient–health care provider relationships are based on responsibilities that provide a foundation for the relationship to grow. Despite acknowledging the benefits AI may have on facilitating SDM, patients continue to expect their health care provider to retain final discretion over treatment plans and monitor their care, as well as to adapt any contribution from the AI intervention to their unique situation [
AI interventions can open up more time for health care providers to spend connecting with their patients; however, they may place the health care provider in a mediator-like role, in which they will be responsible for explaining the AI output to their patients. This can be difficult to achieve, especially when a lack of interpretability and explainability may exist in certain AI models, such as deep learning. This lack of interpretability and explainability can result in a lack of trust and decisional delay or conflict consequently, which are factors that SDM aims to resolve [
One of the principal challenges in the incorporation of modern AI interventions into health care is explainability and interpretability. This refers to the insight an AI intervention gives to clarify its function to an audience; that is,
Despite the promising performance of AI, its implementation in clinical practice remains challenging. Trust in AI is one of the main barriers to its adoption in clinical practice [
In our review, 2 of the included articles [
Moreover, the level of understanding of the explainability and interpretability of AI tools might differ for various stakeholders. For instance, an AI expert trained in this field can understand and interpret the reasoning behind an AI algorithm better and quicker than a nonexpert in AI. Therefore, health care providers and patient education about AI can lead to a better understanding of the algorithm, which leads to a better understanding of the explainability of an AI intervention. In brief, end users’ understanding of the predictions/decisions made by the AI intervention, as well as increased explainability and interpretability of the AI tool, can increase end-user
A lack of trustworthiness is one of the many bioethical barriers that may arise when implementing an AI intervention in health care and SDM; therefore, improving AI literacy in both patients and health care providers, as well as increasing the explainability and interpretability of AI systems, trust can be increased. In addition, there is a discrepancy in the literature regarding the level of explainability required within the health care setting to ensure a proper understanding of and trust in the outcomes provided by the algorithm [
Of the included articles, 3 [
Further efforts are needed, both from the AI and SDM communities, to include health care providers and patients (as end users of the developed AI systems) in the design, development, validation, and implementation of AI-SDM tools. SDM is the core of patient-centered care; thus, patient values and preferences need to be considered in every step defining the process. Ethicists argue that by not using patient preferences or values as input or influencing the output, but rather leaving the
Thus, to ensure that SDM fundamentally occurs when AI interventions are introduced, patient preferences must be incorporated into the design. Termed
In our review, we observed poor reporting of AI interventions in the included studies. Studies that report AI interventions should use validated frameworks and guidelines to report their results. Transparent and complete reporting of AI interventions supporting SDM is important for detecting errors and potential biases and evaluating the usefulness of the intervention [
None of the articles included in this review mentioned adhering to a specific reporting framework or considered reproducibility. This resulted in a lack of clarity in the included articles regarding different aspects, including whether the training data set was representative, how the potential bias (eg, representativeness and algorithmic biases) and missing data were considered, how AI had been used in the clinical setting, and what were the outcomes resulting from AI implementation. In fact, only 1 article [
We did not conduct a quality appraisal of the included articles, although it is not common, nor is it required to include within a scoping review. However, our review sheds light on this important area, and there are some areas for improvement. Our inclusion criteria were quite strict, and only included articles in which AI intervention was used to support the decision-making point in SDM. Therefore, we may have missed work related to other aspects of SDM. Further systematic reviews may be needed in this area to ensure that the results of this review can be applied in policy and practice.
In this scoping review, we demonstrated the extent and variety of AI systems being tested and implemented in SDM, showed that this field is expanding, and highlighted that knowledge gaps remain and should be prioritized in future studies. Our findings suggest that existing evidence on the use of AI to support SDM is in its infancy. The low number of included studies shows that not much research has been conducted to test, implement, and evaluate the impact of AI on SDM. Future research is required to strengthen and standardize the use of AI intervention in different steps of SDM and to evaluate its impact on particular decisions, populations, and settings. Greater focus and effort from the research community needs to be made on addressing the aspects of explainability, interpretability, reproducibility, and human-centered AI, especially when developing an intervention of their own. Finally, future research should further investigate which SDM steps will benefit most from what type of AI and how AI interventions can be applied to enforce the patient–health care provider relationship.
PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist containing the page number where each reporting criterion is addressed.
Detailed data extraction table.
artificial intelligence
electronic health record
Preferred Reporting Items for Systematic reviews and Meta-Analysis extension for Scoping Reviews
shared decision making
Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis
This study was funded by a start-up fund from McGill University (principal investigator: SAR). The authors would like to acknowledge this support. SAR receives salary support, that is, Research Scholar Junior 1 Career Development Award, from the Fonds de Recherche du Québec-Santé, and her research program is supported by the Natural Sciences and Engineering Research Council (Discovery Grant 2020-05246). FL is tier 1 Canada Research Chair in Shared Decision-Making and Knowledge Translation. The authors thank Milad Ghanbari, Sara Makaremi, and Stewart McLennan for their contribution to this work. The authors also thank the Quebec SPOR SUPPORT (Support for People and Patient-Oriented Research and Trials) Unit for their methodological support.
The authors have reported the contributions according to the Contributor Roles Taxonomy. SAR and PP contributed to conceptualization. SAR, RG, PP, HTVZ, and GG contributed to the methodology. SAR and MC contributed to data curation. SAR, YH, PG, and MC contributed to the formal analysis (see the Acknowledgments section). SAR contributed to funding acquisition, project administration, and resources. SAR, YH, and GG contributed to the investigation. SAR and MC wrote the original draft of this paper. SAR, YH, PG, MC, RG, GG, HTVZ, FL, PP, and DP contributed to reviewing and editing the article.
None declared.