Getting the most out of organizational safety work: Accessible considerations for integration of artificial intelligence

Despite best efforts, preventable medical harm still impacts the lives of patients, families, and healthcare workers every day. Decades of research into the root causes of patient safety events, and how to minimize recurrence of these events, support the importance of system-based recommendations rather than solutions targeted at the individual. ¹ Yet, recent works have suggested that our current safety performance has not improved from our performance 20 years ago.^2,3

New safety theories and principles that emphasize system-level solutions rather than individual accountability have emerged, but it has been difficult to move this philosophy from paper to practice. The focus on individual accountability can be observed in safety investigations across many organizations, in which the investigation team asks about the knowledge or training level of the healthcare worker involved in an effort to weed out the “weakest link.” This widespread focus on the individual is reinforced in the recommendations produced from safety investigations, such as recommendations that require re-education of all staff members. While we verbally promote a culture of safety focused on improving the system, rather than blaming the individual, our actions sometimes counter our words. Safety professionals are trained in systems thinking, but it is not uncommon that interventions post-investigation are limited by resource constraints associated with aspects of event review, analysis, reporting, and follow-up. ⁴

The advent of artificial intelligence (AI) has stimulated nearly all sectors to think differently about their current approaches to work and future strategies to maintain a competitive edge. In healthcare, both generative and predictive AI are already being used in many areas related to safety, such as supporting clinical decision-making and improving sepsis detection,^5,6 though the extent to which AI can improve patient safety more broadly is unclear. ⁷

Despite the boom in AI activity in areas across the healthcare space and beyond, there has been little discussion of AI in the work of patient safety at the organizational level. AI tools, such as ChatGPT, may complement our organizational safety work by stimulating further inquiry into system-based approaches and giving pause to the tendency to accept less effective interventions. This viewpoint aims to highlight areas of organizational safety work in which the integration of relatively accessible AI tools may advance practice.

It is important to preface this viewpoint by stating the importance of patient confidentiality. Any AI integration would necessitate careful consideration of organizational policies related to patient confidentiality. There is a need for a greater understanding of how AI tools can be used with privileged and confidential information.

Incident reporting

Incident reporting remains a common mechanism by which organizations identify system risks and patient harm events. ⁸ However, a well-evidenced barrier to incident reporting is the time required to submit an incident report. ⁹ When reports are submitted, organizational safety professionals maintain the responsibility of reviewing all reports and making a judgment call on which events to prioritize without the advantage of a holistic depiction of all system-based risks. In contexts that rely on manually submitted incident reports, AI can ease the reporting process for the reporter automating pre-filled fields or using predictive text. Leveraging AI in existing trigger tool-based adverse event detection systems in electronic health records ¹⁰ may further minimize the manual work of submitting incident reports. Using AI in the context of manual or automated incident reporting can enable a more accurate depiction of the system risks, ultimately helping safety professionals develop a more proportionate response to events. Additionally, its use may enable more timely recognition of themes and trends that have been reported throughout the facility, rather than reports exclusive to one area.

Incident analysis and investigation

Incident analysis and investigation are areas that have much to gain from AI integration, both in terms of facilitating the shift to systems thinking and minimizing the administrative burden that accompanies safety investigations. First, the use of tools such as ChatGPT can facilitate structured extraction of key information from extensive documentation, verbose incident reports, or considerable meeting minutes and transform this information into visualizations or process maps that can be edited and critiqued during multidisciplinary review. Similarly, at the end of a multidisciplinary safety investigation, administrative burden can be further alleviated by using ChatGPT and other AI tools to generate a draft of the final summary report that can be reviewed and finalized by the multidisciplinary team. In the midst of the safety investigation, AI-based tools can stimulate the investigation team to expand the analysis to consider all potential system-based contributing factors, which may provide just enough of a prompt to shift the nature of the discussion. For example, investigation teams might leverage AI as “an extra pair of eyes” to assess the extent to which the identified contributing factors are too targeted toward the individual and should be revisited. Using AI in this way as a third-party reviewer may prompt a more rigorous discussion of how system factors influence individual actions. Finally, ChatGPT can be integrated with evidence-based best practices from the literature to inform the analysis and ensure that the ultimate recommendations align with best practice standards.

However, AI tools are only as good as the material they draw from. As described by Gandhi et al., ¹¹ there may be disparities in the number of events reported by race and ethnicity, demographic data that can be used to identify disparities is often not available in incident reports at all, and health workers themselves may be more or less likely to report incidents based on demographic variables. Without considerable attention to the biases in the sources of information and evidence-based best practices themselves, the use of AI will only perpetuate existing disparities, cognitive biases, and differential treatment.

Incident follow up

It is not uncommon that recommendations from safety investigations are not implemented. ¹² Though resource constraints contribute to this loss, another contributing factor is the follow-up on the actions identified. Currently, the onus of responsibility for follow-up accountability often falls on safety professionals, which can consume a significant amount of time that could be allocated elsewhere. Integration of ChatGPT and other AI tools with organizational platforms for communication, such as email, messaging, and dashboards, could enable safety professionals to automate follow-up on actions identified in safety investigations and hold action item owners accountable without the administrative burden. Considerations for doing so should be informed by Human Factors Science and may include appropriate timing of notifications, adequate context accompanying the notification, and proportionate frequency of notifications based on the importance of the action. However, particular attention should be given to ensure the automated follow-up is sensitive to the nature of the case and does not cause undue emotional stress: the follow-up after a particularly difficult patient mortality will require a different approach than the follow-up about hospital room cleanliness. With any degree of automation, alerts should be tested in their intended environment to minimize over- and under-alerting. In this respect, AI may also be able to facilitate alert fatigue-related data collection to identify trends in and inform decision-making related to over- and under-alerting.

Analysis of trends in historical events

Though safety events happen rarely, when they do happen, they are typically severe. However, there are often indicators of risks or precursor events that do not result in harm to the patient, but highlight key deficiencies that, if not addressed, may result in future harm. Identification and interpretation of weak indicators of risk are aligned with the current safety literature and are discussed as the next step in the future of safety. However, its translation into practice has been minimal. Beyond manual review and organizational memory of previous events, there are few mechanisms for trending these risks, let alone identifying trends that may inform Haddon matrix-like preventive strategies. ¹³

Integration of AI into incident reporting systems can facilitate the analysis of trends in historical events with little manual labor. Ultimately, this analysis would enable more informed decision-making on how to focus safety efforts, particularly those that haven’t yet resulted in patient harm but have significant potential of doing so in the future. If responses to the reported harm event are also documented in the incident reporting system, AI may better inform to what extent the previous attempts to mitigate were effective.

Policy interpretation

It is not uncommon for healthcare organizations to have hundreds of policies that are multiple pages each. However, there is mounting evidence of the difference between work as imagined, as written in policies, and work as done, as performed on the frontline.^14,15 Too often, our policies and education are physically inaccessible to the frontline, are not congruent with daily work, or are too complex for interpretation, all of which perpetuate safety risks. Integration of tools such as ChatGPT via chat-based interfaces can enable easier interpretation of policies in frontline work, ultimately bridging the gap between written policies and daily work. The interface would need to both be accessible to the frontline and integrated into the organization's policy library to ensure that overlapping policies can be reconciled for the user. AI tools may be able to provide examples for the end user based on the policies informing its algorithm. A further step for AI-based tools may be to identify gaps or contradictory information in existing policies. However, with increased AI integration into policy implementation and education, it may be more difficult to intercept an error or misinterpretation of the policy before it becomes widespread.

Conclusion

The integration of AI-based tools into healthcare organizations holds promise for automating incident reporting, facilitating system-focused investigations, and improving policy interpretation. These tools are becoming increasingly accessible and offer an opportunity to offload administrative burdens to ensure quality and safety professionals are practicing at the top of their licenses. However, AI models will have been trained on what has been done in the past, which is not necessarily “best practice” and there is much work to be done to improve the quality of the data AI tools will be trained on. Embracing these advancements is crucial to bridge the gap between safety theory and practice but should be done so with consideration of potential foreseen and unforeseen risks to safety, health equity, and organizational effectiveness.