Staff resistance to any change can often result from a legacy of previous leadership failures. This is especially true when it comes to new or novel technologies. We heard this loud and clear at one of our roundtables, where a hospital CTO said, “For most staff, technology isn’t the solution, it’s the problem. And it’s because of all the bad tech that’s been forced on them in previous years.” Lack of trust in leadership can also manifest as fear or suspicion of technologies (particularly more innovative models that feature audio or video components, such as OR black boxes) being used punitively.  

To address this: health systems must ensure that new solutions or processes enhance clinical practice. Leaders must engage clinicians in the process of developing and implementing new solutions to ensure that they actually make clinicians’ jobs easier and improve patient care. Some leading organizations are developing frameworks to ensure new technologies maximize clinician time spent on patient-facing care, and forecasting the impact on patient-facing time any new technology may cause.

As the AI adage goes, “garbage in, garbage out.” This came up at our roundtables when discussing newer tools. Clinicians reported that one of the biggest barriers to the uptake of new solutions or processes for surgical care, specifically ML-enabled technologies, is mistrust in the quality of the data or lack of context in that data used to develop and train these models. An example here is an Australian health system purchasing a product trained with US DRG code or billing data—the model will probably not work out-of-the-box at the local level for the buyer, and clinicians know this.

To address this: health systems should leverage home-grown data to train new models. Leading organizations are using internal data to develop home-grown models or train imported models. They are also starting on non-clinical support tools before diving into clinical – and potentially thorny – use cases. One example is home-grown OR scheduling optimization tools. These tools, trained on past case and performance data for each surgeon, can optimize a day’s OR schedule blocks based on clinical variables (such as case complexity, co-morbidities, type of procedure, anesthesia speed) and performance (overrun rate, complications, tools used, LOS) that their own clinicians exhibit. Health systems can also partner with tech vendors to train foreign tools on local data to create and commercialize market-specific versions of products in their country or market. 

Another challenge impeding the adoption of ML-enabled innovations in surgical care is clinicians’ inability to see into a tool’s internal mechanisms. Essentially, if a tool doesn’t “show its homework,” users may be skeptical of its outputs, particularly if the output is prescriptive. Context is everything.

To address this: health systems should implement models that contextualize data. Some organizations are developing or implementing models that improve transparency by “showing the work” behind their outputs. They are creating ML-enabled risk prediction and clinical decision support tools that pinpoint which clinical variables (inputs) lead to specific recommendations. Other similar models show the relative weights assigned to each clinical variable and how they contributed to the risk score. Such models improve clinicians’ tolerance for and confidence in AI-generated recommendations by making it clear how and why it arrived at its conclusions.