“If you use bad or limited data and human experts don't correct that bad data or classification, AI can generate inaccurate and incorrect recommendations,” Broderick said. “My biggest concern is when AI is trained on bad data and gives answers that can be harmful.”
Researchers must also develop strict protocols and safeguards to keep patient information used to train models private and HIPAA compliant. This could look like an independent third party, such as the American Heart Association, centrally collecting anonymized patient data before feeding it to an AI model, or training the model using only data from each institution before sharing the learned parameters more broadly.
“Protecting patient privacy is a major challenge when using clinical data to train AI in healthcare, and differences in national data-sharing laws make it even more difficult to share even anonymized data between countries,” the co-authors wrote. “New methods of model development are expected to address some of these privacy concerns.”
Once a powerful stroke AI model is developed and validated by humans, potential applications could include improving the identification of potential trial participants, communicating trial designs to patients in common languages, translating trial information into different languages for patients whose first language is not English, and helping identify the best treatment for individual patients, Broderick said.
“We have been talking about precision medicine for some time, and AI is a huge step forward in achieving this,” he said.
In addition to AI, the authors also discussed new clinical trial designs, such as platform trials that can test multiple research questions at once more efficiently and add new questions as old ones are answered. Another major focus going forward is pragmatic clinical trials aimed at evaluating the effectiveness of treatments when introduced into routine clinical care rather than under ideal conditions.
By comparing existing treatments, incorporating trial procedures into normal clinical workflow, and using data from electronic medical records, researchers and organizations can reduce costs and simplify the infrastructure associated with these types of pragmatic trials. A pragmatic design is expected to increase the likelihood that clinical trials will be completed successfully in a timely and cost-effective manner.
Finally, the stroke research community needs greater community and patient engagement. This should include input from front-line healthcare professionals who enroll and treat stroke patients in clinical trials (EDs, transport and receiving facility physicians, study coordinators).
Common goals for clinical trials should be established to minimize the burden of trial participation on patients and researchers, to expand trial participation to community-based settings whenever possible, and to rapidly disseminate trial results to patients, clinicians, and the general public.
“The future is bright and research will be greatly advanced with these new tools,” Broderick said. “At the same time, the real challenge in our time as AI rapidly expands into our daily lives is recognizing accurate data and truth in a sea of words, images, and videos that can be false, harmful, or inaccurate.”
“Fire can burn down a house just as easily as it can warm you or cook a meal,” he continued. “AI is a rapidly growing flashpoint, but we are only just beginning to learn how best to use it safely and wisely.”
