CHICAGO – AI has the potential to drive early and accurate disease detection, personalized and precision medicine, access to care and population health, clinical decision support, operational efficiency, and improved patient outcomes. involvement of clinicians and patients.
But the reality is, “We don’t see these things coming to fruition,” says Sonja Makhni of the Mayo Clinic Platform.
Speaking before a lively forum on AI and machine learning today at HIMSS23, the AI medical director highlighted the challenges of data standardization, drift in machine learning, accountability, and calculation. feasible in clinical practice.
Due to the inconsistent effectiveness, bias, and other challenges of AI, point solutions have limited real-world application, she said. Suboptimal practices and assumptions lead to systematic errors that skew patient populations.
“If we don’t understand what these things are, then we can’t really address the risks that bias creates,” Makhni said.
But at the same time, even in cases where the most efficient algorithm has minimal bias and performs well, “the algorithm will not be accepted unless it is integrated into the workflow in a smart, durable way. stable and effective.”
Healthcare AI must make sense in clinical practice. Makhni said it was “unacceptable and unsustainable” to expect clinicians to deal with too many pop-ups or have to navigate away from their workflow.
The dynamics of change also lead to monitoring and tracking problems for ML models. “There are also clinical and operational changes that need to be taken into account – like changes in clinical guidelines… How does that affect the algorithm?”
Meanwhile, how math changes over time — aka machine learning drift — needs oversight and accountability mechanisms, she said.
Implement proven algorithms in clinical practice
Makhni demonstrated how platform-based collaboration is developing AI data standards, ecosystem training and validation, cybersecurity, and implementing real-time algorithms at the Mayo Clinic.
On The Clinic’s development platform, collaborators agree to data standards and retain ownership of their data as they work together to train and validate the ecosystem, she explains.
With access to 10.4 million unidentified longitudinal patient records, 12 billion vital signs, 1.3 billion test results, 662 million clinical notes, etc., Mayo Clinic facilitates external innovation at a much faster rate because developers do not jeopardize patient data.
The platform aims to democratize data. With zero-code, low-code and high-code tools, innovators and clinicians can use their expertise and create their own data, says Makhni.
By using frameworks, validation using different data sets across the platform helps AI to be relevant to the clinical audience.
To address the bias problem, presenting bias fields that are easier to understand and understand in practice helps to build intuition.
Makhni says the platform is implementing validated algorithms — such as the EKG algorithm — and integrating them into workflows. She notes that the algorithms implemented provide real-time benchmarking and insights, and presents a roadmap for building a registry that tracks ML drift.
Makhni concludes, regardless of the development approach, there are several key considerations to improve the chances of algorithms being adopted.
While ML collaborators should be committed to values and protecting patient privacy, it is equally important to equip the workforce and empower clinicians to use algorithms correctly and “skills without overload”.
Her final piece of advice is to build trust with transparency, collaboration, and co-creation with stakeholders and decision makers — “even when there’s tension.”
“And never forget the patient. We’re doing this to improve their lives. So we need to find a way to get them involved in the process,” Makhni said.
Andrea Fox is the senior editor of Healthcare IT News.
Email: [email protected]
Healthcare IT News is a publication of HIMSS Media.
