Now Reading
How To Make Algorithms Really Work For Clinical Trials

How To Make Algorithms Really Work For Clinical Trials

“Researchers propose a framework for regulating algorithms that will ensure world-class clinical performance and build trust.”

Machine learning models are considered to gradually take over most of the mundane tasks of clinicians and other medical professionals. In a way, assisting them with insights and cutting the downtime. These models are already getting good at helping radiologists. But how legit are these algorithms? Have they been thoroughly vetted by government bodies like the FDA?

Researchers at Stanford noted a few reasons as to why algorithms are readily accepted into clinical setups:

Register for FREE Workshop on Data Engineering>>
  • Mix up of the diagnostic task with the diagnostic algorithm
  • Not so thorough treatment of the diagnostic task definition 
  • No mechanism to directly compare similar algorithms 
  • Insufficient characterisation of safety and performance elements
  • Lack of resources to evaluate and inherent conflicts of interest

To address these issues, a team from Stanford Institute for Human-Centered Artificial Intelligence (HAI) in their new paper titled “Regulatory Frameworks for Development and Evaluation of Artificial Intelligence”, propose a framework for FDA regulation of AI-based diagnostic imaging software. 

Challenges Of AI Based Diagnostics

(Source: Larson et al.,)

Though in the US, the FDA has a framework in place for evaluating software developed for medical applications, the authors state that these checkpoints cannot be considered for AI-based tools. According to the head of the FDA, AI algorithms continually learn from the medical images they review. So, the traditional approach to reviewing and approving SaMD (software-as-a-medical-device) upgrades may not apply. There is a need to implement a protocol for regulating the entire lifecycle of AI-based algorithms.  

Explaining this, one of the authors cited the example of the early days of the coronavirus pandemic when medical professionals worldwide proposed multiple different scoring systems to categorise lung scans of patients. Though the scoring systems differ, they are corrected over time, and this type of adjustment is common in clinical trials. However, in the case of AI, the authors wrote, that would have been problematic as hard-coded proprietary diagnostic task definitions will make it difficult to compare the performance of algorithms. 

The authors propose that algorithm manufacturers should be required to go through four phases of development and evaluation similar to what the FDA does. They suggest testing for:

See Also

  1. Feasibility of algorithms on a small test set
  2. Capability in a controlled environment simulating real-world conditions
  3. Effectiveness in a real clinic) 
  4. Durability including monitoring and improvement over time

The above flow chart illustrates the proposed linkage of the evaluation of diagnostic algorithm performance from the defined task to local site algorithm performance. Algorithms are developed according to a defined diagnostic task standard. Performance is compared with other algorithms in a controlled environment, which becomes the internal benchmark for general real-world performance and local site performance, which in turn becomes the benchmark for ongoing monitoring.

Key Takeaways

  • Strategies outlined by regulatory bodies address many key aspects to help ensure the safety, effectiveness, and performance of SaMD applications, but a number of gaps remain.
  • Appropriate evaluation and improvement methods should be incorporated into phases of development, similar to what the FDA does. 
  • Algorithms should be thoroughly tested and refined before being deployed in the clinical environment, just as it is expected from other medical devices.
  • Regulatory frameworks should strive to establish conditions that set up a race to the top for consistent excellent algorithm performance at each installed site.

One of the authors likened their proposed regulatory framework to the safety requirements for an aircraft. “… if they stop working, they need to fail in a way that’s not going to hurt people.” In their paper, the authors have listed 12 measures of performance that should be applied to diagnostic algorithms. 

To know more, download the original paper here.

Subscribe to our Newsletter

Get the latest updates and relevant offers by sharing your email.
Join our Telegram Group. Be part of an engaging community

Copyright Analytics India Magazine Pvt Ltd

Scroll To Top