(This post first appeared on Pulse)
Machine Learning and Artificial Intelligence in Healthcare require reliable data with low variability, accurately describing the disease state and progression
If we would go by the ads we see on TV we would conclude that IBM Watson is the perfect husband, wife, boss, coworker. IBM Watson not only knows when the engine on an airline jet needs servicing, it can help with your taxes, play with your kids and even impress Bob Dylan with its knowledge of his lyrics. And although at times terms like “Artificial Intelligence” and “Machine Learning” feel like buzzwords, it is hard to overstate the potential of these approaches especially in healthcare. To put it simply Machine Learning Algorithms can detect patterns in large amounts of data that are too complex to discern with the human eye. One caveat though is that the data used for the machine learning needs to be good, reliable and low variability data. The expression “garbage in, garbage out” very much applies to this type of approach.
In previous posts, I was a strong advocate for a personalized approach to respiratory medicine by assessing lung health rather than lung function alone. Functional Respiratory Imaging or FRI is a suitable enabler of precision medicine as it has the capabilities to yield accurate, regional information on lung structure and function. FRI as a proprietary CT-based approach yields a multitude of parameters related to ventilation, perfusion, lung tissue and even aerosol deposition characteristics and this on a regional, often a lobar, level (Figure 1).
By the vast amount of data the technology provides, it surpasses the conventional lung function measures such as FEV1 in terms of accuracy and clinical relevance. Conventional lung function parameters do not reveal any regional information while disease manifestation and progression happens on a very regional and localized level, especially in the earlier stages of the disease (Figure 2).
The downside of the shear volume of parameters is that it requires a significant effort to analyze them all and put them in perspective.
The low variability of the measurement and the high quality of the FRI parameters makes them highly suited for Machine learning and Artificial Intelligence. Initially applying this approach in clinical trials with FRI parameters obtained before and after a treatment creates a well controlled environment to validate the outcomes of the machine learning with known pathophysiology and treatment effect. It remains important that especially during the initial development human interpretation of the data coincides with increased machine learning.
Within FLUIDDA we have seen some very promising outcomes in our clinical trials when applying AI in the space of COPD responder phenotyping, IPF disease progression and Lung Transplant rejection prediction. All of a sudden the machine summarizes the available FRI information and presents it in a clinically useful manner opening up pathways for extensive clinical utility.
Over the next weeks and months we will use this forum to present a number of applications in more detail where AI and FRI were used with high accuracy to assess and improve Lung Health thereby establishing a powerful triad.