Predicting hospital infections: how AI makes it possible
While difficult to diagnose, through a combination of awareness, better education, and technologies, more people with HAIs are being identified and treated earlier, and fewer of them are dying.
But there’s so much more to be done – and, at long last artificial intelligence (AI) has matured to the point where we can apply its power. It’s a tipping point that has the potential for us to transform healthcare in fundamental ways. We can advance the way care gets delivered by tying disparate pieces of healthcare information together so clinicians can decide how to act on the information. That’s vastly different from “AI 1.0”, which focused on the possibilities of AI without the real-world application in clinical terms.
And patients at risk of HAIs and high-risk conditions like Clostridium difficile (C. Diff) and sepsis – will undoubtedly benefit from our progress. We now not only have real-world clinical use cases thanks to the clinical expertise being applied, but we have the technology building blocks that move beyond the hype and can help us advance more quickly.
There are two major reasons AI has made such inroads in recent years:
- the mass digitalization of health records that makes data accessible to algorithms
- advances in computation power, which makes theoretical algorithms practical, giving deeper knowledge to which types of algorithms and statistical models will yield the best results.
Why AI is a “Good Fit” for Predictive Surveillance
Machine learning, in particular, leverages big data to learn the correlations between very large sets of variables to be more predictive, and to provide results based on these predictors so the clinician can decide how to act.
“When we combine multiple health signals – such as procedures, antibiotic use, or vital signs – we can identify the complex interaction effects between multiple variables that are hard to detect without AI,” said John Langton, PhD and Director of Data Science at Wolters Kluwer, Health.
Through early experiments and the alignment between data science and clinical expertise, we know that detecting HAIs is an ideal use case for AI as it has the ability to monitor all the variables that really mean something and to reduce the noise in the data that can lead to “false alerts” where the monitoring data somehow gets translated into some unwanted action.
But not all AI is created equal cautions Langton. Creating AI models that accurately take into account the nuances of a condition is not an easy process. “Anyone with a computer can now download a machine learning model without having any idea on how the algorithms work,” Langton continued.
“That accessibility is great, but it could lead to a lot of misleading hype and suboptimal models. In the rush to release AI applications in the medical domain, many researchers have used simplistic models. While I believe that one should always start simple, these models do not intrinsically consider ‘time.’
The more simplistic models can be like applying a Swiss Army Knife – slicing and dicing data without regard for the order of events – and depending on aggregated metrics and values. These models apply those events as independent dimensions, losing critical information about the order and timing of clinical events along the way. This information is critical to tracking the changes in a patient to predict outcomes, but is completely lost without more sophisticated approaches. Ultimately, it’s all about time: predicting a condition early enough so that clinicians can respond and change outcomes.