Not all #AI is created equal. The hype around AI has been both good and bad for data scientists. It’s great that any developer can now download a python library and create machine learning models. And I love that so many folks are passionate and interested in getting started in this field. Indeed, many problems can be solved with pretty straightforward approaches (or even “autoML”). But when it comes to tough, real-world use cases, I do believe an understanding of the underlying algorithms makes the difference between an AI solution that works, and one that does not. In a healthcare setting, it is critical that solutions work and provide clinicians the information necessary to make decisions. Black boxes aren’t going to cut it. At Wolters Kluwer, we’ve customized algorithms for class-imbalanced, time-series analysis to predict the onset of hospital-acquired infections. The most common algorithms we’ve seen in research for this use case include random forests and logistic regression. However, these algorithms do not consider the timing or order of features. The timing and order of events during a patient’s hospital stay are critical for practical applications of predictive analytics. Real-world solutions either need to employ algorithms that intrinsically deal with time series data, and/or data scientists must take great care during feature engineering to preserve information about the order and timing of events. I chat about all of these issues and how we predict hospital acquired infections in the article linked below.
https://www.wolterskluwer.com/en/expert-insights/predicting-hospital-infections-how-ai-makes-it-possible