Beyond Image Recognition: Applying Deep Learning to List-Type Medical Data for Risk Prediction

ObjectiveThe purpose of our research was to develop efficient utilization of Deep Learning to the list-type data in which medical characteristics are arranged. Materials and MethodsWe conducted a survey of blood donors, focusing on the rare adverse reaction of falling. In addition to all cases of fainting, we randomly selected a control group of donors who did not fall. This data was then converted into a two-dimensional format suitable for analysis with a convolutional neural network (CNN). We used an under sampling method to randomly divide the dataset into training and testing sets. Finally, we used the CNN and a logistic regression model to predict the probability of fainting, calculate anomaly scores, and rank the risk of falling. ResultsThe convolutional neural network (CNN) identified 3 out of 10 falls among the group with the top 1% of anomaly scores. In contrast, the logistic regression model failed to identify any falls within the same top 1% anomaly score group. ConclusionApplying information converted into two-dimensional data to Deep Learning by using anomaly detection together was useful to narrow down with high-risk group. Although these findings require validation in a larger and more diverse population, the success of this approach in predicting falls after blood donation suggests its potential for predicting other rare adverse events in healthcare, such as adverse drug reactions, complications from medical procedures, or even disease outbreaks.