Get the latest tech news
Low responsiveness of ML models to critical or deteriorating health conditions
Pias et al. evaluate machine learning models designed to predict in-hospital mortality and 5-year cancer survivability. Multiple classification models are unable to recognize critical health conditions or deteriorating patient conditions.
The eICU dataset creation is exempt from institutional review board approval due to the retrospective design, lack of direct patient intervention, and the security schema, for which the re-identification risk was certified as meeting safe harbor standards by an independent privacy expert (Privacert, Cambridge, MA) (Health Insurance Portability and Accountability Act Certification no. These are the neural activation values, calculated after applying the sigmoid function, when the model is fed with test cases varying a single vital, such as g glucose, h diastolic blood pressure, i temperature, and j respiratory rate. When using LSTM gradient ascent to automatically generate multi-attribute deteriorating test cases, we found the resulting test cases all have significantly decreased oxygen saturation and body temperature values in the last 24 h. Because the gradient ascent process follows the shortest path within the loss function space of the model, these findings indicate that (i) oxygen saturation and body temperature are top LSTM features and (ii) the last 24 h (out of the entire 48-h timespan) are important in the model’s decision-making process, which is also consistent with LR feature ranking (Supplementary Fig.
Or read this on Hacker News
