Patients in critical care settings are at significant risk of experiencing severe adverse outcomes, such as stroke, sepsis, coma, infection, and death. There exists great potential for machine learning interventions to help predict patients’ risks for these conditions. Due to the relative rarity of these events, however, available data is poorly suited for training conventional supervised machine learning algorithms, as data of sufficient size would be prohibitively expensive to collect.
Personalized health risk assessment for critical care
An alternative machine learning approach can be used to assess patient risk for adverse events, based on individualized patient parameters. This approach is unsupervised and does not require explicit labeling of patient outcomes to train. As such, the new method is especially well-suited for identifying at-risk patients despite the low relative incidence of adverse events.
- Identification of at-risk patients
- Risk management of critical care patients
- Unsupervised training
- Trainable on data with scarce positive cases