Interpretable Methods for Decision-Making in Intensive Care Units
Harvard's School of Engineering and Applied Sciences
Oracle Fellowship Recipient
Oracle Principal Investigator
Decision-making in intensive care units (ICUs) requires responding quickly to rapidly-changing situations. However, the efficacy of many interventions remains unquantified. The vast amounts of data that are collected in ICUs—vital signs, clinical notes, fluids, medications—suggest an opportunity for more data-driven decision-making about interventions such as ventillation, vassopressor administration, and platlet transfusions. To date, this opportunity has been largely untapped: most machine learning efforts in ICU contexts have focused on mortality prediction—which is of limited value to clinicians who must make decisions about how to treat their patients regardless of their acuity.
We propose to develop machine learning and sequential decision-making tools to provide patient-specific predictions of when an intervention may be necessary, when a patient may be ready to be weaned from an intervention, and whether a specific intervention may be effective. Our work will combine heterogeneous data sources—such as physiological signals, lab values, and terms from the clinical notes—to make patient-level predictions. A core challenge is learning from heterogeneous data sources sampled sporadically at different time scales.