Research Project
9282329
? DESCRIPTION (provided by applicant): Data collected from intensive care units could be used to guide decision-making in real-time, but instead have often led to overwhelmed clinicians trying to uncover the signal buried in the noise. This data deluge is particularly challenging when the data are delayed, as this can lead to events being incorrectly seen as simultaneous or even out of order. Patients' health also changes at different time scales such as due to a new medication or circadian rhythms. Thus as doctors attempt to integrate the many signals to understand a patient's status, their health is a moving target. To transform the data into actionable knowledge, it is also not enough to find correlations. We must be sure that the patterns we find are truly causal to avoid treating symptoms instead of a disease or launching unsuccessful clinical trials. Our prior work, though, has found that ICU data streams can in fact be used to gain insight into recovery from stroke. In particular, we revealed that nonconvulsive seizures may be related to poor outcomes in patients with subarachnoid hemorrhage (SAH). Unlike epileptic seizures, which have a sudden onset, these seizures begin gradually, making them difficult to detect automatically. Further, many SAH patients in our study were unconscious on admission, and it is difficult to frequently and reliably assess consciousness. Therefore while progress has been made, two key barriers to using ICU data to guide treatment are a) a lack of methods for finding gradual changes in a patient's state (which could be used to alert clinicians) and b) finding causal relationships with uncertain data (where the cause may be documented as happening after the effect). To address these challenges, our specific aims are 1) to develop methods for finding timing uncertainty for each variable and using this in causal inference, 2) to develop real-time methods for finding when things change, and 3) to apply these to find when stroke patients have seizures or changes in consciousness, so these can be quickly identified and treated. We propose that by learning the reasons for errors in data, and by developing methods that specifically model their uncertain and changing nature, we will enable better use of large-scale observational biomedical data for real-time treatment decisions.
