BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Flexible deep learning for heterogeneous clinical time series - Ja
 cob Deasy
DTSTART:20200128T130000Z
DTEND:20200128T140000Z
UID:TALK136018@talks.cam.ac.uk
CONTACT:Mateja Jamnik
DESCRIPTION:Extensive monitoring in the hospital\, and in particular the i
 ntensive care unit\, generates large quantities of data which contain nume
 rous trends that are difficult for clinicians to systematically evaluate. 
 Current attempts to address such heterogeneity in Electronic Health Record
 s (EHRs) discard pertinent information and state-of-the-art models are oft
 en the final stage in vast pre-processing pipelines which are left unseen.
  I will discuss a more flexible approach to outcome prediction from EHR da
 ta that attempts to use all uncurated events without variable selection or
  pre-processing. Moreover\, in realistic scenarios\, multivariate timeseri
 es evolve over case-by-case time-scales. This is particularly clear in med
 icine\, where the rate of clinical events varies by ward\, patient\, and a
 pplication. Increasingly complex models have been shown to effectively pre
 dict patient outcomes\, but have failed to adapt granularity to these inhe
 rent temporal resolutions. I will also outline initial research into adapt
 ive prediction timing for clinical time series.
LOCATION:SS03\, Computer Laboratory\, William Gates Building
END:VEVENT
END:VCALENDAR