Documente Academic
Documente Profesional
Documente Cultură
Shameek Ghosh
Shameek.Ghosh@student.uts.edu.au
Supervised By
Associate Prof. Jinyan Li
Outline
Healthcare Analytics
Details of the MIMIC II Database
Some basic statistics
The Research Problem
Healthcare Analytics
Mining of Raw EHR can help develop a
clinical model of the patients history
Aids in the diagnostic process by the
hospital and physician
Complex healthcare systems generating
massive data
MIMIC II WAVEFORM
DATABASE
ECG
Blood Pressure
Heart Rate
Vital Physiological Signals
Case_ID
Will have a set of
Waveforms
Hadm_ID
Subject_ID
Can have
many
Can have
many
Patient timeline
Tables of Importance
The basic information for any given patient is stored in the
table D_PATIENTS
Subject_ID is widely used by throughout MIMIC II to specify to
which patient a given measurement or recording refers to
D_PATIENTS is associated by the subject_id to ICD9 ( codes for
r International Classification of Diseases) and DRGEVENTS (
patient specific diagnosis events)
The ICD-9 table records the ICD-9 codes applied to a particular
patient during a specific hospitalization period.
Diagnosis-related groups (DRGs) are stored in the DRGEVENTS
table and their meanings are stored in D_CODEDITEMS.
Tables of Importance
Caregivers are stored in the D_CAREGIVER table (cgid). Caregivers related
to many other tables such as medevents, noteevents and chartevents and
is used to record the care giver who performed a particular operation,
procedure or event
The Careunits table D_CAREUNITS (cuid), stores information pertaining to
the different ICU rooms in the hospital and whenever a problem occurs or
a chart event is entered, the particular care unit is also recorded
Medication(s) given to a patient are recorded in the medevents,
d_meditems, a_meddurations and additives tables
Patient medical chart data is recorded in the chartevents, d_chartitems,
a_chartdurations and formevents tables
Patient input/output (IO) data is recorded in the ioevents, d_ioitems,
a_iodurations, deliveries, totalbalevents and additives tables
Patient notes are recorded in the noteevents table
Waldin, A et al (2013). In the proceedings of the 30th International Conference on Machine Learning. JMLR
W&CP volume 28.
BP is related to hypotension
Hypotension is related to Acute Kidney Injury (AKI)
AKI can arise in an ICU for cardiopulmonary bypass, major trauma, mechanical
ventilation, burn injuries, sepsis and in many other diseases
AKI results in mortality (loss of life) and morbidity (increased hospital stay causing
chronic conditions)
Need to investigate hemodynamic (blood flow) perturbations to understand the
incidence of AKI
Discriminative sequential patterns could be explored in arterial blood pressure
that might help in predicting stages of AKI
Predicting Acute Hypotensive Episodes (Physionet Challenge 2009)
Predicting an ICU patient's future blood pressure from a recording of his recent
blood pressure history (Waldin, 2013)
Problem Description
Given a continuos mean arterial signal from time - 1t, our
goal is to predict the value of a statistic defined for the sample
over a time period (t+k) to (t+k+a), where k is the lead time
and a is the prediction window
Additional issues that can be handled by algorithm: 1)
Prediction window size 2) Definition of prediction statistic (as
reported by Waldin, 2013)
Conclusion