2. Conflict of Interest
• I have affiliations with, special interests, or have conducted business
with the following companies that in context with this presentation
might possibly constitute a real or perceived conflict of interest:
– I hold two methods patents:
• Patient categorization
• Device utilization following extubation
8. A focus on mechanical ventilation
3% of patients1
7% of days1
12% of cost1
Mechanically ventilated patients in US acute care hospitals
36% of patients2
58% of days2
Mechanically ventilated patients in US intensive care units
Medicare DRG 207 patients
Per patient average revenue:
Average cost of care:
$31,4053
$37,5853
($6,180)
x 200 patients
($6,180)3
Loss per patient: ($1.2 million)
72% of cost2
1 Wunsch H, Linde-Zwirble WT, Angus DC, Hartman ME, Milbrandt EB, Kahn JM. The epidemiology of mechanical ventilation use in the United States. Crit Care Med. 2010;38(10):1947-1953.
2 Dasta JF, McLaughlin TP, Mody SH, Piech CT. Daily cost of an intensive care unit day: the contribution of mechanical ventilation. Crit Care Med. 2005;33(6):1266-1271.
3 Kelly S, Agarwal S, Parikh N, Erslon M, Morris P. US hospital cost vs payment by ventilator status. Critical Care Med. 40(12) Supplement1;1---328, December 2012.
9. An ideal opportunity:
Cost of mechanical ventilation
4 Survey: Seven hospitals with an average of 400 beds and 40 ventilators. CareFusion, Summer 2012.
11. Volume
CO2
Respiratory
Rate
Oxygen
Pressures
Heart Rate
Oxygen in
the blood
Blood
Pressure
Data Contextualization
Cardiopulmonary Classification System
Oxygenation
Status
Ventilation
Status
Safety
VALI
Cardiovascular
Status
Computer Aided Mechanical Ventilation (CAMV)
captures data, analyzes and categorizes a patient status using multiple sources
16. Predicting device utilization following extubation
• Can we predict the unplanned use of NIV using 6 hours of
continuous data
– Ventilator and physiologic data
– 4,320 data points (12 parameters per minute x 360)
– Total Dynamic Compliance (Cdyn), SpO2), FIO2, RRTotal, SpRR, VTexp
specific to ideal body weight (IBW), MAP, HR, PIP, PEEP, etCO2,
VCO2, weight and height
Extubation
Attempt
(A) No Positive Pressure Support
(B) Re-Intubation
(C) Non Invasive Ventilation (NIV)
(C1) Planned NIV
(C2) Un-Planned NIV
17. Machine Learning
• Predictive modeling framework
– Goal 1
• Leaving-one-out cross validation
• 83 extubations were used for training and the 1 left out (out of sample) was used
to validate
• Step-wise forward logistic regression method was used to develop the best
model
– Akaike Information Coefficient (AIC)
– Mallows’ Cp stopping rules
• For robustness we repeated this 84 times
– Goal 2
• Physiological phenotype
– Picked the top five input variables appearing in 70% or more of the 84 predictive
models
• Physiological model
20. Network Diagram
Smart ICU Rooms
• Medical Devices
• Vents
• Monitors
• IV pumps
• Beds
• Cerner CareAware
Bridge
• Spacelab Monitors
Cerner Millennium
EMR
CareAware
CENTRA
ADT
HL7
Labs
Research Team
o Engineering
o Health
Informatics
o Medicine
o Nursing
o Respiratory
Therapy
HTTP/HTTPS
Liberty’s HPC Cluster
LDAP
HL7
HL7
• Demographics
• Medications
• Diagnosis
• Admission
• Discharge
• Transfer
HL7
HL7
21. Towards the Development of Decision Support
Early Event Detection Real-time data
Continuous data: Vital Signs + Ventilator
– Patient categorization (per minute)
– Early detection of poor quality of
mechanical ventilation – value-based
care
– Continuous assessment of readiness to
liberate
– Early determination of high utilization
patients (long length of stay)
Learn from historical patterns to improve care while lowering cost in the most
expensive unit in the hospital.