n. a. loghmanpour 1 , m. k. kanwar 2 , s. h. bailey 3 , r. l. benza 2 ,
DESCRIPTION
Development of a novel predictive model for mortality post continuous flow LVAD implant using Bayesian Networks (BN). N. A. Loghmanpour 1 , M. K. Kanwar 2 , S. H. Bailey 3 , R. L. Benza 2 , J. F. Antaki 1 , S. Murali 2 - PowerPoint PPT PresentationTRANSCRIPT
Development of a novel predictive model for mortality post
continuous flow LVAD implant using Bayesian Networks (BN)
N. A. Loghmanpour1, M. K. Kanwar2, S. H. Bailey3, R. L. Benza2, J. F. Antaki1, S. Murali2
1Department of Biomedical Engineering, Carnegie Mellon University, Pittsburgh, PA
2Department of Cardiology, Allegheny General Hospital, Pittsburgh, PA, 3Department of Surgery, Allegheny General Hospital, Pittsburgh, PA
International Society of Heart and Lung Transplantation Annual Meeting April 12, 2014
Disclosures• N.A. Loghmanpour: None• M.K. Kanwar: None• S.H. Bailey: None• R.L. Benza: None• J.F. Antaki: None • S. Murali: None
Case StudyPatient ACaucasian male60 years oldINTERMACS level 1NYHA class IVOn ventilator and IABP
Patient BCaucasian female70 years oldINTERMACS level 3NYHA class IVChronic renal disease
Motivation• (most) Risk score limitations:– Require a fixed set of data elements • May become outdated or irrelevant
– Typically assume linear relationships between variables – Derived from previous VAD technology, and inaccurate when
applied to newer VADs• Bayesian Network (BN) models provide robust predictions that
correlate pre-operative clinical variables to each other and final outcome.
• Previously demonstrated feasibility of BN to predict 90-day mortality in 2 center study.
Infection
Infection % likelihood Present 50Absent 50
WBC Infection Present Infection Absent
High 90% 4%
Normal 6% 6%
Low 4% 90%
Bayesian Networks
High: >11Normal: 4-11Low: <4
WBC Count
PRESENT
ABSENT
Traditional Statistics
So, what is the difference?
Traditional • Produce binary classifications
– black and white
• Consists of a numerical score
• Incomputable if data is missing – Cannot compute HMRS if no
Albumin recorded for pt
Bayesian• Produce probability
estimates– grey-zone
• Consists of a graphical and a quantitative component
• Robust ability to handle uncertainty and missing data
vs.
DTRS PointsPlatelet ≤ 148*103μl 7Albumin ≤ 3.3 g/dl 5INR > 1.1 4Vasodilator therapy 4mPAP ≤ 25mmHg 3AST > 45 U/ml 2Hct ≤ 34% 2BUN > 51 U/dl 2No IV inotropes 2
(Lietz et al. Circulation, 2007)
Existing Risk Score: HMRS & DTRS
Low risk ≤ 8 Medium risk = 9 - 16High risk = 17 - 19Very high risk ≥ 19
Low risk ≤ 1.58Medium risk = 1.58 - 2.48High risk ≥ 2.48
(Cowger et al. JACC, 2013)
HMRS WeightAge (in decades) 0.0274Albumin -0.723Creatinine 0.74INR 1.136Center LVAD volume 0.807
Study Design• INTERMACS: 8050 patients with continuous flow
LVADs• Inclusion criteria: All adult patients who received
CF LVADs as primary implant• Follow-up data censored for transplant and
device explant• Dependent variable: mortality– 30 day– 90 day– 1 year– 2 year
No. (%)N=111
Demographics: age interval, gender … 21 (19)
Co-morbidities: malnutrition, cancer… 30 (27)
Laboratory: blood type, sodium, INR, albumin… 16 (14)
Hemodynamics: cardiac output, LVEF, RVEF, heart rate… 21 (19)
Medication: ACEI, BB, warfarin, inotrope … 13 (12)
Quality of Life: EuroQoL mobility, pain, anxiety… 10 (9)
Clinical Variable Summary
Inclusion criteria: • >50% completion• Recorded pre-
implant
Total Death30 day 8007 387 (4.8%)90 day 7761 737 (9.5%)1 year 6575 1334 (20.3%)2 year 6099 1667 (27.3%)
Patient Cohort
Cardiac Outcomes Risk Assessment (CORA)
CORA Model Performance
Endpoint Accuracy (%) ROC (%) Kappa30 day 96.3 89.4 0.4490 day 91.4 81.2 0.381 year 84.0 79.4 0.462 year 78.4 80.0 0.43
ROC Curve
1-Specificity
Sens
itivi
ty
Case Study: OutcomePatient AHMRS: low risk
CORA: 66% chance of survival at 30 days44% chance of survival at 90 days
Outcome: died 5 days post-VAD
Patient BHMRS: medium risk
CORA: 99% chance of survival at 30 days 96% chance of survival at 90 days
Outcome: still alive (implant October 2011)
HMRS
CORA
2013 4th Quarterly INTERMACS report
Survival versus INTERMACS level
Limitations• Extensive missing data in many variables• Uneven distribution of outcome• Retrospective bias• Only FDA approved devices included in registry
Conclusion• First application of modern machine learning
algorithms to a LVAD cohort. • CORA models predictive power exhibited
excellent accuracy, sensitivity and specificity.• CORA models have the potential to develop a
reliable risk stratification tool for use in clinical decision making on LVAD patients... – Beta version currently live!
Thank you: Dr. Kirklin, Dr. Naftel and INTERMACS Funding: R41 HL120428-01 and R01HL086918 NIH grants
http://chriss.blenderhouse.com/Username: [email protected]: ishlt2014Contact: [email protected]
Cardiac Health Risk Stratification System (CHRiSS) Demo Site: