Author + information
- Manreet Kanwar,
- Natasha Loghmanpour,
- Yajuan Wang,
- Trevor Snyder,
- Douglas Horstmanshof,
- James Long,
- John Gorcsan,
- James Antaki and
- Srinivas Murali
Patient selection is critical to successful post-LVAD outcomes, but existing risk scores like Destination Therapy Risk Score (DTRS) are limited in scope. Bayesian Networks (BN) is advantageous over traditional analytical methods in using highly efficient & sophisticated algorithms which recognize complex medical relationships, thereby providing reliable predictors of clinical outcomes.
Retrospective data, including demographics, labs, hemodynamics etc. from 144 CF-LVAD pts. at 2 sites from yr 2006-11 was reviewed. Pts. were divided into 2 groups: low risk (alive 90 days post-implant) & high risk (died within 90 days). Several BNs, including Naïve Bayes (NB) & Tree-Augmented Naïve Bayes (TAN) were modeled and results compared to DTRS to assess performance.
Both NB & TAN models were built from datasets using 10 most predictive variables which emerged from feature selection based on their predictive power: HCT, AST, age, HR, TPG, mean PAP, diuretics, PLT count, BUN & HGB. Both BNs consistently outperformed the DTRS (see Table) in their ability to predict outcomes, given their capacity to account for relationships between variables, learn from prior probability & tolerate missing or erroneous data elements.
BNs improve accuracy, kappa, sensitivity & specificity over conventional analytical methods, thereby providing a novel, superior model for risk assessment. Risk stratification models employing BNs should be the focus of a larger, prospective study.
|Naïve Bayes (NB)||Tree-Augmented Naïve Bayes (TAN)||DTRS|
|Predicted Low Risk||Predicted High Risk|
|True Positive (Sensitivity)||False Positive|
|True Low Risk||NB||90.9%||NB||9.1%|
|True High Risk||False Negative||True Negative (Specificity)|
Poster Sessions, Expo North
Sunday, March 10, 2013, 9:45 a.m.-10:30 a.m.
Session Title: Improving Heart Failure Outcomes II
Abstract Category: 28. Quality of Care and Outcomes Assessment
Presentation Number: 1200-106
- 2013 American College of Cardiology Foundation