Naive Bayesian Prediction of Bleeding After Heart By-Pass Surgery Raymond Lister University of Technology, Sydney (presenting author) other authors... Smith, Ray & Hawson The Prince Charles Hospital, Brisbane

Problem Predict Who Will Bleed Excessively After Heart Bypass Surgery.

Physiological Bleeding Not haemophilia –it’s the technology Affects 1 in 8 patients Can cause death, or lifelong disability –e.g. Alan Bond Treatable with drugs and blood products –problem is distinguishing it from surgical error.

Earlier Attempts at Prediction Correlation to single clinical parameter –several inadequately weak indicators found. Multivariate Linear models –Gravlee et al. (1994)... Poor results

Naive Bayes Combines multiple weak evidence e i to calculate probability of hypothesis P(H). It’s simple. Details? See references: –Duda, Gashnig, and Hart (1979) “Prospector” –Shinghal R. (1992) Single pass through data! No magic numbers! –Occam’s razor –c.f. Neural Networks

Our Data N = 83 patients 8 physiological bleeders –“prior probability” = 8/83 ~= pre-operative parameters –including whole blood aggregation (WBAG) –63 possible combinations 8 post-operative parameters –255 possible combinations

Blood Loss v. WBAG

Cross Validation Repeat Construct random 90% subset of “training data” Develop Naive Bayes Model with that data Test that model on remaining “test” data Many times (100?) Report average prediction for each patient, with standard deviation (67% confidence interval)

Pre-op versus Post-op testing A solid pre-operative test strategy ideal, but not possible. Hence a two-stage procedure: 1. Screen all patients pre-op, to cull those at low risk. 2. Take special precautions with remainder and test them post-op.

Parameter Selection For the 63 combinations of pre-op parameters doPerform 100 cross validation runs (N=83). Select best set of pre-op parameters. Also identify the subset of patients judged as being “at risk” by pre-op system (N=29 patients). For the 255 combinations of post-op parameters do Perform 100 cross validation runs, using the subset of “at risk” patients (N=29). (Total computer run time is just a few seconds!)

Best Pre-Operative Predictor Each dot is a patient. 54/83 = 65% patients culled.

Best Post-Operative Predictor Each dot is a patient. All “real bleeders” identified, plus 7% false positives.

Conclusion Results encouraging N=83 not definitive, but cross validation gives confidence. Larger study justified. (Don’t build complex models until simple models are shown to be inadequate.)