COMPUTER EXPERT SYSTEM FOR RISK STRATIFICATION AND THERAPEUTIC DECISION MAKING IN ACUTE CORONARY SYNDROMES
M. Grabowski 1,2, R. Rudowski 1 1 Medical University of Warsaw, Department of Medical Informatics and Telemedicine, Poland 2 Medical University of Warsaw, Department of Cardiology, Poland marcin.grabowski@amwaw.edu.pl Aim: to develop and evaluate the expert system (ES) for risk stratification and therapeutic decision making in acute coronary syndromes (ACS). Material and methods: ES was developed and was integrated with the database. Knowledge base contains rules according to the current guidelines. ES consists of: a) risk voting system which votes on risk according to 4 risk scores (SIMPLE, TIMI, GRACE, ZWOLLE) and on B-type natriuretic peptide levels (BNP); b) module suggesting the type of therapy (invasive vs. conservative); c) module which chooses pharmacotherapy. The system was evaluated on medical data of 149 consecutive patients. Results: BNP showed strong predicting value for 30-day mortality – area under ROC curve was 0,892 (95% confidence interval [CI]: 0,831 - 0,937). BNP at level of 334 pg/ml had 90,0% sensitivity and 88,5% specificity. In multivariate analysis BNP levels had independent prognostic value. Areas under ROC curves for risk scores were for: SIMPLE - 0,859 (95%CI: 0,793 - 0,911), TIMI STEMI - 0,869 (95%CI: 0,804 - 0,918), GRACE - 0,819 (95%CI: 0,748 - 0,877), ZWOLLE - 0,9 (95%CI: 0,84 - 0,943). There was no good agreement between BNP levels and risk scores in risk stratification. In multivariate analysis incorporating BNP and each risk score separately, BNP remained independent risk factor for 30 -day mortality. The addition of BNP into the model increased ROC. Classification error for risk scores and BNP were: for SIMPLE - 8,05%; TIMI - 4,70%; GRACE - 6,71%; ZWOLLE - 6,04%; BNP 11,41%. Developed majority voting system had the lowest classification error - 4,03%. Changes in cut-off number of votes on high risk resulted in sensitivity (from 50% to 90%) and specificity (from 93% to 100%) changes. Full agreement was seen between physician-expert and ES in decision for the need of reperfusion therapy; good agreement in decision for the type of reperfusion therapy (kappa=0,65), good for angiotensin converting enzyme inhibitors (kappa=0,69), beta-adrenolitics (kappa=0,705), nitroglycerin (kappa=0,69), furosemide (kappa=0,72), very good agreement for aspirin (kappa=0,889) and full (kappa =1) for heparin. Conclusions: 1) addition of BNP levels into popular risk scores improves risk stratification in patients with acute ACS; 2) developed voting system significantly decreased classification error to risk groups, changes in cut-off for votes on high risk improved sensitivity or specificity for risk classification dependently on the reference level of cardiologic centre; 3) good and very good agreement was observed between ES and physician-expert in type of therapy (reperfusion, pharmacotherapy) choice. Keywords: acute coronary syndromes, decision making, expert system, risk stratification