Evaluating Patients with Chest Pain Using Classification and Regression Trees

We collected data on 320 patients complaining to their general practitioner of a new episode of chest pain, discomfort or oppression. Relationships were examined between initial signs and symptoms and a follow-up diagnosis after a period of 2 weeks to 2 months. The data were analysed with CART, a statistical decision theory software package. In our first run, the number of misclassifications by CART was 56%. After regrouping of the data and diagnostic categories, there were 37% misclassifications. The most discriminating variable turned out to be pain on palpation. When comparing each of five diagnostic groups to all others, we found a positive predictive value of 27% for gastrointestinal diseases, 72% for cardiovascular disorders, 69% for respiratory diseases, 58% for psychopathology and 73% for chest wall pathology. The CART methodology needs further investigation and testing before any clinical application will be possible in general practice.