AIMS: (i) To determine the value of individual alarm features for predicting cancer in subjects referred to a rapid access upper gastrointestinal cancer service; and (ii) to develop a clinical prediction model for cancer and to prospectively validate this model in a further patient cohort. METHODS: Patient demographics, referral indications, and subsequent diagnosis were recorded prospectively. Logistic regression analyses were employed to determine the predictive value of individual alarm features in an evaluation cohort of 1852 consecutive cases. The potential impact of applying a modified set of referral criteria was then examined in a validation cohort of 1785 patients. RESULTS: Evaluation cohort: mean age was 59 years; cancer prevalence 3.8%; and serious benign pathology 12.8%. Dysphagia (odds ratio (OR) 3.1), weight loss (OR 2.6), and age >55 years (OR 9.5) were found to be significant predictive factors for cancer but the value of other accepted alarm features was more limited. In particular, uncomplicated dyspepsia in those over 55 years was a negative predictive factor for cancer within this high risk cohort (OR 0.1). Validation cohort: the clinical prediction model would have selected 92% of cancer patients for fast track investigation while reducing the "two week rule" workload by 572 cases (31%). CONCLUSIONS: Fast track endoscopy in subjects fulfilling current criteria for suspected upper gastrointestinal malignancy results in a significant yield of cancer ( approximately 4%) and serious benign diseases such as peptic ulceration, strictures, and severe oesophagitis (13%). However, the predictive value of individual features for cancer varies widely. Uncomplicated dyspepsia in older subjects was a poor predictor of cancer. Application of narrower referral criteria for accessing fast track services may reduce pressures while retaining high sensitivity for cancer.