Sepsis, a systemic, deleterious host response to infection that leads to organ dysfunction, is a potentially deadly condition needing prompt identification of the causative organisms and early appropriate antimicrobial therapy. Among non-culture-based diagnostic methods, SeptiFast (SF) can be employed to speed bacterial and fungal DNA detection, but it suffers from poor sensitivity and high cost. The aim of the present study, performed in 285 febrile patients, was to develop a prediction model to restrict the SF assay to clinical cases with a high probability of positive SF results. The prevalence of SF results positive for a pathogen was 17.2 %. Independent predictors of positive results were: blood sampling within 12 h after the onset of fever [odds ratio (OR) 20.03; 95 % confidence interval (CI) 6.87–58.38; P<0.0001]; ≥0.5 ng serum procalcitonin (PCT) ml⁻¹ (OR 18.52; 95 % CI 5.12–67.02; P<0.0001); body temperature ≥38 °C (OR 3.78; 95 % CI 1.39–10.25; P = 0.009); ≤3 g serum albumin dl⁻¹ (OR 3.40; 95 % CI 1.27–9.08; P = 0.014); and ≥13 000 white blood cells mm⁻³ (OR 2.75; 95 % CI 1.09–7.69; P = 0.05). The model showed good calibration (Hosmer–Lemeshow chi-squared 1.61; P = 0.978). Area under the receiving operating characteristic curve was 0.944 (95 % CI 0.914–0.973; P<0.0001). These results suggest that a prediction model based on PCT and a few other routinely available laboratory and clinical variables could be of help in selecting patients with a high probability of SF-positive results.