Haemodynamic performance of bileaflet mechanical heart valves can be severely affected by the formation of thrombotic deposits. Hence, early detection of thrombi is fundamental for a prompt diagnosis and adequate therapy. This article aims at designing a novel diagnostic and prognostic tool able to detect valvular thrombosis at early stages of formation, i.e., before the appearance of critical symptoms in patients who can be effectively treated by pharmacological therapy, preventing re-operation. This approach relies on the acquisition of the acoustic signals produced by mechanical heart valves in the closing phase; the corresponding power spectra are then analysed by means of artificial neural networks trained to identify the presence of thrombi and classify their occurrence. Five commercial bileaflet mechanical heart valves were investigated in vitro in a Sheffield Pulse Duplicator; for each valve six functional conditions were considered, each corresponding to a risk class for patients (one normofunctioning and five thrombosed): they have been simulated by placing artificial deposits of increasing weight and different shape on the valve leaflet and on the annular housing; the case of one completely blocked leaflet was also investigated. These six functional conditions represent risk classes: they were examined under various hydrodynamic regimes. The acoustic signals produced by the valves were acquired by means of a phonocardiographic apparatus, then analysed and classified. The ability to detect and classify thrombotic formations on mechanical valve leaflet would allow ranking patients by assigning them to one of the six risk classes, helping clinicians in establish adequate therapeutic approaches.