G M Podda1, E Grossi2, T Palmerini3, M Buscema4, E A Femia1, D Della Riva3, S de Servi5, P Calabrò6, F Piscione7, D Maffeo8, A Toso9, C Palmieri10, M De Carlo11, D Capodanno12, P Genereux13, M Cattaneo14. 1. Unità di Medicina III, ASST Santi Paolo e Carlo, Dipartimento di Scienze della Salute, Università degli Studi di Milano, Milano, Italy. 2. Centro Diagnostico Italiano, Milano, Italy. 3. Dipartimento Cardiovascolare, Policlinico S. Orsola, Bologna, Italy. 4. Semeion Research Centre, Roma, Italy. 5. Unità Coronarica IRCCS Policlinico San Matteo, Pavia, Italy. 6. Divisione di Cardiologia, Seconda Università di Napoli, Napoli, Italy. 7. Dipartimento di Medicina e Chirurgia, Schola Medica Salernitana, Università di Salerno, Salerno, Italy. 8. Unità di Cardiologia, Servizio di Emodinamica, Istituto Ospedaliero Fondazione Poliambulanza, Brescia, Italy. 9. Divisione di Cardiologia, Ospedale Santo Stefano, Prato, Italy. 10. Ospedale del Cuore, Fondazione Toscana Gabriele Monasterio, Massa, Italy. 11. Dipartimento Cardiotoracico e Vascolare, Azienda Ospedaliero-Universitaria Pisana, Pisa, Italy. 12. Ospedale Ferrarotto, Università di Catania, Catania, Italy. 13. The Cardiovascular Research Foundation, New York, NY, USA. 14. Unità di Medicina III, ASST Santi Paolo e Carlo, Dipartimento di Scienze della Salute, Università degli Studi di Milano, Milano, Italy. Electronic address: marco.cattaneo@unimi.it.
Abstract
BACKGROUND: About 40% of clopidogrel-treated patients display high platelet reactivity (HPR). Alternative treatments of HPR patients, identified by platelet function tests, failed to improve their clinical outcomes in large randomized clinical trials. A more appealing alternative would be to identify HPR patients a priori, based on the presence/absence of demographic, clinical and genetic factors that affect PR. Due to the complexity and multiplicity of these factors, traditional statistical methods (TSMs) fail to identify a priori HPR patients accurately. The objective was to test whether Artificial Neural Networks (ANNs) or other Machine Learning Systems (MLSs), which use algorithms to extract model-like 'structure' information from a given set of data, accurately predict platelet reactivity (PR) in clopidogrel-treated patients. METHODS: A complete set of fifty-nine demographic, clinical, genetic data was available of 603 patients with acute coronary syndromes enrolled in the prospective GEPRESS study, which showed that HPR after 1month of clopidogrel treatment independently predicted adverse cardiovascular events in patients with Syntax Score >14. Data were analysed by MLSs and TSMs. ANNs identified more variables associated PR at 1month, compared to TSMs. RESULTS: ANNs overall accuracy in predicting PR, although superior to other MLSs was 63% (95% CI 59-66). PR phenotype changed in both directions in 35% of patients across the 3 time points tested (before PCI, at hospital discharge and at 1month). CONCLUSIONS: Despite their ability to analyse very complex non-linear phenomena, ANNs or MLS were unable to predict PR accurately, likely because PR is a highly unstable phenotype.
BACKGROUND: About 40% of clopidogrel-treated patients display high platelet reactivity (HPR). Alternative treatments of HPR patients, identified by platelet function tests, failed to improve their clinical outcomes in large randomized clinical trials. A more appealing alternative would be to identify HPR patients a priori, based on the presence/absence of demographic, clinical and genetic factors that affect PR. Due to the complexity and multiplicity of these factors, traditional statistical methods (TSMs) fail to identify a priori HPR patients accurately. The objective was to test whether Artificial Neural Networks (ANNs) or other Machine Learning Systems (MLSs), which use algorithms to extract model-like 'structure' information from a given set of data, accurately predict platelet reactivity (PR) in clopidogrel-treated patients. METHODS: A complete set of fifty-nine demographic, clinical, genetic data was available of 603 patients with acute coronary syndromes enrolled in the prospective GEPRESS study, which showed that HPR after 1month of clopidogrel treatment independently predicted adverse cardiovascular events in patients with Syntax Score >14. Data were analysed by MLSs and TSMs. ANNs identified more variables associated PR at 1month, compared to TSMs. RESULTS: ANNs overall accuracy in predicting PR, although superior to other MLSs was 63% (95% CI 59-66). PR phenotype changed in both directions in 35% of patients across the 3 time points tested (before PCI, at hospital discharge and at 1month). CONCLUSIONS: Despite their ability to analyse very complex non-linear phenomena, ANNs or MLS were unable to predict PR accurately, likely because PR is a highly unstable phenotype.