We propose a novel algorithm to construct binary classifiers, in the spirit of the recently proposed Guaranteed Error Machine (GEM) but with a-posteriori assessment of the “support” instances and without the need for a ternary output. We provide rigorous guarantees on the probability of misclassification; differently from GEM, such guarantees aim to bound the conditional probability of error given the true value of the classified instance. The proposed classifier can be tuned in order to give more importance to one of the two kinds of error, and to balance their ratio also in the presence of unbalanced training sets. Guaranteeing the conditional probabilities of error is crucial in many classification problems, in particular medical diagnoses, where being able to push the trade-off between sensitivity (conditional probability of detecting a “true positive”) and specificity (conditional probability of detecting a “true negative”) towards higher sensitivity is of paramount importance. The application that first motivated our study is the classification of ventricular fibrillation (VF) into cases where restoration of an organized electrical activity is achieved immediately after a defibrillatory shock (“positive”), and cases where prompt resuscitation does not happen (“negative”). We provide experimental evidence that our approach is promising by testing it against three well-known medical datasets, against some data on VF that are available to the authors, and with Monte Carlo simulations.

A New Classification Algorithm With Guaranteed Sensitivity and Specificity for Medical Applications

Care', Algo
;
Ramponi, Federico A.;Campi, Marco C.
2018-01-01

Abstract

We propose a novel algorithm to construct binary classifiers, in the spirit of the recently proposed Guaranteed Error Machine (GEM) but with a-posteriori assessment of the “support” instances and without the need for a ternary output. We provide rigorous guarantees on the probability of misclassification; differently from GEM, such guarantees aim to bound the conditional probability of error given the true value of the classified instance. The proposed classifier can be tuned in order to give more importance to one of the two kinds of error, and to balance their ratio also in the presence of unbalanced training sets. Guaranteeing the conditional probabilities of error is crucial in many classification problems, in particular medical diagnoses, where being able to push the trade-off between sensitivity (conditional probability of detecting a “true positive”) and specificity (conditional probability of detecting a “true negative”) towards higher sensitivity is of paramount importance. The application that first motivated our study is the classification of ventricular fibrillation (VF) into cases where restoration of an organized electrical activity is achieved immediately after a defibrillatory shock (“positive”), and cases where prompt resuscitation does not happen (“negative”). We provide experimental evidence that our approach is promising by testing it against three well-known medical datasets, against some data on VF that are available to the authors, and with Monte Carlo simulations.
File in questo prodotto:
File Dimensione Formato  
CLASS_LCSS2018.pdf

solo utenti autorizzati

Tipologia: Documento in Post-print
Licenza: NON PUBBLICO - Accesso privato/ristretto
Dimensione 236.38 kB
Formato Adobe PDF
236.38 kB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11379/507660
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 28
  • ???jsp.display-item.citation.isi??? 25
social impact