Helena Aidos

Statistical modeling of dissimilarity increments for d-dimensional data: Application in partitional clustering

This paper addresses the use of high order dissimilarity models in data mining problems. We explore dissimilarities between triplets of nearest neighbors, called dissimilarity increments (DIs). We derive a statistical model of DIs for d-dimensional data (d-DID) assuming that the objects follow a multivariate Gaussian distribution. Empirical evidence shows that the d-DID is well approximated by the particular case d=2. We propose the application of this model in clustering, with a partitional algorithm that uses a merge strategy on Gaussian components. Experimental results, in synthetic and real datasets, show that clustering algorithms using DID usually outperform well known clustering algorithms.

Learning Similarities from Examples Under the Evidence Accumulation Clustering Paradigm

The SIMBAD project puts forward a unified theory of data analysis under a (dis)similarity based object representation framework. Our work builds on the duality of probabilistic and similarity notions on pairwise object comparison. We address the Evidence Accumulation Clustering paradigm as a means of learning pairwise similarity between objects, summarized in a co-association matrix. We show the dual similarity/probabilistic interpretation of the co-association matrix and exploit these for coherent consensus clustering methods, either exploring embeddings over learned pairwise similarities, in an attempt to better highlight the clustering structure of the data, or by means of a unified probabilistic approach leading to soft assignments of objects to clusters.

Hierarchical clustering with high order dissimilarities

This paper proposes a novel hierarchical clustering algorithm based on high order dissimilarities. These dissimilarity increments are measures computed over triplets of nearest neighbor points. Recently, the distribution of these dissimilarity increments was derived analytically. We propose to incorporate this distribution in a hierarchical clustering algorithm to decide whether two clusters should be merged or not. The proposed algorithm is parameter-free and can identify classes as the union of clusters following the dissimilarity increments distribution. Experimental results show that the proposed algorithm has excellent performance over well separated clusters, also providing a good hierarchical structure insight into touching clusters.

On the distribution of dissimilarity increments

This paper proposes a statistical model for the dissimilarity changes (increments) between neighboring patterns which follow a 2-dimensional Gaussian distribution. We propose a novel clustering algorithm, using that statistical model, which automatically determines the appropriate number of clusters. We apply the algorithm to both synthetic and real data sets and compare it to a Gaussian mixture and to a previous algorithm which also used dissimilarity increments. Experimental results show that this new approach yields better results than the other two algorithms in most datasets.

Validation of Two US Risk Scores for Percutaneous Coronary Intervention in a Single-Center Portuguese Population of Patients with Acute Coronary Syndrome

INTRODUCTION New scores have been developed and validated in the US for in-hospital mortality risk stratification in patients undergoing coronary angioplasty: the National Cardiovascular Data Registry (NCDR) risk score and the Mayo Clinic Risk Score (MCRS). We sought to validate these scores in a European population with acute coronary syndrome (ACS) and to compare their predictive accuracy with that of the GRACE risk score. METHODS In a single-center ACS registry of patients undergoing coronary angioplasty, we used the area under the receiver operating characteristic curve (AUC), a graphical representation of observed vs. expected mortality, and net reclassification improvement (NRI)/integrated discrimination improvement (IDI) analysis to compare the scores. RESULTS A total of 2148 consecutive patients were included, mean age 63 years (SD 13), 74% male and 71% with ST-segment elevation ACS. In-hospital mortality was 4.5%. The GRACE score showed the best AUC (0.94, 95% CI 0.91-0.96) compared with NCDR (0.87, 95% CI 0.83-0.91, p=0.0003) and MCRS (0.85, 95% CI 0.81-0.90, p=0.0003). In model calibration analysis, GRACE showed the best predictive power. With GRACE, patients were more often correctly classified than with MCRS (NRI 78.7, 95% CI 59.6-97.7; IDI 0.136, 95% CI 0.073-0.199) or NCDR (NRI 79.2, 95% CI 60.2-98.2; IDI 0.148, 95% CI 0.087-0.209). CONCLUSION The NCDR and Mayo Clinic risk scores are useful for risk stratification of in-hospital mortality in a European population of patients with ACS undergoing coronary angioplasty. However, the GRACE score is still to be preferred.

ECG-based Biometrics using a Deep Autoencoder for Feature Learning - An Empirical Study on Transferability.

Biometric identification is the task of recognizing an individual using biological or behavioral traits and, recently, electrocardiogram has emerged as a prominent trait. In addition, deep learning is a fast-paced research field where several models, training schemes and applications are being actively investigated. In this paper, an ECG-based biometric system using a deep autoencoder to learn a lower dimensional representation of heartbeat templates is proposed. A superior identification performance is achieved, validating the expressiveness of such representation. A transfer learning setting is also explored and results show practically no loss of performance, suggesting that these deep learning methods can be deployed in systems with offline training.

Discrimination of Alzheimer's Disease using longitudinal information

Alzheimer’s Disease (AD) is a neurological disorder that leads to a loss of cognitive functioning, affecting older people as well as their families. Although a few treatments are available to slow down the progress of the disease, they are limited in effectiveness and should start at an early stage of the disease. Since an early diagnosis of AD is crucial, to maximize treatment effectiveness and prepare the families for the worsening of symptoms, researchers are studying biomarkers and Computer-aided diagnosis (CAD) systems. Hence, this manuscript proposes a new methodology to obtain an efficient CAD system by relying on [

Semi-Supervised Consensus Clustering for ECG Pathology Classification

^{18}