Search

Diabetes is a chronic diseases that affects nearly half a billion people around the globe, and is almost always associated with a number of complications, including kidney failure, blindness, stroke, and heart attack. An important step towards improved diabetes care is to accurately predict the risk of diabetes complications and to identify the corresponding risk factors associated with the onset of each complication. In this paper, we study the problem of risk prediction and outcome-specific risk factor identification from readily available patient medical record data. We adopt a private-shared multi-task learning (MTL) model, which jointly models multiple complications with each task corresponding to the risk modeling of one complication. The MTL formulation not only boosts prediction performance but also enables identification of outcome-specific risk factors. Specifically, we decompose the coefficient matrix, in which each column (vector) corresponds to the coefficient of one complication risk model, into a shared component and an outcome-specific private component. The shared component is assumed to be low-rank to capture the relationships among complications in terms of overall diabetes health condition. The private component is assumed to be non-overlapping and sparse so that they are discriminative among the different complication outcomes. Further, the shared component and the private component for the same complication are assumed to be orthogonal. Extensive experimental results on a type 2 diabetes cohort extracted from a large electronic medical claims database show that the proposed method outperforms baseline models by a significant margin. Also the identified outcome-specific risk factors provide meaningful clinical insights. The results demonstrate that simultaneously modeling multiple risks through MTL not only improves prediction performance but also enables identification of outcome-specific risk factors.

Topic identification (TID) is the automatic classification of speech messages into one of a known set of possible topics. The TID task can be view as having three principal components: 1) event generation, 2) keyword event selection, and 3) topic modeling. Using data from the Switchboard corpus, the authors present experimental results for various approaches to the TID problem and compare the relative effectiveness of each. In addition, they examine the effect of keyword set size on identification accuracy and gauge the loss in performance when mismatched topic modeling and keyword selection schemes are used. >

This paper describes the development and application of a phonetic recognition system to the task of spoken document retrieval. The recognizer is used to generate phonetic transcriptions of the speech messages which are then processed to produce subword unit representations for indexing and retrieval. Subword units are used as an alternative to words units generated by either keyword spotting or word recognition. We first investigate the use of different acoustic and language models in the speech recognizer in an effort to improve phonetic recognition performance. Then we examine a variety of subword unit indexing terms and measure their ability to perform effective spoken document retrieval. Finally, we look at some simple robust indexing and retrieval methods that take into account the characteristics of the recognition errors in an attempt to improve retrieval performance.

Speech recognition of conversational speech is a difficult task. The performance levels on the Switchboard corpus had been in the vicinity of 70% word error rate. In this paper, we describe the results of applying a variety of modifications to our speech recognition system and we show their impact on improving the performance on conversational speech. These modifications include the use of more complex models, trigram language models, and cross-word triphone models. We also show the effect of using additional acoustic training on the recognition performance. Finally, we present an approach to dealing with the abundance of short words, and examine how the variable speaking rate found in conversational speech impacts on the performance. Currently, the level of performance is at the vicinity of 50% error, a significant improvement over recent levels.

Investigates the fusion of different information sources, with the goal of improving performance on spoken document retrieval (SDR) tasks. In particular, we explore the use of multiple transcriptions from different automatic speech recognizers, the combination of different types of subword unit indexing terms, and the combination of word- and subword-based units. To perform the retrieval, we use a novel probabilistic information retrieval model which retrieves documents based on maximum likelihood ratio scores. Experiments on the 1998 TREC-7 SDR task show that the use of these different information fusion approaches can result in significantly improved retrieval performance.

The authors present a segmental speech model that explicitly models the dynamics in a variable-duration speech segment by using a time-varying trajectory model of the speech features in the segment. Each speech segment is represented by a set of statistics which includes a time-varying trajectory, a residual error covariance around the trajectory, and the number of frames in the segment. These statistics replace the frames in the segment and become the data that are modeled by either HMMs (hidden Markov models) or mixture models. This segment model is used to develop a secondary processing algorithm that rescores putative events hypothesized by a primary HMM word spotter to try to improve performance by discriminating true keywords from false alarms. This algorithm is evaluated on a keyword spotting task using the Road Rally Database, and performance is shown to improve significantly over that of the primary word spotter. The segmental model is also used on a TIMIT vowel classification task to evaluate its modeling capability. >

It is demonstrated that using the proposed probabilistic vector mapping algorithm as a feature preprocessor results in robust performance levels across a wide range of signal-to-noise (SNR) levels. The authors evaluate the algorithm using an HMM word spotting system trained with clean cepstral features and tested with vector mapped noisy cepstra. In addition to robust behavior, it is shown that using the vector mapper results in performance that equals or exceeds that of using matched training and testing. For example, with 10-dB SNR testing speech, word spotting performance with the vector mapping preprocessor and clean training is 15% better than matching training with 10-dB SNR speech. A mapping algorithm based on the method of radial basis functions (RBFs) for mapping noisy speech features into the space of clean features is presented. Performance using this RBF mapper is shown to be comparable to that of the vector mapper. >