A Configurable Deep Network for high-dimensional clinical trial data

Abstract

Abstract-Clinical studies provide interesting case studies for data mining researchers, given the often high degree of dimensionality and long term nature of these studies. In areas such as dementia, accurate predictions from data scientists provide vital input into the understanding of how certain features (representing lifestyle) can predict outcomes such as dementia. Most research involved has used traditional or shallow data mining approaches which have been shown to offer varying degrees of accuracy in datasets with high dimensionality. In this research, we explore the use of deep learning architectures, as they have been shown to have high predictive capabilities in image and audio datasets. The purpose of our research is to build a framework which allows easy reconfiguration for the performance of experiments across a number of deep learning approaches. In this paper, we present our framework for a configurable deep learning machine and our evaluation and analysis of two shallow approaches: regression and multi-layer perceptron, as a platform to a deep belief network, and using a dataset created over the course of 12 years by researchers in the area of dementia.

Variable interactions in risk factors for dementia

“…We hypothesise that these abstract features model the complex dementia risk factor interactions to better predict the outcomes in question. Our approach builds on the outline platform developed in [16] [15], with a new 2-step process to deliver high performing prediction models.…”

“…Neural networks in the context of health analytics have been shown to be at worst, on par with traditional regression approaches [16], [10] and at best outperform these methods [20]. Health benefits accrued from using ANNs in medical interventions have been shown to have significant impact in their application, where notable examples include cervical cytology and the early detection of acute myocardial infarction (heart attack) [9].…”

Detecting Feature Interactions in Agricultural Trade Data Using a Deep Neural Network

“…This amounts to 1 step of Gibbs sampling at each iteration of MSGD [8]. We refer the reader to [13] for (hidden) activation, energy, cost and derivative functions.…”

Forum Jeunes Chercheurs à Inforsid 2016

“…Toute la difficulté pour nous est alors de n'autoriser que des assertions et de détecter les incohérences. Ainsi, contrairement à des approches de data mining (Donoghue et al, 2015), nous ne cherchons ni à nous substituer aux cliniciens ni à optimiser les paramètres de sélection des stimulations.…”