Home » News » Bottom-up: Towards Supporting Personalised Medicine in the Cloud

From birth, a person accumulates healthcare data of multiple forms, including clinical records, medical images, genetic data, and data streams produced via mobile applications and wearable devices.

Bottom-up, evidence-oriented analysis

One application of this data deluge is the development of model-guided personalised medicine. In developing personalised medicine, bottom-up (evidence-oriented) analysis is of fundamental interest. Such analysis attempts to identify latent factors (or disease signatures) that can explain and predict similarities and variabilities in drug therapies and disease evolution. Bottom-up analysis require:

1) data integration from heterogeneous sources

 2)scalable analytics

To this end, our research team is developing a Knowledge Discovery and Data Mining (KDD) platform, called AITION, for biomedical knowledge discovery, feature selection, vertical integration, and semantic modelling under uncertainty. In addition, we utilise Athena Distributed Processing (ADP) middleware providing advanced scaling capabilities, as well as, federated data access that supports a versatile execution of distributed algorithms on ad-hoc clusters and clouds.

The AITION KDD platform

The AITION KDD platform consists of three modules:Data Curation and Validation (DCV), Clustering and Simulation.

The DCV module provides advanced techniques for data validation and preprocessing; checks for inconsistencies, missing values and outliers; computes medical scores; performs attribute discretisation, and more.

The clustering module is a set of advanced clustering and similarity analysis techniques aiming to identify latent factors (disease signatures) and to group homogeneous patients. Following a mixed membership approach, a patient is characterised by a specific distribution (allocation) on multiple, latent disease signatures. Patient similarity is then computed by comparing such allocations using several metrics.nc

The simulation module analyses correlations between variables, discovered latent factors and groups to deliver accurate and reusable predictive statistical simulation models based on Graphical Probabilistic Models (GPMs).

It implements state-of-the-art algorithms for Bayesian Network (BN) Structure & Parameter Learning, Markov Blanket induction and feature selection and real-time inference. The system offers a user-friendly interface to explore the graphical models, as shown in Figure 1. In addition, ontologies and a-priori knowledge can be incorporated, supporting a top-down (model driven) process that complements bottom-up (evidence oriented) analysis, providing a rich ‘natural’ framework for semantic modelling under uncertainty, where a domain expert is able to ‘seed’ the learning algorithm with knowledge about the problem.

Federated data access and scaling utilising ADP

Over the last five years, our group has been building Athena Distributed Processing (ADP), a system for distributed data processing on the cloud.ADP works to harness the power of cloud computing by defining language abstractions to declaratively express complex computation, and by designing an architecture with clear separation into components with well-defined semantics. ADP offers a high-level language called ADP Query Language, which is based on SQL enhanced with user-defined functions and a new syntax that makes them easy to use. Thirty years of database technology has shown that declarative languages are important because they offer data and platform independence.

The functionality of the system can be extended with new user-defined functions, which can be as complex as needed. We offer a rich library of user-defined functions to support the AITION platform, including data import (CSV, XML), statistics (Pearson correlation), and more.

Authors: Harry Dimitropoulos, Anna Gogolou, Herald Kllapi, Omiros Metaxas, Lefteris Stamatogiannakis, Eleni Zacharia, Yannis Ioannidis, and MD-Paedigree EU-Project I.M.I.S., Research Centre “Athena”, Athens, Greece National & Kapodistrian University of Athens, Greece

This article was originally published in Md-Paedigree Newsletter- Isssue 2

md-paedigree twitter page
md-paedigree secure
md-paedigree call
MD-Paedigree Infostructure