IST-2004-027173 Eu Resist : An integrated system for management of antiretroviral drug resistance Francesca Incardona (Informa s.r.l.)
Eu Resist : to support clinicians treating HI V patients The Eu Resist project aims at developing an integrated system for prediction of response to antiretroviral treatment Novel approach: viral genotype data integrated with clinical data. Focus is on genotype - A critical amount of response correlation resistance data is needed. An integrated and comprehensive genotype-response database has been created. Several distinct prediction engines have been developed and combined into the Eu Resist Prediction System. Started: January 1st 2006 Will end: September 30th 2008 Arevir 2008 2
Eu Resist consortium Informa s.r.l. University of Siena Coordinator Scientific coorrdinator Max Plank Institute University hospital of Cologne IBM Haifa Research Lab. Karolinska Institute RMKI (Hungary) Kingston University Roma 3 University Subcontractor Arevir 2008 3
Eu Resist objectives The project specific aims were: � To create the Eu Resist Integrated DataBase by merging three large resistance data sets: ARCA (Italy), AREVIR (Germany) and Karolinska’s (Sweden) � To define a ‘standard datum’ aimed at determining the minimum number of variables that maximise the information � To study different methods to build the predictive engines � To compare and combine the different methods into the final Eu Resist Predictive System To make the final System freely available on the Web Arevir 2008 4
5 Eu Resist System schem a KI Arevir 2008 AREVIR ARCA
The I ntegrated Eu Resist DB Integrates clinical and virological data from multiple sources Figures as of March 2008 (simplified schema) THERAPY AIDS EVENTS PatientID PatientID GENOTYPE Treatment regimen Event PatientID 6 4 .8 6 4 Date of start Date Date 2 2 .0 0 6 Date of stop Sequence Reason for change/stop PATIENTS Method Patient ID CD4 Gender HIV RNA PatientID 1 8 .4 6 7 STATUS Year of birth PatientID Date 3 0 4 .8 3 8 PatientID Country of origin Date CD4/mmc Date 2 4 0 .7 9 5 HCV status Copies/ml CD4% Followed HBV status <LLD (undetectable) Lost Method Died Arevir 2008 6
The I ntegrated Eu Resist DB Integrates clinical and virological data from multiple sources A contribution to health information standards has been defined (HL7) based on Eu Resist DB data structure Arevir 2008 7
Eu Resist “Classical” Standard Datum CD4 Reason for change Viral load Treatment switch Genotype Viral load time 0 to 12 weeks Short-term model: 4-12 weeks Pre-therapy HIV RNA The Forum for Collaborative HIV Research Patient demographics (age, gender, race, route of infection) Past AIDS diagnosis … plus “derived” features (e. g. the Past treatments “genetic barrier” defined as the Past genotypes probability not to develop resistance to the drugs included in the regimen) Arevir 2008 8
The prediction engines An array of independent prediction engines based on different models has been realised: � Instance Based Reasoning (IBR): local fitting procedure which selects compact subsets of predictive variables: large amount of data is crucial � Generative-discriminative engine: global fitting method employs first a generative model that uses all data and then applies Kernel method (or Support Vector Machines) for prediction � Evolutionary model: includes genetic evolutionary information into derived features (not in the SD) and uses different machine learning techniques for prediction � Fuzzy logic: an existing predictor retrained on the Eu Resist IDB to generate derived features (not in the SD) for the IBR engine Arevir 2008 9
Success/ failure prediction performance AUC Accuracy = (1 – Error rate) Train Test Train Test I BM 0.747 (0.027) 0.744 0.745 (0.024) 0.724 Maximal Minimal MPI 0.766 (0.030) 0.768 0.754 (0.031) 0.748 Feature Set feature set I nforma/ performs 0.758 (0.019) 0.745 0.748 (0.031) 0.757 RM3 better than I BM 0.768 (0.025) 0.76 0.752 (0.028) 0.757 minimal Maximal MPI 0.789 (0.023) 0.804 0.780 (0.032) 0.751 Feature Set I nforma/ 0.762 (0.021) 0.742 0.754 (0.030) 0.757 RM3 Arevir 2008 10
The Eu Resist system prediction results Engines com bination : after exploring several methods, simple mean combination has been chosen Results � Mean combiner learns faster than single engines � Performs better than current state of the art (comparison with Stanford HIVDB) Arevir 2008 11
12 The EVE evaluation study Arevir 2008
13 The EVE evaluation study Arevir 2008
14 The Eu Resist Web interface I nput Arevir 2008
15 The Eu Resist Web interface - input I nput Arevir 2008
16 The Eu Resist Web interface Output Arevir 2008
17 The Eu Resist Web interface - Output Arevir 2008
The future � Eu Resist Web based free decision support system on- line � Clustering: Eu Resist is already expanding from the 3 initial databases and entering new collaborations based on reliable and fair rules � Eu Resist network GEIE: A European Grouping to deploy project results Arevir 2008 18
Clustering � Eu Resist Network partner in upcoming CHAIN project (VIIFP Health Programme) � Luxembourg clinic joined the Eu Resist IDB Join us! if you want to collaborate please contact me or visit the web site www.euresist.org � Cooperation with Virolab project (www.virolab.org) Arevir 2008 19 19
Clustering � EuResist provided data to Virolab for projects based on Virolab+ Eu Resist data: Risk factors of accumulation of resistance during failing treatment Influence of primary resistance mutations or substitutions on CD4+ T-cell count evolution among HIV-1 positive patients while naïve to antiretrovirals Evaluation of the predictive performance of fitness landscapes for therapy outcome of baseline estimated fitness and genetic barrier towards resistance Quantification of virological and immunological response of decision support systems Virolab is next to provide data to Eu Resist Arevir 2008 20 20
21 Rules for participation Arevir 2008
Eu Resist Network GEI E � ESTABLI SHED! � A European Grouping (Informa, UniSiena, Max Plank, Karolinska - UniKoeln next to join) to deploy project results, maintain and update the IDB and the prediction system � It will (hopefully!) collect financial support from private companies and/or governmental institutions to carry on Eu Resist activities � The Grouping is a partnership without profit goals � It will give free services to the public Arevir 2008 22
Main results at now Technical The IDB with more than 18.000 patients The Eu Resist prediction system performing better than current state of the art (Stanford HIVDB) Scientific � Clinical, immuno-virologic, therapeutic and socio-demographic features in addition to viral genotype, as well as derived features, improve prediction results � Prediction results seem not to be significantly improved just by further increasing the training data size, given models and features. � Standard datum to be reformulated with long-term model? Strategic � Eu Resist Network GEIE � Clustering Arevir 2008 23
Aknowledgements to: � Maurizio Zazzi (University of Siena) � Andre Altmann (Max Plank Inst.) � Mattia Prosperi (Informa s.r.l.- University of Roma3) � Monica Merito (Informa s.r.l.) � All Eu Resist team Thank you f.incardona@informacro.info – www.euresist.org Arevir 2008 24
Recommend
More recommend
