Using multimodal mining to drive clinical guidelines development Emilie Pasche 1 , Julien Gobeill 2 , Douglas Teodoro 1 , Dina Vishnyakova 1 , Arnaud Gaudinat 2 , Patrick Ruch 2 and Christian Lovis 1 1 SIMED, University of Geneva and University Hospitals of Geneva, Switzerland 2 Bibliomics and Text Mining Group, University of Applied Sciences, Geneva, Switzerland MIE 2011 - Oslo - 29 th of August Oral Presentation Presenter: Emilie Pasche Slide 1 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
DebugIT Detecting and Eliminating Bacteria Using Information Technology European project FP7 (grant #217139) with 14 partners. Disclaimer: this presentation reflects solely the views of the authors and no guarantee or warranty is given that it is fit for any particular purpose. The European Commission, Directorate General Information Society and Media, Brussels, is not liable for any use that may be made of the information contained therein. Slide 2 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Introduction Why we need to create clinical guidelines? Problem Antibiotic resistance is increasing because of inappropriate use of antibiotics Solution Development of clinical guidelines can help to regulate antibiotic prescriptions Slide 3 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Objective: help experts to author clinical guidelines How can we create clinical guidelines ? Without KART With KART Slide 4 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Methods How does KART work? 1. Query Query • Pattern-based query creation 2. Text-Mining TM • Rank answers using question-answering 3. Multimodal-Mining MM • Re-rank answers using source clinical data 4. Evaluation Eval • Evaluate answers using IR metrics (TREC) Slide 5 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Step 1. Query Query Manual creation of a benchmark Query Antibiotic 1 Antibiotic 2 Manual … translation and Query normalization 23x Antibiotic 1 HUG Antibiotic 2 Guidelines Query … Antibiotic 1 72x Antibiotic 2 … 49x Slide 6 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Step 2. Text-Mining TM System architecture of Automatic Question Answering Information Answers Query retrieval extraction Relevant Antibiotic 1 Corpus documents (Medline) Antibiotic 2 (50 docs) … Search engine Terminologies (easyIR, PubMed) (WHO-ATC) Slide 7 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Step 3. Multimodal-Mining MM Multimodal model Query Re-ranking Query Antibiotic 1 Antibiotic 1 Antibiotic 3 Antibiotic 2 … … Costs Costs Costs Costs Resistance Costs (17 subst.) (17 subst.) (17 subst.) (17 subst.) profile s (70 subst.) Slide 8 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Step 3. Multimodal-Mining MM Getting additional features: antibiotic costs Data Data Costs Costs completion normalization Costs Costs Costs (17 subst.) (17 subst.) (70 subst.) (129 prod.) (17 subst.) Arbitrary value (0 – 100) Prescription data (HUG) Slide 9 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Step 3. Multimodal-Mining MM Getting additional features: HUG’s resistance profiles Data Extract Costs Costs completion antibiogram Resistance Resistance (17 subst.) Clinical (17 subst.) profile s profiles Data Repository SPARQL queries Arbitrary value (species - antibiotic) (0 – 1) Slide 10 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Step 4. Evaluation Eval Experimental settings Query Antibiotic 1 Antibiotic 2 … Evaluation Results Benchmark Query Antibiotic 1 Antibiotic 3 … Automatically- TREC-EVAL generated Slide 11 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Results How well does KART perform ? Antibiotic costs: Resistance profile: • EAGLi/easyIR: + 9% • EAGLi/easyIR: + 5.5% • PubMed: - 0.1% • PubMed: + 16% Baseline Baseline Costs Costs Resistance Resistance (easyIR) (Pubmed) (easyIR) (Pubmed) (easyIR) (PubMed) Answers 49/49 32/49 49/49 32/49 49/49 32/49 Top 34.28% 40.37% 43.31% 40.28% 39.86% 56.41% precision Slide 12 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Limits and future works Costs • Currently based on a limited set of costs – HUG costs list (17/70 substances) • We could use broader resources – Swiss Kompendium (all substances) Resistance • Currently based on species-specific antibiograms – E.g. 2 antibiograms for S. pyogenes + clindamycin • We could use aggregated species – E.g. 75 antibiograms for all Streptococcus + clindamycin Slide 13 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Conclusion • Facilitates clinical guidelines development by extracting hypothetical treatments from literature – E.g. Pneumonia and Streptococcus pneumoniae • 4855 publications in MEDLINE • 12 proposed antibiotics in KART • Combining literature-based discovery with clinical data mining can significantly improve authoring of clinical guidelines – 56% of top-ranked answers are correct Slide 14 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Acknowledgments DebugIT, EU-IST-FP7-217139 DebugIT partners EAGL, SNF-325230-120758 • Agfa Healthcare (Belgium) • Empirica (Germany) Infectious disease service (HUG) • Gama Sofia Ltd (Bulgaria) • Angela Huttner • INSERM (France) • Marina Macedo • IZIP (Czech Republic) • Thomas Haustein • Linköping University (Sweden) • Stephan Harbarth • TEILAM (Greece) • University College London (UK) Consultant Physician (Australia) • HUG (Switzerland) • Garry Lane • Freiburg University (Germany) • Geneva University (Switzerland) KART: http://eagl.unige.ch/KART/ • Averbis (Germany) EAGLi: http://eagl.unige.ch/EAGLi/ • MDA (Czech Republic) • HES-SO (Switzerland) Slide 15 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Thank you for your attention emilie.pasche@unige.ch Slide 16 MIE2011 Oslo 29 th of August 2011 Presented by Emilie Pasche
Recommend
More recommend
