Post on 17-Feb-2019
© 2012 IBM Corpora.on
Boaz Carmelia, Paolo Casalib, Esther Goldbraicha Abigail Goldsteena, Carmel Kenta, Lisa Licitrab, Paolo Locatellic, Nicola Restifoc, Ruty Rinotta, Elena Sinib, Michele Torresanib, Zeev Waksa
Cli-‐G: An Evidence-‐based Case Structuring Approach for Personalized Healthcare
a IBM Research – Haifa bFondazione IRCCS - Istituto Nazionale dei Tumori c Fondazione Politecnico di Milano
Oral Presenta+on MIE
Pisa, August 2012
© 2012 IBM Corpora.on
Clinical Decision Support (CDS) § CDS systems have a great poten.al improving health care.
– More treatment op.ons – Complex ‘omic’ data
– Personalized medicine § Most exis.ng CDS rely on domain knowledge – Clinical trials à Clinical guidelines à simple CDS rules
© 2012 IBM Corpora.on
Clinical Decision Support (CDS) § CDS systems have a great poten.al improving health care.
– More treatment op.ons – Complex ‘omic’ data – Personalized medicine § Most exis.ng CDS rely on domain knowledge – Clinical trials à Clinical guidelines à simple CDS rules
§ Penetra.on of EHRs enable data driven CDS systems ▫ Many pa.ents ▫ Diverse popula.on (co-‐morbidi.es, elderly) ▫ New treatments ▫ New Types of data (Gene.c markers)
© 2012 IBM Corpora.on
MLDM techniques benefit from domain knowledge
Data
• Hospitalization records
• Drug approvals
• Drug supply records
• Lab tests
• Prior medications
• Personal information
* not all drugs need approval
* supply dates are unreliable
3 Possible treatments:
• Bevacizumab (Avastin), 5FU
• Bevacizumab (Avastin), 5FU, Oxaliplatin
• Bevacizumab (Avastin), 5FU, Irinotecan
Mapping by prior knowledge
of possible treatment
§ Advanced colorectal cancer: comparing first line treatments
© 2012 IBM Corpora.on
Cli-‐G: Clinical Genomics (Cli-‐G) analy.cs for oncology care appropriateness § Decision support for Treatment Alloca.on
§ Cli-‐G Combines: – Evidence Based Medicine (Guidelines) – Methodical Cased Based Reasoning by Data Mining and Machine Learning Analy.cs over HMO records
§ Since 2010 working with Italian Na.onal Cancer Ins.tute in Milan to develop CDS for SoW Tissue Sarcoma
§ Analyzes ~2000 electronic discharge leYers, (~1000 pa.ents) containing free text and coded fields.
© 2012 IBM Corpora.on
Clinical Problem § Clinical problem is associated with three basic building blocks:
Patient’s Diagnosis/Clinical Status: genomic, clinical, tumor markers
Standard
Outcome: recurrence, survival rate, side effects …
Presentation
Optional Treatments
Outcome
Adult Soft Tissue Sarcoma,
Critical presentations
Individualized Experimental (clinical trial)
Surgery Chemo Radio Surgery+ chemo
Chemo+ radio
E471
© 2012 IBM Corpora.on
Cli-‐G inside: Integra.ng knowledge, data and analy.cs for decision support
Clinical Knowledge Management
Structured data
Evicase Generation & Management
Query
Evidence
Evicase
Clinical & Genomic Data (e.g. EHR)
Free text analysis
public knowledge subject matter expert guidelines
Analytics
Feature Ranking
Patient Similarity
Deviation Analysis
Adherence Analysis
Inconsistency Analysis
Cli-g Mart
Decision Support App
© 2012 IBM Corpora.on
Cli-‐G inside: Knowledge Management
Clinical Knowledge Management
Structured data
Evicase Generation & Management
Query
Evidence
Evicase
Clinical & Genomic Data (e.g. EHR)
Free text analysis
public knowledge subject matter expert guidelines
Analytics
Feature Ranking
Patient Similarity
Deviation Analysis
Adherence Analysis
Inconsistency Analysis
Cli-g Mart
Decision Support App
© 2012 IBM Corpora.on
Declara.ve knowledge – Declara've knowledge
• A set of ontologies – reflecting the presentation, treatments and outcome
© 2012 IBM Corpora.on
Procedural knowledge
Inputs Rules
Outputs
§ Procedural knowledge • A set of rules
– E.g., computing guidelines recommendation, cleansing rules, rules for enrichment etc.
© 2012 IBM Corpora.on
Cli-‐G inside: Data Integra.on
Clinical Knowledge Management
Structured data
Evicase Generation & Management
Query
Evidence
Evicase
Clinical & Genomic Data (e.g. EHR)
Free text analysis
public knowledge subject matter expert guidelines
Analytics
Feature Ranking
Patient Similarity
Deviation Analysis
Adherence Analysis
Inconsistency Analysis
Cli-g Mart
Decision Support App
© 2012 IBM Corpora.on
Integra.ng knowledge and data
What is the recommended treatment by guideline for my patient?
© 2012 IBM Corpora.on
Cli-‐G inside: Analy.cs
Clinical Knowledge Management
Structured data
Evicase Generation & Management
Query
Evidence
Evicase
Clinical & Genomic Data (e.g. EHR)
Free text analysis
public knowledge subject matter expert guidelines
Analytics
Feature Ranking
Patient Similarity
Deviation Analysis
Adherence Analysis
Inconsistency Analysis
Cli-g Mart
Decision Support App
© 2012 IBM Corpora.on 14 Personalized Oncology – Beyond Standard Care March 20, 2012
Most similar patients that were
treated with Ifosfamide
Comparing predictive results
for each treatment based on outcome for similar patients
Most similar patients that were
treated with Trabectedin
Elderly people receive Standard treatment no reference to age is available in guidelines
>70
Adherence Level by Age
Standard Individual Experimental Deviation
© 2012 IBM Corpora.on
Cli-‐G inside: Evicase Genera.on
Clinical Knowledge Management
Structured data
Evicase Generation & Management
Query
Evidence
Evicase
Clinical & Genomic Data (e.g. EHR)
Free text analysis
public knowledge subject matter expert guidelines
Analytics
Feature Ranking
Patient Similarity
Deviation Analysis
Adherence Analysis
Inconsistency Analysis
Cli-g Mart
Decision Support App
© 2012 IBM Corpora.on 16 Personalized Oncology – Beyond Standard Care March 20, 2012
Guidelines Recommendation: Declarative and Procedural knowledge
Statistics for patients with similar presentation: knowledge and data
Features separating between patients receiving standard treatment and deviation: Knowledge, data and analytics
Calculating adherence level: Declarative and Procedural knowledge
© 2012 IBM Corpora.on
Summary
§ Cli-‐G -‐Decision support system combining evidence with cased based reasoning
§ Integrates knowledge and pa.ents data gathered at HCO § Can be applied to different decision points, by changing relevant parts of knowledge model
§ System is currently under evalua.on at the Italian Na.onal Cancer Ins.tute in Milan, many features s.ll under development
© 2012 IBM Corpora.on
Acknowledgements
§ Cli-‐G team (IBM) Boaz Carmeli Esther Goldbraich Abigail Goldsteen Carmel Kent
Zeev Waks
§ Is.tuto Nazionale dei Tumori Paolo Casali Lisa Licitra Elena Sini Michele Torresani
§ Fondazione Politecnico di Milano Paolo Locatelli Nicola Res+fo
© 2012 IBM Corpora.on
Clinical Problem § Decision point is associated with three basic building blocks:
– Pa.ent’s presenta.on – the clinical status – Possible therapies/procedures – Pa.ent’s outcome
Patient’s Diagnosis/Clinical Status: genomic, clinical, tumor markers
Surgery: wide excision Surgery: compartmental resection + Chemotherapy
Outcome: recurrence, survival rate, side effects …
Presentation
Optional Treatments
Outcome
Adult Soft Tissue Sarcoma,
Critical presentations
Surgery: wide excision + Radiation Therapy
. . .
© 2012 IBM Corpora.on
Oncology care appropriateness views
Organiza'on View – a high level view of the pa'ent popula'on
§ Analyze guidelines adherence levels
§ Present actual treatments assigned, and associated outcomes
§ Highlight features informa.ve about treatment alloca.on
§ Highlight features informa.ve about guideline devia.on
Physician View – decision support for a specific pa'ent in a personalized manner
§ Present the guideline recommenda.on this pa.ent according to presenta.on
§ Iden.fy similar pa.ents
§ Analyze treatment distribu.on among similar pa.ents.
§ Predicts outcome for possible treatments, based on results for similar pa.ents
© 2012 IBM Corpora.on
MLDM techniques benefit from domain knowledge
Data
• Hospitalization records
• Drug approvals
• Drug supply records
• Lab tests
• Prior medications
• Personal information
* not all drugs need approval
* supply dates are unreliable
3 Possible treatments:
• Bevacizumab (Avastin), 5FU
• Bevacizumab (Avastin), 5FU, Oxaliplatin
• Bevacizumab (Avastin), 5FU, Irinotecan
Mapping by prior knowledge
of possible treatment
§ Advanced colorectal cancer: comparing first line treatments
© 2012 IBM Corpora.on
Clinical Decision Support (CDS)
Primary Cancer
Grade Grade
Surgery: compartmental
resection
Surgery: wide excision + adjuvant radiation therapy
Depth
Surgery: wide
excision
Surgery: wide excision
Consider adjuvant radiation therapy
Yes No
Low
<5 cm >5 cm
High
Low High
© 2012 IBM Corpora.on
Cli-‐G outputs
Organiza'on View – a high level view of the pa'ent popula'on
§ Present actual treatments assigned, and associated outcomes
§ Analyze guidelines adherence levels
§ Highlight features informa.ve about treatment alloca.on
§ Highlight features informa.ve about guideline devia.on
Physician View – decision support for a specific pa'ent in a personalized manner
§ Present the guideline recommenda.on this pa.ent according to presenta.on
§ Iden.fy similar pa.ents
§ Analyze treatment distribu.on among similar pa.ents.
§ Predict outcome for possible treatments, based on results for similar pa.ents
© 2012 IBM Corpora.on
Knowledge Model for a Specific Clinical Problem § The model comprises of:
– Declara've knowledge • A set of ontologies
– reflecting the presentation, treatments and outcome
– Procedural knowledge • A set of rules
– E.g., computing guidelines recommendation, cleansing rules, rules for enrichment etc.