close window ANNUAL CONFERENCE AND EXHIBITION APRIL 4-8, 2009 / CHICAGO www.himssconference.org View PowerPoint Presentation Print PowerPoint Presentation Roundtable 304 Predictive Informatics: What Is Its Place in Healthcare? Christophe G Giraud-Carrier, PhD Associate Professor Brigham Young University
Predictive Informatics What Is Its Place in Healthcare? ----------- Christophe Giraud-Carrier Brigham Young University
Roundtable Format Introduction Questions to be addressed Preliminary answers and thoughts Discussion Implementation issues Wrap-up
Questions What is predictive informatics? How does it differ from predictive analytics as used in the business world? What can be learned from the longer-standing experience of the business world? Are there unique challenges posed by mining medical/healthcare data and if so, what are they? What are specific ways in which predictive informatics may be used? Are there areas/domains where predictive informatics is not appropriate? How do we implement predictive informatics?
A Working Definition Predictive informatics is concerned with the collection of medical data, and its transformation into actionable knowledge to guide decision-making, with a view to personalized, real-time healthcare
PI Process doctor's office hospital pharmacy...
Differences with Business Data storage practices Long tradition (DB, DWH, etc.) vs. more recent Nature of the data ( genomic Disparate data types (e.g., images, heredity, Different privacy sensitivity Nature and use of the results of analysis Economic vs. life-or-death outcomes Competing actors: insurance, physician, patient Causation vs. correlation Yet, much can be learned... There is a long tradition of storing business data on electronic media and in structured form (i.e., transaction systems, databases, data warehouses), which is not the case in healthcare The kinds of data being manipulated differ significantly, from relatively insignificant transactional information to rather sensitive medical details The way the results of mining are evaluated and used is rather different, with business results being easily quantifiable and generally targeted at economic values (e.g., profit) and medical results having qualitative components and sometimes focusing on less tangible outcomes (e.g., quality of life), including death. Also, different actors have different interests in the results, with different impact in patients (hippocratic oath: never do harm to anyone ). In many situations, correlation is sufficient ground for action in the business world, while causality must generally be strictly enforced in the medical domain
Lessons Learned Involve domain experts throughout the process Proceed with caution Understand the nature (and limitations) of observational studies Keep track of procedures, decisions and results Think of novel ways to frame your problem Do not underestimate data preparation Involve domain experts: extracting knowledge from data requires domain expertise (understanding the domain, asking the right questions, (. etc evaluating the results, understanding implication of findings, Proceed with caution: start small; beware of results that appear too good ( are to be true, they probably Understand the nature (and limitations) of observational studies: correlation may not be sufficient in the medical domain; but it may be helpful to point in the direction of useful RCTs Keep track of procedures, decisions and results throughout the mining process: documenting work is critical to reproducibility, one of the tenets of good (medical) research Think of novel ways to frame the problem: breakthroughs occur when traditional knowledge is challenged Do not underestimate data preparation: up to 80% of the efforts has been reported to go to that activity
Unique Challenges Ethical, legal and social issues Abundance of unstructured data, including text and images Lack of canonical form Need to focus on sensitivity and specificity rather than traditional accuracy Data sharing across healthcare actors Abundance of missing values and their adequate treatment Resistance Ethical, legal and social issues: data ownership, fear of lawsuits, liability to the system, privacy and security Abundance of unstructured data, including text and images: doctor's notes, X-ray, MRI, nurses reports, etc. Lack of canonical form (e.g., heart attack, myocardial infarction, etc.): some attempts at addressing this do exist though, e.g., ICD9/ICD10 Need to focus on sensitivity and specificity rather than traditional accuracy: cost sensitive approaches, over-sampling vs undersampling, etc. Life-or-death outcomes vs. economic outcomes (?) Data sharing: national health record Abundance of missing values and their adequate treatment Resistance: Aristotelian approach (first understand nature of disease) vs. evidence-based approach; Yet: Semmelweis' experience, Berwick's 100,000 Lives Campaign, etc.
Sample Applications General ideas: Literature-based medical discovery Supplement randomized controlled trials Introduce point-of-care information in medical studies Leverage online communities Specific examples Bariatrics Others LBM: Swansons' ABC model of discovery, Raynaud's disease and fish oil, connection between wheezing and diabetes Observational databases are not good sources of information on which treatments are effective and which are not. There are too many confounding factors. They may provide useful clues on what trials are worth considering. Enrich medical studies with information gathered at the point of care Neo-tribe effect: online health communities, more global view of diseases and treatments (beyond single practice), rare symptoms Bariatrics: surgeon preference is main determinant of type of surgery performed, level of success may be predicted Others: TBD
The field is white Discussion Launch Predictive informatics presents unique challenges Predictive informatics offers unique opportunities to improve healthcare How do we promote and implement predictive informatics? Organizational issues Technical issues Etc. The field is white: more systematic data collection, syndication (e.g., EMRs), advances in technology make predictive informatics increasingly applicable to healthcare How do we promote and implement? Are there any domains/areas n which predictive informatics is inappropriate? How do we know? What is most likely to work? What does it mean?