Challenge Proposal: Ensemble Learning of Adverse Drug Reactions With Lab Data and Known Drug Conflic

1
Challenge Proposal Ensemble Learning of Adverse
Drug Reactions With Lab Data and Known Drug
Conflicts

• George Runger, Wade Bannister, Diana Petitti
• ASU Center for Health Information Research
• Yan Zhang, Farooq Azam, and Creed Jones III
• Research Scientists at the Humana Center for
  Health Informatics
• Eugene Tuv
• Senior Staff Research Scientist, Intel, machine
  learning

2
Known Drugs/Conditions Conflicts

• Pharmacovigilance focuses on new, unknown drug
  conflicts
• Little work to identify people where known
  conflicts lead to adverse drug reactions (ADRs)
• Knowledgebase can be applied
• Laboratory data rarely used

3
Conflict Example and Sample Descriptions1

• Conflict criteria (600) based on research
  publications and FDA bulletins
• Example Digoxin and Clarithromycin
• In literature Digoxin toxicity can easily
  develop in patients simultaneously treated with
  clarithromycin because the latter inhibits
  P-glycoprotein, a multidrug efflux pump that
  promotes the renal clearance of digoxin.2
• Iatrogen Conflict 1102 Clarithromycin may
  induce digoxin toxicity by three different
  mechanisms including
• 1. Reduction of renal tubular P-glycoprotein
  secretion of digoxin
• 2. Alteration of intestinal flora
• 3. Inhibition of CYP-450-3A4 in the liver.

1. RxWise criteria, Iatrogen LLC, 2008 2.
Juurlink DN, Mamdani M, Kopp A, et al. Drug-Drug
Interactions Among Elderly Patients Hospitalized
for Drug Toxicity. JAMA. 2003 2891652-1658.
4
ADR and Known Drug Conflicts

• True ADRs are difficult to identify in
  retrospective studiesfrequently not coded
• Enhance a computational method with lab data to
  derive valid ADR information
• ADRs are rare events (even when conflicts are
  present)
• Small fraction of known drug-drug and
  drug-condition conflicts results in an ADR
• Supplement claims data (lab data, patient
  factors) to identify the important features for
  ADR

5
ADRs Expensive and More Prevalent

• Preliminary study of Medicare members (over a
  24-month period)
• Approximately 49 of people experienced a
  conflict identified by a knowledgebase
• 93 were drug-condition or drug-drug-condition
  , 6 were drug-drug, and 1 were drug-other
• Total number of conflicts 15.1M, top 25
  conflicts accounted for 68 of the total (were
  associated with 58 of the ADRs). None of these
  were drug-drug conflicts.
• Few conflicts convert into an ADR
• Number of ADRs detected by the conflict-driven
  methodology 2.4
• For the top 25 conflicts only 2.0 were
  associated with an ADR
• Simply signaling the top 25 conflicts would
  generate approximately 10 million false positives
  
• People carry multiple conflicts
• Among people with any conflict 30 carried 1
  conflict, 32 carried 2-3 conflicts, and 38
  carried 4 or more (and 21 carried 6).

6
ADR Models

• Develop a (machine learning) model for patient
  and/or laboratory data to predict ADRs for
  (myocardial infarction-related) events
• Exemplar for other ADR families
• Model would be used prospectively to provide
  early warning signals to prevent an ADR/signal a
  clinical intervention
• Example conflicts
• Pseudoephedrine and ischemic heart disease
• Cox2 inhibitors and cardiovascular disease
• Serotonin 5-ht1 receptor agonists and ischemic
  heart disease
• Stimulants and advanced arteriosclerosis
• Minoxidil and acute myocardial infarction
• Sulfonylureas and cardiovascular disease
• Risperidone and recent MI
• Thiazolidinediones and cardiac deterioration
• Erythropoesis stimulating agents and serious
  cardiovascular events

7
Share Expertise

• Factors to be considered include
• age, gender, geography, socioeconomic status,
  claims history, co-morbidities, clinical risk
  factors (obesity, etc.), average length of
  hospital stay, poly-pharmacy, poly-physician,
  medication possession ratio, drug switching,
  frailty, alcohol use, depression, number and type
  of procedures, and time-based behavior of each of
  these (velocity, acceleration, volatility), as
  well as lab test results
• Future biomarkers can be added to the strategy
• Share expertise among researchers in health
  informatics, statistics, and machine learning
• Leverage the clinical and domain knowledge and
  analytics experience from industry collaborators
  with machine learning expertise