Approaches in Observational Research Our Start in Constructing Prior Knowledge Current Research in Quantifying the Knowledge Gap Our Future Research on Systematically generating Prior Knowledge
Traditional Epi Approach: An expert defines the study design, including the covariates, and asserts that the results would be unbiased. Expert knowledge driven approach Traditional Epi Approach: An expert defines the study design, including the covariates, and asserts that the results would be unbiased. Expert knowledge driven approach OMOP/OHDSI Approach: Try all designs and covariates, let the data decide. Data driven approach
Could we combine prior knowledge with a data driven approach? Could we combine prior knowledge with a data driven approach? - We do not want to rely on a single expert to provide us the prior knowledge
- The data driven approach may not always be interpretable given covariates or sample size
- There needs to be a systematic way to capture and represent prior knowledge
Drug A causes/does not cause Condition B Covariate C causes/does not cause Condition B Both may be expressed as probabilities
We can construct systematic representation of prior knowledge about causal effects by: We can construct systematic representation of prior knowledge about causal effects by: - Manual construction of reference sets
- (Semi-) Automatic construction by using a statistical model synthesizing available evidence
List of drugs and condition pairs that are either considered: List of drugs and condition pairs that are either considered: - Positive Controls: in an adverse event relationship
- Negative Controls: not in an adverse event relationship
EU-ADR Reference Set AZCERT Dataset
4 conditions (acute kidney injury, acute liver injury, acute myocardial infarction, and gastrointestinal bleed) 4 conditions (acute kidney injury, acute liver injury, acute myocardial infarction, and gastrointestinal bleed) Across the 4 conditions: - 165 positive controls (drugs that are known to cause a condition, ground truth is 1)
- 234 are negative controls (drugs that are known to not cause the condition, ground truth is 0).
Process for Positive Controls: Process for Positive Controls: - Drug product labels that listed condition in the ‘‘Black Box Warning” section, the ‘‘Warnings and Precautions”, or ‘‘Adverse Reactions” sections.
- Independent literature review
- Systematic literature review provided by Tisdale and Miller book “Drug-Induced Diseases: Prevention, Detection and Management”
Process for Negative Controls: - Same process but looking for a lack of evidence across all sources
Exploring and Understanding Adverse Drug Reactions (EU-ADR) Exploring and Understanding Adverse Drug Reactions (EU-ADR) 10 conditions (liver disorder, acute myocardial infarction, renal failure, acute, anaphylactic shock, erythema multiform, mitral valve disease, neutropenia, aplastic anemia, rhabdomyolysis, and gastrointestinal hemorrhage) Across the 10 conditions: - 43 are positive controls
- 50 are negative controls
Conditions considered important from a pharmacovigilance/public health perspective Conditions considered important from a pharmacovigilance/public health perspective Drugs needed enough exposure in the EU-ADR database network Literature review to find associations between drugs and conditions. - Positive Control: 3 drug-ADR association citations found in MEDLINE
- Negative Control: No literature citations and no World Health Organization (WHO) Vigibase mentions
Manual review
Arizona Center for Education and Research on Therapeutics (AZCERT) Arizona Center for Education and Research on Therapeutics (AZCERT) The CredibleMeds group, which manages the AZCERT list, focuses on programs to reduce preventable harm from medication. This DB focused on two condition: - Torsade De Pointes (TDP)
- QT Prolongation.
Only contains positive controls.
First built a model and evaluated performance on individual reference sets EU-ADR and OMOP Second built a model on the combination of EU-ADR and OMOP and evaluated performance on a third reference set AZCERT
Now that we are extracting knowledge, we want to learn what evidence already exists and what evidence does not Now that we are extracting knowledge, we want to learn what evidence already exists and what evidence does not This will start to give us a better feel for how to extract evidence as prior knowledge
Prior knowledge can be used in (at least) two ways: Prior knowledge can be used in (at least) two ways: - As a reference set in an empirical evaluation, to measure how unbiased a method really is (previous work group meeting presentation by Martijn)
- To formulate explicit prior distributions in our models (future research project)
Frequentists Statistics – there is no prior belief, for example about effect size, any result is deemed equally likely. Frequentists Statistics – there is no prior belief, for example about effect size, any result is deemed equally likely. Bayesian Statistics – here there is a prior belief and some results are deemed more likely than others. The evidence generated by a study can modified our prior belief into a posterior.
Experts picking what goes into the model (e.g. Expert says “smoking is a risk factor for lung cancer”) Experts picking what goes into the model (e.g. Expert says “smoking is a risk factor for lung cancer”) Throwing everything into the model (e.g. the model would need to learn that smoking is a risk factor for lung cancer) Still throw everything into the model but use an informed prior to help the model learn (e.g. make the prior for smoking and lung cancer would be given a high probability)
Where we have come from: Not having an automated approach to utilize prior knowledge Where we have come from: Not having an automated approach to utilize prior knowledge What we can do currently: Using an automated approach to identify known drug/outcome relationships (aka negative controls) to calibrate our methods What we are working on now: Improving our evidence extraction by exploring the evidence gap Where we want to go: Using systematically generated prior knowledge to inform prior distributions in our models.
Dostları ilə paylaş: |