The European Society of Cardiology guidelines and strategies such as TRAPID AMI aim to stratify patients into low ("rule-out"), intermediate & high risk ("rule-in") for AMI/ACS based on serial troponins, ECG, risk factors etc. For obvious reasons a >99% sensitivity is the defacto standard for rule-out. What then should be the specificity or ppv be? While ppv will vary according to prevalence, it matters to the cardiologists what proportion of patients they are told high risk actually have the disease. I'd like input, especially from cardiologists, on what they think is an acceptable ppv rate and why? Thanks.
Hi John,
I am a health services researcher, not a cardiologist, so my comment in more general. With a diagnostic test or predictive model, you want to strive to maximize both sensitivity and specificity. Setting these values at some threshold and calling them "acceptable" runs the obvious risk of missclassification for any given patient. The question is how much are you willing to missclassify a patient who really has (or at risk for) AMI, vs missclassifying them when they don't have (or not at risk for) AMI. There is an obvious price to pay for missclassification, and as such, the higher the sensitivity and specificity, the better. If either of these are low, then it makes sense to develop better models/tests that will improve classification.
I hope this helps
Ariel
Thanks Ariel. In this case the trend is to have two thresholds - one which allows a very high sensitivity (~99%) and one a high PPV. I'm interested in the cardiologist view on PPV & obviously your point on willingness to misclassifying those that don't have AMI is the important issue.
John,
As a follow-up, I just remembered that I had written a paper on this issue. You may want to glance at it.
Ariel
Article Measuring diagnostic and predictive accuracy in disease mana...
Thanks Ariel. I had a quick glance. That looks like it is at an excellent level for the med student - I keep a file of papers/references for med students for when they first engage in research, so I'll add that to it. As we are "swapping" papers I've added my paper on the NRI and IDI and risk assessment plots (the form in the paper is my own invention :) ) which I think are a useful addition to the ROC when evaluating if something improves a prediction model.
Article New Metrics for Assessing Diagnostic Potential of Candidate Biomarkers
A must read: Article Sensitivity, Specificity, and Predictive Values: Foundations...
I would like to ask one question! Now I do my research work is MRI image segmentation. When I evaluate the segmentation accuracy, the True positive and True negative rate are high but Predictive positive value is low. Actually, all third conditions must be high for good segmentation. Please, give the suggestion and honorable solutions for my evaluation results.
Aye Min. Your PPV (and NPV) will depend on the prevalence in your study. If you are reporting, then report the Sensitivity and Specificity if you want to discuss the possible transferability of your method to other cohorts. Your PPV could be improved by changing the test threshold that determines what a test positive and test negative is. However, it is possible that will decrease your Sensitivity.
What is the purpose of the test? If it is to be used to determine who receives a treatment that is expense of dangerous then you want to reduce the False Positives. However, if it is to rule-out the disease, then you want to retain a high sensitivity.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis