What is statMed.org?
statMed aims to improve differential diagnosis in those studying medicine.
At it's heart statMed.org is a database of hundreds of conditions. For each condition there is detailed epidemiology data, clinical features and risk factors. For any given presentation (e.g. 50 year old male, several weeks history of abdominal pain and weight loss) it applies Bayesian mathematics to calculate a differential diagnosis, listing the conditions in decreasing order of likelihood.
We are a ".org". statMed.org is free to use and always will be. It is in no way affiliated with any pharmaceutical company or private health provider.
Remind me, what is Bayes' theorem?
Bayes' theorem is a method of revising a probability given additional information.
There are hundreds of articles, YouTube videos etc that do a better job than I can at explaining Bayes' theorem. The following links may be useful:
- Simple Wikipedia explanation
- 'Statistics How To' - includes medical examples with working
- Engaging YouTube video with worked example
Is using Bayes' theorem to generate a differential diagnosis a new concept?
No. For over 50 years doctors and statisticians have understood the potential of Bayesian inference in medical diagnosis. An example paper from 1967 can be seen here
Is statMed.org an example of Artifical Intelligence (AI)?
It really depends on how you define AI. If statMed.org were a commercial product it would likely be marketed as using AI. We personally don't feel this is a particularly helpful descriptor of statMed.org - the results produced are consistent, based on data from human experts and derived from relatively simple calculations. The system certainly doesn't learn ("machine learning") - improvements will only follow human intervention.
What are the advantages of using Bayesian inference to aid medical diagnosis?
We are all subject to a large number of biases in everday life. This is no less true when we are formulating a differential diagnosis. Examples of common biases include:
- Confirmation - Assigning preference to findings that confirm a diagnosis or strategy
- Framing - Assembling elements that support a diagnosis
- Availability - Referring to what comes to mind most easily
- Premature closure - Failing to seek additional information after reaching a diagnostic conclusion
Overall, up to 50 types of bias have been identified in diagnostic formulation. Using techniques such as Bayesian inference can help reduce such biases.
What are some of the potential limitations of the system?
One of the potential limitations is where the inputted features have low levels of independence from each other. An example would be hypotension and tachycardia. Most patients (but not all) who are hypotensive will mount a tachycardia as a compensatory mechanism - they are clearly not independent features. This can result in the algorithm overestimating the likelihood of a condition - sometimes referred to as 'overfitting'.
How can I contribute?
We'd love to know what you think about statMed. What features should be added? Have you spotted any mistakes or bugs? In particular, it's very useful if you let us know about any calculated differential diagnoses that don't include relevant conditions.