How can the process of finding causes be automated?
Which drugs will lead to harmful side effects when taken together?
Randomized trials testing the drugs will not tell us much, since these usually avoid having participants take multiple medications. Simulations can be used to predict some interactions, but they require a lot of background knowledge. We could test some pairs experimentally, but given the cost and time involved, that would be possible for only a small set of possible combinations. Even worse, out of the millions of possible pairings, only a few may interact severely and may only do so in certain populations.
However, after a drug is on the market, suspected adverse events are reported to the FDA by patients, pharmaceutical companies, and healthcare providers and entered into a database.1 So if you start taking a medication for allergies and have a heart attack a few days later, you could submit a report, as could your clinician. Now, these self-reports are not verified. It may be that a particular person’s heart attack was really caused by an unrelated blood clot, but a recent news story about many heart attacks due to the drug just made this explanation seem more salient. There are many ways the data can contain spurious causal relationships. A patient may have had other conditions that led to the outcome (e.g., undiagnosed diabetes), the data could be wrong (e.g., sample contaminated, condition misdiagnosed), and the order of events could ...