Applying the Area of Normal Functioning (ANF) to Pharmaceutical Safety Data

One of the goals of the Area of Normal Functioning (ANF) project is to test whether the same general framework can be applied across very different dynamic systems.

So far, I have been looking at ANF in technical systems, human behavior, sports and other time-dependent data. A natural next question is whether ANF can also be useful in pharmaceutical data.

A particularly interesting opportunity is pharmacovigilance — the analysis of adverse-event reports collected over time.

The basic idea is simple.

For a particular drug, we can construct a monthly time series of adverse-event reports and ask:

What does the normal reporting pattern look like, and when does that pattern change?

Instead of considering only whether a value is “high” or “low,” ANF attempts to characterize the range in which the system normally functions and then detect:

• excursions outside that range;
• persistent shifts;
• change points;
• new operating regimes;
• trends that may eventually move the system outside its established ANF.

A promising first example is montelukast, marketed as Singulair.

The FDA has documented a history of neuropsychiatric safety concerns associated with montelukast. It issued communications beginning in 2008 and later required a boxed warning in 2020. This gives us something especially useful for methodological validation: known historical events that can be compared with change points detected independently from the data.

The proposed experiment would use monthly adverse-event data from the FDA Adverse Event Reporting System, available through openFDA.

Rather than analyze only the total number of reports, I would like to construct at least three related time series:

1. total monthly montelukast adverse-event reports;
2. monthly montelukast reports involving selected neuropsychiatric events;
3. the ratio of neuropsychiatric reports to all montelukast reports.

The third series may be particularly useful because it can partially separate a change in the composition of reports from a general increase or decrease in overall reporting volume.

The main research question would be:

Can ANF and change-point analysis identify a structural change in montelukast adverse-event reporting without being told in advance when FDA communications occurred?

A strong version of the experiment would establish the initial ANF using only data available before 2008 and then process later observations sequentially.

That would let us ask a more interesting question:

At what point would the method first have indicated that the reporting process was no longer behaving within its previously established Area of Normal Functioning?

Detected change points could then be compared with known external events such as the 2008 FDA communication and the 2020 boxed warning.

It is important to be careful about interpretation. FAERS reports do not establish causality or true incidence rates. An increase in reports can result from many factors, including publicity, regulatory attention and stimulated reporting.

Therefore, the purpose of ANF would not be to conclude that a drug caused a particular effect.

The more defensible interpretation is:

ANF detects a change in the normal behavior of the pharmacovigilance reporting system.

Understanding why that change occurred is a separate analytical step.

This distinction is actually one of the reasons the example is interesting. ANF is designed to detect when a dynamic system moves into a different state, without assuming in advance what caused the transition.

If the experiment works well, I would next apply the same methodology to other known pharmaceutical events, such as the withdrawal of ranitidine and historical safety concerns around rosiglitazone.

Together, these examples could help determine whether ANF can provide a useful general framework for monitoring pharmaceutical and healthcare data in addition to technical and human systems.