Can AI Help Predict Drug Shortages?

In 2021, while working as a pharmacist intern on call in the accident and emergency unit of a federal teaching hospital in southeastern Nigeria, I encountered a drug shortage that has stayed with me. A patient with severe preeclampsia needed magnesium sulfate, but the hospital did not have it in stock, and nearby pharmacies were closed. Despite the team’s efforts, we could not provide the medication. As a result, the patient died.

That experience left me with a question I still think about: Could we have identified the supply gap before the patient needed the drug?

Drug shortages look different in the United States. Pharmacy shelves may not be visibly empty, but pharmacy teams absorb disruptions through therapeutic substitutions, compounding, delayed procedures, emergency purchasing, and hours spent contacting wholesalers.

The problem remains substantial. The American Society of Health-System Pharmacists (ASHP) reported 227 active drug shortages at the end of the second quarter of 2026, up for the third consecutive quarter but below the record of 323 in the first quarter of 2024.¹ A 2025 Vizient survey reported that US hospitals spent approximately 20 million hours managing drug shortages in 2023, equivalent to nearly $900 million in annual labor costs.² These figures capture only part of the burden. Shortages can also complicate treatment decisions, disrupt medication selection, and delay care.

Testing What Public Data Can Tell Us

As a pharmacist and MBA candidate studying health care, operations, and business analytics, I examined whether publicly available information could help distinguish drugs associated with shortages from those not listed as being in shortage.

I assembled a dataset covering approximately 3,000 prescription drugs using US FDA drug listings and recall records, alongside shortage information from FDA and ASHP sources. For each drug, I recorded characteristics including dosage form, injectable or oral status, the number of approved products, brand or generic status, controlled-substance status, and how recently a new product had entered the market.

Then, I compared models designed to distinguish drugs appearing on shortage lists from those that did not. A fuller technical report is in preparation.

The initial findings suggested that these characteristics contained some useful information about shortage-list status. However, the signal was modest, and the analysis does not establish whether the models can reliably identify drugs that will enter shortage in the future. Identifying characteristics associated with listed shortages is not the same as forecasting when a new shortage will begin.

Some of the strongest risk signals were familiar to pharmacy professionals. Sterile injectables, controlled substances, and long-established drugs with multiple approved products featured prominently among higher-risk drugs. A relatively simple scoring rule based on these characteristics captured much of the signal identified by more complex models.

In this initial analysis, adding machine learning did not appear to deliver a substantial improvement over a simpler approach. Recall history, which I had expected to be informative, also contributed little.

These findings raise an important question about artificial intelligence (AI): Does a more sophisticated model necessarily produce a better operational decision? In some cases, identifying useful signals and presenting them clearly may matter more than algorithmic complexity.