In late September 2024, patients at Adventist Health in Bakersfield, California, noticed a nurse behaving strangely: walking barefoot through the intensive care unit, talking to herself, acting abrasively. Family members watched her remove IV needles sloppily and were afraid to confront her. A patient in the post-anesthesia recovery unit, in severe pain and believing he was receiving fentanyl and morphine, said the drip was doing nothing for him. "I was white knuckling it," he told federal investigators. When investigators from the Centers for Medicare and Medicaid Services examined the complaint two months later, they concluded the nurse, a travel hire who had been on the job only weeks, was taking drugs from a secured cabinet, using them herself, and documenting them as if they had gone to patients.
The detail that makes the case worth studying is not the theft. It is that the hospital's machine learning software, the kind designed to flag exactly this pattern, had produced alerts, and managers had not acted on them. The STAT investigation by Alexandra Byrne that surfaced the case uses it to make a point about human follow-through. The point goes further than the story says. The alert system at Bakersfield did not merely fail because a human was lazy. It failed because the entire architecture of diversion monitoring concentrates its value on the one step that was never automated, and then leaves that step to whoever is busiest.
The pipeline is automated until the last step
Diversion monitoring software is, by now, sophisticated. The systems pull data from dispensing cabinets, electronic health records, and personnel schedules, then look for the signatures of theft: withdrawals that cluster at shift changes, patients documented as medicated who report pain anyway, nurses whose dispensing patterns deviate from their unit's norm. The analytics are the product of years of research, including machine learning work published in the hospital pharmacy literature, and the vendors sell the capability in exactly these terms: the software finds the needle in the transaction haystack.
Everything the software does ends in the same artifact: an alert. From that point on, the system is no longer a machine. It is a queue, and the queue is processed by people, usually a pharmacist or a diversion specialist, who must open the alert, pull the records, judge whether the pattern is theft or noise, and escalate. That human step is the entire payload. The software does not stop a diversion. It asks a person to stop it. Every dollar spent on detection is an advance payment on the assumption that someone, on some shift, will read the flag and move.
At Bakersfield, per the federal auditors' account, the flags were there and the movement was not. The travel nurse worked for weeks while the queue waited. This is not a rare failure mode. It is the default failure mode of any system built as automation with a manual tripwire, because the tripwire is exactly where the workload concentrates and where no amount of model accuracy helps. A detector that is 99 percent accurate still depends, for all of its effect, on the person who answers the other 1 percent.
The cost of the broken link landed on a patient in pain
It is easy to book this story as an inventory problem, because that is how it shows up in hospital accounting: missing vials, lost drug charges, a nursing incident report. But the cost of the broken link was not paid in inventory. It was paid by the person in the recovery unit, who lay in unrelieved pain believing he was receiving medication that was going into someone else's bloodstream instead. The theft was of drugs, but the harm was to a patient who trusted that the chart meant what it said.
That is the uncomfortable feature of diversion that the monitoring debate tends to skip. A diverted dose is a dose documented as given, which means the record of the patient's care becomes a lie in the exact place it matters most, the pain scale and the medication log. The patient who received nothing while the chart said otherwise is not a bystander to the crime. He is the mechanism of concealment, and the person whose pain went untreated while the system sorted out which alert to believe.
The medical literature has understood diversion as a multiple-victim crime for more than a decade. The victims are the patient whose pain goes unrelieved, the patients exposed to contaminated equipment, the colleagues who worked a shift with an impaired nurse, and the nurse herself. The word "victim" applied to the nurse is not a euphemism. The professional profile of diversion cases is consistent across studies: people who are often well regarded, who do not stand out, who would, in the industry's own framing, be better served by early intervention and treatment than by discipline and prosecution. The monitoring systems are sold partly on that promise: catch it early enough to help.
The people the system is trying to catch are hard to see
The scale of the problem justifies the software, and the scale is easy to understate. Estimates that 10 to 15 percent of health care workers will experience substance use problems at some point in their careers appear consistently across the literature, and the link to access is measurable: a large study of nearly 4,000 nurses found that those with the easiest access to controlled drugs across every workplace dimension were far more likely to misuse prescription-type medications. Some specialties run hotter. Around 10 percent of certified registered nurse anesthetists report current or past misuse of controlled substances, a rate that tracks their proximity to the drugs.
The detection problem is not that diversion is rare. It is that the people who divert are, in the studies' consistent finding, often well respected, competent, and entirely unremarkable in the ways that trigger suspicion. Stigma suppresses reporting from every direction: colleagues hesitate, boards hear only the cases that surface, and the true prevalence is generally assumed to run higher than the measured one. This is the strongest argument for analytics, and it is also the reason the human link matters so much. The software's whole purpose is to find what ordinary attention will not. A hospital that deploys it and then understaffs the review has replicated the detection problem it paid to solve.
The industry bought detectors, then skipped the triage
The gap at Bakersfield is a staffing problem wearing a technology label. A hospital that buys diversion analytics is buying a promise that the transaction record will be watched. But watching is labor, and the labor does not appear on the purchase order. The pharmacist who is supposed to review the alerts is the same pharmacist verifying orders, managing shortages, and precepting. The alert queue competes with everything else and loses, not because anyone decided it was unimportant, but because it is the only task in the pharmacy that can be safely postponed: nobody dies today from an unread diversion flag, until somebody does.
Alert fatigue is a well documented phenomenon in medicine, studied mostly in the context of clinical alarms, and it applies with full force here. A system that emits a hundred low-confidence flags trains its reviewers to dismiss flags. The vendors know this and tune their thresholds to balance sensitivity against noise, but the tuning itself creates the trap. Lower the threshold and the queue floods. Raise it and the next Bakersfield clears quietly underneath. Either way, the bottleneck is not the algorithm. It is the number of trained humans between the flag and the cabinet.
The uncomfortable conclusion is that the software does not reduce the staffing requirement. It reveals it. The value of a diversion program was always in the follow-through, and analytics make the follow-through more effective while making its absence more visible. A hospital that buys the software and staffs the review at a quarter of what the vendor's model assumes has not modernized its diversion program. It has automated the easy part and left the hard part exactly where it was, only now the hard part has a queue.
The point of the system is early intervention, and early is a person's job
The alternative-to-discipline idea, treatment instead of prosecution, has been part of the professional response for decades, and it is the humane half of the diversion industry's promise. Caught early, the nurse at Bakersfield might have been directed into treatment, monitored, and eventually returned to practice under supervision, which is the outcome the profession's own boards have long preferred over simply losing a trained clinician to addiction and discipline. Caught late, the same nurse faces criminal charges, license actions, and the permanent end of a career, and the patients who trusted her care live with what happened in between. The distance between those two outcomes was measured in unread alerts.
That is the argument against reading the Bakersfield case as a technology failure and leaving it there. The technology did its job. The queue did its job. The break occurred at the only point in the chain where a human being had to decide that a flag deserved an afternoon of investigation, and the institution had not made that afternoon possible. Diversion monitoring is sold as a system, but it is delivered as software plus an obligation, and the obligation is the system. Hospitals that want the outcome the vendors describe will have to staff the obligation. The ones that do not will keep discovering, as Adventist Health did, that the most advanced detector in the industry ends at the same place the profession always has: a person, on a shift, with a queue.
None of this argues against the technology. The case for analytics is strong on its own terms: the patterns are real, the stakes are high, and the manual review of millions of dispensing transactions is impossible without them. The argument is that the technology's promise terminates in a human decision, and the industry has been buying the promise without funding the decision. The Bakersfield case is the cost of that omission, paid in a patient's untreated pain and in whatever happened to a nurse whose behavior was visible to everyone in the unit, including the software, before anyone acted.
A diversion monitoring system that ends in an unread queue is not a safeguard. It is a record of what might have been caught, which is a different product entirely, and the difference between the two is measured in exactly one place: whether a trained, funded, protected human opens the alert, and how quickly. The next hospital to buy the software will get the same question Bakersfield got. The software will find the pattern. Whether anything stops it will depend, as it always did, on the person at the end of the queue, and the system has not yet learned to staff that person.
Primary sources
- STAT's investigation by Alexandra Byrne for the Bakersfield case, the CMS investigation, the patient and family member accounts, the ignored alerts, and the framing of AI monitoring with human follow-through.
- Nurses Service Organization's review of substance use and diversion among nurses for the prevalence estimates and the multiple-victim framing drawn from the Mayo Clinic Proceedings literature.
- OregonLive's Medford reporting for the 10 to 15 percent healthcare worker substance use estimate and the detection-difficulty context.