The quiet crisis hiding in medicine cabinets everywhere
Somewhere between the pharmacy and home, a lot of medication stops being taken.
For people managing conditions like HIV, diabetes, high blood pressure, or tuberculosis, skipping doses is not a minor inconvenience. It can mean the disease worsens, becomes drug-resistant, or requires far more expensive treatment down the line.
How big is the problem?
Between 40 and 50 percent of people with chronic diseases globally do not take their medications as prescribed. That is not a small slice of patients — that is roughly half of everyone with a long-term condition.
The consequences are enormous: preventable hospital admissions, disease complications, and healthcare costs that fall on individuals, families, and health systems alike. For HIV, missed doses can allow the virus to replicate and mutate, reducing future treatment options. For tuberculosis, non-adherence is directly linked to the spread of drug-resistant strains that are far harder to treat.
Catching the problem after it has already happened
The traditional approach to non-adherence is reactive. A patient misses a refill. A clinic gets a notification weeks later — if they get one at all. A nurse calls. The patient may or may not respond.
But here's what is different now: machine learning models — a type of artificial intelligence that learns patterns from large datasets — can analyze health records, pharmacy refill histories, demographic information, and even healthcare visit patterns to identify patients at high risk of stopping their medication weeks or months before they actually stop.
Think of it like a weather forecast for treatment drop-off. Instead of waiting for the storm to hit, clinicians could see warning signs forming on the horizon and intervene while there is still time.
What the data actually includes
Researchers reviewing this field found that the most effective AI models drew on multiple types of data simultaneously: electronic health records, pharmacy refill logs, socioeconomic variables (like housing stability and income level), and patterns of healthcare use. Models using that combination of data sources achieved discrimination metrics — a measure of how well the model separates high-risk from low-risk patients — ranging from 0.70 to 0.95 on a scale where 1.0 is perfect prediction.
For context, a score of 0.70 is considered useful and 0.80 or above is generally considered strong for clinical prediction tools.
The diseases where this has been studied
The review covered AI adherence prediction across five major disease areas: HIV, tuberculosis, diabetes, hypertension (high blood pressure), and mental health conditions. Each has its own specific challenges — the social stigma of HIV, the lengthy treatment course for tuberculosis, the daily burden of diabetes management — but AI models have shown meaningful predictive ability across all of them.
That accuracy does not yet mean these tools are ready to use in your doctor's office — and understanding why that gap exists matters.
Why good models have not yet reached patients
The gap between a model performing well in a research study and that model being safely used in a real clinic is significant. Most AI adherence tools have been tested only in the same hospital systems or populations where they were built. When researchers try to apply them elsewhere, performance often drops.
There are also equity concerns. Several studies found that AI models performed worse for marginalized groups — people from low-income backgrounds, racial minorities, or patients in low-resource health systems — precisely the populations where non-adherence is often highest. A tool that works well for the average patient in a well-funded American hospital may fail the people who need it most.
Data privacy, interpretability (can a clinician understand why the model flagged a patient?), and the resources needed to implement these tools in resource-limited settings are all unresolved challenges.
If you manage a chronic condition that requires daily medication, knowing that AI tools aimed at supporting adherence are under development is reassuring — but you should not expect your clinic to have one yet. In the meantime, practical strategies remain the most reliable: pill organizers, phone reminders, open conversations with your pharmacist or doctor about barriers you face, and asking for simplified regimens if your medication schedule feels overwhelming.
Limits worth knowing
This was a narrative review — a summary of existing studies — rather than a pooled analysis or randomized trial. The accuracy figures reported across studies should not be directly compared to each other, since different studies used different methods and different definitions of adherence. External validation (testing models on new patient groups) is still limited across the field.
Researchers are calling for multicenter validation studies, meaning the same AI model tested across many different hospitals and patient populations before it is trusted clinically. They also stress the need for randomized controlled trials that measure whether acting on AI-generated predictions actually improves patient outcomes — not just whether the predictions are technically accurate. If those trials succeed, AI-assisted adherence support could become a routine part of chronic disease management, helping connect the right intervention to the right patient at exactly the right moment.