Integrating multiple streams of data, AI moves medication therapy management from broad, rule-based screening to precise identification of high-risk patients, allowing pharmacists to target interventions more effectively.
By Golda Manuel, PharmD., MS
Every year, medication-related problems send more than a million Americans to the emergency room. Adverse drug reactions, medication nonadherence, and harmful interactions cause many of these events, but many can be prevented.1
Medication Therapy Management (MTM) was designed to prevent medication-related harm before it happens. A critical part of Medicare Part D, MTM is pharmacist-led and helps patients with complex, chronic conditions optimize their medication regimens.2 Yet today, the program faces significant limitations: too many patients, too little clarity about who is at greatest risk, and the added complexity of clinical, behavioral, and social factors. AI makes it possible to sift through vast, complex data to reveal patterns and prioritize the patients who need MTM most, leading to better outcomes in high-risk groups.
On the surface, many patients’ clinical pictures look the same: same age, same health conditions, same number and type of medications. One may be doing just fine, while another is at serious risk, whether from adding an over-the-counter drug that interacts with prescriptions or quietly stopping a medication due to side effects. When we rely only on basic metrics, risks like these often go unnoticed.
This is a fundamental flaw in many of today’s MTM eligibility models. Many health plans and CMS programs still rely on rigid, rule-based criteria, thresholds tied to the number of medications, certain chronic conditions, or projected drug costs. While these metrics offer a starting point, they often miss the dynamic, real-world complexity of medication-related risk.3
Medication risks, like people, are never static. It emerges from the messy, real-life interplay of biology, behavior, access, and circumstance. A patient taking ten medications might be perfectly stable if their regimen is long-standing and well-managed. But someone starting just three medications might be at high risk if the dose is too low, the treatment doesn’t work, and they eventually stop taking it, classic examples of the effectiveness and adherence problems that show up again and again in real-world care.4
Traditional MTM models miss this nuance because they’re not designed to detect the early signals of instability, such as:
Even though these risks are frequent and clinically significant, they often slip past MTM systems.
Rule-based systems rely on rigid if-then logic, like flagging a patient because they’re on eight medications. But risk is often nuanced. It expresses itself in combinations of clinical indicators, behavior shifts, and life changes that only make sense when considered together.
This is where machine learning models excel. Instead of relying on fixed cutoffs, AI can pull together vast streams of structured and unstructured data like electronic health records, pharmacy claims, clinical notes, lab results, and even social factors like neighborhood or education, to detect subtle, nonlinear signals of risk.5
For instance, an AI model might flag a patient who just switched to a GLP-1, missed follow-ups, and shows refill gaps. Add in the fact that they live in a neighborhood with limited pharmacy access, and the risk becomes clear. A simple rule-based system might take far longer to recognize this patient as high-risk.
AI models, especially those trained on large, diverse, real-world datasets, can take in thousands of variables at once. That includes medication lists, lab values, discharge summaries, vitals, and health literacy data. That’s why having these data points interconnected is so critical, one of the key goals of the CMS Health Tech Ecosystem initiative . Using these signals, we can predict risk with surprising precision, identifying patients who are most likely to:
What sets these tools apart from traditional analytics is both speed and nuance. AI and machine learning models can capture nonlinear relationships and context-dependent patterns between variables, such as how a medication change, fluctuation in weight, and recent job change can combine to elevate risk. They process information orders of magnitude faster than human review or rule-based systems, enabling real-time or near-real-time insights.
In a retrospective analysis, patients were ranked for hospitalization risk using either a simple polypharmacy rule or a machine learning model. Among those flagged by the polypharmacy rule, only 16% were actually hospitalized within 90 days, while 84% were not. In contrast, the machine learning model achieved a true positive rate of 56% and a false positive rate of 44%, demonstrating markedly greater precision in identifying patients at high risk.6
Instead of casting a wide net and hoping to catch a few at-risk patients, they could focus on the ones who truly needed attention. For pharmacists already stretched thin, that shift turns overwhelming patient lists into clear priorities, helping care teams stay proactive and avoid burnout.
Predictive models are only as good as the data they’re trained on. If the underlying datasets are biased, incomplete, or fail to represent diverse patient populations, the results can be deeply flawed. Without rigorous design, validation, and ongoing monitoring, these blind spots can generate misleading insights and risk deepening existing health inequities instead of addressing them. In MTM, that could mean high-risk patients never get flagged, while others are overprioritized, leading to wasted resources and missed opportunities to improve care.
This is why patient care still depends on the expertise and judgment of pharmacists and clinicians. AI can scan oceans of data and surface patterns invisible to the eye, but only humans can translate those signals into the realities of a patient’s life. A model may flag likely nonadherence, yet only a conversation uncovers whether the barrier is side effects, cost, or confusion. When paired thoughtfully, AI sharpens clinical insight and frees space for meaningful patient interactions, making care both smarter and more human.
Common questions this article helps answer