Medication adherence data trust: what pharmacy claims miss and wearables catch
Pharmacy claims data tells you a prescription was filled. It cannot tell you whether the pill was swallowed, when it was taken, or what happened to the patient between refills. Wearable sensors and smart packaging now capture the signals that claims miss, but the data they generate carries its own trust problem. Closing the pharmacy claims data gap requires scoring both sources before any AI model acts on them.
Pharmacy claims data confirms one thing: a prescription was dispensed. It says nothing about whether the patient opened the bottle, took the correct dose, or experienced a side effect that made them stop. Yet claims remain the default source for medication adherence measurement across payers, pharma, and health systems. The gap between what claims report and what actually happens at the patient level is not a minor inconvenience. It is a structural failure in adherence monitoring intelligence that distorts clinical AI, inflates quality metrics, and misallocates intervention resources.
What is medication adherence in pharmacy?
Medication adherence in the pharmacy context refers to the degree to which a patient follows a prescribed medication regimen. Pharmacists track adherence primarily through dispensing records: was the prescription filled, was it refilled on time, and does the fill history suggest continuous use? The standard metric is the Proportion of Days Covered (PDC), which divides the number of days a patient has medication on hand by the number of days in a measurement period.
PDC works well for administrative purposes. CMS uses it as a quality measure in Medicare Part D Star Ratings. Health plans use it to stratify risk. Pharma companies use it to estimate real-world drug utilization.
But PDC answers a narrow question. A patient who fills a 90-day supply and leaves it in a drawer scores the same as a patient who takes every dose on schedule. The pharmacy system cannot distinguish between the two.
What are the three categories of medication adherence?
Researchers generally classify medication adherence into three categories: initiation, implementation, and discontinuation.
Initiation measures whether a patient fills the first prescription after it is written. Studies estimate that 20% to 30% of new prescriptions are never filled at all. Claims data can capture initiation if the prescription is transmitted electronically, but if a patient receives a paper script and never brings it to the pharmacy, the event is invisible.
Implementation covers how well a patient follows the dosing regimen over time. This is where the pharmacy claims data gap is widest. A patient may fill monthly prescriptions but skip doses three days per week. Claims see a compliant patient. The clinical reality is partial adherence.
Discontinuation is when a patient stops taking a medication entirely. Claims data can detect discontinuation retrospectively, but only after 60 to 90 days of no refill activity. By then, the clinical consequence may have already occurred: an uncontrolled blood pressure spike, a diabetes complication, or a preventable hospitalization.
The lag between real-world discontinuation and claims-based detection is a well-documented problem. For AI models trained on claims, this delay introduces systematic error. As we have written about in our analysis of claims data lag, a 30- to 90-day reporting delay means that any adherence model running on claims is working with stale signal.
What are the challenges of medication adherence?
The challenges are both clinical and data-structural.
On the clinical side, patients face barriers including cost, side effects, complex regimens, cognitive decline, and lack of perceived benefit. A 2023 WHO review estimated that adherence to chronic disease medications in developed countries averages only 50%. That means half of all prescribed therapy is either partially followed or abandoned.
On the data side, the challenges are equally severe. Pharmacy claims are billing artifacts, not clinical observations. They do not record the moment of ingestion. They do not capture whether a patient splits pills, takes a double dose after missing one, or adjusts timing based on side effects. They do not reflect medication samples provided in-office, which never appear in claims at all.
Electronic health records add context but introduce their own problems. Medication reconciliation during clinical visits is frequently inaccurate. As we have documented in our work on RxNorm drug data integrity, discrepancies between what a patient reports taking and what the EHR lists are common, and these errors propagate into AI training data.
The result is a measurement system built on incomplete, delayed, and structurally biased data. Any AI model that uses this data to predict adherence risk, trigger interventions, or evaluate drug efficacy inherits those limitations.
How to track medication adherence: what wearables and sensors now capture
The past five years have produced a new category of adherence data that operates outside the claims infrastructure entirely.
Smart pill bottles and caps record when a container is opened. Systems like AdhereTech and Pillsy generate timestamped open/close events, creating a dosing log with second-level precision. This eliminates the gap between fill and ingestion, at least at the container level.
Ingestible sensors go further. Proteus Digital Health (before its closure) and current products like Abilify MyCite embed a sensor in the pill itself. When the sensor reaches the stomach, it transmits a signal to a wearable patch, which logs the exact time of ingestion. This is the only technology that confirms a dose was actually swallowed.
Wearable biosensors capture physiological signals correlated with medication effects. Continuous glucose monitors detect the pharmacodynamic response to diabetes medications. Heart rate variability from smartwatches can signal beta-blocker adherence. Accelerometer data from wrist-worn devices can correlate activity patterns with stimulant medication timing in ADHD populations.
Smart packaging and blister packs with embedded circuits track individual dose removal. These are used in clinical trials and are beginning to appear in commercial pharmacy for high-cost specialty drugs.
Each of these sources generates data that pharmacy claims cannot: real-time, granular, patient-level adherence signals tied to specific moments and physiological states.
Key statistics
The scale of the adherence data problem is quantifiable:
The pharmacy claims data gap: five specific blind spots
To understand why medication adherence data trust requires more than claims, consider five concrete scenarios where claims fail.
1. Samples and coupons. When a physician hands a patient starter samples, no claim is generated. When a manufacturer coupon reduces the copay to zero through a hub program, the claim may reflect a fill but misrepresent the patient's cost burden, which is one of the strongest predictors of future discontinuation.
2. Mail-order timing. A patient using a 90-day mail-order pharmacy may have medication on hand but change their dosing pattern midway through the cycle. Claims see a 90-day supply. They cannot see a patient who halved their dose at day 30 due to dizziness.
3. Polypharmacy interactions. Claims record each fill independently. They do not capture that a patient stopped taking Drug A because Drug B caused an interaction that made them feel worse. The causal relationship between fills is invisible in claims.
4. Institutional dispensing. Medications dispensed during inpatient stays, in long-term care facilities, or through 340B programs may not generate standard pharmacy claims. For patients cycling between settings, adherence records have structural gaps.
5. Behavioral context. A patient who searches for "metformin side effects" at 2 AM and then skips their dose the next morning is exhibiting a behavioral adherence signal. Claims capture none of this. Wearable data might capture the missed dose through an unopened smart cap. Search behavior data captures the intent.
What wearables catch that claims miss
Wearable and sensor data fills specific gaps in the adherence measurement chain.
Timing accuracy. A smart pill cap records the exact time a container was opened. For medications with narrow dosing windows, like immunosuppressants after organ transplant, the difference between 8:00 AM and 2:00 PM dosing can affect drug levels and clinical outcomes. Claims data cannot distinguish between the two.
Dose-response correlation. A CGM paired with a smart insulin pen can show whether a dose was administered and whether the expected glucose response occurred. If the pen logs injection but the CGM shows no glucose drop, the dose may have been compromised, expired, or improperly stored. This signal does not exist in any claims database.
Pattern detection. Wearable data is continuous. It reveals adherence patterns over hours and days, not over 30- or 90-day refill cycles. A patient who takes their statin five days per week but skips weekends is 71% adherent by actual dosing but 100% adherent by PDC if they fill on schedule. Only sensor data captures the pattern.
Real-time intervention triggers. When a smart cap has not been opened by a patient's usual dosing time, a notification can fire the same day. Claims-based interventions happen weeks or months after a gap is detected. The clinical difference is enormous.
But wearable adherence data is not automatically trustworthy. As we have analyzed in the wearable data trust problem, consumer-grade devices generate data with variable sensor quality, inconsistent calibration, and no standardized provenance chain. A smart cap that records an "open" event does not confirm ingestion. An accelerometer that infers medication timing from activity patterns introduces algorithmic assumptions that may not generalize across populations.
The trust problem with both sources
The core issue is not that one data source is better than the other. It is that neither source carries a verified trust score.
Pharmacy claims data has known provenance (the PBM system that generated it) but low clinical fidelity. It measures transactions, not behavior. Its recency is poor. Its concordance with actual patient behavior is structurally limited.
Wearable data has high recency and granularity but uncertain provenance. Who calibrated the sensor? Was the device worn consistently? Did the patient share the device? Is the firmware version validated for clinical use? These questions are rarely answered before the data enters a model.
When an AI system fuses both sources without scoring either, it inherits the worst characteristics of both: the latency of claims and the uncertainty of unvalidated sensor data. The result is adherence monitoring intelligence that looks precise but is built on unverified foundations.
This is the problem the Data Trust Index was designed to solve. Every record, whether it originates from a PBM claims feed or a Bluetooth-connected pill cap, needs to be scored across provenance, recency, quality, concordance, and the other trust dimensions before it informs a clinical decision or trains a model.
Building adherence monitoring intelligence that earns trust
A trustworthy adherence intelligence system requires four capabilities that neither claims nor wearables provide alone.
Source-level scoring. Every data element entering the adherence model needs a trust score. A pharmacy claim from a PBM with known data quality issues scores differently than a claim from an integrated health system with real-time adjudication. A smart cap event from an FDA-cleared device scores differently than one from a consumer wellness gadget.
Temporal alignment. Claims data operates on a 30- to 90-day cycle. Wearable data operates in real time. Aligning these timelines requires a system that understands recency as a scored dimension, not just a timestamp. A 90-day-old claims record and a 5-minute-old sensor event should not carry equal weight in an adherence model.
Concordance checking. When claims say a patient filled a prescription on January 15 but the smart cap shows no activity until January 22, that discordance is a signal. It might mean the patient picked up the prescription but did not start taking it for a week. A trust-scored system flags this discordance and adjusts confidence accordingly.
Consent-aware data fusion. Patients consent to pharmacy claims processing through their insurance agreement. Wearable data requires separate, explicit consent. Fusing these sources without a consent governance layer violates the patient's reasonable expectation of how their data will be used. ConsentOS exists specifically to manage this complexity across data sources with different consent tiers.
What this means for pharma, payers, and health systems
For pharmaceutical companies, the adherence data gap directly affects real-world evidence quality. If your RWE studies rely on PDC from claims to demonstrate drug effectiveness, you are measuring refill behavior, not therapeutic adherence. The FDA increasingly expects more granular adherence data in post-market studies, and the gap between what you submit and what regulators expect is widening.
For payers, adherence metrics drive Star Ratings, risk adjustment, and intervention targeting. Models trained on claims-only adherence data misidentify who needs outreach. Patients who fill prescriptions but do not take them consume intervention resources that should go to patients who never fill at all. The cost of this misallocation is embedded in your medical loss ratio.
For health systems, adherence data quality affects readmission risk models, care management workflows, and value-based contract performance. As we have analyzed in readmission prediction model bias, models trained on incomplete adherence data produce biased predictions that disproportionately affect patients with fragmented pharmacy records.
The path forward: scored data, not more data
The solution to the medication adherence data trust problem is not simply adding more data sources. It is scoring every source before it enters an adherence model. More data without trust scoring creates more noise, not more signal.
The DTI framework scores each record across eight dimensions. For adherence data specifically, the most critical dimensions are provenance (where did this adherence signal originate and can we verify its chain of custody?), recency (how old is this signal relative to the clinical decision it will inform?), and concordance (does this signal agree with other available evidence about this patient's behavior?).
A pharmacy claim with a 60-day lag, no concordance with sensor data, and uncertain provenance due to PBM data aggregation might score 35 out of 100. A smart cap event from an FDA-cleared device, captured in real time, concordant with a CGM glucose response, and governed by explicit patient consent might score 82. Both contain adherence information. Only one is trustworthy enough to trigger a clinical action.
SuperTruth's trust layer turns adherence data from a compliance liability into a clinical asset. If your team needs audit-ready provenance for FDA submission, scored adherence data for RWE studies, or trust-verified signals for payer quality programs, contact Louis Simeonidis at louis@supertruth.ai or (215) 918-4140.
Further reading:
Jason Alan Snyder
Co-founder of SuperTruth and Artists & Robots, and an inventor on the Data Trust Index patents. Twenty-plus years building technology inside Interpublic Group. He writes here nearly every day on data trust, provenance, and what AI should be allowed to act on, and publishes essays on his Substack.
About SuperTruth · LinkedIn · Substack · jasonalansnyder.com
See it in practice
Real-world evidence with a trust score.
DTI scores every RWE record before it reaches a model.