Predicting the exact ship day from real consumption, not the calendar
Every adherence-timed refill system begins by replacing an assumption with a measurement. Traditional pharmacy models infer consumption from label directions or refill-purchase timing; connected caps, blister packs, and ingestible sensors instead log the physical act of taking a dose, turning "supposed to take" into a timestamped, queryable event stream.
For decades the best available adherence proxy was the Medication Possession Ratio (MPR) — days' supply dispensed divided by days elapsed, computed from pharmacy claims. It is useful at population scale but structurally blind at the individual level: a filled prescription only proves a purchase, not a swallowed dose. Patients can fill on schedule and still skip doses, double up before a doctor visit, or abandon a bottle in a cabinet.
Self-report is worse. Patients systematically over-report adherence to please clinicians — validated studies comparing pill diaries against electronic monitoring caps consistently find self-report overstates true adherence by 20–30 percentage points. Any refill-timing system built on either signal inherits that error and mistimes shipments in both directions: too early (waste, cost) and too late (stockout, health risk).
A cellular-connected bottle (AdhereTech) or Bluetooth cap (Pillsy) detects the physical opening event, timestamps it, and pushes it to the cloud the moment a connection is available — no smartphone pairing required for cellular models. Robotic multi-dose dispensers (Hero) go further, physically pre-sorting doses by time slot and logging both dispense and retrieval, which distinguishes "cap opened" from "dose actually removed."
At the far end of the spectrum, ingestible sensors go past the container entirely. Abilify MyCite embeds a sand-grain-sized sensor in the tablet itself; stomach fluid activates it on ingestion, and a skin patch relays the confirmed-ingestion signal to a paired app — the only mechanism in this category that verifies the pill actually entered the body rather than merely left the bottle.
Proteus Digital Health — the company behind the ingestible sensor inside Abilify MyCite — raised more than $500 million and shipped the first FDA-cleared digital pill in 2017, then filed for bankruptcy in 2020. The sensing technology worked; the subscription and reimbursement economics around it did not, a cautionary data point for every adherence-hardware business model that followed.
Raw cap-open events are noisy: a patient checking the bottle counts as an open even with no pill removed, and two opens ninety seconds apart are almost certainly one dosing session, not two doses. The ingestion pipeline de-duplicates events inside a short debounce window, cross-references expected regimen frequency, and — where available — combines cap-open timing with pill-count weight sensors (as used in several smart-dispenser designs) to confirm a unit actually left the container.
The output is a per-patient "consumption ledger": a chronological list of confirmed dose events with confidence scores, which becomes the raw material for the empirical daily-rate model in Stage 2 — replacing the label's "take one tablet daily" with what this specific person actually does.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| AdhereTech Smart Bottle | Oral chronic-disease medications | Cellular-connected bottle; real-time cap-open logging plus SMS/IVR reminders | No smartphone pairing required — direct-to-cloud connectivity |
| Pillsy Smart Cap | Consumer supplement / Rx bottles | Bluetooth cap syncs to a phone app, logs opens and reminders | Low-cost retrofit onto any standard bottle |
| Hero Robotic Dispenser | Multi-medication daily regimens | Pre-sorts and dispenses doses on schedule; logs dispense and retrieval separately | Physical dose sorting reduces regimen complexity |
| Abilify MyCite (Proteus sensor) | Aripiprazole (psychiatric) | Ingestible sensor activated by stomach fluid; patch relays signal to app | Confirms actual ingestion, not just container access |
A single number — "adherent" or "not" — hides everything useful. The real signal is a curve: how consumption trails off over weeks and months, where the gaps cluster, and how far the patient's empirical rate has drifted from the label's assumed one-per-day. That curve is what the ship-date algorithm ultimately forecasts against.
The Pharmacy Quality Alliance (PQA) defines Proportion of Days Covered (PDC) as the number of days in a measurement period the patient has drug "on hand," divided by the total days in that period — capped at 1.0 per day so early stockpiling cannot inflate the score. CMS uses PDC, not MPR, for its Medicare Part D Star Ratings, and three PDC-based measures — diabetes medications, RAS antagonists (hypertension), and statins — are triple-weighted in the overall plan score, making adherence one of the single largest levers in a plan's star rating.
A PDC of ≥80% is the conventional adherence threshold used across payer, PBM, and clinical-trial literature; below it, outcome and cost curves in Stage 5 begin to bend unfavorably.
Electronically monitored cohorts show a remarkably consistent pattern: adherence is highest in the first 30 days (often 90%+, sometimes inflated by a "honeymoon" effect around a new prescription), then declines — gently for well-tolerated preventive medications, steeply for asymptomatic chronic conditions where the patient feels no immediate benefit from taking the dose.
Two distinct failure modes matter for refill timing. "Primary non-adherence" is the prescription that is written but never filled at all — estimated at 20–30% of new prescriptions across multiple claims-based studies (Fischer et al. and related work). "Secondary non-adherence" is the gradual drift after an initial fill: doses skipped, gaps widening, refills delayed — this is the pattern a subscription refill system can actually observe and act on, since primary non-adherence never generates a first data point.
Systematic reviews of electronic-monitoring-plus-feedback interventions (building on Vervloet et al. and subsequent Cochrane-adjacent reviews) find that real-time reminders tied to actual dose-monitoring data lift adherence by roughly 15–20 percentage points versus unprompted controls — the exact gap a closed-loop nudge system (Stage 4) is designed to close.
Rather than trusting the label's "one tablet daily," the model fits an empirical daily-consumption rate to each patient's rolling window of confirmed dose events — typically 14–30 days, weighted toward more recent behavior. Confidence intervals widen for patients with sparse data (new starts, irregular openers) and narrow for established, consistent takers.
Fixed-schedule regimens (once-daily statin) and "as needed" regimens (PRN pain or allergy medication) require different models entirely — PRN products cannot be forecast from a fixed daily rate and instead use a rolling average of actual usage frequency, which is precisely the kind of product category (including many subscription supplements) where fixed 30-day autoship cycles are most likely to misfire.
Once a personal consumption rate exists, refill timing becomes a forecasting problem, not a calendar rule. The algorithm has to land a shipment inside a narrow window: late enough to avoid the waste of an unnecessary early refill, early enough that carrier transit time never outruns the pills actually left in the bottle.
Predicted_Depletion_Date = Today + (Pills_Remaining ÷ Empirical_Daily_Rate) Ship_Trigger_Date = Predicted_Depletion_Date − (Shipping_Lead_Time + Safety_Buffer)
Every confirmed dose event nudges Pills_Remaining down and can shift Empirical_Daily_Rate, so the algorithm recomputes the trigger date continuously rather than once at fill time. Shipping_Lead_Time itself is not a fixed constant in mature implementations — carrier tracking APIs feed live transit estimates back in, so a regional weather delay or a carrier network slowdown widens the buffer automatically instead of silently eating into it.
A legacy 30-day autoship cycle assumes 100% adherence to a fixed daily dose. A patient running at 70% PDC on a 30-day supply actually depletes the bottle in roughly 43 days — a fixed day-30 shipment arrives with pills still in the old bottle, producing waste, unnecessary spend, and (for many subscription supplement categories operating under DSHEA's manufacturing framework rather than FDA pre-approval) an accumulating stockpile with no clinical oversight. Run the arithmetic the other direction — a patient exceeding label dosing or with a larger household drawing from one bottle — and the same fixed cycle produces a stockout gap instead.
Both failure modes are avoidable once the ship date is a function of measured consumption rather than assumed consumption.
PillPack — the pre-sorted, by-dose-and-time pharmacy Amazon acquired in 2018 for roughly $753 million — built its entire subscription model around this exact problem: fixed monthly refill cycles were leaving patients with either empty dose packets before the next shipment arrived, or a drawer of unused blister strips from months they had under-consumed.
A naive implementation that recalculates on every single event can thrash — one unusually fast day of consumption fires an early trigger, then the next slow day pushes it back out, generating duplicate or canceled orders. Production systems apply a minimum-order-interval floor, smooth the daily-rate estimate with an exponential moving average rather than reacting to any single day, and fall back to label-directed dosing (with a wider safety margin) whenever sensor data goes stale — a cap left open, a lost packet, a dead battery — rather than trusting a broken data stream at face value.
A forecasting engine that only ships product is half a system. The more consequential branch fires when live PDC crosses the alert threshold: a tiered outreach ladder designed to change behavior before the next depletion forecast, not just react to it after the fact — all while staying inside consumer-protection rules that govern automatic, recurring shipments.
A typical escalation sequence: a low-friction push notification first, then SMS if unopened within a day, then an automated IVR call, then live pharmacist or care-coordinator outreach for patients who remain unresponsive or whose PDC keeps declining despite earlier nudges. Beyond messaging, the loop can also change the product itself — re-sorting a regimen into a simplified blister pack, adjusting reminder timing to the patient's actual historical dose-taking window instead of a generic morning default, or offering an explicit "pause" or "skip this shipment" option when the ledger shows the patient is running behind schedule, not out of stock.
Because refill timing sits on top of a recurring, automatically billed shipment, it inherits consumer-protection law governing "negative option" continuity plans. The Restore Online Shoppers' Confidence Act (ROSCA, 2010) requires clear disclosure of recurring charges and simple, functioning cancellation. The FTC's amended Negative Option Rule, finalized in 2024 (the "click-to-cancel" rule), goes further — requiring that canceling be at least as easy as signing up. For the supplement subscriptions this system targets, DSHEA (the Dietary Supplement Health and Education Act of 1994) additionally means the product itself is not FDA-pre-approved; manufacturing quality instead rests on 21 CFR Part 111 GMP compliance and voluntary third-party marks like the USP Verified Mark or NSF Certified for Sport.
In multiple FTC enforcement actions against auto-ship supplement marketers — brought under Section 5 of the FTC Act and ROSCA — companies were ordered to pay consumer refunds for continuity plans that made cancellation difficult or buried the recurring-charge terms. An adherence-timing engine that quietly reships product without an active, informed opt-in is not just a churn risk; it is a direct regulatory exposure.
Over-alerting has diminishing and eventually negative returns — patients habituate to generic reminders and begin ignoring the channel altogether, including the messages that matter. Mature implementations throttle back messaging frequency for consistently adherent users, personalize nudge timing around the patient's own historical dose window rather than a fixed clock time, and use response/engagement data (open rates, click-through, subsequent dose-event timing) to decide which channel actually moves behavior for that individual rather than applying one ladder uniformly to everyone.
The entire pipeline — sensing, modeling, dynamic shipping, and behavioral nudging — is only justified if it moves a real outcome. Two separate stakeholders converge on the same number: clinicians care that PDC crossing 80% tracks with measurably better biomarkers and fewer hospitalizations, while subscription businesses care that the same intervention is one of the strongest levers against churn.
Across chronic-disease literature the pattern repeats: statin adherence at or above 80% PDC is associated with materially lower LDL cholesterol and fewer cardiovascular events than sub-80% adherence in the same population; oral diabetes medication adherence above the same threshold tracks with meaningfully lower HbA1c and fewer diabetes-related hospitalizations; hypertension medication adherence correlates with better blood-pressure control and reduced stroke risk. None of these are small effects at the tails — they are the difference between a controlled and an uncontrolled chronic condition for a large share of patients sitting near the 80% line.
A frequently cited claims-based analysis by Roebuck and colleagues, published in Health Affairs in 2011, examined five chronic conditions — diabetes, hypertension, hypercholesterolemia, and congestive heart failure among them — and found that each percentage-point increase in medication adherence was associated with a net reduction in total healthcare spending, despite the higher pharmacy costs of taking more medication. The savings were driven overwhelmingly by avoided hospitalizations and emergency-department visits, which cost far more than the incremental pills.
For congestive heart failure patients specifically, Roebuck and colleagues estimated that each 1-percentage-point rise in medication adherence was associated with roughly a 0.3–0.5% reduction in total annual healthcare costs, net of the added medication spend — a rare case where "take the pill" is also the cheaper option for the system paying the bill.
For direct-to-consumer subscription brands, a stockout gap — the patient runs out before the next shipment arrives — is one of the most common triggers for outright cancellation, because it breaks the habit loop the subscription depends on. Vendors that moved from fixed calendar cycles to consumption-based dynamic shipping report churn reductions in the 25–40% range, and pharmacy benefit managers have an added structural incentive: because PDC feeds directly into CMS Star Ratings for Part D plans, adherence-timed dispensing is not just a customer-experience feature but a measurable input to plan-level reimbursement.
The correlational evidence carries a well-known confound: the "healthy adherer" effect, where patients who reliably take medication as directed also tend to differ in other health-protective behaviors, inflating the apparent benefit of adherence itself. Sensor data has its own failure modes — a cap opened is not proof a dose was swallowed, and shared households can generate cap-open events that do not correspond to the tracked patient at all.
Privacy is a separate open question: HIPAA generally does not cover direct-to-consumer subscription apps outside a covered entity or business associate relationship, so granular consumption data from many supplement and DTC pharmacy subscriptions instead falls under FTC jurisdiction — including the FTC's Health Breach Notification Rule, updated in 2024 to explicitly cover health apps and connected devices that were never HIPAA-covered in the first place.