Text Message Medication Adherence Reminders Automated and Delivered at Scale Across Two Institutions: Testing the “Nudge” System: Pilot study

Abstract

Background:

Medication refill behavior in patients with cardiovascular diseases is suboptimal. Brief behavioral interventions called “Nudges” may impact medication refill behavior and can be delivered at scale to patients using text messaging.

Methods:

Patients that were prescribed and filled at least one medication for hypertension, hyperlipidemia, diabetes, atrial fibrillation, and/or coronary artery disease were identified for the pilot study. Patients eligible for the pilot (N=400) were enrolled with an opportunity to opt-out. In phase I of the pilot, we tested text message delivery to 60 patients. In phase II, we tested intervention feasibility by identifying those with refill gap of ≥7 days and randomized them to intervention or control arms. Patients were texted “Nudges” and assessed whether they refilled their medications.

Results:

Of 400 patients sent study invitations, 56 (14%) opted out. In phase I, we successfully delivered text messages to 58 of 60 patients, and captured patient responses via text. In phase II, 207 of 286 (72.4%) patients had a medication gap ≥7 days for one or more cardiovascular medications and were randomized to intervention or control. Enrolled patients averaged 61.7 years old, were primarily male (69.1%) and white (72.5%) with hypertension being the most prevalent qualifying condition (78.7%). There was a trend towards intervention patients being more likely to refill at least 1 gapping medication (30.6% vs 18.0%; p=0.12) and all gapping medications (17.8% vs 10.0%; p=0.27).

Conclusions:

It is possible to set up automated processes within healthcare delivery systems to identify patients with gaps in medication adherence and send “Nudges” to facilitate medication refills. Text message “Nudges” could potentially be a feasible and effective method to facilitate medication refills. A large multi-site randomized trial to determine the impact of text-based nudges on overall CVD morbidity and mortality is now underway to explore this further.

Clinical Trial Registration:

https://clinicaltrials.gov/ct2/show/NCT03973931

Approximately 50% of patients with hypertension (HTN), hyperlipidemia (HLD), diabetes, atrial fibrillation (AFib), and coronary artery disease (CAD) do not take their cardiovascular (CV) medications as prescribed,^1–6 which results in increased morbidity, mortality, and healthcare costs.^4,7,8 Interventions that aim to improve adherence through use of patient education, reminders, pharmacist support, automated telephone communications, digital timers caps, and financial incentives have had mixed results, with many producing small to negative findings long term.^9–17 Despite the variation in results, some interventions that aimed at increasing or automating refills showed increases in adherent patients, and reduced nonadherence.^18,19 Text messages can be automated for delivery at scale for personalized communication between users and healthcare practitioners. Short term text messaging interventions with relatively small samples have shown efficacy for management of diabetes, weight loss, diet, medication adherence, adherence for antiretroviral therapy, and smoking cessation.^{15–17,20,21} Data from meta-analyses suggest personalized messages have greater effects compared to messages that are not specifically tailored for the patient.²¹ One recent meta-analysis by Thakkar et al., showed that mobile phone text messaging approximately doubles the doubles the odds of medication adherence.²² These studies demonstrate potential for text message interventions and their possible impact if taken to scale.²³

Social and behavioral health science theory offers additional approaches to improved adherence interventions. Promoting positive attitudes, improving knowledge, building skills and supporting healthy social norms are consistently effective strategies to support healthy behaviors.^24–26 Using behavioral “nudges” can promote these health intentions. A nudge is defined as a “small change that can alters people’s behaviors in predictable way without forbidding any options or significantly changing their economic incentives.” Mobile and digital technologies for health promotion and disease self-management offer a unique opportunity to adapt behavioral ‘nudges’ using ubiquitous cell phone technology to facilitate medication adherence.^17,27,28

Our goal was to test the feasibility and functionality of a text message intervention with embedded behavioral nudges to improve medication adherence among patients with cardiovascular diseases. The objectives for this pilot study was to determine 1) if we could identify patients with refill gaps using electronic health record pharmacy refill data and 2) if we could successfully send patients automated text message medication refill reminders. The findings of this study will inform a larger pragmatic study to assess the effectiveness of text messaging embedded with behavioral nudges to improve medication adherence and clinical outcomes.

Methods:

Trial design:

The data that support the findings of this study are available from the project manager upon reasonable request. This pilot is a precursor to a larger trial to explore the efficacy of sending text message reminders to facilitate medication refills, and thus we also explored the feasibility of randomizing and enrolling patients in one of four study arms: a usual care arm (patient receives no messages), generic arm (patient receives generic messages, such as “it is time to refill your medication”), optimized arm (patient receives theory-informed messages designed to be personalized and persuasive), and chat bot arm (patient receives the optimized messages and can communicate interactively with an Artificially Intelligent (AI) chat bot about medication refills) (Figure 1). The types of nudges used in the study can be seen in Figure 2. Detailed description of the process employed for message design and development is detailed in a separate paper.²⁹ The design of the study is a multi-arm study for purposes of feasibility for a future randomized control study with an allocation ratio of 1:1. Phase I of the study was used to determine feasibility of sending messages, while Phase II of the study was used to determine feasibility of randomization and implementation of different message arms.

Figure 1:

Consort diagram of phase I and phase II of the pilot study

Figure 2:

Nudges used in the study

Patient identification, eligibility, randomization:

We implemented the pilot in two healthcare delivery systems (HCS): VA Eastern Colorado Health Care System (VA), and Denver Health Medical Center (DH). The VA serves over 9 million enrolled veterans annually,³⁰ while DH servers over 1/3 of Colorado’s population annually and is considered the safety net hospital of Colorado.³¹ Patients were recruited from two clinics in the Denver metropolitan area; one clinic from the VA Eastern Colorado Health Care System (VA) and a second clinic from Denver Health Medical Center (DH). The VA serves as a tertiary referral hospital, providing primary care services to Veterans in Colorado, Wyoming, Montana and parts of Kansas. Denver Health is an integrated health care system that serves as the primary health care safety net for the City and County of Denver, Colorado. Serving an estimated one in four Denver residents, nearly 60% of Denver Health patients identify as belonging to a racial or ethnic minority group. We developed a Nudge study packet for patients that included an introductory letter, an opt out form, and a self-addressed stamped envelope. The study packet and protocol was reviewed and approved by the Nudge study team, the Stakeholder Panel members, and the University of Colorado Multiple Institutional Review Board (COMIRB). A study-specific Stakeholder Panel comprised of 12 patients, primary care physicians, pharmacists, and health administrators representing three healthcare systems and a four-person Protocol Review Committee provided additional input to the Nudge Study Team on study framework and implementation. Next, we developed an algorithm to identify the patient population of interest using International classification of Diseases (ICD)-9, ICD-10 codes, National Drug Codes (NDC), and pharmacy data from each site. Codes were collected from established codes from each institution relating to HTN, HLD, diabetes, AFib, and CAD. Some common mediations tracked included were metformin, atorvastatin, and labetalol. Collected codes were compared for consistency across each site. Patients were eligible if they had 1 or more CV co-morbidity and filled 1 or more of the medications of interest. From each healthcare system, a sample of 200 eligible patients (400 total) were identified and sent study packets if they had a previous history of 7 day medication refill gap. Patients were provided a two-week window to return the opt-out form and those who returned the opt-out form were removed from the study. Patients with opt-out packets returned by the Postal Service due to an incorrect address were contacted by a member of the study team up to two times by phone to verify their address. If a patient did not answer or return the calls, the patient was not included in the study.

In phase I, our goal was to demonstrate that we could successfully send text messages about medication refills to patients. In phase II, our goal was to monitor medication refill, identify patients with a gap, randomize them to intervention or control conditions, and send out messages appropriate to each condition to remind them to refill their medications. Phase I participants were not eligible to participate in phase II activities because prior exposure to the study might have affected the refill behavior of patients. To assess a medication gap, we calculated an anticipated refill date based on previous fill dates within each medication class, the days supply associated with these fills, and the number of outpatient days following the last fill date that would deplete the days supply of medication on hand. This was used to calculate the anticipated refill date. Factors such as hospitalizations (assuming medications were provided during the stay), and cancellation of medications available through the electronic health record (EHR) were also considered when assessing outpatient days that would deplete medication on hand. Patients who had not refilled medications within 7 days following this anticipated refill date were enrolled in phase II activities and randomized into one of four study arms as described above (Figure 3). A study start date was established at each healthcare system following the process of mailing out opt-out forms to patients and allowing adequate time for patients to return the forms. Any patient with a medication gap of 7-days or more at the start of phase II of the study had day 1 of enrollment on the phase II start date. Following this start date, any remaining patients that reached a 7-day gap for a medication was enrolled once the 7-day threshold was met.

Figure 3:

Patient enrollment and randomization

Randomization was stratified based on the health system and the number of medication classes (≤2 vs >2 medication classes) the patient actively received at baseline. Block randomization was used to ensure consistent balance across groups over time performed by the team’s statistician. We assessed the proportion of patients that filled at least 1 medication that they had a gap in refills during the study period. For patients with multiple gapping medications, we also calculated the proportion filling all medications. We restricted analyses to medications gapping on the patient enrollment date and tracked medication behavior for the full enrollment period of the pilot. Enrollment for the pilot began first at the VA and we had 43 days of enrollment prior to our hard stop date to track medication behavior. Because enrollment started later at DH and the study had a hard stop date, we could not extend the length of the enrollment period at DH to match the VA and were limited to 19 days of enrollment and medication tracking. Because our primary objectives were to establish the functionality and accuracy of both identifying patients with a gap in filling their medications and of sending automated text messages, we focused primarily on determining if they were able to fill at least one medication once a gap was established. Primary outcome measures for the main trial will include medication adherence defined by the proportion of days covered (PDC) using pharmacy refill data. Secondary outcomes will include intermediate clinical measures (e.g., BP control), CV clinical events (e.g., hospitalization for myocardial infarction) and procedures (e.g. PCI), medication-associated clinical events (e.g., syncope in patient on anti-hypertensive therapy), healthcare utilization, and costs. To achieve power of 80% or greater, we estimated that we would need to enroll 119 subjects per arm of the study across all healthcare systems for a total of 476 patients.

Message delivery:

We programmed generic, optimized and AI Chatbot messages for delivery on a third party mobile messaging system called Upland mobile.³² We sent text messages to patients included in phase II who had been randomized to one of the three arms that included messaging. Patients were informed that they had a second opportunity to opt-out if they texted “Stop” or the equivalent of stop in either English or Spanish (Figure 3). Some messages invited a response and we tracked patients who responded. Other messages did not require response but patients sometimes responded anyway, another metric we tracked. We followed up all such messages with responses to assist patients in obtaining more information as needed. We followed patients daily thereafter to see if they filled their medication through daily refill data. For example, if a patient responded with confusion we followed with a phone call clarify any questions, and explained the study better. In addition if the patient wanted to speak to clarify their gapping medications, we would connect the patient the appropriate pharmacy to clarify or update changes in medications appropriately. Once they filled the medication, we stopped delivering the text messages. We tracked any errors in sending messages, e.g. delivery of duplicate messages, non-delivery of messages when they should have been delivered. Patients who did not have phones with capability of receiving text message received communications using an interactive voice response (IVR) telephone automated message with 5 patients at the VA and 4 patients at DH by calling patients and reading a script that was identical to the text message that they would receive. Patients received text messages or telephone calls on days 1, 3, 5 and 7.

Results:

Patient identification, eligibility, randomization:

A total of 400 study packets were sent and 60 patients were selected through randomization table for phase I of the pilot study to test feasibility of text message delivery (Table 1). Of these phase I patients, 2 (VA: 1, DH: 1) opted-out of the study (Figure 1). Among the remaining 58 phase I patients, we sent reminders to refill medications, of which some asked for a response and others prompted further communications about barriers to refilling medications. We were able to successfully automate the delivery of pre-programmed messages among 58 patients without errors, such as duplicate messages sent or messages sent at the wrong time or messages sent to the wrong person. Thus, the message delivery system in phase I was deemed functional.

Table 1:

Opt-out consent packets sent out to patients and returned

	Total packets sent	Signed & returned an opt-out forms	Packets returned by USPS
Denver Health	200	13 (6.5%)	6 (3.0%)
VA	200	37 (18.5%)	0
Total	400	50 (12.5%)	6 (2.6%)

Among the remaining 340 patients (400 total patients minus the 60 patients used in phase I), 54 patients (VA: 36, DH: 18) opted-out of the study leaving 286 eligible patients for phase II. Since a medication gap ≥ 7 days was required for phase II enrollment, we tracked medication behavior for these patients and 207 (VA: 92, DH: 115) (72.4%) had a medication gap and were enrolled. We randomized 50 patients into the control arm of the study, 53 into the generic text arm, 52 into the optimized text arm, and 52 patients in the optimized text + chatbot arm, and contacted them based on the method for the arm of study.

Patient characteristics and phase II outcomes

Baseline characteristics and demographics for phase II patients are outlined in Table 2. Most enrolled patients were male (69.1%) and white (72.5%). Of the 207 patients enrolled in phase II, 62 of the 92 (67.3%) VA patients were enrolled on day 1 of the Nudge study enrollment period and were followed for outcomes for the full 43 enrollment days, while 106 of 115 (92.2%) DH patients were enrolled on day 1 and followed for 19 days at this location.

Table 2:

Characteristics of phase II eligible patients in the pilot study.

	Not Enrolled	Enrolled	p
Total N	79	207
DEMOGRAPHICS
Age - Mean (SD)	62.1 (10.9)	61.7 (11.9)	0.810
Male	64.6% (51)	69.1% (143)	0.481
Race			0.050
American Indian, Alaska  Native	1.3% (1)	0% (0)
Asian	0% (0)	0% (0)
Black, African American	24.0% (19)	19.3% (40)
Native Hawaiian, Pacific  Islander	2.5% (2)	0% (0)
White	63.3% (50)	72.5% (150)
Multiple/Missing	3.8% (3)	2.9% (6)
Hispanic	32.9% (26)	44.9% (93)	0.081
QUALIFYING CONDITIONS
AF	6.3% (5)	8.7% (18)	0.631
CAD	13.9% (11)	20.3% (42)	0.238
Diabetes	38.0% (30)	58.0% (120)	0.003
Hyperlipidemia	32.9% (26)	42.5% (88)	0.177
Hypertension	87.3% (69)	78.7% (163)	0.128
BASELINE MEDICATIONS
N Medications at Baseline - Median (IQR)	2 (2, 3)	3 (2, 4)	X
N Medications Gapping at Baseline - Median (IQR)	X	2 (1, 3)	X
Statin	65.8% (52)	74.9% (155)	0.140
ACE/ARB	67.1% (53)	73.0% (151)	0.380
Beta Blocker	20.2% (16)	34.3% (71)	0.022
Other	64.6% (51)	83.1% (172)	0.001

Results are presented as % (N) unless specified. P-values compare enrollment groups using t-tests for continuous variables and Fisher’s test for categorical variables.

Table 2 provides the proportion of patients filling at least 1 or more of their CV medications during the enrollment period stratified by treatment arm. The median number of medication classes gapping at baseline ranged from 1 to 2 across study arms. The highest proportion filling at least 1 medication was in the optimized arm (32.7%) and lowest in the control arm (18.0%). Refill trends across study arms differed by healthcare system, which may have been driven by more follow-up time (Supplemental Table I). There was a trend of patients filled at least 1 gapping medication among patients receiving some form of treatment relative to control patients with an average of 30.6% vs 18.0% respectively (Table 3), although we cannot infer that there was a difference between groups since this result was not statistically significant (P = 0.12). Similarly, treatment patients had a trend of filling all gapping medications relative to control patients (17.8% vs 10.0% respectively), but again, this did was not statistically significance (P = 0.27). Summaries of patient characteristics and phase II outcome results specific to each healthcare system are provided in Supplemental Tables I and II. A power calculation was not preformed.

Table 3:

Medication re-fill rates following phase II enrollment

	Arm 1	Arm 2	Arm 3	Arm 4
Total N	50	53	52	52
N Medications Gapping at Baseline - Median (IQR)	2 (1, 3)	1 (1, 3)	1 (1, 2)	2 (1, 3)
Filled at Least 1 Gapping Medication	18.0% (9)	32.1% (17)	32.7% (17)	26.9% (14)
Filled All Gapping Medications	10.0% (5)	17.0% (9)	21.2% (11)	15.4% (8)

^*Arm1=usual care arm, Arm 2=Generalized messages, Arm 3=Optimized messages, Arm 4= Optimized messages + chatbot

Message delivery and response:

Messages were sent in either English or Spanish, and 10 DH patients requested Spanish after receiving English messages. Overall, 33 patients(11.5%) texted “Done” and 9 patients texted “Stop” during the intervention (Table 4). In addition, we received many additional messages containing questions, concerns, and reasons why some patients didn’t refill. Some messages received could be categorized as either confusion about the study, confusion about which medication they needed to refill, or some type of barrier the patient experience that prevented refills.

Table 4:

Message responses of patients assigned into intervention arms

	Arm 1	Arm 2	Arm 3	Arm 4	Total
N DH (%)	29	29	29	29	116
Responded Stop	-	1	2	2	5
Responded Done	-	6	7	1	15
N VA (%)	22	24	23	23	92
Responded Stop	-	0	2	2	4
Responded Done	-	6	4	8	18

^*Arm1=usual care arm, Arm 2=Generalized messages, Arm 3=Optimized messages, Arm 4= Optimized messages + chatbot

Discussion:

Here we describe outcomes of a pilot study to explore the feasibility of using EHR pharmacy data to identify patients with a gap in refilling their medication and then sending those patients automated text messages. Our goal was to ensure the system for patient identification and automated message delivery worked prior to a trial to establish efficacy of using text messages for medication refill at scale across multiple care delivery systems. In the pilot study, we demonstrated the feasibility of identifying patients with medication refill gaps and delivering text messages with embedded behavioral nudges. While not statistically significant, there was a trends towards higher rates of refilling medications among intervention compared to control patients and these findings are promising for the larger pragmatic trial, and demonstrate feasibility of study design.

We developed theoretically informed text messages with input from patients and stakeholders, and were able to automate their delivery using a third-party software platform. Once programmed into the third-party message delivery system, messages were delivered automatically to the entire sample without identified errors. Patients receiving messages had numerically higher rates of medication refills than those not receiving messages (30.6% vs 18.0%) and they also had numerically higher rates of filling all medications (17.8% vs. 10.0%). Neither of these differences were statistically significant (power analysis not preformed) which is expected, given that this pilot study focused on demonstrating system feasibility and was not powered to assess these exploratory endpoints. Our next step is to explore the efficacy of the text messages in a randomized controlled trial. Extra communication with patients might have a confounding measures on refills and is something to consider to the main trail.

Our focus on feasibility is a critical first step prior to exploring efficacy of text-messaging for medication refills at scale. While the growing body of evidence of text message efficacy suggests the approach has promise, we have very limited evidence that we can deliver messages at scale.^{15,20–22,27,28} Furthermore, we do not know whether messages that are theoretically optimized are superior to generic messages. In order to prepare for research exploring these topics, we first had to design and demonstrate the functionality of systems to target and deliver messages. The evidence presented here shows that there are systems that do function as intended to identify patients with a medication gap and send them messages, allowing us to now move into a larger trial of efficacy.

Our pilot revealed some limitations. We note a particular challenge in the correct identification of patients with a medication refill gap who filled their medications outside the care delivery system, e.g. at CVS, Rite-Aid or Walgreens. In order to ensure completely reliable and valid identification of patients with medication refill gaps, pharmacy data from within systems will need supplemental data from a third party that tracks refills. We have been in communication with such a third-party company, Surescripts, to facilitate their involvement in a larger trial to capture refill behaviors of a larger number of patients. Our pilot was implemented in two care delivery systems where a high proportion of patients refill medication within the system. In systems where higher proportions of patients refill medications outside the system we may face greater challenges in reliable and valid identification of an entire patient population with medication refill gaps unless we have very strong linkages with Surescripts or similar companies. Barriers to entry included two levels of barriers and enablers at the team level and patient level. Most barriers to implementing the intervention included obtaining approvals for contracts consistent across IRB, local compliance teams, and third party vendors. In addition, difficulties were experienced developing an appropriate protocol for screening and messaging, as it would vary somedays for patients who gapped. Minimal difficulties were experienced in implementing different interventions as they were all programed in a similar manor and using the 3^rd party software to deliver text messages stream lined intervention delivery appropriately. We used the same Mobile Messenger platform for all three intervention arms programming the messages was the same for each arm and did not vary by arm. The chatbot had to have specific responses to cue the correct programmed response and we included key words along with the asked response in the programming of the chatbot; for example we asked “Press 1 to respond YES” we programmed 1, YES, and Y to all respond accordingly. We also had difficulties with collecting information from patients once they have opted out due to regulations from our local IRB. For the main trail we plan on adding an optout survey to better gauge the reasons for opting out from the study.

Now that we have established that we can successfully identify patients with a gap in refilling medications within two care delivery systems, and that we can send them text message reminders to refill medication, we are prepared to test the efficacy of messages on increasing timely medication refills.

What is known:

• Cardiovascular medication adherence is poor among patients with cardiovascular disease.

• Poor medication adherence can be attributed to motivation.

• Nudges can be an effective behavior modifier.

• Text messages are used by most of the population during this generation.

What the study adds:

• This study shows that we can identify patients with poor cardiovascular medication refill habits at two major hospital systems through daily refill data updates.

• Automated text messages can be sent to patients with poor cardiovascular medication refill habits at major hospital systems.

• Personalized text messages can be customized through chat bot interfaces for ideal nudges.

• Text messages can potentially identify medication adherence barriers

Non-standard Abbreviations and Acronyms:

CV

cardiovascular

HCS

healthcare delivery systems

VA

VA Eastern Colorado Health Care System

DH

Denver Health Medical Center

AI

Artificially Intelligent

EHR

electronic health record