A remote intervention to improve medication nonadherence guided by a marker of risk derived from the electronic health records of adolescent transplant recipients

Abstract

Improving Medication Adherence in Adolescents who had a Liver Transplant (iMALT), a prospective, block-randomized, single-blind, controlled multisite study in 13 pediatric transplant centers, used the medication level variability index (MLVI) to identify non-adherent adolescents. It compared a two-year remote behavioral telemetric intervention (TI) that focused on adherence and addressed barriers, to standard of care (SOC). The primary endpoint was the composite incidence of centrally-determined (3 masked pathologists) rejection, death, re-transplantation, or consent withdrawal. In 148 participants (TI n=72, SOC n=76), mean age 15.5 years, there were twice as many primary events in SOC (12; 15.8%), as compared to TI (6; 8.3%) (RR 0·57, 95% CI 0.24–1.35, p=0.20). All secondary outcomes improved in TI vs. SOC. The study was underpowered, because iMALT sites used MLVI in clinical care, significantly reducing overall rejections as compared with the previous MALT cohort, where MLVI wasn’t used (rejection observed in 12 / 76 patients in iMALT SOC vs 13 / 25 patients in MALT; p< 0.0001). The effect on the primary outcome, therefore, was not statistically significant; thus, the intervention is not evidence-based. MLVI, an electronic health record-derived behavioral marker, can be used to target interventions to patients with clinically significant nonadherence. ClinicalTrials.Gov, NCT03691220

INTRODUCTION

Non-adherence to medical recommendations is a leading reason for treatment failure, in any population¹. But it is hard to identify patients who are non-adherent. Objective surveillance methods such as electronic monitoring can be cumbersome to use and therefore the most non-adherent patients don’t even use those even when asked to^2–4, leading to inconsistent or non-existent clinical benefits^3,5. Subjective assessments (e.g., patient self-reports) are frequently inaccurate^6,7. Therefore, adherence intervention research frequently consists of delivering interventions (e.g., educational materials) to all patients, even if they do not need that intervention and do not stand to benefit from it³, and clinical benefits are not consistently observed^3,5.

Patients who had a liver transplantation are a special population in this regard, because strict adherence to the prescribed regimen is essential. Pediatric and adolescent recipients face the burden of having to take exactly the right amount of immunosuppressive medications for a lifetime to prevent organ rejection. Taking too little of the immunosuppressant (or taking it inconsistently) endangers rejection, while excessive immunosuppression risks infections and other side effects⁸.

In this population, failure to take prescribed medications (non-adherence), therefore, is a leading cause of organ rejection^9–13. As in other fields, clinicians lack objective, accurate tools to identify non-adherent patients before clinical sequelae (including rejection) occur^7,12.

Over the past two decades, we have developed and validated the Medication Level Variability Index (MLVI)^{12, 14–16}, defined as the standard deviation (SD) of successive immunosuppressant levels over time. MLVI identifies erratic medication-taking behavior (a higher SD indicates less consistent adherence). The prospective multi-site study MALT (Medication Adherence in Children who had a Liver Transplant; NCT01154075)¹⁴ validated MLVI as a behavioral biomarker of posttransplant risk (MLVI ≥2 units predicts increased rejection risk). The marker looks at fluctuation rather than a single blood level value. Effects such as absorption issues, drug-drug interactions, or inconsistent prescription practices may change a blood level but not necessarily lead to excessive fluctuation, as not taking the medication at all has a much more profound effect on the levels. Such minor effects, if they even exist, are eliminated by choosing a sufficiently restrictive cutoff threshold. Theoretical issues besides non-adherence that may affect blood levels were prospectively investigated and shown to have no discernible effect on the marker.¹⁴

MLVI monitors patients’ adherence using information available in Electronic Health Records (EHR) and identifies risky behavior (non-adherence) before adverse events occur. Adherence monitoring in itself is a reasonable goal in clinical practice, but it is unknown whether the use of MLVI could improve transplant outcomes on the basis of enhanced recognition of nonadherence¹⁷.

iMALT [improving Medication Adherence in Adolescent (and young adult) Liver Transplant recipients] was a prospective, randomized, controlled, single-blind trial that delivered a remote intervention to non-adherent (MLVI>2) adolescent and young adult liver transplant recipients^18,19. MLVI was used to identify at-risk patients and also for monitoring during the intervention. The pre-defined primary outcome was the two-year incidence of rejection, evaluated via a masked reading of biopsy slides by three pathologists. Consent withdrawal, re-transplantation, or death were coded as rejection for the purpose of the primary outcome.

METHODS

Study Design

iMALT design was described in detail elsewhere^18,19 (clinicaltrials.gov: NCT03691220, no changes to outcome definitions since inception). This prospective, block-randomized controlled single-blind (masked pathologists) multisite study was conducted in 13 pediatric liver transplant centers affiliated with academic institutions in North America¹⁸ (Icahn School of Medicine at Mount Sinai; University of California, Los Angeles; Ann & Robert H Lurie Children's Hospital of Chicago; Baylor College of Medicine; University of Pittsburgh; University of California, San Francisco; Children's Hospital of Michigan; Children's Hospital of Philadelphia; Emory-Children's Center; The Hospital for Sick Children, Toronto, Canada; Columbia University; University of Miami; Children's Hospital Los Angeles); a 14^th site participated initially but did not recruit any patients. Repeated chart reviews in participating centers identified patients with MLVI>2.0, calculated on outpatient tacrolimus blood levels from 2 years before screening. This metric reflects persistent (rather than fleeting) non-adherence, observed during a period of two years prior to enrollment. Patients who initially did not meet the metric’s requirement were re-screened periodically - if their MLVI went above the threshold later on, they became eligible for inclusion. Patients (12–20 years old at enrollment) were approached in clinic or remotely and randomized to a two-year telemetric intervention (TI) or standard care (SOC). Randomization was determined at enrollment via the internet; personnel had no access to the sequence. Protocol design included feedback from patients and parents, which is cited in the supplementary materials. Table S1 summarizes Inclusion and exclusion criteria.

Outcome Measures

The pre-defined primary outcome was the incidence of biopsy-confirmed acute cellular rejection, as determined by a majority vote of three expert pathologists, masked to patient group assignment, demographics, and disease characteristics. Since non-adherent patients are less likely to follow study procedures and are likely to develop rejection if not followed closely, patients with incomplete follow-up due to death, re-transplant, or consent withdrawal were considered to have experienced rejection for the primary analysis. Sensitivity analyses looked at the site’s biopsy interpretation (local read). Sensitivity and some of the pre-defined secondary analyses looked at individual components of the primary (e.g., rejection) to ensure that any observed effect on the primary was not simply due to a large effect on only one of its components.

Secondary outcomes included analyses of serum alanine aminotransferase (ALT) and gamma-glutamyl transferase (GGT) levels, and MLVI. Exploratory outcomes, including patient-reported outcomes, are not presented in the present manuscript, but will be reported in the future.

Regulatory Oversight

The Icahn School of Medicine at Mount Sinai’s Institutional Review Board (IRB) was the single IRB for the US study sites, and a Research Ethics Board approved the Canadian site (Toronto). Since parents participated in calls, the patient and one parent were each considered subjects, and required consent / assent for minors. A Data Safety and Monitoring Board, chartered by the National Institutes of Health, reviewed data accrual and safety.

Adherence Measure: MLVI

MLVI in this study is the standard deviation of at least three consecutive tacrolimus blood levels; MLVI>2.0 - inconsistent adherence¹⁴ - was required for eligibility.

Intervention

The telemetric intervention (video / audio calls between interventionists, patients, and parents) was iteratively developed by our team over the last decade.^19,20 Interventionists were psychologists, social workers, or psychology trainees, trained in the manualized intervention,¹⁹ which allowed tailoring based on patient preferences. All sessions were audiotaped, and 30% were randomly sampled to assess fidelity (defined as delivering at minimum 80% of the prescribed content.)

The intervention, described in detail elsewhere¹⁹, included a review of medication taking practices, why the medications are important, and of specific barriers to taking the medication including avoidance of medication taking due to distress (which our group previously described^21–23). Following initial review, with possible feedback from parents and clinicians, interventionists assisted patients with overcoming barriers, monitored their progress, and offered praise when improvement was noted. iMALT did not include any mental health intervention component – if a need for a mental health intervention was identified, patients were referred locally through the clinic. Over two years, each patient’s MLVI was re-calculated quarterly using the previous 12 months’ medication blood levels. Calls were conducted weekly for the first quarter and reduced to monthly if MLVI became <2.0. If MLVI subsequently exceeded 2.0 or labs were not drawn for a quarter, weekly calls resumed. In order to ensure that primary outcome data are not skewed in favor of patients who consistently followed the intervention (presumably, those would be the most adherent ones), medical outcome data from participants who asked not to be called at all or who did not consistently attend the session were included, so long as consent to obtain outcome data was not withdrawn. Intent-to-treat outcomes were calculated for all enrolled patients regardless of their propensity to answer the calls.

Controls.

SOC was without restrictions; sites addressed non-adherence per their usual practice and, for ethical reasons, were free to use the measured MLVI as a marker of non-adherence to inform clinical care.

Sample Size

Based on analyses of the MALT cohort,¹⁴ we estimated a rejection incidence of 45% in the control arm and a reduction of rejection incidence by 50% in the intervention arm, aiming at a sample size of 140 analyzable patients, using a two-sided Z-test with 80% power and a 5% type-I error rate. The sample size calculation incorporated an interim analysis after 50% of the patients completed the study using the O’Brien Fleming alpha spending approach²⁴. Assuming 10% attrition, enrollment target was set at 156; since no attrition was observed as the study progressed, the target was revised assuming 5% attrition, resulting in a target of 148 patients, still aiming at 140 analyzable patients.

Statistical Analysis

The NIDDK-CR repository includes iMALT de-identified data that can be accessed with permission. Statistical analysis was supervised by Ravinder Anand. All randomized patients were included in the intention-to treat population to assess the primary endpoint. The primary analysis evaluated the effect of intervention on the primary endpoint using the test-statistic from stratified logistic regression with site being the stratification factor. The same analyses were applied to sensitivity outcomes. No adjustment to multiplicity is required since there is a single primary outcome, as such secondary outcomes are considered exploratory rather than confirmatory.

An interim analysis for efficacy evaluated the primary endpoint once 50% of patients completed 2 years of follow-up. The primary endpoint was not met, and the study was continued as planned.

Safety analyses compared rates of adverse events and clinical laboratory evaluations by group. Events of interest, that were defined a priori, included newly diagnosed conditions which could result from tacrolimus toxicity, including new post-transplant lymphoproliferative disease, diabetes, systemic hypertension, and serum creatinine > 176.8 μmol/L.

Similar to MLVI, the SD of a series of measures of ALT and GGT were computed for each patient, theorizing that higher SD (higher variability) reflects irregular / inadequate immunosuppression.

The primary endpoint was designed to reduce the likelihood of missing data. Missingness beyond the primary outcome definitions was assumed to be missing at random and the complete case analysis was used for secondary endpoints.

Post-hoc analyses compared between age and risk-level matched MALT “adolescents at risk” (MLVI>2) and iMALT SOC cohorts. A Chi-Square test with continuity correction compared the rate of centrally adjudicated rejection during the two years of either study, and baseline demographic characteristics were compared using Kruskal-Wallis or Fisher-Exact tests.

RESULTS

Enrollment and Demographics

iMALT was registered on 2018-09-28, posted 2018-10-01, start date was 2018-11-14, primary (data) completion date was 2025-10-22, and was uninterrupted during the COVID-19 pandemic. 3,017 electronic health records were reviewed, 517 reviews identified potentially eligible patients, 258 patients were approached, 148 consented, 72 block-randomized to TI and 76 to SOC. All 148 participated in the primary analysis (Supplemental Figure S1). The intervention was delivered as planned and fidelity was 91%.¹⁹ Biliary atresia was the most common indication for transplant (Table 1). Mean age at enrollment was 15.5 years, on average 10.0 years after transplant. Patients’ demographics were similar between groups (Table 1).

Table 1:

Baseline Characteristics by Intervention

	Telemetric Intervention (N=72)	Standard of Care (N=76)	Total (N=148)
Randomized Patients
Total	72 (100·0%)	76 (100·0%)	148 (100·0%)
Primary Diagnosis
Biliary Atresia	29 (40·3%)	34 (44·7%)	63 (42·6%)
Other Cholestatic Diseases	10 (13·9%)	9 (11·8%)	19 (12·8%)
Acute Liver Failure	9 (12·5%)	8 (10·5%)	17 (11·5%)
Liver Malignancy	5 (6·9%)	8 (10·5%)	13 (8·8%)
Metabolic Diseases	6 (8·3%)	4 (5·3%)	10 (6·8%)
Autoimmune Hepatitis	5 (6·9%)	4 (5·3%)	9 (6·1%)
Other	8 (11·1%)	9 (11·8%)	17 (11·5%)
Graft Type
Whole liver*	46 (63·9%)	47 (61·8%)	93 (62·8%)
Technical variant deceased	13 (18·1%)	19 (25·0%)	32 (21·6%)
Living donor	13 (18·1%)	10 (13·2%)	23 (15·5%)
Number of Liver Transplants Prior to Study Enrollment
1	67 (93·1%)	69 (90·8%)	136 (91·9%)
>1	5 (6·9%)	7 (9·2%)	12 (8·1%)
Age at Last Transplant (years)
N	72	76	148
Mean (SD)	4·75 (4·39)	6·22 (5·03)	5·51 (4·77)
Median (Q1, Q3)	2·55 (1·29, 7·57)	4·63 (1·22, 10·70)	3·76 (1·27, 9·84)
Age at Enrollment (years)
N	72	76	148
Mean (SD)	15·12 (2·12)	15·94 (2·29)	15·54 (2·24)
Median (Q1, Q3)	14·48 (13·34, 16·67)	15·80 (13·96, 17·84)	15·30 (13·68, 17·57)
Time from Last Transplant to Enrollment (years)
N	72	76	148
Mean (SD)	10·36 (4·43)	9·72 (5·00)	10·03 (4·73)
Median (Q1, Q3)	11·23 (6·16, 13·66)	10·89 (4·07, 13·48)	11·08 (4·95, 13·52)
Sex
Male	31 (43·1%)	28 (36·8%)	59 (39·9%)
Female	41 (56·9%)	48 (63·2%)	89 (60·1%)
Ethnicity
Hispanic or Latino	23 (31·9%)	25 (32·9%)	48 (32·4%)
Not Hispanic or Latino	48 (66·7%)	48 (63·2%)	96 (64·9%)
Unknown	1 (1·4%)	3 (3·9%)	4 (2·7%)
Race
Asian	4 (5·6%)	4 (5·3%)	8 (5·4%)
Black or African American	16 (22·2%)	13 (17·1%)	29 (19·6%)
White	45 (62·5%)	49 (64·5%)	94 (63·5%)
More than One Race	3 (4·2%)	4 (5·3%)	7 (4·7%)
Unknown	4 (5·6%)	6 (7·9%)	10 (6·8%)
Site
Mt. Sinai Medical Center	2 (2·8%)	3 (3·9%)	5 (3·4%)
Ann & Robert H. Lurie Children's Hospital of Chicago	6 (8·3%)	8 (10·5%)	14 (9·5%)
Children's Hospital of Pittsburgh	13 (18·1%)	12 (15·8%)	25 (16·9%)
UCLA Medical Center	9 (12·5%)	11 (14·5%)	20 (13·5%)
Texas Children's Hospital	4 (5·6%)	4 (5·3%)	8 (5·4%)
UCSF Benioff Children's Hospital	4 (5·6%)	3 (3·9%)	7 (4·7%)
Children's Hospital of Philadelphia	7 (9·7%)	6 (7·9%)	13 (8·8%)
University of Michigan	1 (1·4%)	2 (2·6%)	3 (2·0%)
Children's Healthcare of Atlanta	12 (16·7%)	11 (14·5%)	23 (15·5%)
Hospital for Sick Children	3 (4·2%)	3 (3·9%)	6 (4·1%)
Columbia University	3 (4·2%)	4 (5·3%)	7 (4·7%)
University of Miami	4 (5·6%)	4 (5·3%)	8 (5·4%)
Children's Hospital of Los Angeles	4 (5·6%)	5 (6·6%)	9 (6·1%)

^*One patient classified as whole liver received a domino liver transplant.

Analysis of Rejection and Markers of Liver Injury

The primary analysis included 18 end-point events (16 centrally determined rejections, one re-transplantation and one withdrawal of consent, Table 2). Six (8·3%) occurred in the TI group, and 12 (15·8%) in the SOC group (relative risk 0·57, 95% CI 0.24–1.35, p=0.20; primary outcome). The results from sensitivity analyses, which looked at components of the primary end-point, were consistent with the primary analysis: centrally determined rejection occurred in 10·8% overall (16 events in 148 patients): 8·3% of the TI and 13·2% SOC groups (relative risk 0·67, 95% CI 0·27–1·66) (Table 2), with 16·2% overall incidence of locally determined rejection (24 events in 148 patients): 13·9% in the TI and 18·4% in the SOC groups (relative risk 0·79, 95% CI 0·40–1·56) (Table 3).

Table 2:

Primary and Sensitivity Analysis of Rejection

Endpoint	Telemetric Intervention (N=72)	Standard of Care (N=76)	Total (N=148)	Relative Risk or Rate Ratio of Intervention vs Standard of Care (95% CI)
Primary Analysis
Composite endpoint of incidence of central pathology confirmed rejection plus other pre-defined outcomes *
N (%)	6 (8·3%)	12 (15·8%)	18 (12·2%)	0·571 (0·24, 1·35), p=0·20
Sensitivity Analysis
Premature exit due to re-transplantation, withdrawal of consent, lost to follow up, death **
N (%)	2 (2·8%)	3 (3·9%)	5 (3·4%)	0·701 (0·12, 4·09)
Incidence of biopsy-confirmed rejection as read by central pathology ***
N (%)	6 (8·3%)	10 (13·2%)	16 (10·8%)	0·671 (0·27, 1·66)
Number of biopsy-confirmed rejections per patient during follow up as read by central pathology
Mean (SD)	0·1 (0·6)	0·1 (0·4)	0·1 (0·5)	0·952 (0·40, 2·24)
Median (Q1, Q3)	0 (0, 0)	0 (0, 0)	0 (0, 0)
Number of biopsies performed per patient during follow up
Mean (SD)	0·4 (0·8)	0·4 (0·6)	0·4 (0·7)	0·882 (0·52, 1·48)
Median (Q1, Q3)	0 (0, 0·5)	0 (0, 1)	0 (0, 1)

Patients can prematurely exit due to re-transplantation, withdrawal of consent, lost to follow up, or death and have a biopsy-confirmed rejection as read by central pathologist, so the count may not add up to the composite endpoint.

^*Other pre-defined outcomes is defined as other instances cited in protocol, such as re-transplantation, withdrawal of consent, death, premature exit due to loss to follow up

^**These endpoints are classified as rejection for the purpose of primary analyses

^***Patients who met other endpoints comprising the Primary were included in the denominator of this analysis

¹Relative Risk and corresponding asymptotic Wald confidence limit. Site stratification from the Cochran-Mantel-Haenszel test was used for the primary analysis and the sensitivity analysis of biopsy-confirmed rejection as read by central pathology.

²Rate Ratio (Estimated Count per patient during follow up of Telemetric Intervention vs Standard of Care) based on a Poisson regression model with offset set to the log of follow up.

Table 3:

Pre-Defined Secondary Outcome Measures

Endpoint	Telemetric Intervention (N=72)	Standard of Care (N=76)	Total (N=148)	Mean Difference or Relative Risk or Rate Ratio or Hazard Ratio of Intervention vs Standard of Care (95% CI)
Standard Deviation of a Series of Medication Level Variability Index (MLVI) during follow up
N	71	74	145
Mean (SD)	2·1 (1·6)	2·3 (1·6)	2·2 (1·6)	−0·201 (−0·73, 0·32)
Median (Q1, Q3)	1·6 (1·2, 2·5)	2·0 (1·4, 2·6)	1·7 (1·2, 2·5)
Incidence of Locally Determined Rejection by Biopsy
N (%)	10 (13·9%)	14 (18·4%)	24 (16·2%)	0·792 (0·40, 1·56)
Number of Locally Determined Rejections by Biopsy per patient during follow up
N	72	76	148
Mean (SD)	0·2 (0·4)	0·2 (0·5)	0·2 (0·5)	0·743 (0·35, 1·54)
Median (Min, Max)	0 (0, 0)	0 (0, 0)	0 (0, 0)
Time (months) to First Locally Determined Rejection by Biopsy from Enrollment
N	10	14	24
Mean (SD)	5·6 (4·6)	8·3 (4·4)	7·2 (4·6)	0·764 (0·34, 1·70)
Median (Min, Max)	3·9 (2·0, 10·7)	9·4 (5·5, 11·2)	6·9 (3·8, 10·9)
Time (months) to First Centrally Determined Rejection by Biopsy from Enrollment
N	6	10	16
Mean (SD)	6·4 (5·1)	8·0 (4·9)	7·4 (4·9)	0·634 (0·23, 1·73)
Median (Min, Max)	4·3 (3·9, 11·1)	8·7 (5·5, 11·2)	6·9 (3·9, 11·2)
Mean ALT (U/L)
N	72	76	148
Mean (SD)	40·5 (32·4)	45·7 (36·1)	43·2 (34·3)	−5·151 (−16·15, 5·85)
Median (Min, Max)	31·3 (19·8, 49·7)	37·7 (22·3, 54·7)	33·0 (21·5, 51·1)
Number of achieving above predetermined threshold ALT > 150 U/L
N (%)	10 (13·9%)	18 (23·7%)	28 (18·9%)	0·592 (0·29, 1·18)
Maximal ALT (U/L)
N	72	76	148
Mean (SD)	86·6 (100·6)	110·5 (1137)	98·9 (107·8)	−23·961 (−58·39, 10·47)
Median (Min, Max)	51·0 (29·5, 93·5)	67·5 (38·0, 136·0)	58·0 (33·0, 108·0)
Mean gGT (U/L)
N	72	75	147
Mean (SD)	72·8 (126·6)	84·1 (142·3)	78·6 (134·5)	−11·361 (−54·68, 31·95)
Median (Min, Max)	26·6 (16·0, 66·5)	36·0 (17·6, 83·5)	29·0 (16·7, 77·2)
Number of achieving above predetermined threshold gGT > 150 U/L
N (%)	14 (19·4%)	20 (26·3%)	34 (23·0%)	0·742 (0·40, 1·35)
Maximal gGT (U/L)
N	72	75	147
Mean (SD)	142·7 (275·1)	165·5 (283·4)	154·3 (278·7)	−22·871 (−112·60, 66·87)
Median (Min, Max)	36·5 (21·0, 115·5)	63·0 (26·0, 179·0)	47·0 (23·0, 141·0)

Patients in secondary analyses who met an endpoint specified in the primary analysis but not the endpoint of interest in the secondary analysis were included in the denominator.

¹Mean Difference based on a linear regression model.

²Relative Risk and corresponding asymptotic Wald confidence limit. Site stratification from the Cochran-Mantel-Haenszel test was used for the analysis for locally determined rejection by biopsy.

³Rate Ratio (Estimated Count per patient during follow up of Telemetric Intervention vs Standard of Care) based on a Poisson regression model with offset set to the log of follow up.

⁴Hazard Ratio (Rate of the event during follow up of Telemetric Intervention vs Standard of Care) based on a Cox PH model.

ALT and GGT levels in blood were pre-specified as secondary outcomes (Table 3). The maximal ALT per patient was lower with TI (TI 86·6 ± 100·6 vs SOC 110·5 ± 113·7 IU/L, mean ± SD). This pattern was also observed for maximal GGT levels, mean ALT and GGT levels, and incidence of exceeding a pre-specified threshold of 150 IU/L for either parameter: all were consistently lower in the TI group. The SD of ALT and GGT levels in the TI group were also lower; for example, for ALT SD, TI group was 19·6 ± 25·1, vs SOC 27·1 ± 32·1 (Supplemental Table S2).

Figure 1 summarizes treatment effects, which were in favor of TI group for all primary, secondary, and sensitivity analysis endpoints.

Figure 1: Primary and Secondary Outcomes.

All comparisons use the standard of care as the reference group as presented in Tables 2 and 3. Endpoints with estimates using relative risk, rate ratio, or hazard ratio, as appropriate.

MLVI and Timing of Rejection

Figure 2 illustrates quarterly MLVI in TI and SOC groups, showing MLVI and rejection timing in those patients with rejection. In SOC patients, rejection occurred from the 1^st through 8^th quarters; MLVI for those patients was generally greater than 2.0 and above the fitted spline. In contrast, in patients receiving TI, most rejections occurred within the first two quarters and in those that occurred after that MLVI was within or close to the 95% CI for the fitted spline.

Figure 2: Quarterly MLVI Reading.

Fitted with a Penalized B-Spline Curve

Telemetric Intervention Slope (95% CI)=−0·16 (−0·20, −0·12); Standard of Care Slope (95% CI)=−0·13 (−0·16, −0·09); Difference in Slope (95% CI)=−0·03 (−0·09, 0·02)

Patients meeting the primary endpoint of rejection are highlighted in dark blue. The dark blue circle indicates the end of the quarter where the rejection occurred.

Mean MLVI was similar at study end between the TI and SOC, although the rate of reduction in MLVI tended to be faster in the TI (slope TI −0.16 [−0.20, −0.12] vs SOC −0.13 [−0.16, −0.09]) (Figure 2), and the median MLVI was lower (Table 3). Time to rejection tended to be shorter in those receiving the intervention (Table 3).

Adverse Events

We observed no deaths, PTLD, new-onset diabetes requiring insulin, new-onset systemic hypertension requiring medication, or patients with serum creatinine > 176.8 μmol/L. Two patients in standard of care and one in the telemetric intervention underwent re-transplantation (when an earlier rejection was observed before re-transplantation, the re-transplantation was not counted in the primary analysis). Change in creatinine over time was similar between the groups (TI 0.85 (0.61, 1.09) vs SOC 0.55 (0.31, 0.78)) (Supplemental Figure S2 and Supplemental Table S3).

Centrally Adjudicated Rejection: iMALT versus MALT

The centrally adjudicated rejection rates in the SOC arm of iMALT was significantly lower than the population of similar adolescents (MLVI >2) enrolled in the MALT study (12 out of 76 in iMALT vs 13 out of 25 patients in MALT meeting the criteria; chi-square test p-value < 0.0001). These two cohorts were demographically similar (supplemental table S4).

DISCUSSION

This is the first randomized controlled adherence intervention trial in transplant recipients that employed a single-blind design and focused exclusively on clinical, and not behavioral, outcomes¹⁸. It is also the first trial to a priori define and monitor adverse events that could be a consequence of enhanced adherence. The primary outcome of reduced rejection and related events was not met, even though, as predicted, the intervention group experienced half of the number of pre-defined events (mainly, rejection) as compared with controls, and there was a uniform reduction in all pre-defined parameters of worse outcome in the intervention group (figure 1). The study was not adequately powered to show a difference in the primary endpoint, due to reduced overall incidence of rejection compared to the MALT cohort (on which power was modeled). Under the observed rejection rates in the two arms, there is only 21.2% power to detect a statistically significant difference with the enrolled sample size: we would have needed to enroll 636 participants (318 per arm) to detect a statistically significant difference with the observed rejection rates in the two treatment arms at 80% power.

A selection bias against patients at risk for rejection could explain this reduction, but we believe is not likely. iMALT was designed to mitigate selection bias - a common problem in adherence research^3,25. Patients were selected based on risk (i.e. MLVI > 2.0) rather than employing convenience sampling, patients were approached for consent remotely if needed (recruitment did not depend on clinic attendance), and most eligible patients were approached. In addition, our analysis precluded “attrition” from leading to non-inclusion in the primary analysis, thereby preventing bias due to unbalanced attrition (which could be due to worsening non-adherence). We classified loss to follow-up or withdrawal of consent as “negative” outcomes (counted as “rejection” in the primary analysis), because if we removed such patients from the study, we would have biased the results by excluding patients who were non-adherent to study procedures, in a study that purportedly sought to improve adherence.²⁵

There is an alternative, substantive explanation for the paucity of rejections. This study was not blinded to patients or clinicians. For ethical reasons, sites were free to employ whatever methods they deemed clinically appropriate to improve adherence in any patient. A site survey during the trial showed that 10 of the 13 sites evaluated MLVI as a part of their clinical practice in all patients, and all sites acted on MLVI information when available. This was not the case when MALT¹⁴, the original cohort study, was conducted. In addition to “baseline” mitigation of non-adherence, by necessity, inclusion in our trial further flagged “at risk” patients. All families – controls as well as intervention participants - were informed of the increased rejection risk during consent and may have acted on this information, as may have the clinicians. Supporting this interpretation is the fact that most episodes of rejection in the controls occurred later in the trial, almost exclusively in patients whose MLVI remained persistently high, possibly because extra attention to controls faded over time. In contrast, the few rejections in intervention patients, who continued to be monitored and received calls throughout the trial, did not show this pattern (Figure 2) – they were generally observed early on, when the intervention just started and its beneficial effects may not yet have taken effect. Indeed, rejection incidence in the iMALT SOC arm was significantly reduced compared to the matched (and of similar demographics) high-risk subgroup in the previous MALT cohort, supporting the hypothesis that knowledge of MLVI impacts outcome.

The successive MALT/iMALT trials sequentially studied the relationship between accurate identification of immunosuppressant nonadherence and rejection. The incidence of rejection in the iMALT intervention group (8%) is comparable to the incidence of rejection in the “low-risk” group in MALT (6%), which had the same rejection rates as rates observed nationally¹⁴, suggesting that MALT enrollment was not biased. It is likely that the discovery of MLVI, an objective marker of dysfunctional (non-adherent) behavior, which identifies risk prior to rejection, empowered sites’ significant efforts to thwart rejection, i.e. primary prophylaxis, as compared with past practice of intervening only after a rejection has occurred, i.e. secondary prophylaxis. Use of the MLVI marker to flag risk in control arm patients, associated with the knowledge that high-risk patients were enrolled in the trial, reduced “baseline” risk by more than 50% as compared with historical controls. A further 50% reduction was observed in the intervention group. We underestimated the importance of sites’ clinical use of the new MLVI marker to prevent rejection – but clinically, that use is a tremendous step forward for the field.

Besides the primary analysis, there were other critical and novel findings. Sites were able to screen for at-risk non-adherent patients, based on an algorithm tied to EHR data. This abrogates the street-light effect⁸: the use of convenience sampling, which selects against inclusion of non-adherent patients who are less likely to enroll³. In addition, we were able to engage a group of objectively defined non-adherent patients in a long-term intervention by employing remote interactions rather than in-person visits. The specific intervention we used did not lead to statistically significant improvement in our primary outcome. But we believe the study does provide ample evidence that the MLVI itself is of value and can be tied to interventions that seek to reduce non-adherence. Lastly, adverse events related to enhanced adherence (that may cause medication toxicity) were investigated and not found. Most studies do not pre-select for nonadherent patients, and it is quite possible that such sampling would increase risk of toxicity, as occasionally has been observed^26,27. Few adherence-intervention trials conduct robust safety monitoring; our design could inspire future research considerations.

Some authors theorized that MLVI is not necessarily a marker of erratic medication taking behavior, as medication blood levels could also be affected by between-patient differences: for example, drug-drug interactions, or metabolic or absorption patterns. This theoretical consideration was refuted in a prospective trial that found no such effects on the marker¹⁴: not everything that affects blood levels also affects the intra-patient variability of consecutive levels (a stable, for example genetic, difference in metabolism is not expected to lead to more fluctuation between levels), and not taking the medication at all has a much more profound effect on levels as compared with absorption differences or most drug interactions. Empiric results from MALT showed that the effect on the marker is negligible even in with profoundly variable absorption patterns observed in patients with inflammatory disease¹⁴.

Besides being unexpectedly under-powered, we acknowledge several limitations. The study was single but not double-blinded. Awareness of the patients’ non-adherence likely led to attempts to improve it outside of the study intervention. A fully blinded trial would require deception, which is ethically challenging, particularly in minors. Though extensive efforts to mitigate selection bias were used, it is still possible that we missed the most severely non-adherent patients – for example, those who never obtain any bloodwork. Such patients, presumably, are “flagged” at sites anyway, and may require a completely different intervention. Lastly, it is possible that our intervention was not comprehensive enough. A comprehensive intervention which, for example, would have included mental health assessment and treatment and perhaps in-person encounters might be considered in future research – or in practice. We caution that such approaches may make the intervention less accessible, less generalizable, and less likely to engage severely non-adherent patients.

The novel insight that fluctuation in an objective laboratory metric that is measured over time can be used as an indicator of behaviorally-determined risk through repeated electronic health record review is further supported by the observed increased SD of ALT in the SOC vs TI group. Importantly, it is applicable well beyond the transplant field. For example, excessive fluctuation in blood pressure readings over time, measured in the exact same way as the MLVI (the SD of consecutive readings) has been shown to predict poor cardiovascular outcomes independently of the blood pressure values themselves²⁸. Such fluctuation is likely to indicate erratic adherence to medical recommendations and will be expected to respond to a behavioral intervention, as suggested by our recent results²⁹.

We conclude that statistical significance was not demonstrated, and therefore the intervention is not evidence-based. Rejection rates seem to be substantially reduced by employing the marker clinically, and this clinical improvement cannot (and should not) be undone now in subsequent research. Therefore, a much larger study would need to be conducted to have the power to detect a statistically significant intervention effect, which may not be feasible. Given that a non-significant but potentially large effect was detected on all pre-defined clinical outcomes (Figure 1), while the intervention did not meet evidentiary criteria, our general approach seems to be justified: screen for non-adherence, focus only on those patients who are at risk, and employ an intervention strategy that can engage those particularly hard-to-engage patients over time.

The most important conclusion, in our view, is that MLVI emerges as a useful marker of clinically significant non-adherent behavior which increases rejection risk in transplant recipients. The same concept (albeit with different thresholds) could be applied in other circumstances in which serial ambulatory medication blood levels are obtained, and even more broadly whenever an objective outcome metric that is closely tied to adherence to medical recommendations (such as blood pressure) is obtained repeatedly. Following the MLVI marker is likely an extremely cost-effective approach as it focuses the intervention to only a few patients who are as real risk. Further assessment of optimal interventions for nonadherence in transplant recipients will have to account for the fact that leading centers already use primary prophylaxis via MLVI to mitigate the consequences of severe nonadherence (i.e. rejection).

Abbreviations:

ALT

alanine aminotransferase

EHR

Electronic Health Record

GGT

gamma-glutamyl transferase

iMALT

Improving Medication Adherence in Adolescents who had a Liver Transplant trial

MALT

Medication Adherence in children who had a Liver Transplant trial

MLVI

Medication Level Variability Index

NIDDK-CR

National Institute of Diabetes and Digestive and Kidney Diseases – Central Repository

SD

Standard Deviation

TI

Telemetric Intervention group

SOC

Standard Of Care group