Background

Neurodevelopmental disorders (NDDs) are heterogeneous conditions grouped by the DSM-5 as attention deficit hyperactivity disorder (ADHD), autism spectrum disorder (ASD), intellectual disability and learning disorders [1]. The more severe forms often require medicinal interventions but options are currently restricted to symptom suppressing medication. Whereas for ADHD multiple stimulant drugs are registered, for ASD no medication is registered to improve the core defining features but drugs are often prescribed to mitigate associated symptoms such as depression, hyperactivity and irritability.

The advent of genetic animal models of neurodevelopmental conditions has led to the identification of possible mechanism-based treatments, most notably for ASD. One of the most studied options is selective Na+-K+-2Cl− (NKCC1) antagonist bumetanide. Bumetanide is a registered loop diuretic that has been used for almost 50 years in adults and children with a variety of nephrological and cardiac conditions. Bumetanide has a mild side effect profile with diuretic effects such as electrolyte imbalance and hypokalemia that can be safely monitored when kidney function is normal [2, 3]. Blocking NKCC1 chloride import in the brain can lower chloride concentrations and potentially reinstate GABAergic inhibition. In normal development a developmental sequence occurs around birth, which is characterized by dramatic decrease in chloride concentration in neuronal cells. This maturational downregulation of chloride levels causes a shift in the so-called polarity of GABAergic transmission from excitatory (depolarizing) to inhibitory (hyperpolarizing): as referred to as the GABA-shift. The GABA shift is mediated predominantly by a change in the expression of two chloride co-transporters: the Na-K-2Cl cotransporter isoform 1 (NKCC1) importer and K-Cl cotransporter isoform 2 (KCC2) exporter [4, 5], hence the potential for bumetanide to restore inhibitory GABA signaling. Since GABAergic inhibition has an important role in maintaining E/I balance for proper neuronal growth, and synapse and circuit development, alterations in polarity may have wide-ranging consequences. Indeed, in model studies for ASD [6,7,8], epilepsy [9], Rett syndrome [10] and Down syndrome [11], the GABA shift was found to be abolished and excitatory effects of GABAergic signaling were established.

These findings lead to the initiation of bumetanide trials in ASD and several genetic disorders, with varying results [6,7,8]. This is in our opinion, in part the result of ignoring etiological heterogeneity of NDDs, which is likely to result in mechanism-based options not fulfilling a one-size-fits application (as opposed to symptom suppressing treatments). As such, we argue that these treatments will only be effective in a subset of patients with NDDs [2, 12]. Accordingly, we developed a set of trials testing different behavioral neurophysiological and genetic stratifications to evaluate efficacy across different diagnostic classes and to develop strategies for more successful application:

‘Bumetanide for autism medication and biomarker study’ (BAMBI), to replicate effectiveness in ASD on core symptomology and to develop electro-encephalogram (EEG) and cognitive stratification and prediction markers.

‘Bumetanide for the autism spectrum clinical effectiveness trial’ (BASCET), to test effectiveness in a cohort stratified by the presence of sensory reactivity problems across NDDs (ASD, ADHD, epilepsy).

‘Bumetanide for ameliorate tuberous sclerosis complex (TSC) hyper-excitable behaviors’ (BATSCH)’, an open label trial in children stratified by a genetic disorder with previously suggested efficacy of bumetanide.

In these trials, we evaluated multiple outcome levels including: behavior, cognition and neurophysiological changes using questionnaires, neurocognitive testing and resting-state EEG and event-related (ERP) markers, see Tables 1 and 2. The resting-state EEG markers focused on measuring effects on excitation-inhibition (E/I) ratios in line with the putative effect of bumetanide on GABAergic transmission [13]. The main findings included: 1) A superior effect of bumetanide on repetitive behavior in ASD (BAMBI) and TSC (BATSCH) [14, 15], 2) A significant effect of bumetanide on aberrant behaviors across disorders (BASCET and BATSCH), and 3) Enhanced power and excitation-inhibition ratios [13] only in the bumetanide treated group in the BAMBI trial. From the EEG effects, we developed an initial prediction algorithm by incorporating EEG biomarkers and clinical severity scores (RBS-r) [16].

Table 1 Details of BAMBI, BASCET and BATSCH cohorts
Full size table
Table 2 Outcomes of BAMBI, BASCET and BATSCH trials
Full size table

Another important observation was the consistent improvement of several symptoms not captured by the conventional outcome scales. For instance, symptoms such as fatigue, irritability, energy and sleeping problems seemed responsive to bumetanide. To evaluate these symptoms, we have recently developed a ‘patient reported outcome’ (PRO) set that we chose as the primary endpoint in this post cohort study to serve as a potential more personalized method of outcome measuring (van Andel under review).

Here we present the follow-up trial protocol developed to replicate previous bumetanide effects, improve clinical endpoint selection and to validate the treatment prediction algorithm.

Methods/design

The Bumetanide for developmental disorders (BUDDI) study is a post-trial access cohort using Single-Case Experimental Designs (SCEDs) testing bumetanide treatment during 6 months. This type of N-of-1 design is most appropriate for our post-trial access cohort as 1) N-of-1 designs involving placebo treatment periods may not be tolerated, as many participants have already participated in placebo-controlled experiments (i.e. BAMBI, BASCET) and 2) the washout data of BAMBI and BASCET trials suggested prolonged effects of bumetanide treatments, which cause difficulty in placebo versus treatment cross-over designs due to carry over effects. The study will be performed at the N=You neurodevelopmental Precision Center at the Emma Children’s Hospital in the Amsterdam University Medical Center (AUMC), the Netherlands. SCEDs are preferred over the conventional post-trial access design (open label), because of the goals of more individualized effect measurements and improvement of clinical end point selection.

Design

We will use multiple baseline SCEDs (MBD) in which the intervention (bumetanide) is introduced sequentially to different patients with a baseline period ranging from 2 to 12 weeks. The rationale for this multiple baseline is that apart from clinical and response heterogeneity across individuals also symptoms per patient vary over time. In an MBD the variation in the baseline period (A phase) is compared to the variation during the intervention (B phase) on an individual level (i.e. the participant serves as his/her own control). Evidence of such an AB designs is based on demonstrating that the change in behavior only occurs during intervention.

Study population

All participants that participated in the previous studies, as well as a new cohort with matching inclusion and exclusion criteria are eligible for this post-trial access study. See Table 3.

Table 3 In- and exclusion criteria BUDDI trial
Full size table

Recruitment and screening

All previous participants of the BAMBI, BASCET and BATSCH trials will be contacted and informed about this post-trial access study. If interested and eligible, previous participants will receive verbal and written information about the study. We estimate that 50% of the patients that were enrolled in the three previous bumetanide trials will be eligible and motivated to be enrolled in the present study (i.e., 75 patients in total).

Participants of the new cohort will be recruited from the patient population referred to the N=You neurodevelopmental Precision Center at the Emma Children’s Hospital in the AUMC.

Intervention and preparation of study drugs

The intervention constitutes of bumetanide tablets. Bumetanide will be provided at a starting dose of 0.5 mg twice daily and will be increased to the therapeutic dosage of 1.0 mg twice daily at day 7 if there are no signs of dehydration in all participants. In case of limited effects and side-effects and a weight > 45 kg dosage can be increased to 1.5 twice daily. Dose reductions to manage side effects will be allowed at any time. Tablets will be obtained via the research pharmacy of AUMC. Re-labeling will be prepared and applied according to local regulatory requirements (GMP annex 13 guidelines). Participants are instructed to return unused tablets to allow monitoring of drug adherence.

Randomization

A randomization list of 115 baseline periods (2–12 weeks, with an equal distribution between the 11 intervals) is generated using the statistical program SPS. Upon signing informed consent, the participant receives the baseline period corresponding with the next open slot on the list.

Outcomes and measurements

Primary outcome

The primary outcome is a set of patient reported outcome measures (PROMs) containing questions directly or indirectly related to sensory processing difficulties, which will be filled in by caretakers. We chose this outcome over more conventional outcome measures for three reasons. 1) it allows for more personalized method of outcome measuring, 2) the PROMs are selected based on the symptoms not captured by conventional outcome scales (see background), 3) A prerequisite for the primary outcome of a SCED is the frequent and repeated measurement of the target behavior in every phase to address the variability in that behavior during the baseline and the intervention phase. The selected PROM-set asks the respondent to reflect upon the last 7 days, whereas conventional questionnaires often ask for reflection upon a longer time scale, making them less suitable for a SCED design.

The effects on PROMs will be compared to main conventional endpoints used in the original trials (social responsiveness scale, second edition (SRS-2) [17], repetitive behavior scale revised (RBS-r) [18], aberrant behavior scale (ABC) [19] and sensory profile – Dutch version (SP-NL) [20]) as well as accompanying measurements of EEG and neurocognition to further establish effects on brain activity and functioning and to validate predictive markers of treatment response. Individual results will be aggregated to evaluate bumetanide efficacy on a group level.

Secondary outcomes

The secondary outcome measures are divided over three domains: the behavioral, the functional and the translational domain (see Table 4).

Table 4 Outcome measures
Full size table

Behavioral domain

The behavioral domain focuses on clinical outcomes and constitutes of four questionnaires to evaluate core symptomatology. The scales used are consistent with those used in the previous trials: The SRS-2, RBS-r, ABC and SP-NL.

In addition, these questionnaires will be used to validate how well the PROM set captures classically defined core symptomatology.

Functional domain

The functional domain contains neurocognitive and neurophysiological measures. We will use the Emma Tool box, an in-house designed battery of computerized tests, to measure neurocognitive functions in children. Measures will be obtained at 3 time points (baseline and after three and 6 months of treatment). Individual change in domain scores will be analyzed.

We will perform resting state electroencephalography (EEG) to assess neurophysiological functioning. EEG will be recorded by using the 128 channels Magstim/EGI system. EEG data will be processed offline using the Neurophysiological biomarker toolbox (http://www.nbtwiki.net/). Similar to our previous trials we selected five biomarker algorithms that have proven sensitive to the ratio of excitation and inhibition in computational models of neuronal networks generating alpha-band oscillations to quantify EEG. These biomarkers include: Relative and absolute power, central frequency, detrended fluctuation analysis and excitation/inhibition ratio. EEGs will be obtained at baseline, Tmax (1,5 h after first dose) and shall be repeated on a monthly basis.

Translational domain

The translational domain focusses on methods for future translation of (emerging) disease mechanisms and the development of more personalized therapies.

One entry point for personalized therapies are studies in animal models with causal genetic variants for NDDs. Hence, we will perform genetic testing via whole exome sequencing.

In addition, induced pluripotent stem-cell (iPSC) based model systems provide the opportunity to examine disease mechanisms in patient-own neurons and provide the opportunity to test personalized treatment options. Accordingly, we will perform assays with iPSC derived neuronal models.

Safety procedures

Safety will be assessed by the research team under supervision of a child psychiatrist and if necessary, a pediatric nephrologist. The assessment includes checks for the use of other medications, side effects and adverse events. In addition, physical examinations and blood and urine laboratory tests will be performed. See appendix 1 for a schematic overview of the examinations.

Oral potassium supplementation at a dose of 0.25 mmol/kg twice daily will be prescribed via custom pharmacy to all participants in order to avoid hypokalemia. Additionally, adjustments in the dosage of bumetanide are allowed to manage hypokalemia and/or side effects.

Statistics

Power calculations

Without using any kind of data simulation, we estimate that 50% of the patients that were enrolled in the three previous bumetanide randomized controlled trials (RCTs) will be eligible and motivated to be enrolled in the present study (i.e. 75 patients). To estimate sample size requirements and type I and II error magnitudes of a statistical test in a SCED design, the outcome measure that is assessed on the most frequent basis should be used, in our case the PROMs. Since this measure has not been included in the previous RCTs conducted in our patient cohort, we cannot rely on effect size estimates for this particular outcome. As such, due to availability we based the power computations on outcome measures used in the BAMBI study. Specifically, we established the minimal sample size required to detect a significant effect on the secondary BAMBI outcome, improvement on repetitive behavior scale, on which previously, a significant effect was observed.

For this purpose we used an online tool suitable for multiple baseline single case designs [21]. The tool tests a statistically significant effect using a non-parametric, randomization test (see Data Analysis) and the following inputs: effect size to be estimated, number of permutations, number of patients, number of outcome measurements.

In a randomization test, an effect estimate is obtained by drawing a large number of samples from a “randomization distribution”. To overcome the computational difficulties arising from drawing all the possible samples, using Monte Carlo sampling we approximate as accurately as possible the result we would get when drawing all samples. To ensure a compromise between accuracy and computational feasibility, our power simulation uses repeated draws, namely 5 Monte Carlo chains, each drawing 100 samples.

We tested whether, given the current experimental setup, we are able to detect an effect comparable in size to the one obtained in the BAMBI study. For RBS-r, the outcome variable previously shown to be significantly impacted, the effect size of the treatment was estimated to be d = 0.373 (80% C.I. 0.225–0.529), using Cohen’s d for dependent samples. The simulations showed that an effect size d = 0.4 can be detected with a relatively high power (0.71 when N = 40 and 0.73 when N = 50), in the absence of autocorrelation. In practice, measurements which are close in time are related, thus outcome scores at one time point can predict scores at another time point (autocorrelation or serial dependence). With larger sample sizes (N = 60), the power to detect the same effect remains fairly high (0.69) even in the presence of low-medium (r = 0.3) autocorrelation.

Thus, N = 40 is an acceptable minimum sample size to be able to detect the BAMBI effect in a no-autocorrelation scenario and N = 60, a minimum sample size in the presence of low-to-medium autocorrelation. As we intend to enroll a minimum of 75 patients from previous cohorts and a minimum of 40 patients in the new cohort, we expect to have sufficient power to detect a comparable treatment effect.

Data analysis

For the data analysis visual inspection of the data and statistical inference using (interrupted) time series analysis and randomization tests will be performed.

Visual inspection using 2SD-band method

First, the PROMs will be plotted as its own time-series for visual inspection using the 2-SD band method, as described in Hoogeboom et al. [22]. The 2-SD band will be calculated from the baseline data and graphed from the baseline through the intervention and post-intervention phase. If two or more successive data points in the intervention or post-intervention phase fall outside the 2 SDs bandwidth, the result will be considered significant on an individual basis. As autocorrelation can bias the visual inspection, we will check our data in each phase for serial dependence using the lag-1 method. If data are found to be significantly correlated, we will transform the data using a moving-average transformation.

Parametric methods: (interrupted) time series

Single subject measurements are graphed over time in a series based on which we can predict measurements at future time points for that subject.

Due to repeated measurements, the series may exhibit patterns such as autocorrelation, seasonality and trends not explained by the intervention (non-stationarity). Failing to account for said patterns can lead to erroneous effect predictions. Based on visual inspection of the series and residuals, special regression models “ARIMA” (AR auto-regressive, I integrated, MA moving-average) can be tailored to account for said patterns and get more accurate predictions. To obtain group effects, the single-case effect predictions will be aggregated using meta-analysis.

Given our study design we are interested in comparing the (predicted) outcome evolution based on the baseline measurements, to the (observed) evolution based on post- intervention measurements. To that aim we use interrupted time series analysis (ITSA), an extension of classical time series analysis [23].

Non-parametric methods: randomization tests

If the SCED involves a small number of participants, parametric assumptions of data normality and homogeneity of variance might not be met, in which case, tests which do not make parametric assumptions might be more suitable, i.e. Koehler and Levin’s randomization test [24].

The null hypothesis of the randomization test is that, for all possible permutations, the mean difference between baseline and intervention is the same.

Selection bias

There is a likelihood of (patient-induced) selection bias, as it is more likely that previous participants from the responder groups will participate in this post trial access study. Accordingly, there will probably be a larger group of responders than non- responders.

We propose two methods to address this: controlling for covariates associated with selection (modelling-based) and inverse probability weighting. In the first approach, we will include a covariate describing previous responder status and check for significance. The second approach involves computing the probability of selection for each responder category and assigning a weight to each subgroup inverse to the selection probability.

Minimal clinically important difference (MCID)

Another aspect we wish to address is what constitutes a (non)responder. To determine the MCID for our target outcome scales, we intend to use comparisons of care-taker assessments in parent-to-parent conversations (anchor-based method).

Development of new predictors

The newly developed prediction algorithm based on EEG and clinical features will be evaluated and further developed [16]. Analyses will be performed using both regression analyses and supervised machine learning.

Discussion

In conclusion, the BUDDI post trial access study will offer the opportunity to replicate individual and stratified group level effects of bumetanide, to improve clinical endpoint selection and to validate an EEG based treatment prediction algorithm. Forthcoming findings may enhance the applicability of bumetanide in heterogeneous NDD populations.