Study design and patients

The present study was pre-registered in the German Clinical Trials Registry (DRKS00013135) at 04/12/2017 and was approved by the local ethics committee. We conducted a randomized, crossover trial including a total of 8 measurements on two consecutive days for each participant. In addition to using patients as their own control, which was shown to be advantageous for the analysis of psychobiological data [32], crossover trials offer more statistical power, and thus, require a fewer number of participants [33]. Patients participated both in a brief pre-recorded mindfulness session (MI) at one day, and in a resting state control condition (CC) at the second. The order of the two experimental conditions was randomized across patients (computer-based block randomization), and allocation to sequence was concealed by use of sequentially numbered, opaque sealed envelopes. Blinding procedures were not feasible in this study. With regard to the analysis of reactivity, this study design represented a nested data structure, with observations at level-1 (L1), sessions at level-2 (L2), and patients at level-3 (L3). Data on the effectiveness of the MI with regards to other outcomes has been presented in a previous publication, showing that the MI led to reduction in self-rated stress and mean heart rate and to an increase in heart rate variability [34].

We recruited patients from the University Palliative Care Unit at St. Vincentius Hospital, Heidelberg, Germany. Based on an initial patient contact and the medical record, possible participants were screened for eligibility. Patients were included if they 1) received inpatient palliative treatment, 2) were assessed by the treating physician as not being in a final phase of the disease, 3) had no cognitive or hearing impairment, and 4) sufficiently spoke and understood German language.

Procedures and intervention

We provided information about the study goals, benefits, and potential risks, and patients had to sign the informed consent sheet if they were willing to participate. Afterwards, we opened a sealed envelope, which contained information on the treatment sequence. Appointments were made for two sessions on consecutive afternoons between 2 p.m. and 5 p.m. in order to minimize the influence of circadian variation on neuroendocrine outcomes [35]. After a questionnaire-based assessment of quality of life, a photoplethysmography (PPG) sensor was placed on the index finger of the patient’s non-dominant hand to monitor cardiovascular regulation throughout the session. During the following hour, patients were asked to provide a saliva sample and to rate their stress and well-being level every 20 min, leading to a total of four measurement points, respectively (T0-T3, Fig. 1). In the MI condition, patients were invited to listen to a 20-min recording via headphones (between T0 and T1), consisting of a breathing exercise and guided body scan meditation for supine positions, which was found to moderately improve well-being and relaxation in palliative care patients in a previous study [36]. The MI was adapted from the mindfulness-based stress reduction (MBSR) program [37] and was chosen in this study due to its brief, simple, standardized, and safe application and because cortisol was found to be sensitive to change induced by MIs in multiple settings including oncology [38]. The primary purpose of the MI was to defocus the patient's attention from symptom burden by focusing on the breath, the bodily sensations, and the present moment. Hence, it aimed at improving self-regulatory processes by increasing attentional inhibition capacities [39].

Fig. 1

Assessments. Notes: MI = mindfulness intervention, CC = control condition, sCort = salivary cortisol, sAA = salivary alpha-amylase, MQOL-R = McGill Quality of Life Questionnaire – Revised, VAS-S = visual analogue scale “stress”, VAS-W = visual analogue scale “well-being”, mHR = mean heart rate, HFnu = high-frequency heart rate variability in normalized units

From T0-T3 in the CC as well as from T1-T3 in the MI condition, patients were asked to remain in their supine resting position. Hence, assessment procedures were identical in both sessions. Deviations from this protocol (e.g., eating, drinking, examinations, visits) were documented. Figure 1 displays the assessment plan overview. Recruitment and providing of the intervention were carried out by a research assistant who was otherwise not involved in the design and analysis of the study.

MeasuresDemographic and medical data

Both demographic data (e.g., age, sex, diagnoses) and information on the 24-h intake of medication on both assessment days were retrieved from the patient’s medical record. Product name and dose on the eligible day were recorded for each medication. Later, we categorized products to drug classes of analgesic, antidepressant, antipsychotic, cardiovascular, corticosteroid, hormonal, and sedative medication, and created a dichotomous variable for each class (no intake vs. intake).

Clinical scales

For the assessment of overall functional status, we used the physician-rated Karnofsky performance status scale (KPSS) which showed high inter-observer reliability in a validation study [40]. Patients’ performance was rated on a single 11-point scale from 0 (“dead”) to 100 (“normal, no complaints, no evidence of disease”). Moreover, we used the single item on global quality of life (gQOL) from the McGill Quality of Life Questionnaire – Revised (MQOL-R; Cohen, Sawatzky [41]). The original MQOL is a frequently used measure in patients with life-threatening diseases and its shortened and revised version showed acceptable psychometric properties [41]. Patients were additionally asked to rate their perception of acute stress and well-being on two visual analogue scales (VAS-S, VAS-W) four times per session (T0-T3) from 0–10 (with 10 indicating high stress or well-being). The use of single-item VAS was recommended in previous trials in palliative care for the brief and least burdensome assessment of psychological states [42].

Cardiovascular recordings

We aimed to monitor patients’ autonomic, cardiac outflow as the HPA, SAM and parasympathetic nervous system (PNS) are closely and dynamically linked to each other [43]. We used continuous PPG recordings (biosignalsplux, Lisbon, Portugal) to estimate beat-to-beat variations in heart rate in milliseconds, i.e., heart rate variability (HRV), based on a pulse wave peak detection algorithm in Kubios HRV Premium Version 3.3.0 [44]. HRV parameters were calculated for four 5-min segments parallel to the VAS and salivary assessments. We focused on mean heart rate (mHR) as a global and intuitive marker of ANS activity, and the high-frequency band in normalized units (HFnu) as a commonly reported index of vagally-mediated HRV [45] for the subsequent correlation analyses.

sCort and sAA assessments

We intended to collect eight saliva samples from each patient by the use of Salivette© tubes (Sarstedt, Nümbrecht, Deutschland). Patients were asked to chew on the cotton wad for 1 min, which was then placed in the collection tube. If a patient refused or was unable to provide the sample (e.g., due to weakness, nausea or xerostomia), we documented the reason. After the session, all collection tubes were safely stored at -80° C at the stress biomarker lab of the Institute of Medical Psychology, University Hospital Heidelberg. Salivettes were later centrifuged according to the manufacturer’s instructions, and the extracted saliva was aliquoted and stored in polypropylene vials until performance of assays for no longer than nine months.

Cortisol was analyzed using a commercially available enzyme-linked immunosorbent assay (ELISA; DES6611; Demeditec Diagnostics, Kiel, Germany) according to the manufacturer’s protocol. sAA was analyzed using a kinetic colorimetric kit with reagents from Roche (Roche Diagnostics, Mannheim, Germany). The intra-assay coefficient of variation (CV) was 3.94% for sCort and 3.60% for sAA. The inter-assay CV was 8.90% for sAA and 7.79% for sCort.

Patients were asked to refrain from eating, drinking or other activities during sessions, if possible, and reasons for deviations from the protocol were documented. Before and after each session, we assessed possibly confounding variables in a brief interview. These variables were selected based on the recommendations provided in two methodological papers on stress biomarker assessments [35, 46]. The recommended lists of items were shortened and adapted for use in palliative care populations. The final checklist in this study consisted of items assessing either time spans in minutes (time since waking up in the morning/last nap/last meal/last drink/last toothbrush) or binary no/yes variables (caffeine intake, oral injuries, xerostomia, standing up/eating/drinking/other unintended incidents during session, subjective experience of strain due to salivary assessment). Other standard control items on nicotine/alcohol intake, symptomatic allergies or menstrual cycle were also included initially, but were excluded from the analyses due to restrictions in variance and usability in this patient population.

Analytic strategySample size calculations

Sample size calculations with G*Power [47] were adjusted for the change trajectories across conditions (two within-subject factors). As no effect size data for the neuroendocrine reactivity to MIs in palliative care was available, we inspected self-report and HRV data from one of our previous trials on psychosocial interventions in palliative care, and found it reasonable to assume medium-sized effects [36]. G*Power suggested N = 32 as the optimal sample size to detect such an effect in a four (measurements) * two (conditions) within-subjects crossover design (f = 0.25, α = 0.05, (1-beta) = 0.85). Accounting for a drop-out rate of approximately 30%, we aimed at recruiting N = 42 patients in this study.

Evaluation of feasibility and acceptance of sCort and sAA assessments

The evaluation of feasibility of salivary assessments was based on the patient and sampling flow data. Issues addressed in this regard were the percentages a) of patients being able to provide the maximum number of 8 samples, b) of samples not obtained due to canceled sessions or symptom distress (e.g., pain, nausea, xerostomia), c) of samples not assayed due to limited amount of liquid, and d) of samples not analyzed due to outlying values. Acceptance was defined as the percentage of patients disagreeing with the checklist item “Did you experience any difficulties with regard to the saliva sampling procedures? (no/yes)”.

Exploration of associations and confounders

Due to the skewness of data, all psychobiological data were log-transformed. We used a non-imputed per-protocol dataset for the exploration of associations and confounders. Outliers deviating more than three standard deviations (SD) from the sCort and sAA mean were excluded. For both sCort and sAA levels we calculated the mean within-session correlation of successive samples in the CC and the between-session correlation of baseline samples (CC and MI) as indicators of reliability. Associations with other related variables were explored either by bivariate Pearson product-moment correlations and their 95% confidence intervals (CI) in case of continuous data, or by standardized mean differences (Cohen’s d) and CIs in case of dichotomous predictors. Sample-level variables (L1) were the repeated measurements of VAS-S, VAS-W, mHR, HFnu, sCort, and sAA across all sessions. Session-level variables (L2) included checklist items, medication intake, and the sCort /sAA baseline scores (T0) from both sessions. Patient-level variables (L3) encompassed age, sex, KPSS, gQOL, as well as baseline levels (T0) of sCort and sAA in the CC. To provide an emphasis on the association strength of all variables, we preferred to report effect sizes and CIs rather than a multitude of hypotheses tests at this stage of the analyses. Effects that were at least medium-sized (r > 0.30 or d > 0.50, Cohen [48]) were considered relevant for further analyses.

Multilevel modeling of outcome data

To account for the nested data structure (observations at L1, sessions at L2, and patients at L3), multilevel modeling (MLM) was performed in the statistical environment R [49]. Primarily, data was analyzed by intention-to-treat, replacing missing values in sCort and sAA levels by means of multiple imputations. Five imputations were created with the “Amelia II” package [50] and were later pooled into a single dataset. MLM parameters were estimated with the “lme” function of the “nlme” package [51] by maximum likelihood (ML). Random intercepts were added on L2 and L3 to minimize standard errors. Treatment (TREAT) and sequence (SEQ) were dummy coded (0/1) and entered as factors. TIME was coded from 0–3 to assess linear trajectories over time. Repeatedly measured variables were averaged for each participant. Together with all remaining variables (except TIME, TREAT and SEQ), these averages were then centered on their respective grand mean and entered on L3 to obtain pure between-subject estimates [52].

Next, MLMs were built to test the role of confounders identified in the exploratory analyses. The outcomes were sCort and sAA reactivity represented by the repeated measurement of their concentrations on level 1 (L1) over the described time span, which will be referred to as “trajectories” in the following results section (contrasting sCort and sAA “levels” in the beforementioned exploratory analyses). First, full models were built which included TIME, SEQ, and the respective set of covariates. Variables were then removed in a stepwise, backwards deletion process, based on their estimate’s p-values. Successive models were compared with both the likelihood ratio (LR) tests for nested models and the Akaike information criterion (AIC). Given a significant LR-test, we opted to keep the model with the lower AIC. We used the models resulting from this iterative procedure and added a TIME*TREATMENT interaction as focal predictor to test the reactivity of sCort and sAA in response to the MI. Although the direction of effects was not clear to predict, we hypothesized a significantly stronger decrease in sCort and sAA in the MI compared to the CC.

Each final model was graphically assessed for violations of central model assumptions [53]. In this process, we identified skewed and leptokurtic residuals on L1 in all sCort MLMs which were caused by outlying observations from one participant. To resolve this issue, data from this participant were removed and the models affected were rebuilt. Lastly, we performed sensitivity analyses refitting the final models with the non-imputed data to assess the robustness of potentially significant findings with regard to the imputation procedure.