Study design
This was a prospective cohort study with four groups of potential risk factors (clinical, acute psychological, socio-demographic and chronic health). Probable PTSD at three months was the primary outcome while depression, anxiety, and mental and physical quality of life at three months were secondary outcomes.
Participants
The sample consisted of consecutive, highest acuity adult patients who received level three care in a large general ICU at University College Hospital, London, England between November 2008 and September 2009. In the UK, level three patients are those receiving mechanical ventilation for more than 24 hours or patients with two or more organs supported. Patients were recruited in the ICU when physicians determined they were showing signs of recovery; when they had capacity to give informed consent, and were awake, alert and able to communicate. They were not recruited on a specific day of their ICU stay, as patients woke up and became alert at different times. They were excluded if they were not English-speaking; had dementia or remained confused or had a low GCS (Glasgow Coma Scale) until their discharge from ICU; were unable to communicate until their discharge from ICU; had severe sensory impairment; or were deemed terminally ill (for example, were receiving palliative care).
Ethics
The study was approved by the Joint University College London/University College London Hospitals Committee on the Ethics of Human Research.
Procedure
ICU patient lists were checked daily to identify eligible participants who had received level three care during their stay. After being assessed for capacity by a health psychologist (the first author), and giving informed consent, patients completed a psychological questionnaire. Patients found to have current confusion or inability to communicate were recruited later in their stay, if and when these problems had resolved. Clinical and socio-demographic data were collected from electronic patient notes held in the ICU. Three months after discharge from the ICU, patients were sent a postal questionnaire, which included measures of PTSD, depression, anxiety, Health-Related Quality of Life (HRQL) and socio-economic circumstances.
Data collection
Socio-demographic data recorded include age, gender, ethnicity and socio-economic position, measured using the National Statistics Socio-Economic Classification [25]. The NS-SEC is a measure of employment relations and conditions of occupations, and is the most widely used measure of socio-economic positions in official UK statistics. The self-coded version of the NS-SEC used in this study has five classes of occupation: managerial and professional; intermediate; small employers and own account workers; lower supervisory and technical; semi-routine and routine. A sixth unclassified category was added.
Clinical data include: type of admission (elective surgical, emergency surgical, non-surgical), source of admission (theatre, ward, Accident & Emergency, other), acute physiology and chronic health evaluation II score (APACHE II) [26], length of stay (days), days of organ support, type of organ support, an infection biomarker (C-reactive protein) and highest therapeutic intervention (Therapeutic Intervention Scoring System, TISS) score during the admission [27]. The TISS score reflects the type and number of intensive care interventions received, with points added for each intensive care activity. Data on drugs administered included exposure to sleep medications (mainly zopiclone), benzodiazepines, anaesthetic agents (mainly propofol), antipsychotics, inotropes and vasopressors, systemically-administered corticosteroids, and opioids; number of psychoactive drug groups received (0 to 7); and the number of days patients were sedated.
Information on "chronic health" factors (chronic physical conditions, psychological history and alcohol use) was obtained from electronic medical records held in the ICU.
Psychological measures
All acute psychological reactions were assessed once a patient was able to respond to questions. Mood in intensive care was measured with 15 items (on anger, anxiety, depression, positive mood and confusion) from the validated Profile of Mood States [28]. Stress reactions were assessed using a newly developed 18-item intensive care stress reactions scale (ICUSS) as validated stress questionnaires did not contain items relevant to the ICU context. The ICUSS has four subscales: "physical stress" (difficulty breathing, pain, discomfort from tubes, anxiety about breathing), "delirious symptoms" (hallucinations, nightmares, disorientation, agitation), control (communication, control, confidence, information) and support (dignity, emotional support).
Memory items, (on being admitted to the ICU, the ICU stay, and presence and content of early intrusive memories in the ICU), were developed with guidance from Professor Brewin, an expert in intrusive memories and stress. The content of intrusive memories was qualitatively assessed as "factual" (real experiences in the ICU) or "unreal" (hallucinations or delusions experienced in the ICU). The validated Brief Illness Perception Questionnaire (BIPQ) [29] was used to measure patients' subjective illness perceptions including "timeline" (how long they believed their illness would last).
Outcome measures
Three months later, PTSD symptoms were assessed using the Post-traumatic Stress Diagnostic Scale (PDS), a well-validated instrument including a 17-item severity scale [30]. We selected the PDS as it conforms to diagnostic criteria for PTSD [5] and has high diagnostic agreement with the gold-standard Structured Clinical Interview for PTSD. Using a cut-point of 18 (on a scale of 0 to 51), shown to be a highly efficient scoring method [31], the PDS severity scale has sensitivity of 0.86, specificity of 0.87 and an overall efficiency of 0.87. Participants were asked to answer questions in relation to a specific trauma (in this case, admission to intensive care) according to PDS authors' instructions. Symptoms of depression were measured with the 20-item Center for Epidemiologic Studies Depression Scale (CES-D) [32], the most widely used measure of depression in epidemiological studies, validated for intensive care patients [33] and many other populations. We used a cut-point of 19 (on a scale of 0 to 60) rather than the usual 16, as recommended to deal with the effect of somatic items in patients with medical illness [34].
We assessed anxiety at three months using a validated short form of the State-Trait Anxiety Inventory (STAI) [35], a widely used questionnaire in many populations and health conditions. We used a cut-point of 44 (range of scores 0 to 80) as recommended for studies of medically ill patients [36]. The SF-12, extensively evaluated to establish reliability and validity, was used to measure quality of life. It yields mental and physical summary scales, transformed to have a mean of 50 and SD of 10 [37]. The follow-up questionnaire included an item about current or past psychological issues but few patients answered it, so we relied on electronic medical records to obtain details of psychological history. Three months was deemed a suitable time-point to measure outcomes, including acute PTSD [5], and to examine relationships between ICU clinical and stress factors and psychological outcomes.
Power
To obtain an initial estimate of the sample size required, a clinically significant difference in PTSD scores between two groups, defined by a binary risk factor (for example, sex), was deemed to be 10 points on the PDS [30]. For this effect size, 80% power and 5% significance, 34 patients were required in each of the two groups. As the analyses were to be carried out using multiple regression, with both continuous and categorical risk factors, the sample size needed to be inflated. With the initial sample size of 68, a correlation coefficient of 0.3 between a continuous risk factor and outcome could be detected [38]. To detect the same correlation coefficient (0.3) between a risk factor and outcome in a multiple regression model where all other variables in the model explained 30% of the total variation in outcome, calculations indicated that the sample size needed to be inflated by 40% [38]. This yielded a total sample size of 95 patients. A drop-out rate of approximately 30% was estimated on the basis of previous experience, raising the recruitment required to approximately 140 patients. During the study, the drop-out rate was higher than expected (36%) and 17 extra patients were recruited to ensure that the study retained power.
Statistical analysis
All statistical analyses were conducted using SPSS for Windows (version 14) (SPSS Inc., Chicago, Illinois, USA).
Distributions of risk factors were assessed with frequency histograms and statistical tests for normality. Ordinary least squares regression models were used with PTSD and other outcomes treated as continuous variables. Model building was carried out in stages so that highly correlated variables (which confounded each other) were not included in the same model and to ensure parsimony of the final model. To facilitate this, four groups of risk factors (clinical, acute psychological, socio-demographic and chronic health) were pre-defined.
(i) Univariable analysis. In this stage of analysis, each risk factor was related to each outcome to estimate unadjusted associations. Correlations, t-tests and one-way analysis of variance were used with, respectively, continuous, binary and categorical risk factors. Spearman's rank correlation coefficients were used if continuous risk factors were not normally distributed.
(ii) Multivariable analysis. In recognition of the number of potential variables being tested in these analyses and the associated implications for sample size, a two-stage multivariable process was used.
Stage one: Separate multivariable models were built for each outcome from risk factors within each of the four groups (clinical, acute psychological, socio-demographic and chronic health) to identify the "strongest" risk factors from each group. Risk factors included in this first stage of multivariable analysis were those that showed significant unadjusted associations (P <0.05) with outcomes in univariable analysis. This first stage of multivariable analysis was not carried out for a group where two or fewer significant risk factors were identified in the univariable analysis. No more than eight variables were entered into a regression in this stage of multivariable analysis due to the sample size of 100 (a rule of thumb is to have 10 to 15 times more observations than variables).
Stage two: The strongest risk factors from each group identified in the first stage of multivariable analysis (based on an adjusted significance level of P <0.01), were entered in a final series of multiple regressions to assess whether factors from different groups were independent of each other (also based on a significance level of P <0.01). Factors were entered in the following order: socio-demographic, clinical, chronic physical, acute psychological and psychological history (at this stage of analysis, chronic factors were split up into chronic physical and psychological history). Residuals were found to be normally distributed in all multivariable models with no evidence of multicollinearity.