Overview of systematic reviews of therapeutic ranges: methodologies and recommendations for practice

The search strategy returned 339 studies after duplicates were removed. After screening titles and abstracts twenty-two full text articles were assessed for eligibility. Of these, eight studies were excluded as they did not investigate a correlation between serum drug level and clinical outcomes and four studies were not systematic reviews (see Fig. 1).

The ten systematic reviews were published from 2006 to 2016 and studied thiopurine metabolites (n = 3), mycophenolate mofetil (in post-transplant patients) (n = 2), vancomycin trough concentrations (n = 1), amisulpride (n = 1), aripiprazole (n = 1), olanzapine (n = 1), and aminophylline (n = 1). The systematic reviews encompassed a variety of indications including, inflammatory bowel disease, asthma, Staphylococcous aureus infections, psychiatric disorders and patients receiving kidney, cardiac and liver transplants. All ten systematic reviews investigated the effects of serum concentrations on disease activity, while five systematic reviews also assessed the correlation between serum levels and adverse effects (Table 1).

Of the included systematic reviews, two examined paediatric populations [8, 9] and four adult popuations [10‐13]. The remaining four systematic reviews imposed no age restrictions of participants [14‐17].

The results of the AMSTAR assessment are shown in Table 2. The majority of systematic reviews performed a comprehensive literature search and appropriate techniques for data synthesis. A tool for assessment of quality/risk of bias was used in only three reviews. The tools used were i) the National Institute for Health and Care Excellence (NICE) defined criteria for the assessment of case series [9], ii) Meta Analyses Of Observational Studies in Epidemiology (MOOSE) consensus statement [17], and iii) Cochrane tool for the assessment of risk of bias [8]. Though these systematic reviews used standardized assessment tools, the results of these assessments were not formulated into conclusions. The remaining systematic reviews used no standardized assessment tool for the appraisal of quality or risk of bias in individual studies. Three reviews assessed the potential of publication bias across studies using the Egger method [12, 13, 17].

Of the ten included systematic reviews, one followed a predefined protocol [17]. This systematic review followed a generic protocol for the meta-analysis of clinical trials [18]. None of the included reviews followed a specific protocol for systematic reviews of therapeutic ranges.

All included systematic reviews involved PubMed/MEDLINE in their search strategies, five studies searched Embase [10, 11, 14‐16]. Other literature searches included Google Scholar and DirectScience [17], clinical trials registers [11] and the UK national health system database [9]. Two systematic reviews included a grey literature search via the Cochrane Register of Controlled Trials (CENTRAL) [8, 11].

All included reviews analysed more than one study design. Six systematic reviews placed no restrictions on the included study design [10, 12, 14‐17], one systematic review included observational data only [13], and two systematic reviews limited studies to randomised controlled trials, cohort studies and case control studies [9, 11].

Five systematic reviews included the search terms ‘drug monitoring’, [9, 14‐17]. Acronyms such as TDM [10] and AUC/MIC [13] were also used. Other terms included “Blood AND monitoring”, “physiologic monitoring”, [9] or “plasma levels” [10]. One systematic review did not outline the exact search terms used in the methodology [11], and two reviews used no terms specifically relating to therapeutic drug monitoring [8, 12].

For comparing serum drug levels achieved in each systematic review, four reviews extracted an arbitrary threshold level, and measured the beneficial effects of the drug above this threshold [9, 12, 13, 17], and two extracted the data for area under the curve (AUC) or data on single point monitoring [11, 15]. The remaining reviews extracted the mean ± standard deviation of drug level achieved across participants [8, 10, 16]. One review extracted the timing of drug measurement [11], and two reviews performed separate analyses depending on the type of assay used [13, 17].

Of the ten included systematic reviews, four used meta-analyses to investigate the association between low versus high serum drug levels and clinically relevant outcomes (Table 3). One of these reviews assessed the relationship between vancomycin trough levels and treatment failure, defined as a composite endpoint that included mortality and persistent bacteraemia. This review used data from several primary studies, which compared the occurrence of treatment failure in patients with serum levels below 15 mg/l with patients above 15 mg/l. This threshold level was chosen based on a cohort study indicating a two-fold risk of treatment failure when vancomycin levels are <15 mg/l in the treatment of MRSA bacteraemia [19]. This study relied on primary studies using similar threshold levels [13]. Two systematic reviews investigated the relationship between threshold values of tioguanine-6 (TGN-6) and clinical remission in patients with inflammatory bowel disease. In one systematic review, the authors only included studies which compared participants with TGN-6 concentration ≥ 230 pmol/8 × 10⁸^8 RBC with those below, with this figure decided a priori [17]. One study performed a similar analysis, but included studies with different threshold levels in the same meta analysis (either 235 pmol/8 × 10⁸ or 250 pmol/8 × 10⁸) [12]. The fourth systematic review planned a meta analysis investigating the relationship between high and low olanzapine serum concentrations on symptom scores in patients with schizoaffective disorder, however there was insufficient data reported to enable meta analysis [16].

Six systematic reviews did not perform quantitative data synthesis, and instead data analysis was done descriptively. Four reviews planned no meta analysis a priori in their methodologies, it is not clear in these studies why quantitative data synthesis was not performed [9‐11, 15]. One review performed a meta analysis correlating dosage to clinical outcomes, but had insufficient data for a similar analysis correlating plasma concentrations of aripirazole to clinical outcomes [14]. One systematic review planned a meta analysis a priori, but this was deemed inappropriate by the investigators following data extraction due to the heterogeneity of data due to variability in the threshold of drug levels used, and inconsistencies in reporting of drug levels achieved across participants [8].

No included systematic reviews were able to provide a definitive therapeutic range from systematically reviewed evidence. Though a relationship between serum concentration and clinical efficacy was found in three reviews [10‐12], this was not translated into definable upper and lower limits of drug concentrations that conferred maximum efficacy and safety. This is because a definable cut off point reflecting efficacy and safety could not be established at the lower and upper limits respectively. One systematic review found that the established therapeutic range was able to predict toxicity but unable to predict clinical benefit in patients receiving thiopurine for IBD [9], and three reviews found a weak association between serum drug concentrations and clinical outcomes [8, 11, 13].

Most of our included systematic reviews investigated the effect of serum concentrations of a particular drug, for a particular indication, in a particular population. Similar to typical systematic reviews, the research question should be specified a priori. The reviewers should report whether their review is focussed on the lower limit of the therapeutic range, the upper limit, or both; and the population, indication, chronicity and route of the drug. It is important to consider these when deciding the review question as drugs may be used for more than one indication, and in groups of patients with differing characteristics. If the medicine has more than one indication, an appropriate range for each use should be explored as the optimal therapeutic range may vary with age, population and indication. For example, the plasma salicyclate range for the treatment of inflammatory conditions is considerably greater than that required for analgesia [20].

Selected outcomes must reflect beneficial and harmful effects of the drug that are relevant to clinical practice, in context of the population of interest. For example, focussing on surrogate physiological outcomes such as improvements in Forced Expiratory Volume in One Second (FEV₁) in asthma [21] and electroencephalogram (EEG) measurements in epilepsy [3] may not reflect whether the drug has a beneficial clinical effect at a given concentration.

It is unclear how review outcomes were selected in most systematic reviews. One way of identifying important outcomes is to see whether there is an established core outcome set, which should be measured and reported in all clinical trials in a given condition [22]. The Core Outcome Measures in Effectiveness Trials (COMET) Initiative contains a database of existing core outcome sets [23], however this does not comprehensively cover all possible harmful effects of drugs. A framework of outcomes has been developed [24], which may help decide which are most important and relevant for the particular topic under review. In some situations, patients have been involved in identifying which outcomes are directly relevant to them [25], and if such studies are available they should be considered when identifying appropriate outcomes for a systematic review.

With regards to adverse effects of drugs, at particular serum concentrations, it is important to consider whether the scope of the review should be broadened to include outcomes in wider populations and across other indications [26]. This will not only be determined by the time and resources available, but also whether the pharmacological properties of the drug, and the susceptibility of the patient to adverse effects, varies between groups. One example where this may be particularly relevant is in preterm neonates, who are uniquely prone to long term side effects which may not be seen in other populations, for example cerebral palsy from steroid use [27‐29], and problems with cardiovascular adaptation to extra-uterine life from non-steroidal anti-inflammatory drugs [30].

Reviewers must balance the need to capture all relevant data with the time and resources available. Many therapeutic ranges are under researched, and few primary studies specifically aim to investigate the therapeutic range of a drug. Studies of most interest will be those that report both a measure of serum drug concentration and at least one of the selected clinical outcomes. Authors may consider the levels of available evidence they may find prior to searching for relevant studies. Categorising studies in this way helps systematically grade the available evidence. This avoids placing inappropriate weight on studies that are methodologically less robust when formulating conclusions.

The most useful studies are well designed randomised controlled trials (RCTs) comparing two or more therapeutic ranges for the same drug, and reporting clinically relevant outcomes. Such studies have been conducted to examine serum concentrations of lithium in adults with bipolar disorder for prophylaxis of mania [31] and serum concentrations of teicoplanin in adults who are critically ill with gram negative infections [32].

In practice, trials directly comparing therapeutic ranges are uncommon, and we propose that the next level of evidence will incorporate either observational data or indirect comparisons between clinical trials such as RCTs comparing drug to placebo/other treatment or observational studies.

Observational data may be obtained from cohort studies, case-control studies, or case series. Indirect evidence may be gained from identifying RCTs that have compared the drug against a common comparator (likely placebo), and then analysing whether studies in which a higher serum concentration was achieved demonstrated better effects on beneficial and harmful outcomes.

Combinations of indexing terms such as Medical subject headings (MeSH) and textwords should be used by reviewers to form a search strategy that maximizes sensitivity whilst minimizing specificity whilst keeping the number of records to sift manageable within the logistical constraints of the review. If indirect evidence is considered for inclusion, the search strategy must be sufficiently sensitive to return studies that are not necessarily ‘labelled’ as investigations into the therapeutic range of a drug (e.g. RCTs comparing drug to placebo). As Information on therapeutic ranges may be presented using variable terminology, MeSH terms such as “Drug-Related Side Effects and Adverse Reactions”, “Drug monitoring”, “Pharmacovigilance”, “Adverse Drug Reaction Reporting Systems”, “Monitoring, Physiologic”, “Clinical Pharmacy Information Systems” may lead to inclusion of relevant studies in MEDLINE and other relevant indexing terms will be necessary in other databases.

An assessment of quality allows the reviewers to present the degree to which the results in the available literature are likely to be valid and robust, and whether conclusions that impact on clinical practice should be implemented. There are specific considerations when conducting a systematic review of therapeutic ranges.

There are a large number of critical appraisal tools for different study designs, each with varying emphasis on different aspects of quality, although many lack information on how they were developed [33]. These include the Critical Appraisal Skills Programme (CASP) tools, the Graphical Appraisal Tool for Epidemiology (GATE) and the Case Reports (CARE) tool. These tools were not devised for the appraisal of therapeutic drug monitoring studies specifically, and therefore they may not address specific concerns when assessing the quality of included studies in reviews investigating the optimum therapeutic range of a drug.

Risk of bias in RCTs can be conducted using the Cochrane risk of bias (RoB) tool available from http://​handbook.​cochrane.​org/​chapter_​8/​8_​assessing_​risk_​of_​bias_​in_​included_​studies.​htm. This tool was developed for systematic reviews investigating the efficacy of an intervention, and allows a systemic and consistent approach to all included RCTs. Though this tool allows a systemic and consistent approach, it is limited by its inter-assessor variability, and its tendency to judge older studies as a higher risk of bias. Risk of bias in non-randomised studies can be assessed using the ROBINS-I tool [34].

Although a specific systematic assessment tool for TDM studies is yet to be developed, we suggest the following considerations to be integrated into the quality assessment of included studies. Different assays vary in accuracy and this can lead to misleading measurements of serum concentrations and reviewers should consider the appropriateness of the methods for determining serum drug concentrations in individual participants. Furthermore, drug clearance is a dynamic process and the time period between drug administration and serum measurement should be consistent between participants. Selected outcomes should adequately represent harm and benefit, and techniques for measuring adverse effects should be decided a priori as ad hoc measurement of adverse effects carries risk of selective outcome reporting which could misinform the upper limit of a therapeutic range.

Data synthesis in systematic reviews of therapeutic ranges will involve a comparison of outcomes between participants who demonstrate differing serum levels of a drug. Studies vary in how they summarize the serum levels, and this will impact the approach to meta-analysis. For example, studies may compare mean serum levels between those responding to the drug and those not (or experiencing an adverse event or not) (continuous data), in which case a meta-analysis of differences in means can be undertaken to provide a pooled estimate of the difference. Alternatively, studies may compare numbers responding and not responding to a drug (or experiencing an adverse event or not) between those with serum levels below and above a particular threshold (dichotomous data). In such studies it would be common to report the relative risk or odds ratio of the event of interest, in which case a meta-analysis of the effect estimates can be undertaken to provide a pooled effect estimate. In any case, it is important to assess for heterogeneity between studies, and utilising a random effects approach to meta-analysis is recommended where the heterogeneity is found to be large. Potential sources of heterogeneity can also be explored by subgroup analyses and, where possible, meta-regression.

It is important to note that conducting a meta-analysis may not always be possible within a systematic review, and should only be undertaken where there are a sufficient number of studies with comparable outcomes and exposure groups. For example, where patients are grouped according to a drug level threshold, it may only be appropriate to meta-analyse studies which utilize the same, or at least similar, threshold, and some thought would need to be given as to what range of thresholds would be considered similar.

Where a meta-analysis is not appropriate, a descriptive analysis should be provided of the included studies’ findings.

In reviews appraising the evidence around a standard therapeutic range used in practice, the results may have four possible implications. The range may be appropriate, the upper or lower limit should be amended in light of the review results, or there is no evidence on which to make any recommendations (Table 4).

The quality of the evidence around the recommendations of the review should be presented, in context of a summary of the population, indication, and outcomes included in the review.